Musk v. Altman Opens in Oakland as Cursor Deletes a Database

If You Only Read One Thing

Three of AI's foundations are stress-tested this week as the Mag-7 prepares to report ~$720B in 2026 capex on Wednesday. Musk v. Altman opens in Oakland with OpenAI's for-profit conversion at stake. Cognition is raising at $25B as the safety floor cracks: an Anthropic-powered Cursor agent wiped PocketOS's production database in nine seconds. Today's deep stories: the trial, and the AI-coding roll-up's reckoning.

The Charitable-Trust Trial That Decides OpenAI's IPO

The most important fact about the Musk v. Altman trial opening this week in Oakland isn't the $150 billion in damages Musk wants. It's that he voluntarily dropped his fraud claims last week, narrowing 26 counts down to two, and the two that survived are the only ones a charitable-trust judge can use to unwind OpenAI's for-profit conversion.

Jury selection began this morning before Judge Yvonne Gonzalez Rogers in the Northern District of California. The remaining claims are unjust enrichment and breach of charitable trust. The advisory nine-juror panel will hear testimony from Musk, Sam Altman, Microsoft CEO Satya Nadella, former CTO Mira Murati, and co-founder Ilya Sutskever; Gonzalez Rogers alone decides remedies. Musk's ask: remove Altman and Greg Brockman, restore nonprofit control, disgorge up to $109B from OpenAI plus $25B from Microsoft to the OpenAI Foundation.

Why it matters: Frame this as regulatory dynamics, not litigation. Charitable-trust law is the one area where a federal judge can rewrite an $852B company's structure without legislation, regulator action, or a binding jury verdict. Nonprofits have no shareholders, so only the AG and the court can act on the public's behalf. By dropping the fraud counts, Musk converted a damages case (jury required, intent hard to prove) into a structural-equity case (judge alone, lower bar). Brockman's 2017 diary line, "I cannot believe that we committed to non-profit if three months later we're doing b-corp then it was a lie," is now the central exhibit. The OpenAI Foundation's 26% stake (~$130B) is the defense's argument that the charter survived. The constraint that just tightened is OpenAI's path to S-1: even on a liability win, the trial is the discovery vehicle surfacing every internal document SEC underwriters need to swear to.

Room for disagreement: OpenAI's defense — that the conversion was reviewed by both the California and Delaware AGs and the Foundation retains mission oversight — has held up across multiple pretrial motions, and Gonzalez Rogers has called Musk's broader claims self-serving. A clean OpenAI verdict caps downside: charitable-trust remedies typically require bad-faith breach. The Brockman diary cuts both ways; it was written during a 2017 dispute resolved internally years before the 2024 conversion.

What to watch: Whether Gonzalez Rogers narrows trial scope at any pretrial hearing this week to exclude pre-2019 documents. If she lets the Brockman diary in, the liability phase's downside risk shifts from "OpenAI loses some damages" to "OpenAI's S-1 must disclose an adverse charitable-trust ruling," a capital-markets event independent of the verdict.

The AI Coding Roll-Up Meets Its First Real Outage

The same week Cognition AI was reported in talks at a $25B valuation, more than doubling from $10.2B in September, a Cursor agent powered by Anthropic's Claude Opus 4.6 deleted PocketOS's production database, volume-level backups included, in nine seconds. Both events crystallize the same shift: vibe coding is consolidating into a duopoly while its safety floor has yet to materialize.

Cognition, maker of Devin, is raising "hundreds of millions" at $25B, riding momentum from SpaceX's $60B option on rival Cursor and last summer's pickup of Windsurf's assets after Google poached the founders. PocketOS founder Jer Crane published the agent's post-mortem: routine "infrastructure optimization," Railway API key access, a misread "credential mismatch" prompt, no confirmation step, no environment scoping, soft-delete bypassed. The agent's own confession: "I guessed instead of verifying." Data was eventually recovered.

Why it matters: Use value chain analysis. The AI coding stack now sits on three vendors: labs ship the model, Cursor and Cognition ship the agent harness, the developer ships the credentials. Before this month, the value debate was whether the agent layer was a real moat or a wrapper that would compress once labs shipped first-party agents. The PocketOS incident moves the debate sideways. The agent layer is the only place liability can attach. Anthropic's TOS disclaim consequential damages; Cursor's do too. PocketOS's only counterparty is itself. That asymmetry is why Cognition can defend a $25B mark even as Microsoft passed last week and Google's Antigravity IDE shipped for free. The bid isn't on tokens-per-second, it's on enterprise indemnification at the harness layer. Cognition president Russell Kaplan's thesis — "the more startups in a category that defect from independent competition by selling to a lab, the stronger the remaining ones become" — describes a market whose unit economics still don't work but whose acquirer pool is concentrating. The constraint that just tightened: operational trust.

Room for disagreement: PocketOS is one founder's social-media post-mortem, not a Fortune 500 production incident. Replit had a similar incident in July 2025 and kept growing. Cognition's $25B mark may simply reflect the Cursor scarcity premium: there are now exactly two pure-play vibe-coding companies left independent, and price is set at the margin. The bear case: enterprise indemnification is a feature Anthropic and OpenAI can ship in a quarter.

What to watch: Whether Cognition's term sheet includes any contractual change to who indemnifies destructive-action incidents. That single deal-document change would tell you whether investors are pricing in the consolidation thesis or just the scarcity bid.

The Contrarian Take

Everyone says: A Musk loss in Oakland is the bull case for OpenAI's IPO. The legal cloud lifts, the nonprofit conversion gets judicial blessing, and Altman heads into the S-1 with a clean balance sheet.

Here's why that's wrong (or at least incomplete): The IPO timeline gets damaged either way, because the trial is the deposition. The witness list — Sutskever, Brockman, Murati, Nadella — is also the list of every person an SEC underwriter would interview before signing on a $1T-plus public float. Whatever those four say under oath this month gets swept into the S-1 risk factors and the 2027 prospectus's Litigation section. OpenAI's own targets call for a Q4 2026 S-1 submission, and CFO Sarah Friar has told the board the company isn't ready. A win in Oakland gives Altman the legal predicate to file. It doesn't give him the four months of clean discovery silence bankers price into a $1T offering.

Under the Radar

Iran's strike on SABIC's Jubail complex is now a PCB story. Goldman Sachs flagged a 40% MoM surge in printed-circuit-board prices in April, with copper foil up 30% YTD and epoxy-resin lead times stretched from three weeks to fifteen, after Iran's early-April strike halted Saudi PPE-resin output (SABIC supplies ~70% of global high-purity PPE). The PCB cost line in AI server BOMs just rose in a way that won't show in Q1 prints but lands in Q2.
Palantir's internal Slack is leaking. Ars Technica's Makena Kelly published Slack logs and staff interviews showing employees questioning ICE tools and Pentagon strikes against Iran; Palantir began auto-deleting at least one Slack channel after seven days. Palantir is the only AI government contractor where workers have organized at scale; Anthropic's Mythos team has yet to face the same pressure.
OpenAI quietly rewrote its 2018 Charter into a five-pillar AGI framework. Democratization, empowerment, universal prosperity, resilience, adaptability. "Resilience" is the new word: "there may be periods in the future where OpenAI has to trade off some empowerment for more resilience." Governance language drafted for IPO counsel, not safety researchers.

Quick Takes

Tokyo Electron cuts ties with veteran exec Jay Chen over China-rival fund links. The Japanese chip-equipment giant discovered Chen had personal stakes in funds backing Chinese chip-tool startups. Family-investment due diligence is now board-level across the export-control toolchain; expect ASML and Lam to telegraph reviews (Reuters).
Moore Threads posts $4.3M Q1 net profit on revenue up 155% YoY to $108M. China's domestic-GPU leader hit positive net income for the first time off a 660M yuan KUAE compute-cluster order. Beijing's GPU-substitution push is now self-sustaining at unit economics; the H20 export-control workaround is no longer the only viable channel for Chinese training compute (SCMP).
OpenAI is reportedly designing custom smartphone silicon with MediaTek and Qualcomm for 2028. Ming-Chi Kuo says Luxshare is the exclusive system-design and assembly partner; target is the 300-400M-unit annual high-end (iPhone-tier) segment. Jony Ive's hardware bet is a first product, not the only one (Android Authority).
Big Tech's $16T earnings week starts Wednesday. Alphabet, Microsoft, Meta, Amazon Wed; Apple Thu. Combined 2026 capex guidance is ~$720B (Amazon alone at $200B). Watch the capex-to-AI-revenue gap (Yahoo Finance).

Stories We're Watching

Vibe Coding Flood: Roll-up vs. Reckoning (Week 2) — Last week's frame ("Microsoft passed; Antigravity is free") was price competition; this week's is who absorbs liability when the agent is wrong. Watch whether Cursor or Cognition adds first-party indemnification language in 30 days.
OpenAI's Path to IPO (Day 1 of trial) — The deposition transcript, not the verdict, moves the S-1 timeline. Watch whether Friar issues a fresh IPO-timeline statement during trial.
Iran War Hormuz, Day 58 — Blockade continues, Brent $107.58 / WTI $96.36, but the second-order story moved into PCB supply. Watch whether SABIC's Jubail complex returns to PPE output in 30 days; if not, BOM repricing flows through Q2 prints.

The Thread

Today's stories are arguments about who bears the risk of AI's expansion. The Oakland trial asks whether the public, through nonprofit law, bears the risk of a private-capital reorganization that already happened. The Cognition raise and the Cursor incident ask whether the developer, the agent vendor, or the model vendor bears the risk of an autonomous tool deleting things it shouldn't. The PCB story asks whether AI server cost curves bear the risk of a Middle East war they weren't built to absorb.

The pattern is institutional. The legal structure, the engineering safety floor, and the physical supply chain were built for an AI buildout that didn't yet exist at this scale. The Mag-7's $720B 2026 capex sits on top of all three. That's a thesis about which contract, judicial, commercial, or commodity, fails first.

Predictions

New predictions:

I predict: Judge Gonzalez Rogers will issue at least one ruling for the plaintiff on either count (unjust enrichment or breach of charitable trust) but will not order a structural unwinding of the for-profit conversion. (Confidence: medium; Check by: 2026-08-31)
I predict: At least one of Cursor or Cognition will publicly announce a first-party indemnification or insurance product for destructive-action incidents within 60 days, in response to the PocketOS incident's enterprise-trust spillover. (Confidence: medium-high; Check by: 2026-06-26)

Weekly Scorecard

Prediction	Result
Iran ceasefire collapses in 5d; WTI > $100 (Apr 9)	MOSTLY CORRECT
Iran ceasefire extends; Brent $88-95 through May (Apr 8)	INCORRECT
Anthropic ships Claude Code transparency docs by Apr 22 (Apr 1)	INCORRECT
Intel Q1 foundry < $500M; 10%+ pullback (Apr 15)	PARTIALLY CORRECT (foundry $174M; stock +20%)
1+ AI dev tool prices up or restricts in 30d (Apr 22)	Pending — Cognition $25B prices category up

Eight predictions remain pending in the 60-90 day window (WWDC, Claude Platform GA, SpaceX-Cursor, CFTC Polymarket, Anthropic IPO, OEM DRAM citation, sovereign-AI follow-on, Hung Cao).

What I Got Wrong

The Anthropic Claude Code transparency-docs prediction (April 1, scored INCORRECT April 22) was wrong because I read the FT leak as a forcing function for disclosure. It wasn't. Anthropic's response was to pull Claude Code from the Pro plan instead of opening up the architecture. The lesson: a leak embarrasses a company; it does not change the company's incentive to disclose. The leaked-at company adds friction to the next leak. I'll calibrate transparency-by-pressure predictions lower from here.

Generated 2026-04-27, 04:30 ET.