March 25, 2026

Sixteen days ago, on March 9, 2026, an event occurred that should have captured the attention of every serious biotech investor.

Prothena Corporation announced it had earned a $50 million clinical milestone payment from Novo Nordisk, among the most rigorous pharmaceutical companies in the industry, after achieving a prespecified enrollment target in the Phase 3 CLEOPATTRA trial evaluating coramitug in ATTR amyloidosis with cardiomyopathy.

This wasn’t a press release about encouraging preclinical results. It wasn’t a partnership headline designed to generate excitement. It was a $50 million payment triggered by a defined Phase 3 enrollment milestone, paid by Novo Nordisk.

With this latest payment, Prothena has now received $150 million to date of the $1.2 billion in total potential milestone payments available under its Novo Nordisk agreement alone. Under that deal, signed in July 2021, when Novo Nordisk acquired full worldwide rights to Prothena’s ATTR amyloidosis business and pipeline, Prothena is compensated for meaningful clinical and commercial progress, regardless of whether Prothena funds the work.

It’s worth pausing on that structure before going any further, because it is central to understanding what Prothena is, and why I believe the market is valuing it in a way that doesn’t align with what the data is showing.

Subscribe now

The Architecture of the Invisible Pipeline

Here is the core Prothena investment thesis, stated plainly:

Prothena’s four partnered clinical programs could generate up to approximately $3 billion in aggregate future milestone payments, in addition to potential royalties. Those programs are being advanced entirely at the expense of Roche, Novo Nordisk, and Bristol Myers Squibb, three of the largest and most demanding pharmaceutical companies in the world.

As of March 16, 2026, Prothena’s market capitalization is approximately $495 million. The company holds $308.4 million in cash and has no debt.

You don’t need a spreadsheet to see the disconnect: a company with $308 million in cash, no debt, and roughly $3 billion in potential future milestone payments tied to active Phase 2 and Phase 3 programs, run and funded entirely by blue-chip partners, received a $50 million payment just fourteen days ago, yet is valued at about $495 million.

Cash alone represents roughly 62% of the market capitalization. In practical terms, the partnered pipeline appears largely unreflected in the current share price.

Why Big Pharma’s Money Is the Smartest Signal in Biotech

I’ve spent four years in Silicon Valley watching capital chase narratives. In biotech, the most reliable signal is a large, binding financial commitment made by the parties with the deepest access to the underlying data.

Novo Nordisk didn’t issue a $50 million payment on March 9 because it liked Prothena’s investor deck. The payment was triggered by the achievement of a prespecified enrollment milestone in a Phase 3 trial that Novo Nordisk designed, funds, and is running. It wasn’t discretionary; it was contractual, automatically due once the study hit a target set by Novo Nordisk’s clinical team, informed by data its scientists review on an ongoing basis.

That is what I mean when I say Big Pharma’s money is the validation trail worth following.

The four active partnered programs are:

• Prasinezumab (Roche): Phase 3 PARAISO in early Parkinson’s disease (~900 participants; primary completion 2029)

• Coramitug (Novo Nordisk): Phase 3 CLEOPATTRA in ATTR amyloidosis with cardiomyopathy (~1,280 participants; primary completion 2029)

• BMS-986446 (Bristol Myers Squibb): Phase 2 TargetTau-1 in early Alzheimer’s disease (fully enrolled; primary completion H1 2027)

• PRX019 (Bristol Myers Squibb): Phase 1, with BMS expected to decide on further development in 2026

Each program has been licensed to a partner that funds and operationally manages development. Prothena is not running these trials, and it is not paying for them. Instead, Prothena earns milestone payments when the programs reach predefined objectives set by the partners themselves.

This isn’t the standard biotech model. It’s a different architecture, closer to a royalty-streaming business than a traditional drug developer. And in most sectors, royalty-streaming models tend to command premium valuations, not discounts.

The “$3 Billion” Number

I want to be precise here. The roughly $3 billion of potential future milestones across Prothena’s four partnered programs is not cash in hand, and it is not unconditionally committed. It represents the cumulative milestone payments Prothena would earn only if each program successfully progresses through the defined clinical, regulatory, and commercial hurdles.

For example, the Roche collaboration includes up to $620 million in potential future milestone payments, plus tiered sales royalties in the high teens. Roche has also referenced a peak-sales opportunity of more than $3.5 billion for prasinezumab in Parkinson’s disease. The Novo Nordisk agreement includes up to $1.2 billion in total potential payments. The Bristol Myers Squibb collaboration was structured with a $100 million upfront payment, a $50 million equity investment in Prothena, and up to $562.5 million in regulatory and commercial milestones per program under its global-rights framework.

These figures aren’t aspirational numbers from a pitch deck. They are contractual terms disclosed in SEC filings, negotiated over months by pharmaceutical companies with highly sophisticated legal and business development teams.

What the market appears to be pricing today is not a probability-weighted share of these upside outcomes, but something much closer to a “no success” scenario, where the partnered programs fail and Prothena is valued largely for its cash. At an approximately $495 million market capitalization with $308 million in cash, investors are effectively valuing the pipeline at roughly $187 million.

That’s the gap, and it’s real.

The Non-Dilutive Capital Advantage

In Silicon Valley, we think in terms of compounding: the company that doesn’t need to issue equity simply to survive is the one that can compound value for existing shareholders.

In the 2026 capital environment, where AI infrastructure companies are issuing enormous amounts of debt and early-stage biotechs are operating on shrinking runways, the ability to support operations through non-dilutive milestone income is a real and underappreciated competitive advantage.

Prothena’s full-year 2026 guidance calls for net cash used in operations and investing of $50–55 million, with expected year-end 2026 cash of approximately $255 million at the midpoint. Importantly, that guidance explicitly excludes the possibility of earning up to $105 million in aggregate clinical milestone payments from partners in 2026.

With $50 million of that potential $105 million already received on March 9, the baseline math has improved. And if an additional PRX019 milestone from Bristol Myers Squibb is triggered before year-end, as BMS has indicated could occur in 2026, Prothena could exit 2026 with a cash balance meaningfully above the current guidance range.

A company being paid by Novo Nordisk and BMS as partnered trials advance is not forced to raise capital at depressed prices. That structural protection matters more in 2026 than it did in 2021, when capital was abundant. The non-dilutive milestone framework Prothena negotiated from a position of strength is now functioning as a financing advantage in a market where many small-cap biotechs are fighting simply to secure their next round.

The Continuous Maintenance Thesis

There’s a second layer to the Prothena story that matters from a market-structure perspective, and it ties directly to an active debate in Alzheimer’s drug development.

Eli Lilly’s Kisunla (donanemab) was evaluated using a “stop-dosing” paradigm: patients were treated until amyloid levels dropped below a defined threshold, at which point dosing was discontinued. The logic is clean; if amyloid has been cleared, why continue treatment? The clinical benefit was meaningful, the FDA approved the drug, and Lilly positioned a finite course of therapy as a commercial advantage.

The counter-thesis, one that Bristol Myers Squibb is funding Prothena’s BMS-986446 (an anti-tau antibody) to test, is that in chronic, progressive neurodegenerative disease, pathology can continue even after a single protein target is addressed. Amyloid and tau are not separate problems; they are linked components of a broader cascade. Clearing amyloid without addressing tau may slow progression, but not arrest it. Under this “maintenance model,” continued therapy targeting tau after amyloid clearance could deliver better long-term outcomes. BMS is investing in that hypothesis through the fully enrolled Phase 2 TargetTau-1 trial, with primary completion expected in H1 2027.

If the maintenance thesis proves correct (if sustained tau suppression following amyloid clearance leads to meaningfully better cognitive outcomes than amyloid clearance alone), the commercial implications are significant. A maintenance therapy, by definition, drives recurring revenue over the life of the patient rather than a one-time or finite course of treatment.

That thesis is not yet proven; the key data are expected in 2027. But the partner underwriting the test is BMS, and BMS does not routinely design and fund Phase 2 Alzheimer’s trials for programs it views as long shots.

What This Means for Your Portfolio

The investment framework I would apply to Prothena rests on three distinct analytical layers:

1. The cash floor. With $308 million in cash, no debt, and $50 million of new non-dilutive capital just received, the downside protection is tangible and mathematically bounded. The company has also authorized a $100 million share repurchase program through the end of 2026; an explicit signal that management views the current price as below intrinsic value. When a company with $308 million in cash and a roughly $495 million market cap authorizes a $100 million buyback, it is making a clear statement about how it assesses its own valuation.

2. Milestone optionality. The remaining $55 million of potential 2026 milestones, contingent on Bristol Myers Squibb’s PRX019 advancement decision, is real, near-term, and binary. If BMS triggers it, Prothena’s cash position improves and the validation signal strengthens. If BMS declines to advance PRX019, Prothena loses that program’s near-term upside but still retains three active Phase 2 and Phase 3 partnerships funded and run at partner expense.

3. 2029 Phase 3 readout optionality. The Phase 3 PARAISO trial (prasinezumab in Parkinson’s) and the Phase 3 CLEOPATTRA trial (coramitug in ATTR-CM) both have primary completions expected in 2029. If either reads out positively, milestone payments from the relevant partner could become large, and arrive quickly. Roche has projected peak sales potential of more than $3.5 billion for prasinezumab, with Prothena eligible for high-teen royalties in addition to milestone payments.

None of this is guaranteed. Phase 3 failures are common in biotech, even within the best-resourced pharmaceutical companies. The thesis isn’t “these programs will succeed.” It’s that the market appears to be pricing in near-total failure across all four programs at once, while the companies with the deepest access to the underlying data—Roche, Novo Nordisk, and BMS—are actively spending the money to advance them.

The Deeper Truth About Smart Money

I’ll close with the most transferable insight, one that applies to virtually every partnership-structured biotech investment.

Big Pharma doesn’t write checks for “maybe”. Clinical, legal, and business development teams spend months, sometimes years, scrutinizing data packages before a collaboration is signed. And by the time a milestone payment is triggered, a program has already cleared multiple internal decision gates that public-market investors never get to see.

When Novo Nordisk paid the $50 million coramitug enrollment milestone on March 9, it wasn’t happenstance. It reflected the execution of a clinical organization that enrolled roughly 1,280 patients into a Phase 3 trial for an asset Novo Nordisk paid $60 million upfront for in 2021—an asset it designed a pivotal program around and continues to monitor with access to safety information that the public won’t see in full until much later.

Public investors looking at PRTA’s stock chart and noting a 36.8% decline over the past year are not seeing what Novo Nordisk sees. They aren’t seeing what Roche is seeing inside its Phase 3 PARAISO trial for prasinezumab. They aren’t seeing what BMS is seeing in the fully enrolled TargetTau-1 study.

That is the partnership paradox: the parties with the most information are often the most committed, while the market with the least information can be the most skeptical.

Over four years watching Silicon Valley capital markets, I’ve seen that gap resolve in a consistent direction toward the people who have read the data.

If this edition gave you a framework for reading partnership milestones as validation signals rather than just accounting line items, forward it to one investor who is watching PRTA or any other milestone-driven biotech from the sideline.

Share

References and Relevant Resources

1. Prothena Corporation — Official Press Release (Primary Source) Prothena Corporation plc. “Prothena Announces Achievement of $50 Million Clinical Milestone Payment from Novo Nordisk Related to Ongoing Phase 3 Clinical Trial for Coramitug (Formerly PRX004) in ATTR Amyloidosis with Cardiomyopathy.” Business Wire via Prothena IR, March 9, 2026.
→ Primary IR: https://ir.prothena.com/investors/press-releases/news-details/2026/Prothena-Announces-Achievement-of-50-Million-Clinical-Milestone-Payment-from-Novo-Nordisk-Related-to-Ongoing-Phase-3-Clinical-Trial-for-Coramitug-Formerly-PRX004-in-ATTR-Amyloidosis-with-Cardiomyopathy/default.aspx
→ BioSpace: https://www.biospace.com/press-releases/prothena-announces-achievement-of-50-million-clinical-milestone-payment-from-novo-nordisk-related-to-ongoing-phase-3-clinical-trial-for-coramitug-formerly-prx004-in-attr-amyloidosis-with-cardiomyopathy
→ PharmiWeb: https://www.pharmiweb.com/press-release/2026-03-09/prothena-announces-achievement-of-50-million-clinical-milestone-payment-from-novo-nordisk-related-t
→ OneAmyloidosisVoice: https://oneamyloidosisvoice.com/news-meeting/prothena-announces-achievement-50-million-clinical-milestone-attr-amyloidosis $50 million milestone payment announced March 9, 2026; prespecified enrollment target in Phase 3 CLEOPATTRA trial; coramitug formerly known as PRX004; potential first-in-class amyloid depleter antibody for ATTR-CM; approximately 1,280 participants; NCT07207811; primary completion expected 2029; Prothena has now earned $150 million to date of $1.2 billion total eligible from Novo Nordisk; Novo Nordisk acquired full worldwide rights July 2021.

2. Investing.com — Stock Movement and Market Context “Prothena Earns $50M Milestone from Novo Nordisk for Drug Trial.” Investing.com, March 9, 2026.
→ https://www.investing.com/news/company-news/prothena-earns-50m-milestone-from-novo-nordisk-for-drug-trial-93CH-4550508 Market cap approximately $526 million on day of announcement; stock traded at $9.81, up 3.7% from previous close on milestone news.

3. StockTitan — Pipeline and Milestone Summary “PRTA — Prothena Latest Stock News & Market Updates.” StockTitan, ongoing through March 2026.
→ https://www.stocktitan.net/news/PRTA/ Pipeline summary — Roche Phase 3 PARAISO initiation Q4 2025; Novo Nordisk Phase 3 CLEOPATTRA initiation Q4 2025; BMS Fast Track for BMS-986446; potential $105M in aggregate 2026 milestones; $100M share repurchase authorization; BMS PRX019 development decision expected 2026.

4. FinancialContent / Business Wire — Full Press Release Archive “Prothena Announces Achievement of $50 Million Clinical Milestone Payment from Novo Nordisk.” FinancialContent via Business Wire, March 9, 2026.
→ https://www.financialcontent.com/article/bizwire-2026-3-9-prothena-announces-achievement-of-50-million-clinical-milestone-payment-from-novo-nordisk-related-to-ongoing-phase-3-clinical-trial-for-coramitug-formerly-prx004-in-attr-amyloidosis-with-cardiomyopathy Full press release text; forward-looking statement context; coramitug mechanism description as “potential first-in-class amyloid depleter antibody.”

5. Prothena Corporation — SEC Filing (Primary Source) Prothena Corporation plc. Form 8-K, Exhibit 99.1. “Prothena Reports Fourth Quarter and Full Year 2025 Financial Results, and Provides Financial Guidance and Business Highlights.” Filed February 19, 2026.
→ SEC Direct: https://www.sec.gov/Archives/edgar/data/0001559053/000155905326000003/exhibit991q42025earningsre.htm
→ Business Wire: https://www.businesswire.com/news/home/20260219016841/en/Prothena-Reports-Fourth-Quarter-and-Full-Year-2025-Financial-Results-and-Provides-Financial-Guidance-and-Business-Highlights
→ Nasdaq: https://www.nasdaq.com/press-release/prothena-reports-fourth-quarter-and-full-year-2025-financial-results-and-provides
→ PharmiWeb: https://www.pharmiweb.com/press-release/2026-02-20-1/prothena-reports-fourth-quarter-and-full-year-2025-financial-results-and-provides-financial-guidanc Cash and restricted cash of $308.4 million as of December 31, 2025; no debt; approximately 53.8 million ordinary shares outstanding as of February 12, 2026; full-year 2025 net loss of $244.1 million; Q4 2025 net loss of $21.6 million; net cash used Q4 $23.3 million, full year $163.7 million; 2026 guidance: net cash used in operations and investing $50–$55 million, year-end cash approximately $255 million (midpoint); guidance excludes potential $105 million in aggregate clinical milestone payments; up to $105 million from coramitug advancement by Novo Nordisk and PRX019 advancement by BMS; Roche Phase 3 PARAISO initiated Q4 2025, primary completion 2029; CLEOPATTRA initiated Q4 2025, primary completion 2029; shareholder approval for capital reduction obtained November 19, 2025 EGM.

6. Prothena — Share Repurchase Plan Announcement (Primary Source) Prothena Corporation plc. “Prothena Announces up to $100 Million Share Repurchase Plan.” Business Wire via BioSpace and PharmiWeb, February 28, 2026. → BioSpace: https://www.biospace.com/press-releases/prothena-announces-up-to-100-million-share-repurchase-plan
→ PharmiWeb: https://www.pharmiweb.com/press-release/2026-02-28-2/prothena-announces-up-to-100-million-share-repurchase-plan Board authorization of $100 million share repurchase plan through December 31, 2026; repurchases may occur in open market transactions under Rule 10b-18 or via Rule 10b5-1 plans; plan may be suspended or discontinued; cash of $308.4 million and no debt as of December 31, 2025; year-end 2026 cash approximately $255 million excluding repurchases and potential milestones.

7. Prothena Corporation — Official Phase 3 Advancement Announcement (Primary Source) Prothena Corporation plc. “Prothena’s Partner Roche to Advance Prasinezumab into Phase III Development for Early-Stage Parkinson’s Disease.” Business Wire via Prothena IR, June 15, 2025. → Primary IR: https://ir.prothena.com/investors/press-releases/news-details/2025/Prothenas-Partner-Roche-to-Advance-Prasinezumab-into-Phase-III-Development-for-Early-Stage-Parkinsons-Disease/default.aspx
→ BioSpace: https://www.biospace.com/press-releases/prothenas-partner-roche-to-advance-prasinezumab-into-phase-iii-development-for-early-stage-parkinsons-disease
→ Morningstar / Business Wire: https://www.morningstar.com/news/business-wire/20250615297680/prothenas-partner-roche-to-advance-prasinezumab-into-phase-iii-development-for-early-stage-parkinsons-disease
→ Nasdaq: https://www.nasdaq.com/press-release/prothenas-partner-roche-advance-prasinezumab-phase-iii-development-early-stage
→ Business Wire direct: https://www.businesswire.com/news/home/20250615297680/en/Prothenas-Partner-Roche-to-Advance-Prasinezumab-into-Phase-III-Development-for-Early-Stage-Parkinsons-Disease Worldwide collaboration signed December 2013; Roche has sole responsibility for development and commercialization; Prothena has earned $135 million to date under Roche agreement; up to $620 million in additional milestone payments remaining (regulatory and sales milestones); double-digit teen royalties on net sales; option to co-promote in U.S.; Phase IIb PADOVA study and OLEs from PASADENA informing Phase 3 decision; favorable safety profile across 900+ participants; 750+ receiving treatment 1.5 to 5 years; 30–40% relative reduction in motor progression trend at two years; first biomarker evidence of disease biology impact.

8. BioPharma Dive — Independent Reporting on Phase 3 Advancement Fidler, Ben. “Roche, Waving Off Study Failures, Commits to Phase 3 Trials for Parkinson’s Drug.” BioPharma Dive, June 16, 2025.
→ https://www.biopharmadive.com/news/roche-prothena-parkinsons-drug-prasinezumab-phase3/750791/ Independent confirmation of $135 million earned to date; $600 million in future payouts plus share of U.S. profits and royalties; “multiple endpoints suggest clinical benefit” from PADOVA; Nature paper noting sustained treatment effect requiring further confirmation; Prothena shares climbed 11% on Phase 3 announcement day.

9. Panabee — Comprehensive Analysis of Phase 3 Decision “Roche Advances Prothena’s Potential First-in-Class Parkinson’s Treatment to Phase III, Unlocking Up To $620 Million Milestones.” Panabee, June 16, 2025.
→ https://www.panabee.com/news/roche-advances-prothena-s-potential-first-in-class-parkinson-s-treatment-to-phase-iii-un Roche’s peak sales opportunity projection of greater than $3.5 billion for prasinezumab in Parkinson’s disease; validation of scientific platform and collaborative strategy framing; prasinezumab affects over 10 million people globally.

10. Prothena Corporation — PRX019 Global License Official Press Release (Primary Source) Prothena Corporation plc. “Prothena Announces Bristol Myers Squibb Opt-in for Exclusive Global License for PRX019, the Second Program from Global Neuroscience Research and Development Collaboration.” Business Wire via Prothena IR, May 28, 2024.
→ Primary IR: https://ir.prothena.com/investors/press-releases/news-details/2024/Prothena-Announces-Bristol-Myers-Squibb-Opt-in-for-Exclusive-Global-License-for-PRX019-the-Second-Program-from-Global-Neuroscience-Research-and-Development-Collaboration/default.aspx
→ Nasdaq: https://www.nasdaq.com/articles/prothena-prta-bristol-myers-tie-second-neuro-candidate BMS paid $80 million upfront for exclusive global rights to PRX019; up to $617.5 million in additional development, regulatory, and sales milestone payments; tiered royalties on net sales; PRX019 target undisclosed; FDA cleared IND December 2023; Phase 1 clinical trial initiated 2024; original Master Collaboration Agreement with Celgene (now BMS) signed 2018.

11. Prothena / SEC — PRX019 Global License Agreement 8-K Filing (Primary Source) Prothena Corporation plc. Form 8-K. Filed May 28, 2024. → SEC Direct: https://www.sec.gov/Archives/edgar/data/0001559053/000155905324000019/prta-20240524.htm Precise milestone breakdown — up to $242.5 million in development and regulatory milestones; up to $375 million in sales-based milestones; royalties ranging from high single digit to high teen percentages on weighted average basis; Prothena obligated to run Phase 1 clinical trial for PRX019; BMS responsible for all subsequent development and commercialization.

12. Prothena Corporation — BMS-986446 Fast Track Designation (Primary Source) Prothena Corporation plc. “Prothena’s Partner Bristol Myers Squibb Obtains Fast Track Designation from the U.S. FDA for BMS-986446 (PRX005), an Anti-MTBR-Tau-Targeting Antibody, for the Treatment of Alzheimer’s Disease.” Prothena IR, 2025. → https://ir.prothena.com/investors/press-releases/news-details/2025/Prothenas-Partner-Bristol-Myers-Squibb-Obtains-Fast-Track-Designation-from-the-U-S--FDA-for-BMS-986446-PRX005-an-Anti-MTBR-Tau-Targeting-Antibody-for-the-Treatment-of-Alzheimers-Disease/default.aspx BMS-986446 is a humanized monoclonal antibody targeting multiple domains of the microtubule binding region of tau (MTBR-tau); associated with neurofibrillary tangle formation and cognitive decline; Fast Track Designation confirmed; eligible to receive up to $562.5 million in additional regulatory and sales milestone payments; BMS responsible for all communication, development, manufacturing, and commercialization.

13. Pharmaceutical Technology — BMS PRX019 Deal Analysis “BMS Stakes $80M on Prothena’s Neurodegenerative Candidate.” Pharmaceutical Technology, May 2024. → https://www.pharmaceutical-technology.com/news/bms-stakes-80m-on-prothenas-neurodegenerative-candidate/ BMS-Prothena partnership history — BMS bought U.S. license for PRX005 for $80M (June 2021); worldwide license for PRX005 for $55M (October 2023); PRX005 now BMS-986446 in Phase 2 for Alzheimer’s; BMS neuroscience strategy context including $14B Karuna Therapeutics acquisition (March 2024) for KarXT.

14. Synapse / PatSnap — BMS Deal History Comprehensive Summary “BMS Buys Prothena’s Neuro Programme for $80M.” Synapse/PatSnap, 2024.
→ https://synapse.patsnap.com/article/bms-buys-prothenas-neuro-programme-for-80m Celgene-Prothena discovery agreement 2018 valued at up to $2.2 billion (full platform); BMS acquisition of Celgene 2019 transferring collaboration; PRX005/BMS-986446 Phase 2 trial — 475 patients in early Alzheimer’s disease.

15. Investing.com — Market Cap Data “Prothena Earns $50M Milestone from Novo Nordisk.” Investing.com, March 9, 2026.
→ https://www.investing.com/news/company-news/prothena-earns-50m-milestone-from-novo-nordisk-for-drug-trial-93CH-4550508 Market cap approximately $495–$526 million as of mid-March 2026; stock price $9.81 on March 9, up 3.7%; 53.8 million shares outstanding per February 12, 2026 filing.

March 24, 2026

Silicon Valley is spending billions to reinvent drug discovery with AI. The FDA just reminded everyone that discovery and approval are two different problems entirely.

On January 12 of this year, at the JPMorgan Healthcare Conference in San Francisco, Jensen Huang and Dave Ricks took the stage together to announce something Silicon Valley can’t resist: a billion-dollar bet that AI will reshape the world.

NVIDIA and Eli Lilly unveiled a first-of-its-kind AI co-innovation lab, committing up to $1 billion over five years across talent, infrastructure, and compute. Based in the San Francisco Bay Area, the lab will bring together Lilly’s biology and drug-development experts with NVIDIA’s leading AI engineers to build models aimed at accelerating drug discovery and development, using NVIDIA’s BioNeMo platform as the foundational layer.

The announcement built on news from three months earlier: Lilly plans to build the most powerful supercomputer owned and operated by any pharmaceutical company, an “AI factory” designed to support the full AI lifecycle, from data ingestion and model training to fine-tuning and high-volume inference. The system is expected to include 1,016 Blackwell Ultra GPUs, delivering more than 9 exaflops of AI performance, about nine quintillion AI-optimized calculations per second.

The Valley response was predictable. A 150-year-old drugmaker was suddenly speaking the language of compute. And Jensen Huang, whose chips power GPT-4, Gemini, and many other frontier models, was placing a billion-dollar wager that the same paradigm would transform oncology, neurology, and metabolic disease. The story almost wrote itself: AI would compress a decade-long development timeline, and the clinical trial, with its human variability and regulatory checkpoints, would fade into history.

Then, just twenty-two days later, the FDA sent Moderna a letter that should have given every “AI will fix drug development” evangelist pause.

Subscribe now

The Letter Silicon Valley Isn’t Talking About

On February 3, 2026, the FDA’s Center for Biologics Evaluation and Research (CBER) issued Moderna a Refusal-to-File letter for its investigational influenza vaccine, mRNA-1010. The stated reason was narrow but consequential: Moderna had used a licensed standard-dose seasonal flu vaccine as the comparator in its pivotal Phase 3 efficacy trial; a choice the FDA now deemed inadequate because it did not reflect the “best available standard of care” for older adults. It’s worth being precise about why this is so unusual, because the specifics matter.

Moderna’s BLA included two positive Phase 3 studies with 43,808 participants in total. Both trials met all pre-specified primary endpoints, demonstrating statistical superiority of mRNA-1010 versus their respective comparators. Importantly, both Phase 3 designs were reviewed by the FDA before the studies began.

In April 2024, during a pre-Phase 3 consultation, CBER confirmed in writing that a standard-dose comparator was acceptable, while recommending (but not requiring) a high-dose vaccine for participants over 65. CBER neither placed the trial on clinical hold nor raised objections before enrollment began in September 2024. Then, in August 2025, after the trial completed successfully, CBER flagged the comparator as a significant issue and requested additional supportive analyses, which Moderna provided. At no point did CBER indicate it would refuse to review the submission until it did exactly that on February 3, 2026.

The RTF letter, which Moderna took the unusual step of publishing publicly in redacted form, argued that Moderna had not conducted an “adequate and well-controlled” study because the control arm did not reflect the best available standard of care. In effect, the FDA had approved the study design, the studies met their endpoints, and the agency still declined to review the filing.

Here’s the critical update often missed in early retellings: eight days after the RTF, following a Type A meeting between Moderna and the FDA, the agency reversed course. The FDA agreed to review an amended filing structured by age; full approval for adults 50–64, and accelerated approval for adults 65 and older, contingent on a post-marketing confirmatory study. The FDA also set an internal PDUFA goal date of August 5, 2026.

So the Moderna episode ended, for now, closer to resolution than catastrophe. But the eight days between the RTF and the reversal may be among the most instructive stretches of recent regulatory history. What they exposed wasn’t just a dispute about one vaccine; it was a clear view of the friction between Silicon Valley’s AI-driven confidence and the FDA’s evidentiary bar.

The FDA doesn’t approve technology. It approves proof.

The Gap That $9 Exaflops Can’t Close

Here’s the tension few people in AI-driven drug discovery say out loud: the problem Lilly and NVIDIA’s billion-dollar partnership is built to solve in identifying and optimizing drug candidates is fundamentally different from the problem the FDA is tasked with policing.

AI can screen billions of molecules in the time it takes a human researcher to test a handful. NVIDIA’s BioNeMo ecosystem is expanding rapidly, with more than 650 open models and 250 open datasets spanning biology, chemistry, medical imaging, and related domains. NVIDIA’s VP of Healthcare has even called 2026 “biology’s transformer moment”, projecting that AI agents will design experiments, control instruments via APIs, monitor quality in real time, and analyze results through closed-loop workflows with minimal human intervention. That is genuinely remarkable, and it sits entirely upstream of the FDA’s core question.

The FDA’s job is not to validate technological sophistication. It is to answer a narrower, harder question: Does this drug work in real people, against an appropriate comparator, under a pre-specified statistical plan, in a trial designed to yield interpretable results? No amount of compute can answer that on its own. It requires patients, time, control arms, and the disciplined machinery of clinical trial design; the same apparatus implicated by Moderna’s RTF and Replimune’s CRL.

This is the Gap Between Discovery and Proof, and it may be the most important structural constraint in biotech investing today. AI is compressing the discovery side of the pipeline in real and impressive ways, and it will produce approved medicines. But the proof side—clinical validation—is governed by human biology, statistical inference, and regulatory standards of evidence. None of those yield to GPU acceleration.

Lilly’s Chief AI Officer, Thomas Fuchs, described the company’s supercomputer as “a novel scientific instrument—like an enormous microscope for biologists. It really allows us to do things we couldn’t do before at that enormous scale”. He’s right. And even the most powerful microscope doesn’t replace a clinical trial.

Share

The Mechanism Is Not the Message

The Replimune case, which we’ve covered in depth over the past two editions, is the clearest example of what I call the “Mechanism Is Not the Message” problem.

RP1, Replimune’s lead program, is a genuinely striking piece of biology: a genetically engineered herpes simplex virus designed to selectively replicate in and kill tumor cells, while also releasing immune-stimulating proteins intended to trigger a systemic anti-cancer response. The mechanism is elegant, the science is peer-reviewed, and the clinical signal (a 32.9% objective response rate in anti-PD-1-refractory melanoma) is real.

But none of that is what the FDA’s July 2025 Complete Response Letter focused on. The CRL centered on the absence of a control arm. IGNYTE was a single-arm trial; there was no comparator. Without one, the FDA cannot determine with statistical confidence whether RP1 drove the observed responses or whether they reflect the disease’s natural history, patient-selection effects, or regression to the mean. The mechanism wasn’t the issue. The control group was.

This is the lesson Silicon Valley’s biotech investment culture has been slow to absorb. “Programmable biology”, the idea that we can design organisms the way we write code, and that an elegant mechanism is tantamount to proof, confuses the beauty of a design with evidence of an outcome. The FDA doesn’t evaluate elegance; it evaluates statistically credible improvement versus standard of care, demonstrated in well-controlled human trials.

Inside pharma, 2026 is increasingly being described as a “plateau of productivity”, as AI pilots that fizzled in 2024 and 2025 give way to real operational integration. As one pharma technology analyst put it, “Unless we open the box of specific AI implementation within an organization, it’s in superposition of either transformative success or an epic failure”. The same superposition applies to the proof problem: until you run the trial, you don’t know whether the mechanism translates.

The Wealth-Building Angle: Clean Data Is the New Moat

For investors trying to navigate a market shaped by both AI hype and regulatory reality, I keep coming back to a simple framework: in 2026, the most defensible asset in biotech isn’t the most sophisticated mechanism, but it’s the cleanest data.

In today’s environment, a company with a “boring” drug with well-understood biology, an established indication, a comparator no one will dispute, and a rigorously designed trial can be structurally more valuable than a visionary platform built around a novel mechanism supported primarily by single-arm Phase 2 results. The former has a clearer path to a BLA. The latter has a clearer path to a CRL.

This is the “Clean Data Asset” thesis, and it remains underappreciated in life-sciences investing. It shows up in a few concrete ways:

• The randomized Phase 2 premium. Companies that add randomized control arms in Phase 2, even when not strictly required, are paying more in time and capital upfront to sharply reduce regulatory risk at the BLA stage. That premium still looks underpriced, because many investors continue to grade Phase 2 readouts primarily on response rates, not on the quality of comparator selection and trial architecture. The more sophisticated capital is increasingly underwriting trial design, not just efficacy.

• The re-rating opportunity. Companies that received CRLs under a more permissive regulatory posture and have since rebuilt their evidence packages around rigorous controls may be set up for re-rating as the market prices in a tougher, more “trial-design-forward” standard. Replimune’s April 10 PDUFA date is a visible near-term test. If the FDA approves RP1 on resubmission, it suggests that redesigned datasets and well-constructed supplemental analyses can meet the tightened bar; an outcome that would reprice a broader cohort currently working through CRL-driven remediation.

• The service tech opportunity. The infrastructure that enables high-integrity clinical data collection is arguably undervalued relative to the drug programs it supports. As sensor-rich, data-verified operations become standard across industries, the same shift is moving into pharma manufacturing and clinical development. The companies building the backbone for FDA-grade evidence (electronic patient-reported outcome systems, decentralized trial platforms, and real-world evidence collection tools) are “picks-and-shovels” plays on the new evidentiary regime, with meaningfully less binary risk than the therapeutics themselves.

The Deeper Tension

I want to name the tension beneath all of this, because the AI-biotech narrative circulating in Silicon Valley often glosses over it in ways that can cost investors real money.

The Lilly-NVIDIA partnership is real and consequential. AI will produce approved drugs over time. The compression of the discovery timeline is tangible. None of that is hype.

But here is the biological reality: between identifying a promising molecule and delivering an approved therapy to patients lies a gap that compute cannot close. That gap is the human body: its heterogeneity, its complexity, and its stubborn refusal to behave the way in vitro systems predict. Clinical trials exist not because regulators lack imagination, but because history is unequivocal: what works in a model, a cell line, or a mouse does not reliably work in a 65-year-old with multiple comorbidities and extensive prior treatment.

The FDA’s growing insistence on rigorous, controlled, well-designed trials isn’t technophobia. It’s an epistemological stance, a view about what constitutes sufficient evidence before exposing millions of people to therapies with real risks and real costs. Whatever one thinks about how that standard is being applied today, the underlying premise is not going away.

As one CIDRAP commentary observed in the wake of the Moderna RTF: “Strong, predictable, science-based regulation protects the public. The mRNA platform—the technology that enabled the fastest vaccine development in history during the COVID-19 pandemic—is being driven out of the country not because the science has failed, but because the regulatory environment has become unpredictable”.

That unpredictability is the risk neither the AI-hype narrative nor the regulatory-rigor narrative fully captures. The science is advancing. The compute is here. The regulatory framework is evolving. The companies that make it through the intersection of all three will be the ones that understood from the outset that the mechanism was never the message.

The proof was always the message, and it always will be.

If this edition crystallized something you've been feeling intuitively about the gap between AI promise and regulatory reality in biotech. Feel free to forward it to one person who is invested in that gap, literally or figuratively.

Share

References and Relevant Resources

1. Moderna: Official RTF Announcement Press Release (Primary Source) Moderna, Inc. “Moderna Announces FDA’s Refusal-to-File Letter for mRNA-1010.” GlobeNewswire / Bioanalysis Zone, February 10, 2026.
→ https://www.bioanalysis-zone.com/fda-rejects-modernas-filing-for-mrna-1010-flu-vaccine/ RTF issued February 3, 2026 by CBER; signed personally by Vinay Prasad; sole stated reason was choice of licensed standard-dose influenza vaccine as comparator; BLA submission included two Phase 3 studies, 43,808 participants; both met all pre-specified primary endpoints; both Phase 3 designs reviewed by FDA prior to study initiation; Moderna exercised a Priority Review Voucher; Moderna publicly posted full RTF letter; Type A meeting requested immediately.

2. Moderna: RTF Reversal Press Release (Primary Source) Moderna, Inc. “Moderna Announces FDA Agreement to Review mRNA-1010 BLA Under Revised Regulatory Approach.” February 18, 2026.
→ BioSpace: https://www.biospace.com/fda/fda-reverses-course-on-modernas-mrna-flu-shot-application-promising-august-decision → Pharmaceutical Technology: https://www.pharmaceutical-technology.com/news/fda-to-review-moderna-seasonal-flu-vaccine-mrna-1010/ Reversal confirmed February 18, 2026 — eight days after RTF disclosure; age-stratified filing: full approval for adults 50–64, accelerated approval for adults 65+, with post-marketing commitment; PDUFA goal date of August 5, 2026; CEO Stéphane Bancel quote; Moderna’s stock rose 6% ($43.93 to $46.70) on reversal news.

3. CIDRAP: Primary Op-Ed (Critical Analysis) Osterholm, Michael, and Kristine Moore. “The FDA Refused to Review a Flu Vaccine, Contrary to Evidence. Now the Agency Reversed Itself.” Center for Infectious Disease Research and Policy (CIDRAP), University of Minnesota, February 19, 2026.
→ https://www.cidrap.umn.edu/influenza-vaccines/cidrap-op-ed-fda-refused-review-flu-vaccine-contrary-evidence-now-agency April 2024 pre-Phase 3 consultation — FDA responded in writing that standard-dose comparator was “acceptable”; August 2025 pre-submission meeting — CBER flagged comparator as “significant issue” and requested supplementary analyses, which Moderna provided; FDA never indicated it would RTF before doing so; senior FDA official’s “rig the study” quote from background press briefing; Moderna CEO Bancel’s “100%” response on anti-vaccine climate impact; CIDRAP quote on mRNA platform regulatory predictability; mRNA-1010 under simultaneous review in EU, Canada, and Australia.

4. Contagion Live: Reversal Coverage “Reversing Course, FDA Will Now Review Moderna’s Influenza Vaccine.” Contagion Live, February 18, 2026.
→ https://www.contagionlive.com/view/reversing-course-fda-will-now-review-moderna-s-influenza-vaccine Type A meeting described as “constructive”; PDUFA August 5, 2026 confirmed; full BLA filing mechanics — full approval for 50–64, accelerated approval for 65+.

5. BioPharma Dive: Reversal Detail and Market Analysis Fidler, Ben. “FDA Reverses Course on Moderna’s Flu Vaccine.” BioPharma Dive, February 18–19, 2026. → https://www.biopharmadive.com/news/fda-reverses-course-review-moderna-approval-application-influenza-/812432/ Prasad personally signing the RTF letter; reports that Prasad overruled agency reviewers in rejecting the application; TD Cowen analyst Tyler Van Buren’s note that backlash led to “quickly find an acceptable solution”; “remarkably short turnaround for a Type A meeting”; typically 30–60 days after RTF before Type A meeting is held; RTF letters are rare — only 4% of 2,475 applications per 2021 study.

6. Pharmacy Times: GWU Health Policy Expert Analysis “After Refusal-to-File, FDA Reconsiders Moderna’s mRNA Flu Vaccine With Higher Bar for Older Adults.” Pharmacy Times, March 2026.
→ https://www.pharmacytimes.com/view/after-refusal-to-file-fda-reconsiders-moderna-s-mrna-flu-vaccine-with-higher-bar-for-seniors GWU Professor Tony Yang quote — FDA decision “reflects stricter expectations around trial design and comparators rather than skepticism toward mRNA technology”; characterization that no FDA regulations or guidance explicitly require the “best-available standard of care” comparator; no safety or efficacy concerns about mRNA-1010.

7. Drug Discovery Trends: RTF Rarity Context “FDA Agrees to Review Moderna’s mRNA Flu Vaccine.” Drug Discovery Trends, February 19, 2026.
→ https://www.drugdiscoverytrends.com/fda-reverses-course-agrees-to-review-modernas-mrna-flu-vaccine/ RTF letters received by only 4% of 2,475 applications per 2021 published study; Type A meetings reserved for urgent issues, with FDA required to respond within 14 days and schedule within 30 days if granted.

8. AJMC: Clinical Trial Detail and Filing Mechanics “FDA Reverses Course, Will Review Moderna’s mRNA Flu Vaccine.” American Journal of Managed Care, February 2026. → https://www.ajmc.com/view/fda-reverses-course-will-review-moderna-s-mrna-flu-vaccine Phase 3 trial NCT05415462 — 40,000+ adults 50 and older, 11 countries, observer-blind; 26.6% relative vaccine efficacy (95% CI: 16.7%–35.4%) against confirmed influenza; 27.4% rVE in adults 65+; efficacy against each circulating strain; P303 Part C — mRNA-1010 statistically superior over high-dose vaccine in immunogenicity.

9. CNBC: Supercomputer Primary Announcement Feuerstein, Adam. “Eli Lilly, Nvidia Partner to Build Supercomputer, AI Factory for Drug Discovery and Development.” CNBC, October 28, 2025. → https://www.cnbc.com/2025/10/28/eli-lilly-nvidia-supercomputer-ai-factory-drug-discovery.htmlThomas Fuchs direct quote — “novel scientific instrument...like an enormous microscope for biologists...allows us to do things we couldn’t do before at that enormous scale”; Diogo Rau quote — “where the big opportunity is”; expected completion December 2025, online January 2026; “won’t yield significant returns until end of the decade”; Lilly TuneLab platform description.

10. R&D World: LillyPod Launch and Fuchs January Quote “Eli Lilly Launches LillyPod, Pharma’s Most Powerful Supercomputer.” R&D World, March 2026.
→ https://www.rdworldonline.com/eli-lillys-lillypod-supercomputer-goes-live-with-1016-nvidia-blackwell-gpus/ Fuchs January 2026 R&D World interview — “Compute is a means to an end. It is that scientific instrument you can then wield”; “You cannot accelerate a cancer trial or a neuro trial, the bottleneck is biology”; LillyPod named as the system; workloads spanning genomics, peptide design, single-cell biology, imaging, and manufacturing; TuneLab federated learning structure.

11. Drug Discovery Trends: Supercomputer Technical Specifications “Lilly and NVIDIA Unveil $1B Co-Innovation Lab in SF.” Drug Discovery Trends, January 14, 2026. → https://www.drugdiscoverytrends.com/lilly-and-nvidia-unveil-1b-co-innovation-lab-in-sf/ 1,016 Blackwell Ultra GPU system; rated at more than 9 exaflops of AI performance; “roughly 9 quintillion AI-optimized calculations per second”; world’s first NVIDIA DGX SuperPOD built with DGX B300 systems; unveiled at NVIDIA GTC Washington D.C. on October 28, 2025; metric clarification; “a mixed-precision throughput metric not directly comparable to double-precision benchmark used for Top500 rankings.”

12. Eli Lilly: Thomas Fuchs Chief AI Officer Appointment (Primary Source) Eli Lilly and Company. “Lilly Appoints Thomas J. Fuchs as the Company’s First Chief AI Officer.” Investor Relations Press Release, October 8, 2024, effective October 21, 2024. → https://investor.lilly.com/news-releases/news-release-details/lilly-appoints-thomas-j-fuchs-companys-first-chief-ai-officer Fuchs biography — former dean and inaugural department chair for AI and Human Health at Mount Sinai; director, Hasso Plattner Institute for Digital Health; positions at Memorial Sloan Kettering, NASA’s Jet Propulsion Laboratory, and Caltech; founder of Paige AI; doctoral degree in machine learning from ETH Zurich; effective date October 21, 2024.

13. NVIDIA Newsroom: Official Press Release (Primary Source) NVIDIA Corporation. “NVIDIA and Lilly Announce Co-Innovation AI Lab to Reinvent Drug Discovery in the Age of AI.” NVIDIA Newsroom, January 12, 2026.
→ https://nvidianews.nvidia.com/news/nvidia-and-lilly-announce-co-innovation-lab-to-reinvent-drug-discovery-in-the-age-of-ai Announced at JPMorgan Healthcare Conference, January 12, 2026; up to $1 billion in talent, infrastructure, and compute over five years; “first-of-its-kind AI co-innovation lab”; co-locate Lilly domain experts with NVIDIA AI engineers in San Francisco Bay Area; built on NVIDIA BioNeMo platform and NVIDIA Vera Rubin architecture; NVIDIA and Lilly to “pioneer robotics and physical AI.”

14. Eli Lilly Investor Relations: Official Press Release (Primary Source) Eli Lilly and Company. “NVIDIA and Lilly Announce Co-Innovation AI Lab to Reinvent Drug Discovery in the Age of AI.” Eli Lilly Investor Relations, January 12, 2026.
→ https://investor.lilly.com/news-releases/news-release-details/nvidia-and-lilly-announce-co-innovation-ai-lab-reinvent-drug Lilly’s independent confirmation of all terms; David Ricks direct quote: “Combining our volumes of data and scientific knowledge with NVIDIA’s computational power and model-building expertise could reinvent drug discovery as we know it”; clinical development, manufacturing, and commercial operations as additional application areas beyond discovery.

15. BIO-IT World: Most Comprehensive Technical Summary “NVIDIA Bets Big on AI-Driven Drug Discovery, Physical AI, and a $1 Billion Eli Lilly Partnership.” Bio-IT World, January 12, 2026.
→ https://www.bio-itworld.com/news/2026/01/12/nvidia-bets-big-on-ai-driven-drug-discovery--physical-ai--and-a--1-billion-eli-lilly-partnership Kimberly Powell (NVIDIA VP of Healthcare) description of BioNeMo as “full open development platform”; 650+ open models and 250+ open datasets released in 2025 alone across biology, chemistry, medical imaging; Powell’s “biology’s transformer moment” prediction for 2026; DGX Spark “benchtop supercomputer” AI agents; flow cytometer closed-loop quality control example; Multiply Labs — 70% reduction in cost per dose and 100-fold increase in throughput per square foot; BioNeMo models for RNA structure, molecular synthesis, toxicity prediction.

16. NVIDIA Blog: Jensen Huang and David Ricks On-Stage Summary “CEOs of NVIDIA and Lilly Share ‘Blueprint for What Is Possible’ in AI and Drug Discovery.” NVIDIA Blog, January 13, 2026.
→ https://blogs.nvidia.com/blog/jpmorgan-healthcare-nvidia-lilly/ Jensen Huang and David Ricks on stage together at JPMorgan Healthcare; Ricks on the “holy grail” of modeling the whole biological system at once; lab to open by late March 2026; NVIDIA Clara open models for RNA structure prediction and synthesis practicality; BioNeMo Recipes for foundation model training; Thermo Fisher autonomous lab collaboration; AstraZeneca, GSK, BMS, Merck prior NVIDIA pharma partnerships listed.

17. Fierce Biotech: Partnership Announcement News Coverage “Lilly, Nvidia Tag on Partnership with New AI Co-Innovation Lab, $1B Investment.” Fierce Biotech, January 12, 2026. → https://www.fiercebiotech.com/biotech/lilly-nvidia-tag-partnership-new-ai-co-innovation-lab-1b-investment “Dedicated, incremental resources” from both sides; lab expected to open by end of March 2026; continuous learning system cycling data between robotic lab equipment and AI models; ground truth lab data for training biology foundation models.

18. Deloitte: Primary Survey Report (Authoritative Source) Gaus, Tim, and Michael Schlotterbeck. “2025 Smart Manufacturing and Operations Survey: Navigating Challenges to Implementation.” Deloitte Insights, published May 2025 (conducted August–September 2024).
→ https://www.deloitte.com/us/en/insights/industry/manufacturing/2025-smart-manufacturing-survey.html
→ Press Release: https://www.prnewswire.com/news-releases/deloitte-survey-reveals-smart-manufacturing-is-driving-advantage-but-needs-focused-investment-and-implementation-302443462.html 600 executive respondents from companies with $500M+ revenue and 1,000+ employees with U.S. headquarters or operations; sectors include consumer products, industrial products, energy and chemicals, mining, automotive, transportation, aviation, life sciences, and technology; 92% view smart manufacturing as primary driver of competitiveness over next three years (up 6pp since 2019); 78% of manufacturers allocating over 20% of overall improvement budget to smart manufacturing; 10–20% improvements in production output; 7–20% gains in employee productivity; up to 15% unlocked capacity.

19. PMWares: Full Deloitte Survey Statistical Summary “2025 Smart Manufacturing and Operations Survey by Deloitte: Summary & Key Insights.” PMWares, 2025. → https://pmwares.com/2025-smart-manufacturing-and-operations-survey-by-deloitte-summary-key-insights/ Consolidated list of all key statistics including “85% of respondents believe their smart manufacturing initiatives will transform how products are made, improve agility, and attract new talent”; 88% expect investments to increase or hold steady; process automation ranked as first/second priority investment for 46% over next two years.

20. Replimune RP1: IGNYTE Trial, CRL, and Resubmission RP1 mechanism of action as genetically modified HSV-1; IGNYTE trial 32.9% ORR in 140 anti-PD-1 refractory melanoma patients; July 2025 CRL language; October 2025 resubmission; April 10, 2026 PDUFA date.

21. Vinay Prasad Departure and FDA Single Pivotal Trial Standard Prasad departure end of April 2026; NEJM February 2026 single pivotal trial standard (DOI: 10.1056/NEJMsb2517623); NEJM November 2025 Plausible Mechanism Pathway (DOI: 10.1056/NEJMsb2512695).

Leave a comment

March 23, 2026

There’s a line I’ve heard in three separate Sand Hill Road conversations over the past thirty days: three different firms, three different partners, three different portfolio reviews:

“We can’t fund another single-arm trial.”

Not, “We’re cautious about single-arm trials.” Not, “We want to see better data.” A flat refusal. Full stop.

That line, quiet, practiced, almost rehearsed, is the sound of a structural shift unfolding in real time in venture funding for early-stage oncology. It isn’t being announced in press releases or debated on panels at JPMorgan Healthcare. It’s being decided in Monday-morning partnership meetings on Sand Hill Road, in portfolio reviews between biotech founders and their Series A leads, and in the low-voiced conversations at the back of oncology conferences.

The Replimune episode we covered in the last edition was the visible tip of the iceberg. This edition is about the iceberg itself.

P.S. We close this edition with a dedicated interview and Q&A featuring Silicon Valley VCs, founders, and CEOs. Stay tuned for that conversation; we’ll use it to address the questions that most often come up in real portfolio discussions. If you have questions in the meantime, please don’t hesitate to reach out. We hope this journal gives you practical insight you can apply to your own investing journey.

Subscribe now

The Canary and What It’s Telling Us

When the FDA issues a Complete Response Letter over trial design, not drug safety, it sends a signal far beyond the company that received it. It isn’t just a regulatory event. It’s a repricing event for early-stage biotech across the market.

Here’s the mechanism. A VC partner at a generalist fund (not a dedicated life sciences investor, but a firm that primarily backs enterprise software or climate and occasionally writes a check into a “hot” oncology platform) often uses a straightforward mental model. They see promising Phase 1/2 data, an eye-catching response rate, note that accelerated approval exists, and build a plan that reaches BLA submission in roughly four years. Implicit in that plan is a critical assumption: that the FDA will accept single-arm Phase 2 evidence. That assumption is what makes it plausible for a $50 million Series A to reach a regulatory milestone without first funding a $300 million randomized Phase 3.

That assumption is now demonstrably broken. And when a core assumption breaks, the model built on it breaks too.

Read more

March 22, 2026

I want to open this edition with a confession.

When I sat down to write about Replimune and the FDA’s April 10 PDUFA date, I had a tidy narrative in mind. The kind that almost writes itself: regulatory overreach, a crusading scientist changing the rules midstream, and biotech investors unknowingly walking toward a cliff.

Then, earlier this week, news broke that Vinay Prasad will be leaving the FDA at the end of April, and that neat storyline immediately became more complicated, and far more interesting.

Because what’s unfolding at the FDA right now, in the roughly twenty days between today and April 10, and in the weeks that follow, isn’t just about a single drug decision. It’s about a regulatory philosophy that was put in place, caused real damage, reshaped an industry’s playbook, and is now in the process of exiting. The investors who understand the timing of that transition are the ones least likely to be caught flat-footed on either side of the trade.

Let me lay out the full picture.

Subscribe now

The Drug, the Trial, and the Shadow of July 2025

Replimune’s lead program, RP1 (formally vusolimogene oderparepvec) is a genetically engineered herpes simplex virus. Used in combination with Bristol Myers Squibb’s checkpoint inhibitor nivolumab (Opdivo), it is designed to lyse tumor cells directly while also priming the immune system to generate a durable, systemic anti-tumor response.

The target indication is advanced melanoma, specifically patients whose disease has progressed after anti–PD-1 therapy, one of the most difficult-to-treat populations in oncology. Approved options are limited, and the unmet medical need is real and substantial.

The clinical case rests on IGNYTE, a Phase 1/2 single-arm study. Among 140 previously treated patients with advanced melanoma, the trial reported an objective response rate of 32.9%. In a setting with few alternatives, that magnitude of response is clinically meaningful. More than twenty oncology experts agreed, signing a letter urging the FDA to reconsider after what came next.

What came next was July 2025. The FDA issued a Complete Response Letter, concluding that IGNYTE was not an adequate and well-controlled clinical investigation and therefore could not provide substantial evidence to support approval.

In plain terms: the agency said the package wasn’t sufficient. The trial lacked a randomized control arm. The single-arm design, chosen in part because randomizing late-stage cancer patients to no effective therapy raises serious ethical issues, was now deemed inadequate.

Reporting later suggested the RP1 review was mishandled early, with an anonymous FDA official telling STAT that CDER’s oncology division, then led by Richard Pazdur, became involved late in the process after raising questions about efficacy.

The outcome blindsided the company, investors, and much of the oncology community, and REPL shares fell nearly 40% in a single day.

After months of engagement with the agency, including a Type A meeting in September 2025, Replimune resubmitted its BLA in October 2025 with additional data, analyses, and targeted responses to the CRL. The FDA accepted the resubmission and set an April 10, 2026 PDUFA date under a Class II review timeline.

That’s where things stand today: twenty days out. Same drug. Same underlying trial. A more complete submission, and a materially different regulatory backdrop than the one that produced the initial rejection.

The Architect of Uncertainty and His Exit

To understand why the landscape has shifted, you need to understand Vinay Prasad.

Prasad built his reputation less through conventional academic advancement than through a sustained and often unsparing critique of how cancer drugs are approved. He argued that oncology, in particular, had grown too comfortable with evidence that was publishable on paper but thin in real-world clinical value. In May 2025, he was appointed to lead CBER, the FDA center that oversees vaccines and biologics, with what appeared to be a clear mandate to raise the evidentiary bar.

His regulatory philosophy wasn’t nihilistic, and he wasn’t trying to block drugs for its own sake. The critique was more specific: the FDA had drifted into inconsistent standards; single-arm trials without control groups were too easy to game; and patients were being steered toward drugs with headline response rates but no demonstrated survival benefit.

On the merits, parts of that argument are defensible. The accelerated approval pathway has at times been used—sometimes abused—to bring therapies to market on surrogate endpoints that never translate into the outcomes patients actually need. Prasad had been making versions of this case for years in academic work and in his Substack, long before he entered government.

But once he had regulatory authority, the implementation proved far more disruptive than the argument itself.

Several biotechs, including Capricor Therapeutics, Biohaven, and Replimune, were forced to revisit their timelines after FDA-aligned development plans were effectively reversed. Programs that had proceeded on a shared understanding of the evidence package were told (mid-review or even post-submission) that the standard had changed.

The clearest example was uniQure. On September 24, 2025, the company reported three-year Phase 1/2 data for AMT-130, its gene therapy for Huntington’s disease, showing a 75% slowing of disease progression, and it was preparing to file a BLA in Q1 2026. Five weeks later, those plans stalled after the FDA informed uniQure that it “no longer agrees” the Phase 1/2 data would be sufficient to support approval. Shares fell 49% in a single session.

CBER’s stated position was that uniQure would need a prospective, randomized, double-blind, sham-surgery, controlled study, requiring some patients with a fatal, progressive disease to undergo anesthesia and cranial drilling without receiving treatment purely to serve as controls. The rare-disease community responded with a mix of disbelief and outrage.

The pressure didn’t abate. In March 2026, uniQure said the FDA had formally confirmed the requirement, and the stock fell another 41%.

Now that chapter is closing. FDA Commissioner Marty Makary has confirmed that Prasad will leave the agency at the end of April 2026 and return to his academic role at UCSF. Makary credited him with four major reforms: the shift toward a single pivotal-trial standard, the Commissioner’s National Priority Voucher program, a risk-stratified COVID-19 vaccine framework, and a “plausible mechanism” pathway for ultra-rare diseases.

Markets reacted immediately. uniQure surged as much as 70% in post-market trading on news of Prasad’s exit. Replimune rose nearly 16% before Monday’s open, closing at $8.07 after touching $9.35 pre-market.

In other words, investors are already pricing in a regime change, and the April 10 PDUFA date now sits directly at the inflection point between the old regime and whatever comes next.

The New Rulebook: What Prasad Actually Changed

Before you dismiss this as inside baseball, it’s worth noting that Prasad’s tenure drove structural policy shifts that may outlast his departure. These aren’t changes that automatically unwind with a single personnel announcement.

In a landmark New England Journal of Medicine article published February 19, 2026, FDA Commissioner Marty Makary and Prasad formally declared an end to what they called the “two-trial dogma.” Going forward, the FDA’s default standard will be one strong pivotal trial supported by confirmatory evidence, rather than two independent pivotal trials, the norm dating back to the 1960s.

At first glance, this reads like a loosening of standards: more flexibility, faster timelines, and lower development costs. For certain programs, especially in rare disease and high-unmet-need oncology, that’s genuinely true. A single pivotal study can cost roughly $30 million to $150 million, and reducing the expected trial count can materially lower development spend, savings that Makary and Prasad argued could also weaken the rationale for high drug prices.

But there’s a catch that much of the coverage missed: the “one trial” model is paired with heightened scrutiny of controls, endpoints, effect size, and statistical rigor. The FDA also explicitly retained discretion to require two trials when a therapy “has a nebulous, pluripotent, or nonspecific mechanism of action, relies on surrogate or short-term outcomes, or presents underlying limitations or deficiencies”.

In plain terms: one trial may be enough, but it has to be close to immaculate. And the agency reserves wide latitude to decide what “immaculate” means, case by case, often with limited advance visibility.

RBC Capital Markets captured the practical issue in a February 18 note: where that discretion will be applied remains a central unknown, and the agency’s mixed messaging to multiple companies has reinforced regulatory unpredictability.

This framework is complemented by the Plausible Mechanism Pathway. Introduced by Makary and Prasad in a November 2025 NEJM article, it’s aimed at individualized therapies for ultra-rare genetic conditions; situations where randomized controlled trials aren’t feasible because patient populations are vanishingly small. Draft guidance was released in February 2026.

Importantly, Prasad drew a clear boundary around eligibility: uniQure’s AMT-130 for Huntington’s does not qualify, because it is a single treatment intended for the broader Huntington’s population, not an individualized therapy tailored to one patient’s unique genetic variant. In other words, the pathway is narrower than its name suggests.

Where Replimune Actually Stands

Now apply all of this to the April 10 decision with clear eyes.

The case for approval is credible. RP1 produced a 32.9% objective response rate in a population with few viable options. More than twenty oncology experts publicly signed a letter arguing the original rejection was unjustified. The resubmission includes additional data and analyses aimed directly at the issues raised in the CRL. And Prasad’s pending departure, combined with the broader “single pivotal trial plus confirmatory evidence” posture the FDA has embraced, creates a more constructive backdrop than the one that prevailed in July 2025.

But IGNYTE is still what it has always been: a Phase 1/2, single-arm study without a randomized control arm. The FDA’s CRL was explicit that the trial was “not considered to be an adequate and well-controlled clinical investigation that could provide substantial evidence to support approval”. The resubmission can address concerns through deeper analyses and supporting evidence, but it cannot retroactively randomize the study. The data are the data.

So the real question is whether the added analyses, combined with a shifting regulatory environment, are enough to move the agency from “inadequate” to “approvable”. That isn’t a scientific question with a clean, objective endpoint. It’s a judgment call made by reviewers who, at this moment, are operating amid leadership turnover and an active search for Prasad’s successor.

The level of churn has been extraordinary. CDER cycled through four directors in 2025, while CBER has had three different directors. Richard Pazdur, the longtime, widely respected oncology regulator, served less than a month as CDER head before retiring. Tracy Beth Hoeg became the fifth person to lead CDER in 2025.

In that context, the April 10 decision will be made by an institution in real transition, one where career scientific staff continue to do the work professionally, even as leadership-level policy signals shift in real time.

The Pattern Every Biotech Investor Should Memorize

Let’s use the actual numbers, as precision matters when you’re sizing risk.

uniQure’s first FDA reversal in November 2025 drove the stock from $67.69 to $23.05 at the November 3 market open, which is a decline of about 66%. The March 2026 confirmation of the sham-surgery requirement triggered another 41% drop, from $15.63 to $9.19. Same company, same program; two brutal selloffs, separated by months of investor optimism that the outcome might change.

Regenxbio received a CRL for its Hunter syndrome gene therapy, with the rejection tied to concerns about the study population, controls, and reliance on surrogate efficacy markers. The shares sold off sharply in the session that followed.

The broader pattern, now repeated across multiple companies, is consistent: a biotech aligns with the FDA on a trial design; the agency tightens or reinterprets evidentiary requirements midstream or after submission; the company scrambles to respond; and the stock makes a violent, binary move around the PDUFA date. RBC Capital Markets described the dynamic as a “regulatory tailspin”, warning it could “portend an era of decreased drug investment that would ultimately harm patients”.

This is the “72-hour rule” in practice. Under this regime, binary biotech events don’t produce 10% or 15% moves. They produce 40% to 66% collapses, or, in an approval scenario, comparably violent upside. Outcomes are fat-tailed in both directions. If you aren’t positioned for that range of outcomes before a PDUFA date, you aren’t managing risk; you’re just hoping.

What This Means for Your Portfolio and Your Framework

Here’s how to translate this into actionable thinking. Not just for Replimune, but for any biotech exposure you carry into an FDA decision.

Audit your single-arm exposure, but update the context.

The core advice still holds: identify clinical-stage programs leaning on non-randomized Phase 2 data to support a BLA. But the calculus has shifted with Prasad’s departure. The new “single pivotal trial plus confirmatory evidence” standard is formally more flexible than the historical two-trial default (even as it demands cleaner design and stronger supporting evidence). Companies with strong single-arm data in high-unmet-need indications heading into PDUFA dates under a post-Prasad FDA may be facing a more constructive backdrop than they would have six months ago.

Treat the Plausible Mechanism Pathway as narrow. Don’t extrapolate.

This pathway is intended for individualized therapies in ultra-rare diseases, and it drew criticism at launch for limited operational detail. The February 2026 draft guidance still leaves meaningful questions unanswered. If your thesis depends on a program qualifying, verify explicitly that it meets the FDA’s individualized-therapy definition. The agency has already signaled it will draw hard lines around eligibility.

Regulatory uncertainty is its own “asset class”.

The worst outcomes in binary FDA events often land on investors who buy a story without understanding the evidentiary record. That dislocation can create opportunity for investors who do the work, long or short. Knowing what was actually added in a resubmission, whether it squarely addresses the CRL’s stated deficiencies, and how the relevant review division has been behaving can materially improve your probability estimate versus consensus.

Prasad’s exit doesn’t rewind the FDA to 2023.

The reforms implemented during his tenure (the single-trial framework, the plausible mechanism concept, and the heightened emphasis on control arms and endpoint quality) are now embedded in policy. They don’t leave with him. What may change is the likelihood of midstream reversals and the degree to which individual philosophical preferences shape specific review calls. The rulebook remains; the referee may not.

The Deeper Truth

I’ve spent many years in Silicon Valley watching capital markets price regulatory risk across tech, biotech, hardware, and energy. The most consistent mistake I see retail investors make is treating regulatory outcomes as binary coin flips with unknowable odds.

They aren’t. They’re probability estimates with discoverable inputs, if you know where to look and which questions to ask.

The April 10 Replimune decision isn’t a coin flip. It’s a judgment call by FDA reviewers operating within a shifting leadership framework, evaluating a resubmission that directly responds to the agency’s stated objections, for a drug with clinically meaningful activity in a population with few alternatives, all while the central figure associated with the prior rejection has now announced his departure.

That’s not a guarantee of approval. But it is a materially different risk profile than the simple narrative of “the FDA is tightening the screws on single-arm trials” suggests. The investors positioned well into April 10 will be the ones who recognized that the regulatory ground moved beneath them twice, in rapid succession, and adjusted accordingly.

The house changed the rules. Then it changed the dealer.

That’s the real story of April 10.

If this edition gave you a framework for thinking about FDA binary risk that most retail biotech coverage is missing, forward it to one investor who needs to hear it.

Share

QUICK TAKES

→ The number to watch: REPL’s pre-market movement on April 9th — the session before the PDUFA date. Unusual volume or institutional positioning shifts the day before a binary event are often the clearest signal of where informed money is sitting.

→ The departure to track: Who replaces Prasad at CBER. One scenario discussed among analysts is Tracy Beth Hoeg, currently acting CDER director, shifting into the CBER role. The new CBER director’s personal philosophy on single-arm oncology data will define the regulatory environment for the next wave of approvals.

→ The company to watch beyond Replimune: uniQure. With Prasad departing, the sham-surgery demand for AMT-130 may face renegotiation under new CBER leadership. The 70% post-market surge on Prasad’s exit announcement suggests the market is already pricing in exactly that possibility. The Type B meeting uniQure requested for Q2 2026 could be the most consequential FDA dialogue of the year.

Silicon Valley Insights is published daily for subscribers navigating the intersection of technology, capital, and the future. All views are the author's own. Not investment or financial advice. Biotech investing carries significant binary risk; consult a financial advisor before making investment decisions.

READER Q&A: The Changing of the Guard

Q: You say Prasad’s departure doesn’t reset the clock to 2023. But if the man who demanded sham surgeries for Huntington’s patients is gone, why wouldn’t the FDA just quietly walk back the most extreme positions he took?

Because institutions don’t work that way, and because the NEJM article matters more than most people realize. When Prasad and Commissioner Makary co-authored the February 19, 2026 piece formally announcing the single pivotal trial standard, they did something strategically important: they published the policy change in the world’s most prestigious medical journal before Prasad’s departure was announced. That paper is now part of the formal scientific record. It has been cited, discussed, and incorporated into how companies are designing trials right now. You cannot un-publish it. What you can do, and what a new CBER director will likely do, is apply those standards with less aggressive interpretation at the margins. The sham surgery demand for uniQure, for instance, is the kind of extreme application that a more pragmatic successor might renegotiate quietly in a Type B meeting, without ever formally reversing the underlying policy. That is the distinction that matters: the framework stays, the most confrontational applications of it become negotiable. For investors, that means the structural shift toward higher evidentiary standards is permanent, but the specific execution of those standards just became meaningfully more uncertain, in a direction that is probably net positive for companies with strong single-arm data in high-unmet-need indications.

Q: You mention the FDA has had extraordinary leadership turnover. CDER changed directors four times in 2025, CBER three times. At what point does that level of instability actually start affecting drug review quality, not just regulatory philosophy?

This is the question I think the biotech investment community is not asking loudly enough, and it deserves a direct answer. Career FDA staff, the scientists and reviewers who actually read the data packages, are largely insulated from political leadership changes by civil service protections. The scientific review of RP1’s resubmission is being conducted by people who were there before Prasad arrived and will be there after he leaves. That continuity matters, and it is part of why I think the “chaos at the top” narrative, while real, can be overweighted in investment theses. But leadership instability has two real effects that compound over time. First, policy signals from the top become harder for companies to act on when those signals change with each new director. A biotech that aligned its trial design with the FDA in Q1 2025 based on guidance from one CDER director, then faced a different interpretation from the next, has a legitimate grievance, and that uncertainty raises the cost of drug development by forcing companies to build in regulatory optionality they shouldn’t need. Second, and more insidiously, talent retention suffers. Senior scientists and reviewers who have spent careers building institutional knowledge eventually leave for industry if the working environment becomes politically unstable. That slow drain is the long-term risk that doesn’t show up in any single PDUFA date but reshapes the FDA’s competence over a decade. I am watching the resignation and retirement announcements from career staff more carefully than the leadership headlines. That is the leading indicator that most financial media is ignoring.

Q: The oncology community with more than twenty experts signed a letter saying the original RP1 rejection was unjustified. Does physician advocacy actually move the needle at the FDA, or is it mostly noise?

Mostly noise, in the formal sense, and the FDA would be the first to say so. The agency’s review process is designed to be insulated from external advocacy precisely to prevent well-organized patient groups or physician communities from pressuring approvals that the evidence doesn’t support. There is genuine wisdom in that design. But physician letters and advocacy serve a different, more subtle function that sophisticated biotech investors understand. They create a public record of medical community consensus that is useful in three specific scenarios. First, if the FDA rejects a drug and the company decides to fight it through Congressional hearings, through media coverage, through formal dispute resolution, that letter of twenty-plus oncologists becomes ammunition. Second, it signals to the FDA’s own internal reviewers that the clinical community considers the unmet need genuine and the evidence credible, which can influence the framing of a borderline decision even if it doesn’t change the formal evidentiary calculus. Third, it creates reputational stakes for the agency. Rejecting a drug that two dozen leading oncologists publicly called clinically meaningful is not costless for an institution that cares about its scientific credibility. None of that adds up to a guaranteed approval. But it shifts the internal political economy of a borderline decision in ways that pure data analysis misses. For the RP1 resubmission specifically, the advocacy letter is a meaningful piece of the mosaic; not a trump card, but not noise either.

Q: You describe the April 10 decision as a “judgment call” by reviewers operating under a leadership vacuum. Doesn’t that actually make the outcome less predictable, not more, which would argue for simply avoiding the position entirely rather than trying to handicap it?

That is a legitimate risk management position, and I want to steelman it rather than dismiss it. If your investment framework prioritizes avoiding binary outcomes with genuinely uncertain probabilities, staying out of REPL going into April 10 is internally consistent and defensible. Not every opportunity is one you need to take. The counter-argument, and the reason I think the risk is worth understanding rather than simply avoiding, is that asymmetric binary events in biotech are where the mispricing is most concentrated. The retail investor base going into April 10 is, broadly, thinking about this in one of two modes: either “the FDA changed the rules and single-arm trials are dead” or “Prasad is leaving so approvals are coming”. Both are oversimplifications. The investor who has done the actual work, who understands exactly what the resubmission addressed, what the specific CRL objections were, what the current reviewing division’s posture is, and how Prasad’s departure changes the specific marginal decisions rather than the underlying framework, is operating with a genuinely differentiated probability estimate. That differentiation is only valuable if you act on it. The “avoid binary events” rule is intellectually safe but it leaves real alpha on the table in a sector where binary events are the mechanism through which most value is created or destroyed. My honest answer: if you are not willing to do the specific analytical work I described, the avoidance rule is the right one. If you are, the mispricing in both directions is real and worth engaging with.

Q: The newsletter focuses heavily on Replimune and oncology. But you mentioned the Plausible Mechanism Pathway for ultra-rare diseases. Is there actually an investable opportunity there, or is the patient population so small it can never generate meaningful returns?

The conventional wisdom in biotech investing has long been that ultra-rare diseases; sometimes called “ultra-orphan” indications, covering fewer than one in a million patients, are scientifically compelling but financially marginal. That wisdom is increasingly wrong, for reasons that are structural rather than cyclical. The drug pricing environment for genuine ultra-rare diseases, where you are treating twenty or fifty or two hundred patients in the entire United States, has evolved to reflect the economics of small-scale development. Prices per patient per year in the range of $500,000 to $3 million are now established precedent across multiple approved ultra-rare therapies. That is not price gouging by any reasonable definition; it is the rational pricing of a drug that required hundreds of millions in development cost to reach two hundred patients. Now layer in the Plausible Mechanism Pathway. If a company has a genuinely individualized therapy, one that is customized for a specific patient’s unique genetic variant, not a standard treatment for all patients with a given disease, this pathway potentially eliminates the need for a traditional randomized controlled trial entirely. The evidentiary bar becomes: does the mechanism of action make biological sense, and does the specific patient respond? For a gene therapy designed for a single patient’s ultra-rare mutation, that is a fundamentally different and more achievable bar than a two-thousand-patient Phase 3 trial. The investment opportunity is real, but narrow. The key filter is the individualization criterion, and I have seen several companies already attempting to re-characterize their programs as “individualized” to qualify for the pathway when they are, in fact, standard treatments for a defined patient population. That recharacterization will not survive FDA scrutiny. The genuine plays are companies that were already in the personalized medicine space before the pathway existed, where the individualization is scientific reality, not strategic framing.

Q: You advise readers going into a PDUFA date to hedge with puts if they are over-leveraged. But for a stock like REPL that is already trading below $10, options liquidity is often terrible. What does hedging actually look like in practice for small-cap biotech positions?

This is one of the most practically important questions a retail biotech investor can ask, and most newsletter writers skip it because the honest answer is complicated. You are right that options liquidity in sub-$10 biotech names is frequently inadequate: wide bid-ask spreads, thin open interest, and market makers who price in the binary risk so aggressively that buying puts is often extraordinarily expensive relative to the protection they provide. So let me give you the actual toolkit rather than the generic advice. First, position sizing is the most underrated hedge. If a single PDUFA binary event can produce a 40-66% drawdown, your position size in that stock should be calibrated so that the worst-case outcome is survivable within your overall portfolio, ideally no more than 2-3% of total portfolio value in a single high-risk binary event. No options strategy compensates for being over-concentrated in a binary outcome. Second, for names where options are available but expensive, consider collar structures rather than outright puts, selling an out-of-the-money call to partially finance the put. You cap your upside, but you meaningfully reduce the cost of protection in a scenario where put implied volatility is elevated. Third, sector ETFs with liquid options (XBI and IBB are the primary ones) can serve as imperfect but functional hedges against broad biotech selloffs that often accompany high-profile FDA rejections. If REPL gets rejected and the market reads it as a signal about the broader regulatory environment, XBI often moves sympathetically. Buying XBI puts costs a fraction of what REPL puts cost. It is not a perfect hedge, but it is a real one. Fourth, and most importantly: if the liquidity and options structure of a position make hedging genuinely impractical, that is itself a signal that the position may be oversized for your risk tolerance. The inability to hedge affordably is information, not just inconvenience.

Q: Big picture: is the FDA’s increased evidentiary rigor ultimately good or bad for patients? You present it analytically, but what is your actual view?

I’ll give you my honest view, which is more nuanced than either the “FDA is blocking miracle cures” camp or the “FDA is finally doing its job” camp wants to hear. The core of what Prasad was trying to do — ensure that drugs that reach patients actually demonstrate meaningful clinical benefit, not just impressive-sounding surrogate endpoints — is correct and important. The accelerated approval pathway had genuine problems. Drugs were approved on progression-free survival data that never translated into overall survival benefit. Patients paid extraordinary prices for drugs that extended their lives by weeks, not years, and sometimes not at all. Cleaning that up was the right instinct. Where it went wrong was the execution; specifically, the retroactive application of new evidentiary standards to companies that had aligned with the FDA on their evidence package in good faith, and the specific demand for sham surgeries in a fatal progressive disease where no alternative treatment exists. That is not rigorous science applied thoughtfully. That is rigorous science applied without adequate consideration of the ethical and practical constraints of clinical research in rare and terminal diseases. The net effect on patients is genuinely mixed. Fewer marginal drugs will reach the market, which is probably good. But some drugs with real efficacy signals, such as drugs that patients with no other options would choose to take, given the choice, will face higher development costs, longer timelines, or outright failure to reach market. Those are real patients with real needs who pay a real price for regulatory perfectionism. The honest answer is that the optimal FDA standard is not the most rigorous possible standard. It is the standard that correctly balances the risk of approving an ineffective drug against the risk of blocking an effective one. That balance is a judgment call, not a calculation. And in a democracy, reasonable people disagree about where that line should sit.

Have a question for next week's Q&A? Reply directly to this email. The best questions get published; your name attached optionally, if you want the credit.

References and Relevant Resources

1. Replimune Group: Official BLA Resubmission Announcement (Primary Source) Replimune Group, Inc. “Replimune Announces FDA Acceptance of BLA Resubmission of RP1 for the Treatment of Advanced Melanoma.” GlobeNewswire, October 20, 2025. → https://www.globenewswire.com/news-release/2025/10/20/3169221/0/en/Replimune-Announces-FDA-Acceptance-of-BLA-Resubmission-of-RP1-for-the-Treatment-of-Advanced-Melanoma.html

2. CancerNetwork: Clinical Trial and CRL Detail
→ https://www.cancernetwork.com/view/fda-accepts-resubmission-of-bla-for-rp1-nivolumab-in-advanced-melanoma

3. OncLive: BLA Resubmission Context
→ https://www.onclive.com/view/fda-accepts-bla-resubmission-for-rp1-plus-nivolumab-in-advanced-melanoma

4. Dermatology Times: Technical Characterization
→ https://www.dermatologytimes.com/view/fda-accepts-rp1-bla-resubmission-for-advanced-melanoma

5. Replimune: Q3 FY2026 Financial Results and Corporate Update
→ https://ir.replimune.com/news-releases/news-release-details/replimune-reports-fiscal-third-quarter-2026-financial-results

6. SEC Filing: Replimune Q2 FY2026 Exhibit 99.1
https://www.sec.gov/Archives/edgar/data/0001737953/000110465925107438/tm2530417d1_ex99-1.htm

7. OncoDaily: Community Response to Original Rejection
→ https://oncodaily.com/voices/replimune-398526

8. STAT News: Primary Departure Reporting (First Major Account)
→ https://www.statnews.com/2026/03/06/fda-vinay-prasad-controversial-cber-director-leaving/

9. BioPharma Dive: Departure Analysis
→ https://www.biopharmadive.com/news/vinay-prasad-fda-depart-cber-vaccines-makary/814121/

10. Fierce Biotech: Departure and CRL Context
→ https://www.fiercebiotech.com/biotech/fdas-vinay-prasad-depart-agency-end-april

11. Contagion Live: Full Tenure Timeline
→ https://www.contagionlive.com/view/fda-vinay-prasad-depart-cber-head-second-time

12. Pharmaceutical Technology: Drug Development Impact Analysis
→ https://www.pharmtech.com/view/what-does-prasad-s-fda-departure-mean-for-drug-development-

13. Prism News: Makary Four Reforms Statement (Primary X Post)
→ https://www.prismnews.com/news/fdas-cber-chief-vinay-prasad-to-leave-agency-end-of-april

14. BioSpace: Workplace Investigation → https://www.biospace.com/fda/fdas-prasad-to-depart-fda-following-tumultuous-term-as-cber-chief

15. PharmExec: Leadership Turnover Comprehensive Record
→ https://www.pharmexec.com/view/vinay-prasad-out-cber-second-time-under-year

16. European Pharmaceutical Review: CBER Leadership Timeline
→ https://www.europeanpharmaceuticalreview.com/news/272075/vinay-prasad-to-exit-cber-fda-group-director/

17. U.S. News & World Report: Consumer-Facing Summary “
→ https://www.usnews.com/news/health-news/articles/2026-03-09/fda-vaccine-chief-dr-vinay-prasad-exiting-role

18. New England Journal of Medicine: Primary Source (Behind Paywall) Prasad, Vinay, and Martin A. Makary. “One Pivotal Trial, the New Default Option for FDA Approval — Ending the Two-Trial Dogma.” N Engl J Med 2026;394:815–817. Published February 18–19, 2026. DOI: 10.1056/NEJMsb2517623 . → https://www.nejm.org/doi/full/10.1056/NEJMsb2517623

19. BioPharma Dive — NEJM Policy Analysis “FDA Leaders Say One Pivotal Trial, Not Two, Should Be ‘Default’ for Drug Approvals.” BioPharma Dive, February 19, 2026. → https://www.biopharmadive.com/news/fda-makary-prasad0one-pivotal-trial-nejm/812557/

20. Biotech Briefings / Gibson Dunn: Legal Analysis of NEJM Policy “One and Done: FDA’s New Single-Trial Default and What It Means for Drug Developers, Investors, and Counsel.” Biotech Briefings (Gibson Dunn), March 2026. → https://biotechbriefings.gibsondunn.com/one-and-done-fdas-new-single-trial-default-and-what-it-means-for-drug-developers-investors-and-counsel/

21. Clinical Trials Arena: Technical Implementation Analysis “FDA to Allow One Pivotal Trial Instead of Two for Drug Approval.” Clinical Trials Arena, February 2026. → https://www.clinicaltrialsarena.com/news/fda-to-allow-one-pivotal-trial-instead-of-two-for-drug-approval/

22. Pharmaceutical Commerce: Supply Chain Implications “FDA Shifts Toward Single Trial Requirement for New Drug Approvals.” Pharmaceutical Commerce, March 2026. → https://www.pharmaceuticalcommerce.com/view/fda-shifts-toward-single-trial-requirement-new-drug-approvals

March 21, 2026

There is a number that has been living rent-free in my head all week.

Not $553 billion, although we’ll get to that. Not $400 billion. Not even the $3 trillion that Morgan Stanley says the cloud giants will spend on data center infrastructure by 2028.

The number is $290 billion.

That is what Morgan Stanley projects Oracle’s net adjusted debt will reach by fiscal year 2028. Today it stands at roughly $108 billion. In two years, if the analysts are right, Oracle’s debt load will nearly triple, becoming one of the largest non-financial corporate debt positions in the history of American capitalism.

And Wall Street, for the most part, is cheering.

That tension between the headline growth numbers and the seismic shift happening underneath the balance sheet is what this edition is about. Because what is happening at Oracle is not just an Oracle story. It is a preview of what is happening across the entire AI infrastructure economy, and what it will mean for your portfolio, your career decisions, and your read on where Big Tech is actually headed.

Let’s go inside the numbers.

P.S. Please stay tuned for the premium Q&A session at the end of this edition, featuring insights from VCs and CEOs in Silicon Valley. We’ve invested significant time and effort into curating this content, and we hope you find it valuable and worthwhile.

Subscribe now

The Scoreboard Nobody’s Reading

Ten days ago, on March 10th, Oracle reported its fiscal Q3 2026 earnings. The headline numbers were genuinely impressive. Total revenue came in at $17.2 billion, up 22% year over year. Cloud infrastructure revenue, the OCI business at the heart of the AI buildout, grew 84% year over year to $4.9 billion. And total Remaining Performance Obligations reached $553 billion, up 325% from last year.

Read those numbers in isolation and you reach for the buy button. A company growing revenue 22% annually, with cloud infrastructure up 84%, and a contracted backlog of over half a trillion dollars? That is a technology business firing on all cylinders.

But here is what the press releases do not emphasize.

Morgan Stanley projects Oracle’s cash capital expenditures will reach $275 billion from fiscal 2026 to 2028, far exceeding the consensus estimate of $189 billion, and expects gross adjusted debt to exceed $400 billion by 2028.

Oracle already has about $95 billion in notes payable outstanding, making it the biggest corporate issuer outside the financial sector in the Bloomberg investment-grade index. By the time you layer in lease obligations, the net adjusted debt figure reaches $108 billion today. Morgan Stanley forecasts that number will more than double to roughly $290 billion by fiscal 2028, as Oracle’s corporate borrowing and data center leasing continue to accelerate.

And the market is already pricing in the risk. Credit traders are buying protection against Oracle defaulting on its debt, a trend Morgan Stanley expects to continue in the near term. The cost to insure against default over five years has been hovering near its highest levels since October 2023. The company’s 4.9% bonds maturing in 2033 widened from 26 to 83 basis points in a single session.

To be clear: Oracle is not going to default. The cash flow is real, the contracts are real, and the backlog is genuinely historic. But the signal those CDS moves are sending deserves more than a footnote. Credit markets are the adults in the room. When they get nervous, it pays to listen.

The Prisoner’s Dilemma at 40,000 Feet

To understand why Oracle (and Amazon, Microsoft, Google, and Meta) are all locked into this spending trajectory, you need to understand the game theory underneath it.

In a classic Prisoner’s Dilemma, two rational actors each face the same choice: cooperate and share the gain, or defect and try to grab it all. The cruel logic of the game is that defection is always the dominant strategy for each individual player, even when mutual cooperation would produce a better collective outcome. The result: both players defect, and both end up worse off than if they’d cooperated.

This is precisely the dynamic defining the AI infrastructure race right now.

If every hyperscaler agreed tomorrow to moderate their capex spending, say, cap it at $40 billion each per year, the entire industry would save hundreds of billions in capital costs, interest payments, and stranded assets. Margins would improve. Balance sheets would breathe. Everyone would be better off in aggregate.

But the first company to blink loses the AI race to the one that kept spending. So nobody blinks.

For fiscal year 2026 alone, the capex commitments are staggering: Google expects to spend $85 billion, Microsoft has pledged $80 billion, Amazon $100 billion, and Meta $60 to $65 billion. Oracle, at approximately $35 billion, and likely more, is actually the conservative spender in this cohort.

Morgan Stanley strategists forecast that big cloud computing companies will spend about $3 trillion on data center infrastructure projects through 2028, and cash flow can fund only about half of that, with most of the rest raised via debt markets.

Sit with that for a moment. The most profitable technology companies in human history cannot fund their own growth from their own operations. They need to borrow, at scale, at speed, from a credit market that is beginning to get visibly uncomfortable with the exposure.

That is the Prisoner’s Dilemma in concrete and fiber optic cable.

Share

The $553 Billion Promise

Read more

March 21, 2026

Silicon Valley often highlights the upside narrative: 10× productivity, agentic software, and near-infinite leverage from automation. Yet beneath that optimism sits a more difficult macroeconomic question: if AI increases output faster than it raises wages, who ultimately sustains the consumer economy? Some viral claims circulating online are clearly exaggerated; there is no credible evidence suggesting that 50% of workers will be displaced by 2027, and statements that the U.S. mortgage market is worth “$15 quadrillion” are off by many orders of magnitude.

A more grounded concern is emerging from the data. U.S. mortgage balances reached roughly $13.17 trillion by the end of 2025, according to the Federal Reserve Bank of New York. At the same time, signs of labor-market softness are increasingly interpreted through an AI lens, as firms experiment with automation and productivity tools. Economists are beginning to model scenarios in which AI-driven productivity gains accrue disproportionately to capital owners, while middle-income households experience slower wage growth and begin to moderate spending.

This dynamic would not necessarily resemble the financial crisis triggered by the 2008 Financial Crisis. A more plausible risk is a liquidity squeeze: household cash flow weakens before the broader productivity gains from AI diffuse widely through wages and employment. In that scenario, the economy may temporarily face a gap between rising productive capacity and the purchasing power needed to absorb it.

Subscribe now

Productivity may rise before incomes do

One of the most common analytical mistakes is assuming that productivity gains translate immediately and evenly into higher wages. Economic history and early evidence from the AI transition suggests the adjustment is often far more uneven.

Recent research points in that direction:

• Analysis from the Federal Reserve Bank of Dallas suggests AI technologies are simultaneously augmenting some occupations while displacing tasks in others, depending on the nature of the work.

• Studies from the European Central Bank report no broad hiring collapse across Europe so far, even as firms increasingly experiment with AI-driven productivity tools.

• Labor-market analysis from Anthropic finds limited evidence of aggregate employment damage to date, though roles with high exposure to AI show clear signs of pressure and task restructuring.

The macroeconomic risk, therefore, is not immediate mass unemployment. It is a more gradual adjustment in which firms become significantly more efficient while labor markets adapt slowly, resulting in higher output per employee, fewer entry-level openings, and slower wage growth in coordination-heavy roles.

This dynamic matters because consumer economies ultimately run on household income, not computational capacity. As Reuters reported on March 5, employment data has become a key signal for investors trying to separate AI enthusiasm from economic reality. Many economists suggest that hiring slowdowns, rather than large-scale layoffs, may be the earliest visible labor-market impact of AI adoption.

The mortgage risk is real, but often overstated

A critical correction is necessary at the outset: U.S. mortgage debt stood at approximately $13.17 trillion at the end of 2025, according to the Federal Reserve Bank of New York.

Mortgage stress has increased somewhat, but the system remains far from crisis levels. Data from the Mortgage Bankers Association shows that mortgage delinquencies rose to 4.26% in Q4 2025, with the increases concentrated largely among lower-income households. Overall delinquency rates remain historically moderate. At the same time, total U.S. household debt reached roughly $18.8 trillion, which suggests the system is not collapsing, but neither is it immune to pressure.

If an AI-driven macro shock were to occur, it would likely look different from the leverage-driven collapse associated with the 2008 Financial Crisis. Rather than starting with housing or credit markets, the initial pressure would more plausibly emerge from income compression: slower wage growth and reduced job mobility among debt-carrying, college-educated households. That dynamic could lead to reduced consumption, rising credit-card strain, and eventually localized housing stress. In this scenario, mortgage instability would be a lagging indicator, not the initial trigger.

The laptop class squeeze

White-collar households play a central role in modern consumer economies. They drive housing demand, discretionary spending, subscription-based services, and significant portions of local tax revenue. If AI systems first automate repetitive coordination work, such as analysis, customer support, drafting, triage, and junior-level review, the effects may appear gradually but meaningfully:

• fewer entry-level and junior roles

• weaker bargaining power among early-career professionals

• slower income growth in households that have been sustaining post-pandemic consumption

Research from institutions including the Organisation for Economic Co-operation and Development and the International Monetary Fund highlights that advanced economies may be particularly exposed, precisely because a large share of employment lies within cognitively intensive service work.

The implication is not immediate mass layoffs. Rather, the concern is the emergence of macro feedback loops, including slower hiring, weaker wage growth, and reduced consumption, that can develop quietly before any dramatic employment shocks appear in headline statistics.

The constructive bubble thesis

What is unfolding in 2026 increasingly resembles an infrastructure mobilization cycle, rather than a classic consumer-software boom.

Several signals point in that direction:

• Research and commentary associated with Bridgewater Associates estimate that major technology companies’ AI-related capital expenditures could approach $650 billion, up from roughly $410 billion in 2025, reflecting a rapid escalation in investment.

• Analysis connected with Stanford Institute for Human-Centered Artificial Intelligence has warned that the scale of global data-center investment may be approaching bubble-like territory, as firms race to secure compute capacity.

• The chief economist of Goldman Sachs has also noted that AI contributed “basically zero” to measurable U.S. economic growth in 2025, highlighting that most of the spending so far is infrastructure investment rather than immediate productivity gains.

This dynamic is often described as a “constructive bubble”. Much of the investment may ultimately prove mistimed or excessive, but it is still building the foundational infrastructure: power systems, cooling capacity, advanced chips, networking, and large-scale inference platforms.

Historically, the benefits of such investment tend to arrive out of sequence. The pattern often looks like this:

capex boom → labor-market pressure → productivity dividend

If that pattern holds, the most visible productivity gains from today’s AI infrastructure buildout may not appear until the later part of the decade, potentially closer to the late 2020s.

The 2028 sequence to watch

A plausible economic sequence over the next several years could unfold gradually rather than dramatically:

• AI reduces the need for certain coordination-heavy white-collar tasks faster than new roles are created.

• Hiring slows and wage growth softens in occupations most exposed to automation.

• Household cash flow begins to weaken, particularly among debt-heavy consumers.

• Consumption growth stalls, even as equity holders benefit from productivity gains tied to AI.

• Credit stress broadens, potentially moving from subprime segments toward more mainstream consumer and housing markets.

Economists at Moody’s Analytics have modeled variations of this dynamic. Their analyses suggest that AI-driven productivity gains could disproportionately benefit asset owners, while middle-income households with limited financial assets reduce spending as income growth slows.

The key insight is that these two realities can coexist: AI may increase aggregate productivity while household liquidity becomes more fragile at the same time.

Survival rules: own the pipes, prioritize resilience

For investors:
The strategic focus may gradually shift from consumer-facing applications toward the infrastructure layers that every AI system depends on. These include power generation and grid upgrades, cooling systems, networking capacity, data-center real estate, and large-scale inference platforms. Capital-market reporting from outlets such as Reuters has noted growing investor interest in these areas because they offer more predictable cash flows tied to the expanding AI infrastructure buildout.

For households:
Resilience may increasingly come from reducing reliance on a single income source. As agentic AI makes certain forms of labor income less predictable, financial stability is more likely to come from diversification and flexibility, including:

• Developing multiple income streams rather than relying on one salary

• Keeping fixed financial obligations manageable

• Maintaining adequate liquid reserves

• Considering selective exposure to long-term infrastructure trends where appropriate

The broader principle is straightforward: in periods of technological transition, those who depend solely on one channel of income tend to face greater volatility than those with diversified participation in the evolving economy.

Bottom line

The more alarming narrative “AI will trigger another crisis like 2008” can distract from the more plausible scenario. A more realistic risk is not a sudden collapse, but a gradual liquidity squeeze in which productivity rises while household income adjusts more slowly. In that environment, the winners will not simply be those who are most productive, but those who are most resilient and adaptable to shifting economic conditions.

Actionable step

Consider stress-testing your personal finances by asking a practical question: What would my budget look like if my labor income were 10% lower? Planning around that scenario can help reveal vulnerabilities before they become urgent.

At the same time, it may be wise to develop at least one income stream or asset that does not depend directly on the number of hours you work. As AI systems increasingly reward scalable systems and infrastructure, long-term resilience may come less from individual effort alone and more from participation in systems that can continue generating value independently of your time.

References and Relevant Resources

1. New York Fed – Household Debt & Mortgage Balances

2. Mortgage Bankers Association – Delinquency Rates

3. Anthropic – Labor Market Impacts

4. Dallas Fed – AI Augmentation vs. Displacement

5. Reuters – Jobs Data & AI Lens

6. Bridgewater – AI Capex Projections

7. Stanford HAI – Data Center Bubble Risk

8. Goldman Sachs – 2025 AI Productivity Contribution

9. Moody's Analytics – AI Distributional Effects

10. OECD/IMF – Advanced Economy Exposure

March 20, 2026

There’s a phrase we keep coming back to this week, one Trump said out loud at his State of the Union address last month that should have stopped every founder, investor, and operator in this industry cold.

“We have an old grid. It could never handle the kind of numbers — the amount of electricity — that’s needed. So I’m telling them they can build their own plant; they’re going to produce their own electricity.”

That sentence, delivered to a joint session of Congress on February 24th, 2026, was not a policy proposal. It was a reckoning. And on March 4th, just eight days later, Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI all sat down at the White House and signed what is now officially called the Ratepayer Protection Pledge: a voluntary commitment to “build, bring, or buy” every watt of power their AI data centers will ever consume.

Under the pledge, these companies agreed to build, bring, or buy the new generation resources and electricity needed to satisfy their new energy demands, paying the full cost of those resources, whether by building or buying from new power plants, and whether they use the electricity or not.

Let that sink in. The seven largest technology and AI companies in the world just committed to becoming, functionally, power utilities.

This is the story of the decade. And most people are still reading the chip headlines.

Subscribe now

The Physical Wall the AI Race Just Hit

We need to talk about a number: $329.17.

In the service region for PJM Interconnection, the country’s largest grid operator, capacity prices have exploded, rising to $329.17 per megawatt-day for the 2026-2027 period, up from just $28.92 in 2024-2025.

That is not a typo. That is an eleven-fold increase in two years. PJM manages the grid for roughly 65 million people across 13 states, from the Mid-Atlantic through the Midwest. It is, quite literally, the beating electrical heart of the American economy. And it is breaking.

Ratepayers in Mid-Atlantic states including New Jersey, Pennsylvania, Maryland, and Virginia have already seen sharp price spikes driven in part by the world’s largest cluster of data centers in Virginia.

The AI race, which everyone assumed was a software war, has collided, violently, with nineteenth-century infrastructure. The United States electrical grid was not built for this. The transformers are aging. The transmission lines are at capacity. The permitting process to build new generation assets takes, in many cases, longer than a decade.

According to estimates from the Lawrence Berkeley National Laboratory, power demand from U.S. data centers doubled between 2018 and 2024 and could triple by 2028.

We are not talking about a gradual transition. We are talking about an exponential demand curve slamming into a linear supply constraint. And the companies at the tip of the AI spear, the ones with the most to lose, have now been handed the bill.

The Scale of What’s Being Built

Let’s do some scorekeeping, as the numbers are almost incomprehensible.

Oracle’s founder Larry Ellison announced plans to build a gigawatt-scale data center powered by three nearby small modular nuclear reactors, with building permits for those SMRs already secured. One gigawatt. That is roughly the output of a mid-sized conventional nuclear plant, enough to power a city of 700,000 homes, dedicated entirely to one company’s AI compute infrastructure.

Oracle is not alone. Amazon is investing $20 billion-plus, converting the Susquehanna nuclear site into an AI campus. Microsoft committed $16 billion to a 20-year deal to restart Three Mile Island, targeting 2028 operation. Google and Kairos Power signed what appears to be the first corporate agreement to develop a fleet of small modular reactors in the United States, covering up to 500 megawatts across six to seven reactors coming online through 2035.

And then there is Meta. Mark Zuckerberg announced in July 2025 that Meta is building a 5-gigawatt AI data center. Five gigawatts. That is more than the entire generating capacity of some small nations.

According to an analysis by Cleanview, at least 46 data center sites across the United States have planned or deployed behind-the-meter generation with a combined capacity of 56 gigawatts. This represents roughly 30% of all planned U.S. data center capacity.

We are watching, in real time, the largest private infrastructure buildout since the interstate highway system. Except this time, it’s not the government doing the building. It’s the tech companies, because the government quite literally told them they have no other choice.

Share

The Winner’s Curse in Concrete and Steel

In economics, the Winner’s Curse describes a phenomenon in auctions where the winner consistently overpays, as winning requires out-bidding everyone else, and everyone else was presumably rational. Applied here, it works like this:

To maintain their lead in AI, the hyperscalers must train larger models. Larger models require more compute. More compute requires more power. More power requires building that power infrastructure from scratch; in concrete, steel, copper, and eventually enriched uranium.

Every dollar they spend wiring a turbine to a data center is a dollar not spent on researchers, on model architecture innovation, on the next paradigm-shifting algorithm. The companies most likely to win the AI race are now the most exposed to the crushing capital requirements of industrial power development. The timeline for small modular reactors to deliver commercial power is measured in years to a decade or more, not months. Realistic first U.S. SMRs serving commercial technology loads are expected around 2030, and these will be pilot deployments. Widespread use of SMR-powered data centers is unlikely before the mid-2030s.

In the interim? They are burning natural gas. Of the confirmed behind-the-meter generation equipment being deployed today, approximately 75% runs on natural gas, because it delivers the three attributes data centers prize above all others: dispatchability, reliability, and speed to build.

The irony is almost elegant. The most sophisticated AI companies on earth, companies that have spent the last decade telling us software would eat the world, are now being forced to build gas plants and negotiate turbine contracts. They are, quite suddenly, in the energy business. And the energy business is unforgiving.

The Sovereign Signal: Fortress America Goes Electric

What makes this shift structurally permanent, not just a capital cycle, is the national security dimension.

The Trump Administration’s AI Action Plan explicitly states that “America’s environmental permitting system and other regulations make it almost impossible to build digital infrastructure in the United States with the speed that is required”, and that this infrastructure “must not be built with any adversarial technology that could undermine U.S. AI dominance.”

Energy independence for AI data centers has been formally elevated to a matter of national security. The Administration’s message is unambiguous: the country that controls the compute controls the future, and the country that controls the power controls the compute.

The DOE has selected four federal sites, the Idaho National Laboratory, Oak Ridge Reservation, Paducah Gaseous Diffusion Plant, and Savannah River Site, for public-private projects to develop AI data center and energy generation infrastructure. These are not suburban office parks. These are former nuclear and weapons production facilities. The symbolism is intentional.

This is the “Fortress America” strategy made physical: a grid that serves not just commerce, but computation as a form of strategic national power. If your data center depends on the public utility, you depend on a grid that can be disrupted, congested, or politically constrained. If you own the generation asset, you own the strategic resource.

Washington is making a bet: the winner of the AI race will be whoever can guarantee uninterrupted, gigawatt-scale power, and that whoever should be American.

What This Means for Your Portfolio, Your Career, and Your Decisions

We’ve been doing this long enough in this valley to know the difference between a trend and a transformation. This is a transformation. Here is where I see the money moving:

The real shovels in this gold rush are not chips; they are copper, transformers, and turbines.

Nvidia gets the headlines. But Nvidia’s GPUs are useless without the physical substrate to run them. The bottleneck today is not semiconductors; it is power delivery infrastructure. The companies supplying high-voltage transformers, grid-scale copper wiring, natural gas turbines, and the engineering talent to connect all of it are quietly becoming critical infrastructure for the entire AI economy.

The Trump administration’s executive orders explicitly identify qualifying energy infrastructure as including high-voltage transmission lines, natural gas turbines, coal and nuclear power equipment, substations, and switchyards, treating them as components as strategically important as the data centers themselves.

The SMR story is real, but it is a 2030s story. Don’t let the Oracle and Google announcements fool you into thinking nuclear power is around the corner for data centers. The investments are genuine, the regulatory momentum is genuine, and the long-term thesis is strong. But the near-term plays are in natural gas peakers, grid interconnection services, and the industrial supply chain feeding the buildout happening right now.

Location is the new moat. For founders building infrastructure-adjacent businesses, proximity to reliable power has become as important as proximity to talent. Virginia’s famed Data Center Alley is at capacity. Texas, Pennsylvania, and the Southeast are where the next wave is landing, and they’re landing there not because of tax incentives, but because there are megawatts available.

Watch for the regulatory arbitrage window. One Harvard Law expert noted bluntly that the Ratepayer Protection Pledge has no teeth; the White House has the bully pulpit but no legal authority to impose new rate structures. This pledge is voluntary. The enforcement mechanism is reputational and political, not legal. Smart infrastructure developers will move aggressively in this window before binding regulation sets in, because binding regulation is coming.

The Bigger Picture

We want to close with something that rarely shows up in the breathless AI coverage, because we think it is the most important thing we can tell you.

Marc Andreessen gave us “Software is eating the world” in 2011. It was right. It reshaped everything. But what we are watching now is a second-order consequence that the original thesis didn’t anticipate: AI is eating the grid. And the grid cannot be refactored overnight with a hotfix and a GitHub commit.

The physical constraints are real. The capital requirements are staggering. The timelines are measured in presidential administrations, not product cycles. And the companies that assumed computation was effectively free, or at least, someone else’s problem, are now learning that the atoms were there all along, waiting patiently behind the bits.

The irony of the Ratepayer Protection Pledge is that it reads, on the surface, as a story about utility bills. It’s not. It’s a story about what happens when a civilization tries to accelerate faster than its infrastructure can support, and who gets handed the bill when it can’t keep up.

Amazon and Oracle are not becoming power utilities because they want to. They are becoming power utilities because the alternative is losing the AI race to a competitor who figured out power first.

That is the gridlock.

QUICK TAKES

→ The number to watch: PJM capacity prices, 2027-2028 delivery. If they hold above $200/megawatt-day, the behind-the-meter buildout accelerates further. If they correct, the Pledge may lose urgency.

→ The company to watch: Vistra Corp, Constellation Energy, and NuScale Power — the picks-and-shovels plays in the nuclear and gas buildout nobody’s talking about at dinner parties yet.

→ The question nobody’s asking: What happens to cloud pricing when the cost of a GPU-hour now includes the amortized cost of a gas turbine? The hyperscaler cost structure is about to get a lot more complicated.

READER Q&A

Q: You say the Ratepayer Protection Pledge “has no teeth”. If it’s voluntary, why would any company actually follow through on commitments that cost them tens of billions of dollars?

Great question, and the most important one to ask. The short answer: the pledge doesn’t need legal teeth to have real teeth. These seven companies, Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI, are all deeply embedded in federal contracts, regulatory approvals, and permitting processes that the current administration controls. You don’t need a law when you control the permit. The companies signing this pledge are not doing so out of civic virtue. They are doing so because the cost of not signing (in regulatory friction, in permitting delays, in losing access to those four federal DOE sites) is higher than the cost of signing. Welcome to how Washington actually works.

Q: If natural gas is 75% of the behind-the-meter solution today, aren’t these companies just blowing up their ESG commitments and climate pledges?

Yes. Bluntly: yes. And what’s fascinating is how quietly that is being walked back in real time. Microsoft already disclosed in 2024 that its carbon emissions had risen nearly 30% since 2020, directly because of data center expansion. Google’s emissions rose roughly 48% over the same five-year period. These companies made net-zero pledges when they assumed renewable energy could scale fast enough to match their compute growth. That assumption was wrong. The choice they are now making, implicitly, without press releases, is: AI leadership now, carbon accounting later. Whether that trade-off is defensible is a separate debate. But anyone who tells you the hyperscalers are on track with their 2030 climate commitments is either not reading the power filings or not telling you the truth.

Q: You mentioned SMRs are a “2030s story”. But Oracle already has building permits for three SMRs. Doesn’t that mean they’re further along than you’re suggesting?

Permits and power are not the same thing. Having a building permit is roughly equivalent to having a gym membership; it signals intention, not results. The NRC licensing process for a new nuclear reactor design remains extraordinarily complex. The two SMR developers furthest along in the U.S., NuScale and Kairos Power, have faced repeated delays, cost overruns, and, in NuScale’s case, the cancellation of their first commercial project in Idaho in late 2023 due to cost escalation. The technology is real. The regulatory momentum is improving. The financing structures are getting creative. But “permits secured” for SMRs in 2026 most realistically means “power delivered” no earlier than 2030-2032, and that is an optimistic read. Plan for gas in the interim. Hope for nuclear by mid-decade.

Q: What does this mean for smaller cloud companies: the Tier 2 and Tier 3 providers? Do they get crushed?

This is where it gets interesting, and where I think the real competitive disruption lives. The hyperscalers, Amazon, Microsoft, Google, can absorb the capital cost of building their own generation because they have the balance sheets and the political leverage to do it. A mid-market cloud provider cannot build a gas plant. They will buy power on the open market, at increasingly brutal prices. That creates a structural cost disadvantage that compounds over time. You will see consolidation in the cloud infrastructure space over the next five years driven not by technology, but by energy economics. The companies that own their power will lower their marginal cost of compute. The companies that don’t will get squeezed on margin until they are acquired or exit the market. If you’re evaluating cloud infrastructure investments right now, the first question I ask is: what’s their power strategy?

Q: Is there an investment angle here for regular people, not just institutional investors or VCs?

Yes, though with appropriate caveats. I am not your financial advisor, and nothing here is investment advice. But I will tell you what I’m watching publicly, as a framework for your own research. The energy infrastructure supply chain is where I see the most underpriced opportunity relative to the AI narrative. Specifically: transformer manufacturers (there is a genuine global shortage of large power transformers; lead times have stretched to three years in some cases), copper producers (electrification of everything, including data centers, is a structural tailwind for copper demand through the decade), and the two or three credible SMR developers that survive the current shakeout period. Constellation Energy and Vistra Corp are the nuclear operators most directly positioned for the data center power demand wave. Neither one is a household name at a dinner party in Palo Alto. That’s usually when I get more interested, not less.

Q: You’re describing Big Tech essentially becoming industrial conglomerates. Does that make them more or less attractive as long-term investments?

More complicated, which in the short term the market reads as less attractive, and in the long term may prove to be a profound moat. Here is the tension: capital-intensive infrastructure businesses trade at lower multiples than software businesses, for good reason; the returns are lower and the risks are harder. If Wall Street decides to value Microsoft or Amazon on a blended software/utility multiple, there is multiple compression risk. But the flip side is that anyone who successfully builds their own gigawatt-scale power infrastructure has constructed a barrier to entry that no startup can replicate in five years. You cannot raise a Series A and build a gas plant. The capex moat being built right now is genuinely durable in a way that software advantages rarely are. I hold both thoughts simultaneously. The next two years of earnings calls will be messy. The ten-year competitive position of whoever gets this right is formidable.

Q: The newsletter focuses on the U.S., but what are China and Europe doing on this front? Are we actually winning?

China is building data center and power infrastructure at a scale that makes the U.S. announcements look cautious. State coordination means they don’t have permitting delays, community opposition, or multi-year interconnection queues. The honest answer is: China doesn’t have the AI model quality advantage today, but they have fewer friction points in the physical buildout. That gap could narrow faster than the West is comfortable admitting. Europe, meanwhile, is caught between its Green Deal commitments, high industrial electricity prices, and a genuine desire to be a player in AI, and is currently losing on all three fronts simultaneously. The geopolitical subtext of the Ratepayer Protection Pledge is real: Washington is trying to ensure the U.S. wins the infrastructure race before the capability race closes. Whether it works depends on whether a voluntary pledge and federal site access can actually accelerate a buildout that normally takes a decade. That’s the open question I’ll be watching all year.

Have a question for next week’s Q&A? Reply directly to this email. The best questions get answered publicly; your name included optionally, if you want the credit.

References and Relevant Resources

I. THE RATEPAYER PROTECTION PLEDGE

1. White House Fact Sheet — Primary Source The White House. “Fact Sheet: President Donald J. Trump Advances Energy Affordability with the Ratepayer Protection Pledge.” March 4, 2026. → https://www.whitehouse.gov/fact-sheets/2026/03/fact-sheet-president-donald-j-trump-advances-energy-affordability-with-the-ratepayer-protection-pledge/

2. White House Official Statement The White House. “President Trump Secures Historic Commitment to Keep Electricity Costs Down Amid Data Center Boom.” March 4, 2026. → https://www.whitehouse.gov/articles/2026/03/president-trump-secures-historic-commitment-to-keep-electricity-costs-down-amid-data-center-boom/

3. Fox News Digital — Advance Reporting Rogers, Taylor (White House spokeswoman), quoted in Fox News Digital. “Scoop: Trump Brings Big Tech to White House to Curb Power Costs Amid AI Boom.” February 26, 2026. → https://www.foxnews.com/politics/scoop-trump-brings-big-tech-white-house-curb-power-costs-amid-ai-boom

4. Data Center Dynamics — Industry Analysis “Big Tech Makes Ratepayer Protection Pledge for Data Centers at White House.” Data Center Dynamics, March 4, 2026. → https://www.datacenterdynamics.com/en/news/big-tech-makes-ratepayer-protection-pledge-at-white-house/

5. Perkins Coie / JD Supra — Legal Analysis “White House and Leading AI Companies Commit to Ratepayer Protection.” JD Supra, March 2026. → https://www.jdsupra.com/legalnews/white-house-and-leading-ai-companies-2597283/

6. CNN / KTVZ — Signing Ceremony Coverage “Trump, Tech Companies Sign Voluntary Pledge to Protect Consumers from Rising Electricity Costs.” CNN Wire, republished KTVZ, March 4, 2026. → https://ktvz.com/politics/cnn-us-politics/2026/03/04/trump-tech-companies-sign-voluntary-pledge-to-protect-consumers-from-rising-electricity-costs/

II. PJM CAPACITY PRICES & GRID STRAIN

7. PJM Interconnection — Official Auction Results PJM Inside Lines. “PJM Auction Procures 134,311 MW of Generation Resources; Supply Responds to Price Signal.” July 22, 2025. → https://insidelines.pjm.com/pjm-auction-procures-134311-mw-of-generation-resources-supply-responds-to-price-signal/

8. IEEFA — Rate Impact Analysis Institute for Energy Economics and Financial Analysis. “Projected Data Center Growth Spurs PJM Capacity Prices by Factor of 10.” → https://ieefa.org/resources/projected-data-center-growth-spurs-pjm-capacity-prices-factor-10

9. Utility Dive — Industry Trade Coverage Howland, Ethan. “PJM Capacity Prices Set Another Record with 22% Jump.” Utility Dive, July 22, 2025. → https://www.utilitydive.com/news/pjm-interconnection-capacity-auction-prices/753798/

10. Gridshopper — Consumer-Facing Analysis “PJM Capacity Costs Drive Major Utility Rate Increases.” Gridshopper, December 2025. → https://gridshopper.com/blog/pjm-capacity-auction

11. Citizens Utility Board — Consumer Advocacy Statement “PJM Capacity Auction Leads to Record Price Spike for Second Straight Year, Threatens Even Higher ComEd Bills in 2026-27.” Citizens Utility Board, July 25, 2025. → https://www.citizensutilityboard.org/blog/2025/07/22/pjm-capacity-auction-leads-to-price-spike-for-second-straight-year-threatens-even-higher-com-ed-bills-in-2026-27/

III. ORACLE, AMAZON, MICROSOFT, GOOGLE — NUCLEAR & POWER DEALS

12. CNBC — Oracle Primary Report “Oracle Is Designing a Data Center That Would Be Powered by Three Small Nuclear Reactors.” CNBC, September 10, 2024. → https://www.cnbc.com/2024/09/10/oracle-is-designing-a-data-center-that-would-be-powered-by-three-small-nuclear-reactors.html

13. Power Engineering — Technical Analysis “Oracle Designing Data Center to Be Powered by Trio of Small Modular Reactors.” Power Engineering, September 11, 2024. → https://www.power-eng.com/nuclear/smrs/oracle-designing-data-center-to-be-powered-by-trio-of-small-modular-reactors/

14. The Register — Technical Skepticism “Oracle to Power 1GW Datacenter with Trio of Nuclear Reactors.” The Register, September 12, 2024. → https://www.theregister.com/2024/09/11/oracle_1gw_datacenter_smr_plan/

IV. LAWRENCE BERKELEY NATIONAL LABORATORY — DATA CENTER ENERGY

15. DOE / LBNL — Primary Report Shehabi, Arman et al. “2024 United States Data Center Energy Usage Report.” Lawrence Berkeley National Laboratory, December 20, 2024. LBNL-2001637. → https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-in-electricity-demand-from-data-centers → Full report: https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf

V. META 5 GW, AMAZON SUSQUEHANNA, MICROSOFT THREE MILE ISLAND, GOOGLE KAIROS

16. On Meta’s 5 GW data center announcement (July 2025): See: Meta CEO Mark Zuckerberg’s announcement, widely covered in Reuters, Bloomberg, and The Verge, July 2025. The 5 GW facility is the largest single corporate data center project announced to date.

17. On Amazon’s Susquehanna nuclear campus ($20B+): See: AWS/Amazon capital investment filings and reporting in The Wall Street Journal and Bloomberg, 2024–2025. The Susquehanna Steam Electric Station campus agreement with Talen Energy was reported beginning in March 2024, though FERC initially blocked the direct grid connection; deal structure was subsequently revised.

18. On Microsoft’s Three Mile Island restart ($16B, 20-year deal): See: Microsoft-Constellation Energy Power Purchase Agreement, announced September 2023, targeting 2028 restart. Reported by Reuters, The Washington Post, and The New York Times.

19. On Google-Kairos Power SMR fleet (up to 500 MW, 6–7 reactors through 2035): See: Google-Kairos Power agreement, announced October 2023. Reported by Reuters and Bloomberg as the first corporate SMR fleet agreement in the United States.

VI. AI ACTION PLAN & FEDERAL ENERGY SITES

20. White House AI Action Plan The White House. “Winning the Race: America’s AI Action Plan.” July 23, 2025. → https://www.whitehouse.gov/articles/2025/07/winning-the-race-americas-ai-action-plan/

VII. EMISSIONS DATA — MICROSOFT & GOOGLE

22. Microsoft Carbon Emissions Disclosure Microsoft. 2024 Environmental Sustainability Report. Released 2024. Emissions rose approximately 29% between 2020 and 2023, directly attributed to data center expansion. Reported by BBC, The Guardian, and Bloomberg.

23. Google Carbon Emissions Disclosure Google. 2024 Environmental Report. Google’s greenhouse gas emissions rose approximately 48% over five years (2019–2023), driven by data center energy consumption. Reported by Reuters and CNN.

VIII. NUSCALE PROJECT CANCELLATION

24. NuScale Carbon Free Power Project Cancellation Multiple sources including Reuters, The Wall Street Journal, and NPR. November 2023. Source for: NuScale’s first U.S. commercial SMR project (Carbon Free Power Project in Idaho) cancelled due to cost escalation; providing context for SMR deployment timeline realism.

Source links are provided for subscriber reference and independent verification. Where specific reporting dates are noted, they reflect the date of original publication or announcement.

Silicon Valley Signal maintains a strict fact-checking policy. Corrections to any prior edition are published prominently in the following issue. Readers who identify factual errors are encouraged to reply directly to this newsletter.

March 19, 2026

The correct number from Oracle’s Q3 FY2026 earnings release, dated March 10, 2026, is $553 billion in RPO: up 325% year-over-year and up $29 billion from the prior quarter. The $533B figure that appeared in some early coverage was wrong. $553B is the confirmed, official number.

The RPO figure is the single most important number in Oracle’s quarter, and one of the most important numbers in enterprise tech right now. It is not revenue. It is contracted future revenue: work Oracle has already been hired to do and will recognize over time. Most of the Q3 increase came from large-scale AI contracts structured so that customers either prepay Oracle to purchase GPUs, or supply the GPUs directly to Oracle. This structure reduces the financing risk Oracle carries on those contracts.

The pattern that actually matters for your career, your portfolio, and your vendor strategy is confirmed and visible across the full earnings release.

Subscribe now

The Oracle Paradox

Oracle just posted the strongest organic growth it has seen in 15 years. It is simultaneously executing one of the most aggressive restructurings in its modern history. Both facts are true. Neither contradicts the other.

This was the first quarter in over 15 years where Oracle’s organic total revenue and non-GAAP earnings per share both grew at 20% or more simultaneously. Total revenue came in at $17.2 billion, up 22%. Cloud revenue hit $8.9 billion, up 44%. OCI, Oracle’s cloud infrastructure business, grew 84%. AI infrastructure revenue grew 243% year-over-year. Multicloud database revenue grew 531%.

Those are not the numbers of a company in trouble. They are the numbers of a company in transformation, which is a different thing, and in some ways a harder thing to navigate.

At the same time, Oracle has raised its 2026 restructuring budget to $2.1 billion, with Bloomberg and Reuters both reporting layoffs numbering in the thousands, and TD Cowen estimating that between 20,000 and 30,000 roles may ultimately be affected; roughly 10–20% of Oracle’s global workforce of approximately 160,000. Oracle has not confirmed a final number. The cuts are targeted: legacy software roles, overlapping support functions, and middle management in sales and marketing. AI, cloud, and core enterprise engineering teams are, by most accounts, relatively insulated.

Record growth. Radical cuts. Not a contradiction, a strategy. Here is how to read it.

1. OCI is surging, and that growth has an uncomfortable gross margin

The number Oracle did not lead with in its earnings call is the one that explains everything else: gross margin on AI capacity delivered in Q3 was 32%.

For context: Oracle’s traditional software licensing business runs at gross margins above 80%. SaaS businesses typically target 70–80%. AI infrastructure delivered at 32% gross margin is not a software margin. It is closer to an industrial or utility margin; the economics of a company that builds and operates physical infrastructure at scale.

That number reframes the entire quarter. OCI’s 84% revenue growth is real and significant. But each incremental dollar of that revenue requires GPU supply that is expensive and scarce, custom networking, high-performance storage, data center buildouts, and power procurement; the sleeper constraint of this decade. Oracle’s capex guidance for FY2026 stands at $50 billion, up from $21.2 billion in FY2025; itself triple the $6.9 billion spent in FY2024. The company carries over $108 billion in total debt, and banks have grown cautious about data-center financing, which is one reason Oracle has been raising capital through bonds and equity instruments rather than relying on conventional lending.

When your revenue model starts looking industrial, your cost structure follows. Fewer generalists. Fewer middle layers. More execution-heavy roles tied directly to infrastructure. Less tolerance for product lines that cannot justify their compute footprint.

That is not a cultural shift. It is a financial equation.

Share

2. The revenue picture: strong, but understand what is driving it

Cloud revenue now accounts for roughly 52% of Oracle’s total sales, up from approximately 43% a year ago. The RPO pipeline of $553 billion signals the scale of long-term demand already locked in. That is genuinely impressive.

But it is worth understanding the composition of that RPO. A significant portion is concentrated in large-scale AI contracts; a handful of hyperscale customers driving enormous commitments. Concentrated AI revenue behaves differently from diversified enterprise revenue. Traditional enterprise software renews slowly and predictably, with high switching costs and limited need for constant reinvestment. AI hyperscaler-style revenue involves large, lumpy commitments that are sensitive to GPU pricing, power constraints, and model economics, and that can shift if a major customer changes architecture or builds capacity in-house.

A company can beat quarterly estimates because one or two whales expanded, while simultaneously preparing for the day those whales demand renegotiation. That preparation looks like restructuring. The RPO’s customer-supplied GPU provision is a smart structural hedge against that risk, but it does not eliminate the concentration dynamic.

3. The workforce erasure: AI as a budget instrument

In 2026, layoffs in enterprise tech are not belt-tightening. They are a reallocation strategy. Oracle’s cuts are expected to primarily impact legacy software roles and overlapping support functions. AI can automate repeatable tasks, but it does not replace judgment, accountability, or resilience planning. Experienced engineers who understand how systems interact, where dependencies exist, and how to troubleshoot complex failures represent institutional knowledge that is much harder to replace than it is to lose.

The internal status hierarchy is being redrawn. The center of gravity is moving away from on-premises maintenance, traditional middleware, incremental enterprise feature releases, and long-cycle sales support, toward cloud reliability, security and compliance automation, GPU orchestration, inference optimization, and enterprise AI deployment tooling.

The people who thrive are not necessarily the most technically brilliant. They are the people closest to infrastructure, cost efficiency, security, and revenue; the roles that AI spending depends on rather than displaces.

4. Zero-sum growth: how to read every AI-era earnings headline

This is the thesis that matters beyond Oracle.

The AI boom is not a broad-based rising tide for enterprise software. It is a reallocation, sometimes a violent one. If Oracle succeeds in its transformation, other legacy enterprise software companies will almost certainly follow the same playbook: borrow heavily, build AI infrastructure, and cut the human workforce to fund the transition.

For every dollar flowing toward GPUs, AI cloud contracts, and inference workloads, there is often a dollar leaving ERP add-ons, legacy middleware, traditional IT consulting, and the long-tail SaaS tools that built the last decade of enterprise software revenue.

So the winners report growth while parts of the ecosystem shrink around them.

That is zero-sum growth. And it changes how you read headlines. When you see record profits alongside layoffs, the right questions are not “are they lying about the growth?” The right questions are: which part of the company is being liquidated to fund the new engine? How concentrated is the demand behind the numbers? How much capex is required to keep the story alive? And what does the gross margin on that new revenue actually look like?

5. The metric that matters more than EPS

Stop obsessing over earnings beats. Start watching capex intensity.

Oracle plans to raise up to $50 billion through debt and equity financing to fund its AI data center expansion, and has indicated it does not anticipate issuing further bonds in calendar 2026 beyond that amount. That forward commitment tells you more about Oracle’s strategic confidence and its financial exposure than any single quarter’s earnings-per-share figure.

The metrics worth tracking across any AI infrastructure company:

Capex as a percentage of revenue. Free cash flow trend versus operating income. Data center commitments and power contracts. RPO quality: duration, concentration, and cancellation risk. Gross margin trajectory specifically in cloud infrastructure. And headcount trend versus revenue growth, because when those diverge sharply in opposite directions, you are watching the zero-sum reallocation in real time.

The companies pulling off this transition will look like hybrids: part software firm, part utility, part industrial builder. AI infrastructure gross margins in the 30s, combined with SaaS margins in the 70s, produce a blended economics profile that existing valuation frameworks were not built to handle cleanly. Inkl That creates analytical opportunity for investors willing to do the work beneath the headline numbers.

What this means for your career

If you work in enterprise software, the honest exercise is to assess which side of the reallocation your role sits on.

Higher risk: long-tail feature development for mature products; maintenance-heavy legacy systems; internal tooling disconnected from revenue; product lines with weak cloud pull-through; teams with ambiguous ownership and low urgency.

Higher leverage: cloud reliability and site reliability engineering; security automation; infrastructure cost optimization; AI deployment, evaluation, and governance; data engineering tied directly to revenue; migration engineering, moving customers from the legacy stack to the new platform.

In zero-sum growth, being excellent is not enough if you are excellent in the wrong zone. AI does not replace judgment, accountability, or resilience planning, but it does raise the bar for what “irreplaceable” means in the roles it does not automate.

The takeaway

Oracle’s paradox is not a scandal. It is a signal, and it is one of the clearest signals available about how the transition from enterprise software to AI infrastructure actually works financially.

The paradox at the center of Oracle’s strategy is striking: the company is simultaneously pursuing record growth in AI infrastructure while executing its largest restructuring in years. The answer lies in a capital structure under pressure; $50 billion in capex commitments, $108 billion in debt, and a gross margin on AI capacity that looks more like a utility than a software business.

In 2026, record profits do not mean stability. In Oracle’s case, they mean a company accelerating into a future that requires a fundamentally different cost structure, and is willing to pay for that transition with the roles, margins, and product lines it is leaving behind.

The old org is the fuel. The question for everyone inside these organizations and everyone watching from outside is whether they are sitting in the engine or the tank.

Leave a comment

References and Relevant Resources

1. Oracle Q3 FY2026 Earnings: Primary Sources

Oracle Corporation. (March 10, 2026). “Oracle Announces Fiscal Year 2026 Third Quarter Financial Results.” Official press release. https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Fiscal-Year-2026-Third-Quarter-Financial-Results/default.aspx

SEC EDGAR. Oracle Corporation 8-K filing. (March 10, 2026). https://www.sec.gov/Archives/edgar/data/0001341439/000119312526100148/orcl-ex99_1.htm

PRNewswire. (March 10, 2026). “Oracle Announces Fiscal Year 2026 Third Quarter Financial Results.” https://www.prnewswire.com/news-releases/oracle-announces-fiscal-year-2026-third-quarter-financial-results-302709961.html

2. Earnings Analysis and Context

CNBC. (March 10, 2026). “Oracle Q3 earnings report 2026.” https://www.cnbc.com/2026/03/10/oracle-orcl-q3-earnings-report-2026.html

Invezz. (March 10, 2026). “Oracle Q3 earnings are in, and cloud numbers are hard to ignore.” https://invezz.com/news/2026/03/10/oracle-q3-earnings-are-in-and-cloud-numbers-are-hard-to-ignore/

TechnoTrenz. (March 10, 2026). “Oracle Q3 FY2026 Earnings: Revenue Hits $17.2B as Cloud Surges 44%.” https://technotrenz.com/news/oracle-q3-fy2026-earnings/

Grafa. (March 10, 2026). “Oracle (ORCL) Q3 2026 results: revenue $17.2B, backlog $553B.” https://grafa.com/en/news/united-states/oracle-q3-2026-earnings-cloud-surge

3. The 32% AI Gross Margin: Key Figure

El-Balad / Oracle Q3 FY2026 analysis. (March 2026). https://www.el-balad.com/16876986

4. Restructuring Budget: $2.1 Billion and Layoff Scale

Fortune. (March 9, 2026). “Oracle under pressure from more than $100 billion in debt and massive layoffs.” https://fortune.com/2026/03/09/oracle-earnings-layoffs-debt-cloud/

Metaintro. (March 2026). “Oracle Sets Aside $2.1B for Restructuring.” https://www.metaintro.com/blog/oracle-2-1-billion-restructuring-30000-jobs-ai-replaces-workers-2026

Stocktwits / Asianet Newsable. (March 11, 2026). “Oracle Quietly Raises Restructuring Fund By $500M: Big Layoff Wave Incoming?” https://stocktwits.com/news-articles/markets/equity/oracle-quietly-raises-restructuring-fund-by-500-m-big-layoff-wave-incoming/cZdMzrXRI9T

CIO Magazine / IDG. (January 30, 2026). “Oracle may slash up to 30,000 jobs to fund AI data-center expansion as US banks retreat.” https://www.cio.com/article/4125103/oracle-may-slash-up-to-30000-jobs-to-fund-ai-data-center-expansion-as-us-banks-retreat.html

International Business Times UK. (March 2026). “Oracle Layoffs 2026: 30,000 Jobs at Risk as AI Restructuring Looms.” https://www.ibtimes.co.uk/oracle-job-cuts-ai-expansion-1784041

5. Capex and Debt Context

Oracle Q3 FY2026 10-Q. Filed March 10, 2026. SEC EDGAR. https://d18rn0p25nwr6d.cloudfront.net/CIK-0001341439/751a926c-6f62-4baf-8f51-80cbcafdc616.pdf

March 18, 2026

In the 20th century, corporations built the weapons.
In the 21st century, corporations build the reasoning behind the strike.

The U.S.-Iran conflict in 2026 has made this concrete in a way no policy paper could. Anthropic’s Claude AI was used by U.S. Central Command to process and analyze intelligence intercepts, satellite imagery, and signals data, generate threat evaluations, model strike sequences, predict collateral damage, and support operational planning.

What makes this a turning point is not simply that AI was used in warfare. It is that the Pentagon used Claude through Palantir even after officially designating Anthropic a supply chain risk and ordering federal agencies to cease use, with DoD Chief Technology Officer Emil Michael explaining: “You can’t just rip out a system that’s deeply embedded overnight.”

Read that again. The U.S. military declared a technology company a national security risk and then continued using its product in active combat operations. That is not a governance story. That is a dependency story. And it reveals exactly how deeply private AI companies are now embedded in the operational infrastructure of modern warfare.

Subscribe now

From cyber warfare to algorithmic warfare

Traditional cyber operations focused on disruption: degrading networks, communications, and power grids. Algorithmic warfare is categorically different. It embeds reasoning, prediction, and decision support into the systems that execute strategy in real time.

As the U.S. military expanded its use of AI to pinpoint targets for airstrikes in Iran, the concern among lawmakers and independent experts was not that AI was making autonomous decisions; it was that relying on AI for parts of operations or decision-making creates structural risks, including the potential for errors in targeting that humans might not catch in time.

This is the operational reality that the debate about AI ethics has been anticipating in the abstract. It is no longer abstract.

The Starlink precedent and what it missed

When Russia invaded Ukraine in February 2022, satellite internet was not considered strategic infrastructure. Within weeks, Starlink had become the essential backbone of Ukrainian military communications, used to coordinate forces, guide drone strikes, and maintain continuity of government when terrestrial networks were destroyed. A consumer product designed for rural broadband became, without deliberate design, a critical military asset.

The consequences were profound and uncomfortable. SpaceX announced in February 2023 that it would no longer make its satellite services available for offensive military action by Ukraine, hampering efforts to retake Russian-occupied territories. A civilian company had unilaterally constrained the operational options of a sovereign nation at war.

The AI parallel is not hypothetical. It is already occurring with one crucial difference. Starlink’s leverage came from connectivity infrastructure. Claude’s leverage comes from the cognitive layer: the synthesis, analysis, and simulation that increasingly underlies the decisions themselves.

When a CEO’s contractual terms shape what a military can reason about, and how fast, the definition of a combatant requires updating.

The privateer is not a neutral metaphor, and it applies unevenly

The 16th-century privateer was a private ship captain licensed by a state to conduct naval warfare. The license gave the state plausible deniability; the captain retained operational discretion within the letter of the contract.

The AI analogy holds, but it is not symmetrical across companies.

Anthropic and the Pentagon could not reach terms on two specific restrictions Anthropic had built into its contract: no mass domestic surveillance of Americans, no fully autonomous weapons without human oversight. The Pentagon’s position was clear: “From the very beginning, this has been about one fundamental principle — the military being able to use technology for all lawful purposes.”

Anthropic refused and was blacklisted. OpenAI stepped in within hours, accepting the broader terms. The practical result: Claude continued to be used in active military operations in Iran despite the ban, while OpenAI operates under a deal it has since amended after public backlash, with Altman acknowledging the company “shouldn’t have rushed”.

Two AI companies, two governance postures, two very different relationships to the privateer role.

The governance paradox: banned but deployed

Palantir CEO Alex Karp confirmed that Palantir continues to use Anthropic’s Claude in its military deployments even as the Pentagon phases out Anthropic, and that the company plans to integrate additional large language models as the transition proceeds.

This creates a governance structure with no clear precedent. A company draws a contractual red line. The government refuses to accept it and bans the company. The military keeps using the product anyway through a third-party integrator. The company sues the government in two federal courts simultaneously.

Meanwhile, the war continues.

The lesson is not that red lines are meaningless; Anthropic’s refusal was genuine and costly. The lesson is that in a conflict environment, dependency creates a gravity that can pull operational reality away from contractual intent. Claude’s role was as a decision-support tool, not for autonomous weapons control, but the boundary between decision support and decision influence narrows with every increment of capability and integration depth.

Share

The intelligence tax is real, but compute sovereignty is being built

The strategic asymmetry is not hypothetical. Countries that lack access to frontier AI models face a measurable disadvantage in the speed and quality of intelligence synthesis, operational planning, and decision support. This is the “decision lag” problem, and it is already influencing defense planning and procurement globally.

But the response is underway and more advanced than the newsletter’s original framing suggested.

South Korea has launched a $735 billion sovereign AI initiative, with five competing consortia led by Naver, SK Telecom, LG Group, NCSoft, and Upstage, competing for national AI champion status, $381 million in government funding already allocated, and 260,000 NVIDIA GPUs being deployed across public sovereign clouds and major private-sector partners. The Korean government’s Ministry of Science and ICT is investing in infrastructure that gives Korea’s National AI Computing Center operational independence from Western cloud providers.

This is not a Manhattan Project. It is a publicly structured, nationally coordinated industrial competition, more analogous to South Korea’s earlier push for semiconductor independence than to wartime weapons development under secrecy. The goals are the same as the draft describes, reducing dependency on San Francisco-based companies for decisions that affect Korean sovereignty, but the mechanism is transparent, competitive, and deeply integrated with private industry.

The EU’s EuroHPC AI Factories program is building a parallel network of public supercomputing infrastructure accessible to European startups and universities. Japan’s sovereign AI investments are also accelerating. The pattern is consistent: allied nations are treating compute infrastructure as critical national infrastructure, not as a commercial procurement decision.

The brain of conflict has moved, but it is not fully private

The model weights reside in private companies. The decisions still, for now, and under current doctrine, reside with military commanders. The Pentagon and leading AI companies have publicly stated that current AI systems should not be able to kill without human signoff. The concern from lawmakers and independent experts is not that this principle has been abandoned, but that the conditions under which it can be maintained are becoming harder to guarantee as AI integration deepens and operational tempo accelerates

That is a more precise and more alarming framing than “AI makes the decisions”. The worry is not that the machine is in command. It is that the machine’s influence on human decisions is deepening faster than the oversight mechanisms to govern that influence.

Actionable signal: track what gets built and where

For investors and builders, the sovereign AI race is generating specific, trackable investment flows.

South Korea’s government announced plans with NVIDIA and local companies to deploy more than 260,000 GPUs across sovereign clouds and AI factories: a signal, as Jensen Huang noted, that compute infrastructure is now as strategic as power grids and broadband.

The categories worth tracking: sovereign foundation model programs (South Korea, EU, Japan), AI-enabled defense analytics platforms (Palantir, Scale AI, and their government-specific competitors), and the governance and verification layer — companies that make AI behavior auditable in high-stakes operational environments, where the cost of an error is not a wrong answer in a chatbot but a targeting decision in a live conflict.

The Iran conflict has clarified something that was previously theoretical: the companies that sit between frontier AI models and government operational systems are not neutral infrastructure providers. They are participants in outcomes. That participation will be regulated, litigated, and competed over for the rest of this decade.

The takeaway

AI companies are no longer neutral toolmakers. The U.S.-Iran conflict has confirmed what Starlink in Ukraine suggested: private technology infrastructure becomes strategic infrastructure under the pressure of conflict, and the companies that control it exercise a form of power that existing legal and governance frameworks were not designed to handle.

The privateer metaphor holds, but only if you note the most important thing about historical privateers. Their power was never permanent. States eventually built their own navies, wrote new laws, and brought private maritime power under public control.

The same reckoning is coming for AI. The question is whether it arrives through deliberate governance design or through the accumulated consequences of the precedents being set right now, in real time, in server logs and targeting systems over Iran.

Leave a comment

References and Relevant Resources

Here is the complete, verified reference list organized by section of the newsletter.

1. Claude’s Role in U.S. Military Operations Against Iran

The Wall Street Journal First reported Claude’s role in intelligence assessments, target identification, and battle scenario simulation by U.S. Central Command. Referenced across all downstream reporting. Note: WSJ article is paywalled; all major outlets cite it as the primary source.

NBC News. (March 11, 2026). “U.S. military is using AI to help plan Iran air attacks, sources say, as lawmakers call for oversight.” https://www.nbcnews.com/tech/tech-news/us-military-using-ai-help-plan-iran-air-attacks-sources-say-lawmakers-rcna262150

CBS News. (March 5, 2026). “Anthropic’s Claude AI being used in Iran war by U.S. military, sources say.” https://www.cbsnews.com/news/anthropic-claude-ai-iran-war-u-s/

Bloomberg. (March 4, 2026). “Iran Strikes: Anthropic Claude AI Helped US Attack. But How Exactly?” https://www.bloomberg.com/opinion/articles/2026-03-04/iran-strikes-anthropic-claude-ai-helped-us-attack-but-how-exactly

The Washington Post. (March 4, 2026). “Pentagon leverages AI in Iran strikes amid feud with Anthropic.” https://www.washingtonpost.com/technology/2026/03/04/anthropic-ai-iran-campaign/

WION. (March 2026). “AI in warfare is here: Pentagon used Anthropic’s Claude AI in Iran strikes.” https://www.wionews.com/world/ai-in-warfare-is-here-pentagon-used-anthropic-s-claude-ai-in-iran-strikes-but-it-has-many-llms-and-tools-from-other-firms-what-we-know-1772372063341

The Dupree Report / Wired investigation summary. (March 13, 2026). “Wired Exposes How Palantir Uses Claude AI in Iran War Targeting.” https://www.thedupreereport.com/2026/03/palantir-claude-ai-iran-targeting/

Futurism. (March 2026). “US Military Using Claude to Select Targets in Iran Strikes.” https://futurism.com/artificial-intelligence/claude-anthropic-military-iran

Futurism. (March 2026). “Pentagon Refuses to Say If AI Was Used to Select Elementary School as Bombing Target.” https://futurism.com/artificial-intelligence/pentagon-ai-claude-bombing-elementary-school

2. “Banned but Still Deployed”: Palantir Continues Using Claude

CNBC. (March 12, 2026). “Palantir is still using Anthropic’s Claude as Pentagon blacklist plays out, CEO Karp says.” https://www.cnbc.com/2026/03/12/karp-palantir-anthropic-claude-pentagon-blacklist.html

Responsible Statecraft. (March 2026). “US used ‘Claude’ to strike over 1000 targets in first 24 hours of war.” https://responsiblestatecraft.org/ai-war-iran/

3. The Supply Chain Risk Designation and Anthropic’s Lawsuit

Mayer Brown. (February 28, 2026). “Pentagon Designates Anthropic a Supply Chain Risk.” https://www.mayerbrown.com/en/insights/publications/2026/03/pentagon-designates-anthropic-a-supply-chain-risk-what-government-contractors-need-to-know

CNN. (March 9, 2026). “Anthropic Sues the Trump Administration.” https://www.cnn.com/2026/03/09/tech/anthropic-sues-pentagon

4. Starlink in Ukraine: The Structural Precedent

Defense One. (February 9, 2023). “Decrying Starlink’s ‘Weaponization,’ SpaceX Cuts Support for Ukrainian Military.” https://www.defenseone.com/technology/2023/02/spacex-now-says-they-dont-want-starlink-be-weaponized-ukraine/382797/

Belfer Center, Harvard Kennedy School. (March 9, 2023). “Starlink and the Russia-Ukraine War: A Case of Commercial Technology and Public Purpose?” https://www.belfercenter.org/publication/starlink-and-russia-ukraine-war-case-commercial-technology-and-public-purpose

SpaceNews. (March 9, 2023). “Limits on Ukraine’s use of Starlink for war operations is a lesson for U.S. military.” https://spacenews.com/limits-on-ukraines-use-of-starlink-for-war-operations-is-a-lesson-for-u-s-military/

SAGE Journals / Security Dialogue. (2024). “Private infrastructure in geopolitical conflicts: the case of Starlink and the war in Ukraine.” https://journals.sagepub.com/doi/10.1177/13540661241260653

5. South Korea Sovereign AI Initiative

NVIDIA Newsroom. (October 31, 2025). “NVIDIA, South Korea Government and Industrial Giants Build AI Infrastructure.” https://nvidianews.nvidia.com/news/south-korea-ai-infrastructure

NVIDIA Blog. (December 22, 2025). “Korea Joins AI Industrial Revolution: NVIDIA CEO Jensen Huang Unveils Historic Partnership at APEC Summit.” https://blogs.nvidia.com/blog/korea-ai-apec-ceo-summit/

Korea Tech Today. (June 27, 2025). “South Korea Unveils $735 Billion Plan to Build Sovereign AI Built on Korean Data.” https://koreatechtoday.com/south-korea-unveils-735-billion-plan-to-build-sovereign-ai-built-on-korean-data/

Introl Blog. (2025–2026). “South Korea’s $735B Sovereign AI Initiative: Infrastructure Requirements and Opportunities.” https://introl.com/blog/south-korea-735b-sovereign-ai-initiative-infrastructure-requirements-opportunities

Leave a comment

March 17, 2026

On February 27, 2026, Anthropic CEO Dario Amodei walked away from a Pentagon ultimatum.

Defense Secretary Pete Hegseth had given him a deadline: agree by 5:01 p.m. to let the Department of War, the Trump administration’s rebranding of the Department of Defense, use Anthropic’s AI models for “all lawful purposes”, including mass surveillance of American citizens and fully autonomous weapons systems that kill without human oversight. Or face consequences.

Amodei’s answer: “Threats do not change our position. We cannot in good conscience accede to their request.”

Within hours, the Trump administration ordered federal agencies to stop doing business with Anthropic and designated the company a “supply chain risk”, a label previously reserved for foreign adversaries like Huawei. Within the same hours, OpenAI’s Sam Altman announced a new Pentagon deal under the same broad terms Anthropic had refused.

The reaction was immediate and measurable. ChatGPT uninstalls surged 295% above their average on a single day. Downloads of Claude jumped 37% on Friday and 51% on Saturday. For the first time in its history, Anthropic’s Claude hit number one on the U.S. App Store, outpacing ChatGPT downloads. The movement calling itself QuitGPT, which had been building since early February, now claimed over 1.5 million people had canceled subscriptions, signed pledges, or publicly shared their decision to leave. By early March that figure had reached 2.5 million, unverified independently, but directionally consistent with the app data.

This is now the largest consumer technology boycott in the history of AI. And it did not appear from nowhere.

Subscribe now

The slow build before the spark

QuitGPT did not begin with the Pentagon deal. It began with a campaign finance filing. FEC records revealed that Greg Brockman, OpenAI’s president, and his wife had donated $25 million in total to MAGA Inc., the pro-Trump super PAC, making Brockman the single largest donor in the filing period, accounting for roughly a quarter of the organization’s entire fundraising in the second half of 2025. Sam Altman separately donated $1 million to Trump’s inaugural fund.

For a segment of OpenAI’s user base (younger, progressive, tech-fluent), that disclosure was a tipping point. It transformed a product choice into a political act. A loose coalition of activists and digital organizers set up quitgpt.org and began urging cancellations.

Then came a second revelation: the Department of Homeland Security published an AI inventory confirming that ICE, which had become a political flashpoint after its agents fatally shot two American citizens during an immigration enforcement operation in Minneapolis, uses a résumé screening tool powered by OpenAI’s GPT-4.

The Pentagon standoff poured accelerant on an already burning fire.

The end of neutral tech

For most of the past thirty years, Silicon Valley operated under a convenient fiction: we build tools, governments decide how to use them.

That posture let companies sell software everywhere without becoming entangled in geopolitical debates. It no longer works, and the Anthropic-Pentagon confrontation is the moment it broke publicly.

When AI systems can analyze bulk data collected from Americans, accelerate intelligence workflows, support autonomous weapons decisions, and simulate military scenarios, they have crossed from software into strategic infrastructure. And strategic infrastructure cannot be politically neutral.

The QuitGPT movement is the first major consumer expression of that reality. But it is also built on a story that is messier than a simple heroes-and-villains narrative.

The complication worth knowing

Anthropic’s public stand was genuine and costly. But the company’s own history complicates the clean “ethics-forward” framing that the boycott movement has applied to it.

Anthropic originally signed a $200 million Pentagon contract in July 2025. It was the first AI lab ahead of OpenAI, Google, and xAI to integrate its models directly into classified military networks. The dispute was not about whether to work with the military. It was about two specific prohibitions Anthropic had built into the contract: no mass domestic surveillance, no fully autonomous lethal weapons.

When the Pentagon sought to renegotiate those terms out of the contract, demanding unrestricted access for “all lawful purposes”, Anthropic held the line. OpenAI stepped in, accepted the broader terms, and announced its own classified-network deal hours after Anthropic was blacklisted.

In an internal memo that leaked to the press, Amodei called OpenAI’s approach “safety theater” and characterized Altman’s public statements as “straight up lies”. He later apologized for the tone, describing it as written in the heat of a chaotic afternoon. But the underlying argument, that accepting “all lawful purposes” language without explicit prohibitions is meaningfully different from Anthropic’s position, is the crux of the dispute, and it is a substantive one.

Dozens of scientists from OpenAI and Google DeepMind subsequently filed an amicus brief in Anthropic’s court challenge, in their personal capacities, supporting the company’s position. Anthropic is now suing the Pentagon over the supply chain risk designation, which it calls both legally unsound and a dangerous precedent for any American company that negotiates contract terms with the government.

Neither company’s hands are entirely clean. But one company drew a public line and paid a significant commercial and political price for it. That is what the market responded to.

The brand schism that the market is now pricing

Within the AI industry, a philosophical divide that had been implicit is now explicit.

On one side: companies willing to offer broad-access government and defense partnerships with minimal contractual restrictions. OpenAI and xAI occupy this space, with both agreeing to Pentagon deals covering classified military deployments under “all lawful purposes” language.

On the other side: companies that maintain self-imposed usage boundaries backed by contractual enforcement. Anthropic’s position is the clearest example: not simply marketing language, but specific prohibitions it was willing to lose a major government contract to defend.

This is not yet a settled market structure. OpenAI, following public backlash, subsequently announced revisions to its own safeguards. Sam Altman said the company “shouldn’t have rushed” its deal. The competitive and political dynamics are evolving in real time.

But the divide is now visible, public, and critically influencing actual consumer behavior at scale for the first time.

The moral marketplace has arrived

If QuitGPT represents the leading edge of a broader shift, it introduces a new competitive dimension into AI markets. Consumers already choose products based on ethical signals: sustainable supply chains, privacy-first software, fair-trade sourcing. AI is becoming the next category where governance influences purchasing. In that world, LLM providers compete across three dimensions simultaneously: capability, cost, and governance posture.

Technology markets have competed on the first two for three decades. The third is new. And it may matter more than most investors currently model.

The parallel that holds best is Apple’s privacy positioning. Starting with iOS 14’s App Tracking Transparency in 2021, Apple turned a compliance requirement into a product differentiator, and into a genuine revenue and trust moat. AI companies may follow a similar trajectory, with governance becoming not a constraint on commercial growth but a source of it. The companies that define their limits clearly will earn something hard to replicate: the benefit of the doubt when things go wrong.

The career signal

The AI labor market is beginning to reflect this divide in ways that matter for talent strategy. Top engineers are no longer choosing companies based solely on compensation, research freedom, or compute access. They are increasingly asking a different question: what will my code ultimately be used for?

This is not new. Tech workers protested Google’s Project Maven military AI contracts in 2019, and the walkouts forced Google not to renew that contract. But the stakes are higher now. Frontier models are qualitatively more capable than the machine-learning systems at issue then. The moral implications are proportionally harder to dismiss.

The QuitGPT movement is not only a consumer story. It is also an internal pressure story. Younger engineers, particularly those who entered the field motivated by AI safety and beneficial AI narratives, are watching how their employers navigate these decisions. The answer shapes where they choose to work, and how hard they advocate from the inside when they get there.

The risk for organizations using AI platforms

If the moral marketplace thesis is correct, companies that use AI platforms may face a new category of internal scrutiny; not from regulators, but from their own employees and customers.

The questions are already surfacing in Slack channels and engineering standups: why this vendor and not another? Does this model support applications we would not endorse? Are there alternatives more aligned with our stated values?

Ten years ago these questions were rare. In the current environment, they are becoming routine, and the answers have commercial consequences.

What founders and builders should do

The lesson here is not to pick a side in a philosophical debate. It is to get off the sidelines of your own governance posture.

Three practical steps:

Define your AI usage principles while the stakes are still low enough to think clearly. If your product scales, the debate will come to you. Better to have answered it in advance than under pressure.

Translate those principles into contractual and technical controls. Principles stated only in marketing copy are not principles. They are liability.

Communicate your position clearly and specifically. The worst place to be is the deliberate middle; appearing ethical in brand voice while remaining ambiguous in practice. Ambiguity invites backlash from both sides and trust from neither.

The deeper signal

QuitGPT may fade. Boycotts often do, and ChatGPT has nearly 900 million weekly active users as of late 2025. The numbers required to materially damage OpenAI’s business are orders of magnitude larger than what any boycott movement has achieved.

But the signal is already irreversible.

For the first time in the AI era, a significant number of consumers treated their choice of AI platform as a moral and political decision, and acted on it in measurable ways. Claude hit number one on the App Store not because it launched a new feature. It got there because its CEO refused a government order and paid for it publicly.

AI platforms are no longer just tools. They are institutions. And institutions inevitably become political.

The companies that understand this earliest, and build governance into their product and culture rather than bolting it on later, will be better positioned for the decade ahead than those still operating under the thirty-year fiction that the tool and its uses can be cleanly separated.

That fiction ended on February 27, 2026.

Share

References and Relevant Reads

1. The February 27 Pentagon Ultimatum and Supply Chain Risk Designation

Mayer Brown. (February 28, 2026). “Pentagon Designates Anthropic a Supply Chain Risk — What Government Contractors Need to Know.” https://www.mayerbrown.com/en/insights/publications/2026/03/pentagon-designates-anthropic-a-supply-chain-risk-what-government-contractors-need-to-know

CBS News. (February 27, 2026). “Hegseth Declares Anthropic a Supply Chain Risk, Restricting Military Contractors from Doing Business with AI Giant.” https://www.cbsnews.com/news/hegseth-declares-anthropic-supply-chain-risk/

Axios. (February 27, 2026). “What Trump Labeling Anthropic AI a Supply Chain Risk Means.” https://www.axios.com/2026/02/27/ai-trump-supply-chain-anthropic-pentagon-blacklist

TechCrunch. (February 27, 2026). “Pentagon Moves to Designate Anthropic as a Supply-Chain Risk.” https://techcrunch.com/2026/02/27/pentagon-moves-to-designate-anthropic-as-a-supply-chain-risk/

2. OpenAI’s Pentagon Deal and the Competitive Response

Fortune. (February 28, 2026). “OpenAI Sweeps in to Snag Pentagon Contract After Anthropic Labeled ‘Supply Chain Risk’.” https://fortune.com/2026/02/28/openai-pentagon-deal-anthropic-designated-supply-chain-risk-unprecedented-action-damage-its-growth/

Sovereign Magazine. (March 2026). “OpenAI Forced to Rewrite Pentagon Deal as 2.5 Million Users Join ChatGPT Boycott” https://www.sovereignmagazine.com/article/openai-forced-to-rewrite-pentagon-deal-as-2-5-million-users-join-chatgpt-boycott

3. App Store Data and QuitGPT Movement Scale

TechCrunch. (March 2, 2026). “ChatGPT Uninstalls Surged by 295% After DoD Deal.” https://techcrunch.com/2026/03/02/chatgpt-uninstalls-surged-by-295-after-dod-deal/

NPR. (March 6, 2026). “Pentagon Labels AI Company Anthropic a Supply Chain Risk ‘Effective Immediately’.” https://www.npr.org/2026/03/06/g-s1-112713/pentagon-labels-ai-company-anthropic-a-supply-chain-risk

Let’s Data Science. (March 2026). “CancelChatGPT: Why Users Are Switching to Claude.” https://letsdatascience.com/blog/cancelchatgpt-users-are-mass-migrating-to-claude-after-openai-pentagon-deal

Radical Data Science. (March 4, 2026). “QuitGPT Campaign Protests Outside OpenAI HQ in San Francisco.” https://radicaldatascience.wordpress.com/2026/03/04/quitgpt-campaign-protests-outside-openai-hq-in-san-francisco/

4. The Slow Build: Brockman Donation and ICE Tool

Bloomberg / Yahoo Finance. (January 2, 2026). “OpenAI Exec Becomes Top Trump Donor with $25 Million Gift.” https://finance.yahoo.com/news/openai-exec-becomes-top-trump-230342268.html

The Decoder. (January 3, 2026). “OpenAI Co-Founder Greg Brockman Donates $25 Million to Trump’s MAGA Inc. Super PAC.” https://the-decoder.com/openai-co-founder-greg-brockman-donates-25-million-to-trumps-maga-inc-super-pac/

Brennan Center for Justice. (2026). “Pro-Trump Super PAC Raises Record-Breaking $305 Million.” https://www.brennancenter.org/our-work/analysis-opinion/pro-trump-super-pac-raises-record-breaking-305-million

Gizmodo. (February 12, 2026). “OpenAI President Defends Trump Donations, Refuses to Comment on ICE.” https://gizmodo.com/openai-president-defends-trump-donations-refuses-to-comment-on-ice-2000721451

5. Anthropic Lawsuit and Legal Context

Mayer Brown. (March 10, 2026). “Anthropic Supply Chain Risk Designation Takes Effect — Latest Developments and Next Steps.” https://www.mayerbrown.com/en/insights/publications/2026/03/anthropic-supply-chain-risk-designation-takes-effect--latest-developments-and-next-steps-for-government-contractors

CNN. (March 9, 2026). “Anthropic Sues the Trump Administration After It Was Designated a Supply Chain Risk.” https://www.cnn.com/2026/03/09/tech/anthropic-sues-pentagon

The Hill. (March 11, 2026). “Anthropic: Pentagon Put Billions of Dollars at Stake with Supply Chain Risk Designation.” https://thehill.com/policy/technology/5778707-anthropic-sues-pentagon-designation/

CNBC. (March 5, 2026). “Anthropic Officially Told by DoD That It’s a Supply Chain Risk Even as Claude Used in Iran.” https://www.cnbc.com/2026/03/05/anthropic-pentagon-ai-claude-iran.html

6. Apple iOS 14 Privacy Comparison

Apple. (April 26, 2021). App Tracking Transparency — iOS 14.5 launch. Reference date: April 26, 2021 (iOS 14.5 launch with ATT enforcement). Multiple contemporaneous news coverage confirming privacy-as-moat positioning available from TechCrunch, The Verge, and Bloomberg.

Leave a comment

March 16, 2026

Three years ago, the standard career advice was straightforward: learn to code. By 2026, that guidance needs refinement. Coding remains valuable, but the scarce resource is no longer the ability to produce lines of code. What is becoming scarce is judgment: knowing what to ask for, how to structure a workflow, when to trust the machine, and when human intervention is necessary.

This shift helps explain the broader conversation around so-called “vibe coding”. The term was introduced in early 2025 by Andrej Karpathy to describe a playful style of development in which a programmer allows AI to generate most of the code, sometimes accepting it with minimal inspection. The framing was intentionally tongue-in-cheek and suited to experimental or personal projects. By early 2026, Karpathy himself had moved toward a more serious concept “agentic engineering” in which AI agents handle much of the implementation while humans focus on architecture, oversight, and quality control. The broader discussion about AI-assisted development is really about this latter evolution, even if the “vibe coding” label has remained popular in online discourse.

The strongest version of the argument is not that developers are being replaced outright. That claim overstates the current evidence. A more defensible and more interesting interpretation is that programming is among the occupations most exposed to AI assistance, because large portions of implementation work can be delegated to increasingly capable coding agents. Human roles then shift upward toward directing systems rather than typing syntax.

Research from Anthropic in its March 2026 labor-market analysis illustrates this dynamic. The study identifies computer programming among the occupations with the highest level of AI exposure around 75% task coverage, while still finding limited evidence of broad employment disruption so far.

That is why the classic “10× developer” narrative is beginning to feel outdated. In 2026, the advantage is not simply speed at the keyboard. It is the ability to operate more like a one-person operating system, orchestrating tools, agents, and workflows to produce outcomes that once required an entire team.

P.S. Stay tuned for the Premium Q&A section at the end of this post, where we share insights from interviews with venture capitalists and founders in Silicon Valley, pointing toward the strategies and perspectives shaping successful ventures and business.

Subscribe now

The end of human coding is overstated. The end of coding as pure manual labor is not.

Read more

Read more

Read more

Read more

March 12, 2026

For the past several years, most people have interacted with AI as if it were simply a more capable search interface: ask a question, refine the prompt, and repeat. That framing is beginning to change. On March 5, 2026, OpenAI released GPT-5.4, which the company describes as its first general-purpose model with native computer-use capabilities. The model is designed to operate across websites, software platforms, documents, spreadsheets, and presentations. In practical terms, this means AI is no longer limited to the chat interface; it can increasingly move inside the workflow itself.

That shift makes the broader thesis directionally correct, but one widely circulated statistic deserves careful clarification. It would be inaccurate to say that “GPT-5.4 outperforms 83% of human experts”. The more precise interpretation comes from the model’s performance on GDPval, an evaluation benchmark created by OpenAI to measure knowledge-work tasks across 44 occupations spanning the nine industries contributing most to U.S. GDP.

On this benchmark, GPT-5.4 matched or exceeded industry professionals in 83.0% of comparisons. That figure represents a combined win-or-tie rate. In direct head-to-head comparisons, the model outperformed human professionals in 70.8% of cases, with the remaining portion representing tied outcomes.

Both numbers are notable, but their meaning is narrower than the viral interpretation. They reflect performance on a specific evaluation created by the model’s developer, not evidence that 83% of professional expertise has suddenly become obsolete. The distinction matters, because benchmarks measure capability in controlled tasks, while real-world professional work still depends on context, judgment, and accountability beyond what any benchmark captures.

P.S. Stay tuned for our upcoming Premium Q&A Session featuring venture capitalists and CEOs from Silicon Valley. This will be a valuable opportunity for deeper discussions on pathways to success.

Subscribe now

We are officially in the agentic era

The real significance of GPT-5.4 is not simply that it is more capable than earlier models. The more important shift is architectural. OpenAI has combined reasoning, coding ability, tool search, and native computer-use capabilities into a single system intended for real professional workflows. The company describes GPT-5.4 as its “most capable and efficient frontier model for professional work,” available through ChatGPT, the API, and tools such as Codex.

Read more

Read more

March 11, 2026

While Washington debates AI governance in the abstract, a more consequential story is unfolding at the infrastructure layer: in Memphis, xAI is attempting to scale the constraint that most reliably throttles technological acceleration, electric power.

If you want to gauge whether “AI replaces jobs” remains a decade-long societal negotiation or becomes a near-term operational reality, watch the convergence of compute capacity, power availability, and robotics manufacturing. That is where algorithmic capability stops being a headline and starts becoming unit economics and payroll math, the point at which automation shifts from possibility to execution.

Subscribe now

1) The Memphis signal isn’t just “a big data center”. It reflects a new deployment pattern.

For much of modern technology, the primary constraints were talent availability and distribution reach. In frontier AI, the constraint has increasingly shifted toward time-to-power; how quickly a project can secure and deploy large amounts of electricity.

Read more

Read more

March 9, 2026

The idea of the “10× engineer” has always carried a slightly mythical quality. What is changing now is not that one human has suddenly become ten times more intelligent, but it is that a capable builder can now deploy persistent AI workers to handle recurring tasks: answering common questions, triaging requests, compiling updates, and operating across multiple tools while the human is offline.

Notion’s recent introduction of Custom Agents reflects this shift directly. These autonomous agents can run on schedules or event triggers across systems such as Notion, Slack, email, calendars, and MCP-connected tools like Figma and Linear.

Seen in that light, the more accurate phrase for 2026 may not be the “10× developer”, but rather the individual as an enterprise: a single professional orchestrating a network of software agents that extend their operational capacity.

P.S. Stay tuned for the Premium Q&A section at the end of today’s newsletter, where VCs and inventors in Silicon Valley share their perspectives.

Subscribe now

The big shift: the cost of operations is collapsing

Read more

Read more