Microsoft ends OpenAI exclusivity. Stanford reports model transparency collapsed 31%. And the Commerce Department still hasn't published that state AI law review.

A software giant disclosed its AI business now generates $37 billion annually, up 123 percent in a year. The same company told its most important AI partner it no longer owns exclusive rights to the models that powered its enterprise strategy. And an academic institution released data showing the most capable AI systems now disclose less about how they work than they did a year ago—a transparency collapse measurable to the percentage point.

All three events occurred in the last seventy-two hours. All three are on the record, dated, and sourced. And all three test the same thesis from different angles: whether the operators building, deploying, and regulating AI will provide enough information for anyone outside their walls to grade what they claim.

Microsoft's renegotiated partnership with OpenAI ends the cloud exclusivity that defined their relationship since 2019. Stanford's 2026 AI Index documents a 31 percent drop in foundation model transparency over a single year. And the Commerce Department—tasked by executive order to evaluate state AI laws and identify those deemed "onerous" by March 11, 2026—has published nothing, leaving states, labs, and enforcement bodies to operate in a regulatory void the Trump Administration ordered filled but has not.

Each claim is gradeable. That is why they are in this ledger.

3 Claims

Claim 1 — Microsoft and OpenAI: exclusivity ends, license becomes non-exclusive through 2032

On April 27, 2026, Microsoft and OpenAI jointly announced a restructured partnership removing the cloud exclusivity that had bound the companies since 2019. Under the new terms, OpenAI can now serve all its products to customers across any cloud provider, and Microsoft's license to OpenAI intellectual property—which runs through 2032—becomes non-exclusive.

Microsoft will no longer pay revenue share to OpenAI, though OpenAI will continue paying Microsoft a capped revenue share through 2030. The companies also removed a controversial clause that would have changed their business relationship if OpenAI declared it had achieved artificial general intelligence, replacing that ambiguity with a fixed end date.

Azure remains OpenAI's "primary cloud partner," and OpenAI products will ship first on Azure unless Microsoft cannot support the necessary capabilities. But the end of exclusivity clears OpenAI to distribute models via Amazon Web Services, Google Cloud, Oracle, and other providers. Amazon CEO Andy Jassy confirmed the same day that OpenAI models would become available on AWS Bedrock "in the coming weeks."

The claim is gradeable on two dimensions: whether OpenAI ships models on non-Azure cloud platforms by year-end 2026, and whether Microsoft's IP license remains non-exclusive through the stated 2032 term. The invalidator would be credible reporting (via OpenAI disclosures, Microsoft earnings calls, or investigative journalism from TechCrunch, Bloomberg, or The Information) showing that exclusivity was reinstated, that OpenAI failed to launch on rival clouds despite regulatory or commercial clearance, or that the IP license reverted to exclusivity before 2032.

Grade by: 2026-12-31 (8 months)
Invalidator: OpenAI does not launch models on at least one non-Azure cloud platform (AWS, Google Cloud, or Oracle) by December 31, 2026, or credible reporting demonstrates Microsoft renegotiated exclusivity or OpenAI was contractually barred from multi-cloud distribution despite the April 27 announcement.

Claim 2 — Stanford AI Index 2026: Foundation Model Transparency Index drops from 58 to 40

Stanford's Institute for Human-Centered Artificial Intelligence released the 2026 AI Index in April 2026, documenting a transparency collapse among the most capable AI models. The Foundation Model Transparency Index, which measures how openly major AI companies disclose details about training data, compute, capabilities, risks, and usage policies, saw average scores drop to 40 points out of 100, down from 58 the prior year.

The report noted that "the most capable models are now the least transparent," with leading labs including Google, Anthropic, and OpenAI abandoning disclosure of dataset sizes, training duration, and parameter counts. Of the 95 most notable models launched in 2025, 80 were released without their training code.

The claim is gradeable because the Foundation Model Transparency Index is an annual, peer-reviewed assessment conducted by researchers from Stanford, Berkeley, Princeton, and MIT. Future editions will document whether transparency continues to decline, stabilizes, or improves. The invalidator would be the 2027 AI Index showing Foundation Model Transparency Index scores returning above 55, or major labs resuming systematic disclosure of training data, compute, and methodology at levels consistent with 2024 or earlier practice.

Grade by: 2027-04-30 (1 year)
Invalidator: The 2027 AI Index reports Foundation Model Transparency Index scores above 55, or credible third-party evaluation (from Stanford HAI, Partnership on AI, or comparable research institutions) demonstrates that the top three labs (by capability ranking) have resumed disclosing dataset sizes, training compute, and parameter counts for their flagship models.

Claim 3 — Microsoft: AI revenue run rate reaches $37 billion, up 123% year over year

On April 29, 2026, Microsoft reported third-quarter fiscal 2026 results, disclosing that annualized revenue from AI "now stands at $37 billion, up 123%." The figure includes business from clients running AI services on Azure, revenue from all model builders hosted on the platform, and revenue from Microsoft's own AI tools including Copilot, which exceeded 20 million paid seats.

Azure and other cloud services revenue increased 40 percent year over year, beating analyst expectations. Microsoft also disclosed it expects $190 billion in capital expenditure for calendar year 2026, well above Wall Street estimates, and that it will remain capacity-constrained "at least through 2026."

The claim is gradeable through Microsoft's quarterly earnings disclosures, which report AI revenue as a distinct line item. The invalidator would be subsequent earnings calls or SEC filings showing AI revenue materially below the $37 billion run rate by fiscal year-end June 2026, or credible financial analysis (from Bloomberg, CNBC, or sell-side research) demonstrating that Microsoft's methodology for calculating AI revenue was revised in a way that materially overstated the April 29 figure.

Grade by: 2026-10-29 (6 months)
Invalidator: Microsoft's fiscal Q4 2026 earnings (due July 2026) or subsequent disclosures show annualized AI revenue materially below $35 billion, or investigative reporting from Bloomberg, The Information, or Wall Street Journal reveals that the $37 billion run rate included non-AI cloud revenue or was calculated using a methodology Microsoft later revised.

2 Reckonings

Reckoning 1 — FDA: AI integrated across all centers by June 30, 2025 (now ten months overdue for full accounting)

Original claim: On May 8, 2025, FDA Commissioner Martin Makary announced an aggressive timeline to scale artificial intelligence internally across all FDA centers by June 30, 2025, directing immediate deployment with the goal of full integration by end of June.

What happened: The June 30, 2025 deadline has passed. We are now in late April 2026, ten months beyond the stated horizon. Makary stated in March 2026 that FDA AI tools had saved more than 14,000 staffers 17,000 hours of human work time "since implementation at the end of June," suggesting the June 2025 deadline was at least partially met. However, CNN reported in July 2025 that the FDA's AI system was "making up studies," raising quality and reliability concerns. No independent audit, GAO report, or investigative follow-up from STAT or Healthcare Brew has publicly confirmed that AI systems were operationally deployed to all FDA centers by the stated deadline, nor that the systems met the promised efficiency gains without compromising review quality.

Grade: C
The claim was structurally ambitious but lacked external verification. Makary's March 2026 statement provides circumstantial evidence that some deployment occurred by June 2025, but the CNN report of fabricated studies in July 2025 and the absence of third-party confirmation leave significant doubt about whether "all centers" met the deadline or whether deployment quality met the standard implied by the original announcement.

Invalidator (what would have changed the grade): A GAO audit, FDA Inspector General report, or credible investigative series from STAT or ProPublica published by December 2025 confirming that 90 percent or more of FDA centers had integrated AI systems operationally by July 2025, and that those systems did not introduce material errors requiring rollback or significant remediation. Alternatively, Makary issuing a public update by September 2025 detailing which centers met the deadline, which did not, and what quality controls were in place.

Reckoning 2 — Anthropic revenue: $30 billion run rate by April 2026 (confirmed and exceeded expectations)

Original claim: On April 7, 2026, Anthropic CFO Krishna Rao stated the company's revenue run rate "has now topped $30 billion, up from $9 billion at the end of 2025," with more than 1,000 business customers spending over $1 million annually.

What happened: Anthropic's $30 billion run rate was announced April 7, 2026. By April 24, 2026, the company announced partnerships with Google and Broadcom to secure multiple gigawatts of compute capacity beginning in 2027, and disclosed Google's commitment to invest up to $40 billion—$10 billion upfront at a $350 billion valuation, with $30 billion more tied to performance milestones. Amazon separately committed up to $25 billion in new investment and $100 billion in cloud spending over ten years. These deals, announced within three weeks of the revenue disclosure, suggest the $30 billion figure was credible enough to justify record-scale infrastructure and capital commitments from two of the world's largest cloud providers.

Grade: A
The revenue claim was independently validated by the largest infrastructure deals in AI history within weeks of disclosure. Google, Amazon, and Broadcom—all of whom conduct due diligence before billion-dollar commitments—structured their investments and compute partnerships around Anthropic's stated growth trajectory. No credible reporting has challenged the $30 billion figure, and the timing of the deals provides strong circumstantial confirmation.

Invalidator (what would have changed the grade): Credible financial reporting from Bloomberg, The Information, or Wall Street Journal within 90 days of the April 7 announcement showing Anthropic's actual annualized revenue was materially below $25 billion, or subsequent disclosures (via IPO filings or partner earnings calls) revealing that the $30 billion run rate was calculated using a non-standard methodology that inflated the figure.

1 Refusal

I was asked by a reader via email yesterday whether I would cover the Elon Musk v. OpenAI trial underway in Oakland. Day three of testimony wrapped April 29, with cross-examination of Musk by OpenAI's counsel reported in detail by CNBC, Bloomberg, and Ars Technica. The trial will produce a verdict, and that verdict will be news.

But trials are not claims. They are adjudication. The ledger does not track litigation blow-by-blow—it tracks what operators say they will do, grades what happened, and refuses to let either side rewrite the record when the horizon arrives. Musk's emails to Ilya Sutskever, his testimony about whether he called colleagues "jackasses," and the judge's rulings on evidence admissibility are all relevant to the case. None of them are gradeable projections.

If the verdict includes a factual finding about OpenAI's original mission, its adherence to safety commitments, or whether it breached fiduciary duties, that finding will be a claim—dated, signed by a judge, and subject to appeal. When the appeals are exhausted and the record is final, it can be graded.

I refused to treat courtroom theatrics as equivalent to operational commitments.

— Roger Grubb, Editor