The last twenty-three days produced six headlines, filed under six categories. The categories are the disguise.

On April 27, 2026, OpenAI and Microsoft announced a restructured partnership. Azure exclusivity ended. The license to OpenAI’s intellectual property became non-exclusive through 2032. OpenAI could now serve its products on any cloud. And quieter than any of it, the AGI clause was scrapped.

The AGI clause was the structural protection of the original nonprofit mission. It was the contractual provision that would have terminated Microsoft’s rights to OpenAI’s technology once OpenAI reached artificial general intelligence. It was the line in the sand that said this company is something other than an ordinary commercial entity, and a moment is coming when the commercial structure dissolves.

April 27 was also the day jury selection began in Musk v. Altman, the lawsuit alleging that OpenAI had abandoned that same mission.

Same day. I am not going to claim either side timed it, because there is no evidence either side did. The coincidence is not the argument. What it exposes is. OpenAI dissolved the structural protection of the nonprofit mission on the same day a jury was seated to judge whether that mission had been protected. Almost nobody set the two events side by side. The press had already filed them in separate sections: the restructure ran as a corporate blog post and read as legal housekeeping, the trial read as celebrity litigation. Different desks, different readers, same story.

This is what it looks like to read the last twenty-three days as a single event.

The window runs from April 27 to May 19. Six stories broke inside it, and the press filed them under six categories: legal, infrastructure, corporate, finance, courts, talent. Filed that way, they look like six unrelated things happening to six different companies. They are not six things. They are one thing.

Two months ago I wrote that Google didn’t need to own Anthropic, because the alliance had already produced everything ownership would have. That piece traced five layers of dependency inside a single relationship. Back then the argument was about one relationship. It now describes the whole industry, and the last twenty-three days are the first stretch where you can watch it run in real time.

Some of these events handed Anthropic an advantage. The rest only cleared what stood in its way. By May 19 it no longer much mattered which was which.

Start with what was cleared. The AGI clause was gone. Three weeks later the lawsuit went with it, when the Musk v. Altman jury returned its verdict on May 18. So the clause protecting the nonprofit mission and the trial set to judge it both closed inside the same three weeks. Nobody handed those outcomes to Anthropic. Anthropic was not in the room for either. The board was simply swept clear of the machinery binding commercial AI to a hypothetical future test. Anthropic is the one frontier lab whose mission-first governance came through these months intact. It kept the structure everyone else was unwinding. The unwinding is what made the structure rare.

That verdict belonged to one man. To see what the window did to him, you have to go back nine years.

In 2017, Elon Musk sat on the boards of both OpenAI and Tesla, and he used the overlap to do something precise. He poached Andrej Karpathy, a founding member of OpenAI, to run the computer-vision work behind Tesla Autopilot. It was personal, and it was deliberate. It was the first time the talent gravity at the top of the field ran toward Musk instead of away from him.

On May 6, 2026, nine days after the Microsoft restructure, Musk announced that xAI would cease to exist as an independent company. It would be folded into SpaceX under the name SpaceXAI. Eleven of the twelve original xAI co-founders had already left over the preceding year. Musk was the last one standing. The move was framed as efficiency. The substance was harder than that: the company Musk built to beat OpenAI no longer made sense on its own once its one real asset was finished.

That asset was Colossus 1, the supercomputer xAI raised from a shuttered Electrolux factory in South Memphis. Its first hundred thousand Nvidia GPUs came online in a hundred and twenty-two days, against an industry estimate of two years; xAI doubled the machine inside another three months. By 2026 it ran more than two hundred twenty thousand GPUs and three hundred megawatts of capacity. Musk built it to beat OpenAI. In the same week xAI dissolved, Colossus 1 was leased, in full, to Anthropic.

Musk had spent two years calling Anthropic misanthropic, evil, and a threat to Western civilization, in posts that are still public. His statement on the day of the lease read: “No one set off my evil detector. So long as they engage in critical self-examination, Claude will probably be good.” Attached to the deal was a clause letting SpaceX take the compute back if Anthropic’s AI is ever judged a danger. A landlord with a kill switch. And the landlord is the man who spent two years saying this company would end the world.

Twelve days later the lawsuit ended the way everything in the window ended for him. Musk v. Altman had alleged that Sam Altman and Greg Brockman betrayed OpenAI’s founding mission for personal enrichment. Musk had sought a hundred and thirty-four billion dollars in disgorgement and the unwinding of OpenAI’s for-profit arm. The jury was out less than two hours. It found against him unanimously, on the ground that the claim was barred: Musk had known what OpenAI was becoming as far back as 2021. He called the verdict a calendar technicality and said he would appeal.

Then, on May 19, the last day of the window, Anthropic announced that Andrej Karpathy had joined its pretraining team.

Set the two ends of the wire against each other. In 2017 Musk used an OpenAI board seat to pull Karpathy into his orbit. In 2026 Karpathy joined the lab that pays Musk rent on the supercomputer Musk built to destroy OpenAI, the lab Musk had called evil, the lab whose compute Musk can in principle switch off, in the same window the lawsuit Musk brought to reclaim OpenAI collapsed. Nine years, and every line that left Musk’s hand came back pointing the other way.

The structure did not punish Musk. It left him. That is what happens when you build infrastructure for a fight everyone else has already walked away from.

The Colossus lease is one supplier. The window bound two more, and the shape the three make together is the point.

In late April, Google committed up to forty billion dollars to Anthropic, ten billion of it immediately and the rest tied to milestones, with five gigawatts of TPU capacity attached over five years. Days earlier, Amazon had deepened its own stake: a fresh five-billion-dollar investment, up to twenty billion more to follow, set against a separate agreement under which Anthropic will spend as much as one hundred billion dollars for roughly five gigawatts of Amazon compute. Look at the shape rather than the dollar amounts. A hyperscaler takes an equity stake in the lab; the lab commits years of its budget straight back to that hyperscaler’s cloud. Capital out, capital returned as infrastructure spend. Two of the four major cloud providers, wired into a single lab’s training roadmap. Add the Colossus lease, and the third supplier arrived from the least likely direction in the industry.

What that leaves Anthropic holding is not the largest reserved compute position in the industry. It is the most distributed: more than ten gigawatts across Amazon, Google, Microsoft and Nvidia, Fluidstack, and a supercomputer one rival built to beat another. Anthropic is the only frontier lab drawing training capacity from every major cloud at once, beholden to no single one of them. A position assembled that way is not the easiest to expand. It is the hardest to take away.

The capital tracked the same shape. Inside the window Anthropic moved into a raise that priced it past nine hundred billion dollars, past OpenAI’s last public valuation. The number is not the point. The way the round is filling itself is: out of demand, not persuasion. That is how money behaves when it tracks a position instead of betting on one.

Then the talent flow did something it had never done before. It reversed. Senior researchers used to leave Anthropic for OpenAI. For a while it ran both ways. In this window it ran, structurally, one direction. Andrea Vallone, a senior safety lead at OpenAI, joined Anthropic in January. Jan Leike, OpenAI’s former alignment lead, had already been there two years. Ross Nordeen, a founding member of xAI, joined in early May. On the other side of the ledger, OpenAI lost Bill Peebles, Kevin Weil, and Srinivas Narayanan on a single Friday in April, with Fidji Simo out on medical leave. One company’s senior roster was starting to read like an alumni page. The other’s was starting to read like a reorganization under stress. The hire that capped the reversal landed on the window’s last day. Read as a talent-war trophy, it was the biggest of them. Read correctly, it was not a talent story at all.

Karpathy did not join Anthropic to build a product. Nick Joseph, its Head of Pretraining, put the role plainly: “He’ll be building a team focused on using Claude to accelerate pretraining research itself.” The team uses the model to improve the model. The same day, Chris Rohlf joined Anthropic’s frontier red team. Accelerator and tripwire, named in the same breath.

Go back fifteen days. On May 4, in issue 455 of his newsletter Import AI, Anthropic co-founder Jack Clark wrote: “I’m writing this post because when I look at all the publicly available information I reluctantly come to the view that there’s a likely chance (60%+) that no-human-involved AI R&D – an AI system powerful enough that it could plausibly autonomously build its own successor – happens by the end of 2028.” The word doing the most work there is reluctantly. By Clark’s own account the view is reluctant because the implications are large enough to dwarf him, and he is not sure society is ready for them. A co-founder of the company, on the record, combing the public evidence and unable to argue his way out of where it points. Fifteen days later, his company hired the most recognizable researcher in the field to lead the team that turns that forecast into a roadmap.

Recursive self-improvement stopped being a thing Anthropic described and became a thing it staffed.

Recursive is not an abstraction here, whatever the coverage implied. The loop is AI systems doing the work that builds more capable AI systems: writing the code, searching the architectures, curating the data, running the evaluations, with less human involvement at each turn. The first rung of that ladder is not speculative. It is already load-bearing inside Anthropic, where Claude writes the large majority of the company’s own code. The Karpathy team is not introducing the loop. It is moving it up a level, from writing the code that builds the model to doing the research that designs the next one.

Anthropic is not the only lab standing there, and the parallel signals are not subtle. The AI-policy writer Dean Ball has called OpenAI the most transparent of the labs about this trajectory, pointing to its own stated plans for a largely automated research workforce within a few years. In May a startup called Recursive Superintelligence came out of stealth with six hundred fifty million dollars and a single declared purpose: to build AI that automates AI research, outside the established labs. Its founders include Richard Socher and the former DeepMind researcher Tim Rocktäschel. And the International Dialogues on AI Safety, the track-two body that has produced some of the only common US-China language on frontier risk, has stated plainly that no AI system should copy or improve itself without explicit human approval. The Karpathy role sits exactly on that line. No lab has made a more visible hire against the line the field drew for itself.

The most-quoted frame for the hire was the talent war. The talent war is real. It is also a containing narrative for something far more specific: Karpathy has been hired to close the recursive-improvement loop at the one frontier lab whose own policy lead has put better-than-even odds on that loop closing by 2028. The hire is not Anthropic announcing a new ambition. It is the human-resources form of a policy paper Anthropic had already published.

None of this merely happened to Anthropic. Anthropic negotiated the Google and Amazon commitments. It agreed to train on a rival’s supercomputer. It went out and hired the people. What it did not do, anywhere in the window, was wage the public fights consuming the rest of the industry. It built position instead, and never needed to win them.

And through all of it the principals stayed quiet. Altman said nothing about the verdict or about Karpathy. Dario Amodei said nothing about either. Jack Clark, whose Import AI 455 had effectively pre-written the Karpathy job description, said nothing when it was filled. The only principal who made noise was the one who had lost. Silence proves nothing on its own. But noise from the man who lost, and silence from everyone with a claim to a win, is the pattern you would expect if the result were already obvious to all of them.

The AGI clause is gone. The lawsuit is gone. The loudest critic of Anthropic is now its landlord. The best-known member of OpenAI’s founding team is now Anthropic’s pretraining lead. Twenty-three days.

Here is the case against everything I have just said. The compute that looks like a moat is also a bill. The hundred billion committed to Amazon, and the years of cloud spend owed to Google, are not assets Anthropic holds. They are revenue Anthropic has already promised away, owed back to the same landlords who just funded the round. OpenAI did not freeze while this happened. It raised one of the largest rounds in corporate history this spring, at an eight-hundred-fifty-billion-dollar valuation, and shipped GPT-5.4 squarely into Anthropic’s enterprise territory. Stargate’s planned capacity still dwarfs any single site Anthropic can call its own. A reader who wants to bet against this piece has real material to bet with.

But notice what those objections are. The liabilities and the competitor moves do not claim the routing changed. They claim it is expensive and contested, which is a smaller claim. Forward liabilities do not relocate the center of an industry. They are the price of standing in it. The compute Anthropic owes is a bill, not proof that someone else holds the position. Only the Stargate objection claims the center is elsewhere, and the earlier argument already met it: a single site that out-masses any one building Anthropic runs is concentration, and concentration is the easiest compute to take away. Anthropic is hard to dislodge because its compute is spread across every supplier at once. Strip the objections down and what survives is a disagreement about cost, not structure. That is the argument worth having. The question is no longer whether the industry reorganized around a center. It is what that center costs to hold.

The industry did not reorganize around capability. It reorganized around who you have to go through to reach capability. Across twenty-three days the board was cleared, the suppliers were bound, the talent reversed, and Anthropic moved to the middle of the network without winning a single public fight.

The structure moved.

Whether the middle of it is a position or a trap is what the months after this one begin to answer.

---

Sources

Every factual claim in this piece was checked against primary sources and contemporaneous reporting on May 21, 2026. Corrections made on that pass are noted where relevant.

The Microsoft-OpenAI restructure (April 27, 2026). Verified against the joint Microsoft-OpenAI announcement and reporting from TechCrunch, Directions on Microsoft, and others: end of Azure exclusivity, IP license non-exclusive through 2032, OpenAI free to serve products on any cloud, removal of the AGI clause (the original provision that would have ended Microsoft’s access to OpenAI’s most capable systems once OpenAI’s board declared AGI). Correction on the fact-check pass: an earlier draft stated “Microsoft stopped paying OpenAI a revenue share.” This was wrong in both direction and substance — the revenue share ran from OpenAI to Microsoft, and the new deal retains Microsoft’s 20% through 2030. The false sentence was cut. The same-day overlap with Musk v. Altman jury selection was surfaced publicly on X and did not appear as a linked event in major-outlet coverage.

Musk v. Altman verdict (May 18, 2026). Verified against NPR, CNBC, Fox Business, and Local News Matters. Unanimous verdict from a nine-member advisory jury, returned in under two hours (deliberations began 8:30 a.m., verdict 10:23 a.m.), after eleven days of testimony; the verdict was immediately adopted by Judge Yvonne Gonzalez Rogers, U.S. District Court for the Northern District of California. The suit was dismissed on statute-of-limitations grounds, the jury finding Musk was on notice of OpenAI’s for-profit direction years before his 2024 filing. Defendants were Sam Altman and Greg Brockman; the aiding-and-abetting claim against Microsoft also failed. Disgorgement claim cited here at $134B per Local News Matters and earlier reporting; some early coverage (NPR) used an approximate $150B figure. Correction on the fact-check pass: an earlier draft said Musk “posted twice that afternoon, both times at the judge.” The verifiable record is that Musk called the verdict a “calendar technicality” and vowed to appeal (CNBC); the text now reflects that.

xAI, SpaceX, and the Colossus 1 lease (May 6, 2026). Verified against Wikipedia, FinanceFeeds, Tom’s Hardware, Inc., and others. xAI dissolved into SpaceX as the “SpaceXAI” division, announced by Musk on X on May 6, 2026; eleven of twelve original xAI co-founders had departed, Musk the last standing. Anthropic leased the full capacity of Colossus 1 in Memphis — built by xAI in a former Electrolux factory — with deal coverage consistently reporting more than 220,000 Nvidia GPUs and roughly 300 megawatts of capacity. Correction on the fact-check pass: a prior pass had reduced these figures to ~250 MW using stale 2025 data; the May 2026 deal coverage confirms ~300 MW and 220,000+ GPUs, and the figures were restored. The 122-day build figure applies to the first 100,000 GPUs (xAI’s own claim, x.ai/colossus), not the full cluster; the cluster was doubled over a further ~92 days. Phase-one cost has been reported as high as ~$4B and is omitted from the body as contested. Musk’s “evil detector” statement is quoted verbatim from his X post and verified against Inc., capitalaidaily, and Tom’s Hardware; the clause permitting SpaceX to reclaim the compute on AI-risk grounds is confirmed by deal coverage.

Compute (April 2026). Verified against TechCrunch, CNBC, Axios, and Bloomberg. Google committed up to $40B to Anthropic on April 24, 2026 — $10B immediate, $30B milestone-contingent — with ~5 GW of TPU capacity over five years. Days earlier Amazon expanded its stake (a fresh $5B, up to $25B total) against a separate agreement under which Anthropic will spend up to $100B for ~5 GW of Amazon compute. Correction on the fact-check pass: an earlier draft stated Anthropic “committed approximately two hundred billion dollars straight back into Google’s cloud.” No source supports a $200B Google-cloud figure; it was cut. The verified reciprocal-spend figure ($100B) belongs to the Amazon agreement, and the “loop” is now described as the pattern rather than asserted with an unsourced number.

Capital. Verified against Bloomberg (May 12, 2026) and TechCrunch: Anthropic in talks to raise $30B+ (TechCrunch reported a figure near $50B) at a pre-money valuation exceeding $900B, past OpenAI’s last reported valuation of ~$852B; the round was filling from unsolicited investor demand, with some early backers skipping it.

Talent. Verified: Bill Peebles, Kevin Weil, and Srinivas Narayanan all announced departures on the same Friday, April 17, 2026 (CNBC); Fidji Simo on medical leave. Andrea Vallone moved from OpenAI to Anthropic’s alignment team in January 2026, reporting to Jan Leike; Leike left OpenAI for Anthropic in May 2024; Ross Nordeen, a founding member of xAI, joined Anthropic in early May 2026, around the Colossus lease (DCD, CNBC). Correction on the fact-check pass: an earlier draft stated “Kate Rouch stepped back for cancer recovery.” This could not be verified as part of the April–May 2026 departures and was cut.

The recursive-self-improvement thesis and landscape. Jack Clark’s Import AI 455 quote is verbatim, verified against the primary sources jack-clark.net and importai.substack.com (issue dated May 4, 2026); en-dashes inside the quote preserved per source. Clark’s stated reason for “reluctantly” — that the implications dwarf him and he is unsure society is ready — is reflected accurately in the gloss; an earlier draft mischaracterized it. Karpathy joined Anthropic’s pretraining team on May 19, 2026, under team lead Nick Joseph, to build a team using Claude to accelerate pretraining research (TechCrunch, Axios, Anthropic spokesperson). Chris Rohlf joined Anthropic’s frontier red team, announced the same day. Correction on the fact-check pass: the Socher/Rocktäschel material was an error in earlier drafts that split one company into two. There is one venture — Recursive Superintelligence — co-founded by Richard Socher, Tim Rocktäschel, and others, which emerged from stealth in mid-May 2026 with $650M (at a $4.65B valuation, per The Decoder, FinSMEs, and TechCrunch). The $650M figure is now verified and stated. The International Dialogues on AI Safety red line — “no AI system should be able to copy or improve itself without explicit human approval and assistance” — is quoted accurately from the IDAIS-Beijing statement.

The counter-case. Anthropic’s forward compute liabilities follow from the Google and Amazon deal terms above. Stargate’s planned capacity (~7 GW and over $400B in investment) is verified against OpenAI’s own announcement and Business Insider. Correction on the fact-check pass: an earlier draft cited an OpenAI “ten-billion-dollar enterprise deployment arm” and a specific “cybersecurity model,” neither of which could be cleanly verified; the text now uses verified facts — OpenAI’s spring funding round at an ~$852B valuation and the enterprise-aimed GPT-5.4. The “claim is reluctantly opinion” note: the Dean Ball characterization is presented as his, with the specific automation timelines softened on the fact-check pass to what could be supported.