The deal era went to court. The nine-person jury in Musk v. Altman started deliberating this morning at the Ronald V. Dellums courthouse in Oakland. After three weeks of testimony, what got entered into the public record is bigger than the verdict will be: Microsoft’s CEO calling OpenAI’s pre-2023 governance “amateur city,” Ilya Sutskever confirming under oath that Anthropic was approached about taking over OpenAI during the Altman firing, and Sam Altman explaining that Microsoft was never even his biggest investor.

While the lawyers were closing, the labs were opening — three new fronts each, all of them about owning the layer above the model. OpenAI launched a $4B deployment arm, a cybersecurity initiative, and a personal-finance product that connects to your bank. Anthropic is raising at a near-trillion-dollar valuation and pushed Claude into small business and legal practice. Microsoft canceled internal Claude Code licenses mid-trial. Google’s threat intel team said it identified the first AI-built zero-day before its operators could deploy it. Apple’s lawyers got their first warning shot.

The trial closes one chapter. The labs already started the next.

🚨 Closing Arguments

Three weeks of testimony, four bombshells, jury deliberates today

Closing arguments in Musk v. Altman wrapped Thursday. Judge Yvonne Gonzalez Rogers sent the nine-person jury home for the weekend; deliberations begin this morning. The jury’s verdict is advisory — Gonzalez Rogers makes the final liability call. The remedies phase also opens today. Musk wants Altman and Brockman removed, OpenAI’s October 2025 recapitalization unwound, and “wrongful gains” returned to OpenAI’s nonprofit.

What landed on the public record over the last two weeks doesn’t depend on the verdict.

Satya Nadella, May 11. Microsoft’s CEO called OpenAI’s pre-2023 governance “amateur city” — the “jealousies” framing he gave the jury about why Altman was actually fired. Lawyers walked him through an April 2022 internal email where he wrote “I don’t want Microsoft to become IBM while OpenAI becomes Microsoft” — the company that helped Microsoft up off the floor in the 80s and watched Microsoft eat its lunch in the 90s. Nadella testified that Microsoft has invested $13B in OpenAI since 2019. Asked whether Microsoft demanded Altman’s reinstatement during the firing: never. Microsoft didn’t pick the CEO. The board flinched.

Ilya Sutskever, May 11. OpenAI’s co-founder, whose stake is now worth around $7 billion, took the stand and gave the day’s most consequential testimony. Asked under oath whether he had ever promised Musk that OpenAI would remain a nonprofit, Sutskever “rejected the framing entirely”: “I made no such promise.” On the November 2023 firing, he testified he had been collecting documentation on Altman for over a year. And the bomb — within 48 hours of Altman’s removal, calls began about merging OpenAI into Anthropic, with Dario Amodei taking over the combined entity. The deal died on “practical obstacles.” Sutskever said he was “very unhappy about it.” Helen Toner, named as the merger’s most-supportive board member, publicly disputed Sutskever’s recollection on X the same day.

Sam Altman, May 12. Altman took the stand to make a specific point on the partnership question: Microsoft was never his biggest investor. Asked by a Microsoft attorney whether SoftBank’s $30B was bigger than Microsoft’s $13B, Altman: “About 2.5 times larger. I tripped up on that.” He named Amazon ($50B), Nvidia ($30B), and SoftBank ($30B) as each having invested more than Microsoft.

Zico Kolter, May 12. OpenAI’s board safety committee chair followed Altman to the stand. Kolter testified that his committee has formal authority to delay model releases — and that it has used it. He didn’t name the models.

9 jurors

decide today whether OpenAI was ever really a charity.

The verdict is advisory. Judge Gonzalez Rogers decides liability. The remedies phase — which the jury doesn’t see — also opens today and runs concurrently. If Musk wins on remedy, Altman and Brockman could be removed from OpenAI and the October 2025 recapitalization unwound. ChatGPT has 700 million weekly users. The Information has reported OpenAI internal projections of roughly $14 billion in 2026 losses.

The TNG read: the testimony does more damage to the partnership era than the verdict possibly can. Sutskever’s testimony explains why Anthropic exists in the form it does. Altman’s testimony explains why no single hyperscaler can claim the company anymore. Nadella’s “amateur city” testimony explains why Microsoft spent the same week cutting internal Claude Code licenses. And Kolter’s testimony — quiet, technical, deeply uncomfortable — explains why everyone keeps shipping anyway.

Hype vs. Reality: 9/10 — whatever Gonzalez Rogers rules in two weeks won’t change what’s now discoverable in every future contract negotiation.

📡 OpenAI Goes Vertical

Three product moves, three different markets, same playbook

On Monday May 11, OpenAI announced what amounts to a new company strategy in a single press cycle. Three launches in five days, all targeting the same idea: stop selling tokens, start owning the workflow.

OpenAI Deployment Company (May 11). A majority-controlled subsidiary, $4 billion in initial capital, valued at $14B post-money per Axios. 19 investment and consulting partners — TPG lead; Advent, Bain Capital, Brookfield co-leads; Goldman, SoftBank, Warburg Pincus, McKinsey, Capgemini in. Same-day acquisition: Tomoro, a London applied-AI consultancy, bringing ~150 forward-deployed engineers on staff. Fidji Simo, running the unit, told staff that Anthropic’s enterprise-services lead last week was “a wake-up call. We cannot miss this moment because we are distracted by side quests.”

Daybreak (May 11). OpenAI’s answer to Anthropic Mythos and Project Glasswing. Three-tier model access: GPT-5.5 (general), GPT-5.5 with Trusted Access for Cyber (verified defensive workflows), and GPT-5.5-Cyber (red team / pen test). Codex Security is the agentic harness — it scans a repo, builds a threat model, runs scoped exploits in isolated environments, proposes patches. Unlike Mythos, which is Glasswing-gated to a closed consortium, Daybreak is open to request — any organization can apply for a scan, though OpenAI hasn’t committed to broad general availability. TAC partners: Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, NVIDIA, Oracle, Palo Alto Networks, Sophos, Zscaler. Enterprise side: BoA, BBVA, BlackRock, BNY, Citi, Goldman, JPMorgan. OpenAI said its prior cyber model, GPT-5.4-Cyber, has helped fix more than 3,000 vulnerabilities since April.

ChatGPT + Plaid (May 15). ChatGPT Pro users in the US can now connect their bank accounts, credit cards, brokerages, and loans. Plaid handles the auth, 12,000 institutions supported — Schwab, Fidelity, Chase, Robinhood, AmEx, Capital One, Citi. ChatGPT reads balances, transactions, investments, liabilities. It can’t see full account numbers or move money. The Hiro Finance team OpenAI acquired in April — founded by Ethan Bloch, who sold Digit to Oportun — shipped it. Intuit integration is coming for tax flows and credit-card approval odds. OpenAI claims 200M+ users already ask ChatGPT finance questions monthly. Worth noting: Perplexity shipped its own Plaid integration one day earlier, on May 14. Less press, same playbook.

Translation: the model lab is now the implementation vendor, the security vendor, and the personal-finance assistant. The bottleneck stopped being model intelligence months ago. It’s who owns the workflow, who connects the data, who carries the deployment risk. OpenAI just made a $4B+ bet that the answer is them.

Hype vs. Reality: 8/10 — each is a real product with real partners, not a press release. The risk is execution: the forward-deployed-engineer business is brutal at margin, and a personal AI that sees your bank account is one OAuth misconfiguration away from a Senate hearing.

💰 Anthropic’s Counter

$900B, the SMB push, Japanese megabanks, and the Gates Foundation

Anthropic is raising at a valuation that would put it ahead of OpenAI. Bloomberg reported May 12 that Anthropic is in early talks for a round of at least $30B at a $900B+ valuation, with a target close at end of May. The deal isn’t finalized, no term sheet signed. Reporting elsewhere this week put Anthropic’s annualized revenue at roughly $44B (up from $9B a year ago), with the round positioned as the final private financing before an October IPO target. If it closes, it surpasses OpenAI’s $852B.

What the money is buying is on the record across the same week:

Claude for Small Business (May 13). A toggle inside Claude Cowork, 15 pre-built agentic workflows, 15 reusable skills. Integrations with QuickBooks, PayPal, HubSpot, Canva, DocuSign, Microsoft 365, Google Workspace. A 10-city US training tour kicked off in Chicago May 14 — stops include Dallas, Salt Lake City, San Jose, Indianapolis, and Birmingham. Anthropic is going where OpenAI Deployment Company can’t easily go: the SMB market doesn’t pay for forward-deployed engineers.

Claude for Legal (May 12). Twelve new legal-practice plugins shipped — “commercial counsel,” “employment counsel,” “litigation associate,” “law student” among them. Integrations: Thomson Reuters Westlaw and CoCounsel, Harvey, Box, Everlaw, DocuSign. The legal stack, fully wired into Claude.

Mythos to Japan’s megabanks (May 12–13). US Treasury Secretary Scott Bessent met with MUFG, SMBC, and Mizuho in Tokyo on Tuesday and signaled US support for granting them access to Mythos. Nikkei broke it Wednesday; Bloomberg confirmed. Onboarding targeted for end of May. Japan’s FSA stood up a public-private working group — first meeting Thursday May 14 — chaired by Mizuho’s CISO, including the megabanks plus BoJ and Finance Ministry officials. This is the first time a Japanese company has been added to Anthropic’s restricted Glasswing rollout. JPMorgan was the only bank in the original cohort. Last week we covered the Anthropic financial-services JV; this week it crossed an ocean.

Gates Foundation $200M (May 14). Five-year partnership applying Claude to global health, agriculture, and education in low-income countries. Small money relative to Anthropic’s run rate, big strategic signal: a foundation built by a Microsoft founder is now publicly partnered with the company Microsoft just told its engineers to stop using internally.

The TNG read: OpenAI’s verticalization is top-down — Deploy Co, Daybreak, Plaid hit consulting, security, and consumer at the same time. Anthropic is going bottom-up — SMB workflows, legal practice, foreign megabanks, a global-health foundation. Different routes to the same destination: a workflow you can’t replace without replacing them.

Hype vs. Reality: 7/10 — $900B is a number, not a close. The SMB and legal verticals are real but unproven at margin. Mythos-to-Japan is the most concrete win, and it’s the one nobody covered.

🔬 The Zero-Day That Almost Got Loose

Google GTIG put a body on the AI-cyber arms race

On the same day OpenAI launched Daybreak, Google’s Threat Intelligence Group said it had identified the first zero-day exploit it believes was developed with AI — and that its proactive counter-discovery may have prevented a planned mass exploitation event.

The target was a popular open-source admin tool. The vulnerability was a 2FA bypass — not a memory corruption bug, not an input-sanitization failure, but a semantic logic flaw stemming from a hardcoded trust assumption. The kind of bug a fuzzer doesn’t find but an LLM that can read developer intent does.

GTIG identified the exploit as AI-generated by fingerprint: textbook docstrings, structured Python with the clean _C ANSI color class (a tell — humans don’t reach for that), and the giveaway — a hallucinated CVSS severity score embedded directly in the code. The model invented a number. The attacker built it into the payload anyway. Then GTIG found it, worked with the vendor on a patch, and disrupted the operation before it deployed.

GTIG’s chief analyst John Hultquist: “the race has already begun.” The same report flagged North Korean APT45 running thousands of AI-generated exploit probes, China-linked groups deploying Strix and Hexstrike against Japanese targets, an Android malware family called PromptSpy that calls Gemini at runtime, and Russian actors padding payloads with AI-generated decoy code (CANFAIL, LONGSTREAM, HONESTCUE) to confuse reverse engineers.

Translation: the cybersecurity arms race got a body the same week the two largest labs shipped weapons. Daybreak on Monday. Mythos to Japan on Tuesday and Wednesday. GTIG report on Monday. This is what that week looks like.

Hype vs. Reality: 8/10 — one near-miss isn’t a flood. The hallucinated CVSS score is darkly funny but the underlying capability is real. Defender iteration speed needs to step up by an order of magnitude, this quarter.

⚡ Microsoft Picks Its Own Stack

Claude Code is out internally. Claude is still in everywhere else.

On May 14, The Verge’s Tom Warren reported that Microsoft is canceling most internal Claude Code licenses. Microsoft’s Experiences + Devices division — the engineers behind Windows, Microsoft 365, Outlook, Teams, and Surface — must move to GitHub Copilot CLI by June 30, the close of Microsoft’s fiscal year.

EVP Rajesh Jha, who’s running the division until his July retirement, called it tool unification. Sources cited by Warren added two more reasons: external license cost-cutting ahead of the new fiscal year, and — most pointed — Claude Code’s internal adoption was undermining Copilot CLI. Microsoft was paying Anthropic for thousands of seats while its own product went neglected by the engineers who were supposed to use it.

The relationship isn’t dead. Claude models remain available through Copilot CLI alongside OpenAI and Microsoft’s own models, M365 Copilot consumer features still use Claude for specific tasks, the Foundry agreement is intact, and Cowork features are unaffected. Microsoft is narrowing how and where Claude shows up — not whether.

Translation: Nadella’s “I don’t want to be IBM” testimony arrived in court Monday. Microsoft’s “we’re going to shape our own tools” memo arrived on Thursday. Same week, same message.

🚨 The Apple Crack

OpenAI is working with outside counsel on legal options against Apple. Bloomberg’s Mark Gurman broke May 14 that OpenAI lawyers are actively working with an outside legal firm on a range of options that could include sending Apple a formal breach-of-contract notice — short of filing suit, for now. Any action would wait until after the Musk-Altman trial concludes.

The grievance has two layers. The first is money: OpenAI thought ChatGPT integration into Siri, iOS, iPadOS, and macOS would generate “billions per year in subscription revenue,” and the actual number “hasn’t come close.” The second is positioning: deeper integration across Apple apps, prime placement within Siri, real promotion. Apple delivered none of it. The third-party extensions reveal we flagged last week — letting users pick Claude or Gemini as the default — apparently isn’t the issue, because the partnership was never meant to be exclusive.

Hype vs. Reality: 6/10 — outside counsel ≠ filed complaint. But the leak is the message: Apple, you have leverage problems, and the storefront is not the contract. WWDC is in three weeks.

👀 Quick Signals

The $1.5B Anthropic copyright settlement is hanging. Judge Martinez-Olguin (who took over after Alsup’s retirement) held the fairness hearing May 14 and did not grant final approval — she pressed for more on the $187.5M attorney fee request and lead-plaintiff payments. 92%+ of 480,000 works are claimed; class members get ~$3,000 each. The day before, Dave Eggers, Vendela Vida, and 25 other opt-out writers filed a new complaint. Largest known US copyright settlement, still unsigned.

DeepMind’s union recognition deadline hits Tuesday May 19. UK DeepMind staff voted 98% to unionize on May 5; Google has 10 working days to voluntarily recognize the Communication Workers Union and Unite the Union, or trigger UK Central Arbitration Committee proceedings. Tuesday is the cutoff. A yes would be the first frontier-lab unionization in the West.

Google I/O is Tuesday. This week’s Android Show I/O Edition (May 12) previewed Gemini Intelligence on-device and Googlebook — a new Gemini-native laptop category. The main keynote drops May 19. Whatever ships, we’ll cover next week.

🎯 The Playbook

What to do this week

  1. Audit your single-vendor exposure. Microsoft just demonstrated how fast a hyperscaler can cancel licenses for an upstream tool its own engineers preferred. If your stack depends on Claude Code, Codex, or Cursor and your buyer is a procurement officer at a $1T enterprise, the procurement officer just got a playbook.
  2. Consider requesting a Daybreak scan. If your security policy permits third-party scanning of your repos, this is worth a look — compare results against your current review stack. Mythos isn’t on offer to outside organizations. The asymmetry is the actual signal: the lab that ships its security AI to everyone is making a different bet than the lab that gates it. Treat that as a procurement input.
  3. Watch the Plaid pattern. OpenAI shipped a financial OS in a week. The opportunity isn’t building a competitor — it’s building the audit layer, the redaction layer, the “what did the AI see” layer that every regulated industry will require before agents touch their data.
  4. If you’re in Birmingham, Alabama: the Claude for Small Business tour is coming here. SMB owners — show up. Builders — the next AI meetup is on the calendar, and the free audit is still open.

The verdict comes today. The labs already shipped.

— Matt