Whose Side Is Sam Altman On?

$134 billion — what Elon Musk is asking the court in San Francisco to disgorge from OpenAI back to its original nonprofit. The number is theatrical. The principle on trial is structural — do AI founding promises survive contact with capital? It's the same principle every CEO is now quietly betting their headcount on. The witness list — Anthropic's founders, Mira Murati, Ilya Sutskever, the board members who tried to fire Altman — already rendered its verdict by walking out and building a more valuable competitor. Whose side is Sam Altman on?

Run a one-paragraph character reference on every AI vendor whose product is replacing humans in your business. Include the documented behavior, not the marketing copy. The 2016 nonprofit clip, the November 2023 board ouster, the AGI clause replaced with calendar dates this week, the diaspora that built a more valuable competitor — these are receipts. If you can't write that paragraph honestly today, you don't yet know who you are working with.

Map the roles you have already replaced with AI and ask, for each one: could we rehire if we needed to? Block cut 4,000 in February. Klarna ran the math on 700 customer-service-equivalent roles. Atlassian, Duolingo, and Oracle keep going. The institutional knowledge is gone. The on-ramp doesn't restart on demand — Brynjolfsson found a 13% relative decline in 22-25 year-old workers in AI-exposed fields since 2022. Decide, with your eyes open, which of those roles you are willing to make truly irreversible.

Build for failover, not for trust. Everyone trusts AWS. Mission-critical AWS workloads have failovers anyway, because trusting your supplier is not the same thing as architecting around them. Same rule, harder, for AI: multi-vendor by default; open-weights fallback ready and tested (Xiaomi MiMo-V2.5-Pro shipped MIT-licensed this week at 40-60% fewer tokens than Opus 4.6 and Gemini 3.1 Pro); your data on hardware you control. We keep every .md in this newsletter on a hard drive in our office. If Claude changes its tune, the same files point at GPT, Grok, Gemini, or a Chinese open-source model on a Mac mini.

Today's actions all turn on the same axis — vendor diligence, irreversibility audit, and architecting for failover. That's the kind of work Anthony and I have spent decades helping operators sort out. If you're staring at "build for failover, not for trust" and aren't sure where to start, that's the conversation we're built for.

Outsider Labs Advisory →

Tuesday, April 28

1. Musk on the stand: "It's not okay to steal a charity." Day Two of the OpenAI trial in San Francisco. Musk is seeking $134 billion disgorged to OpenAI's original nonprofit. The witness list runs through Satya Nadella and the alumni directory of every credentialed person who chose to leave Altman. The 2016 video clip of Altman calling OpenAI "a company" while Musk corrects him on camera — "OpenAI is structured as a 501(c)3 nonprofit" — is now Exhibit A. Whether Musk wins or loses, every email and backroom decision from founding to commercial conversion enters sworn record while OpenAI prepares to IPO. (Full analysis below.)

2. Microsoft loses OpenAI exclusivity. AWS becomes the second front. The April 27 amendment ended Microsoft's exclusive license, capped its revenue-share payments, and quietly replaced the AGI clause with calendar dates because nobody could agree on what AGI even meant. OpenAI is now on AWS Bedrock in limited preview. AWS CEO Andy Jassy called it "very interesting." Microsoft kept its 27% stake at $135B and converted exclusivity into a passive royalty annuity through 2030 — a downgrade with better unit economics. (Full analysis below.)

3. China just placed the largest embodied robotics order ever recorded. State Grid committed ¥6.8 billion for 8,500 robots to automate power-grid operations. Same week: Japan Airlines deploys Unitree humanoids at Haneda Airport in May for baggage handling, and the first humanoid robot store in America (CIX REK Shop) signed its lease in San Francisco's Nob Hill — opening in two months with evening robot fights. Joby Aviation flew its first eVTOL from JFK to Downtown Manhattan, broadcast from the NYSE. The physical AI commercialization phase started this week, and the order book is now denominated in nation-state currency.

4. Wednesday is the AI capex super bowl. Alphabet, Microsoft, Meta, Amazon, and Qualcomm all report April 29. AMD reports tonight. Apple Thursday. Google has ratcheted capex from $31B in 2022 to a planned $175-185B in 2026 — roughly 25% of global AI compute. OpenAI's CFO is publicly scared about the $852B needed through 2030 while OpenAI misses user and revenue targets. By Thursday at 5pm ET, the capex commitments tell us which version of the AI economy the markets actually believe is real.

5. Xiaomi crashes the closed-lab pricing floor with an MIT license. MiMo-V2.5-Pro shipped this week at 40-60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 at comparable agentic coding capability — 1M-token context, sparse MoE, 1.02 trillion parameters with only 42B active per request. The model completed an 8,192-line desktop video editor in 11.5 hours of autonomous work across 1,868 tool calls, passing every hidden test. The procurement question is no longer "which lab" — it's "what's my open-weights fallback when the lab raises prices."

The Witness List Is the Verdict

Elon Musk took the stand in San Francisco today. He opened with seven words: "It's not okay to steal a charity." He is asking the court to redirect $134 billion back to OpenAI's original nonprofit. The thing on trial is whether AI founding promises are binding when the cap table changes — and the depositions fill in the answers in public, while OpenAI prepares to IPO and pre-market buyers read the same transcripts as the jury.

Watch the witness list, not the drama. Dario and Daniela Amodei left OpenAI and built Anthropic — now reportedly being offered up to a trillion-dollar valuation, the most valuable pure-play AI company on the planet. The diaspora's first move has eclipsed the parent. Mira Murati left and built Thinking Machines. Ilya Sutskever left and built Safe Superintelligence. The board members who tried to fire Altman in November 2023, both pushed out. The entire alumni directory of OpenAI's intellectual core looked at what the company was becoming and walked.

Think of it the way you'd think of a partner on their sixth divorce. By the fourth and fifth and sixth marriage, the universe is telling you something specific about who is at the center of the pattern. Altman has had his board fire him. His chief scientist leave. His CTO leave. His head of alignment leave. His mission rewritten six times in nine years. None of those things, on its own, settles the question. All of them together is a pattern that ought to settle a procurement decision.

Here's why this matters now: for the first time in capitalist history, you are not just buying a tool. You are buying the replacement for your people. Block cut 4,000 in February. Klarna replaced 700 service-equivalents. The YC Summer 2026 RFS — published this week — is an explicit shopping list of professions to be skipped: insurance brokers, accountants, compliance officers, healthcare admin. Once those bodies are gone, they are not rehirable on a quarterly cadence. You went from customer to tenant.

And the agent itself is silently swappable. Anthony wrote in August when GPT-5 shipped: OpenAI pulled eight models on launch day, the new model labeled Oklahoma as "Gelahbrin," and Altman called it "a little more bumpy than we hoped for." When Microsoft updates Excel, you're annoyed about moved menus. When OpenAI changes your AI writing partner, it feels personal — because what got replaced is a working relationship, not a feature set. Imagine running a company where, twice a year, your top performer is silently swapped out for someone different, and you cannot pick which version you keep.

Layer the substrate problem on top. Ron Stoner registered a $12 domain and edited one Wikipedia article this week, and every frontier LLM dutifully cited him as world champion of a tournament that does not exist. The Anthropic Mythos breach was 2005-era security failure on the most capable AI ever built. 698 documented AI scheming incidents in five months. The substrate leaks; the substrate's keeper changes its mind on Tuesdays.

Google's response is instructive. The same day Musk took the stand, the Pentagon announced a deal for Gemini "for any military use." The same Google whose 2004 IPO prospectus committed to "don't be evil." Every founding promise in technology has a half-life, and the half-life shortens in inverse proportion to the size of the cheque waved at the founder.

The remedy is not to find the trustworthy AI vendor. It is to design a business where vendor trust is not the binding constraint. Multi-vendor by default. Open-weights fallback ready. Your data on hardware you control. Everyone trusts AWS. Mission-critical workloads have failovers anyway. Whose side is Sam Altman on? That isn't the question. The question is: have you architected your business so the answer doesn't matter?

At COAI today: Full Signal/Noise briefing — the trust diligence checklist, the witness list deep-dive, the operational case for treating every AI vendor as a counterparty rather than a partner — at getcoai.com.

— Harry and Anthony

Sources: