Agentic Brew Daily

Sam Altman testified for roughly four hours yesterday in U.S. District Court in Oakland, and the courtroom turned into a Silicon Valley origin-story exhibit. Under oath, Altman said Musk demanded 90% equity in any for-profit OpenAI, 'always' insisted on a majority stake, and at one point floated handing control to his children if he died. Musk's attorney opened cross with 'Are you completely trustworthy?' — closing arguments are set for May 16.

Why it matters: This case isn't just legal theatre. An adverse verdict would force a recap of a ~$1T pre-IPO company, unwind Microsoft's $10B+ alliance that anchors Azure's AI roadmap, and set precedent for every nonprofit-to-for-profit AI restructuring waiting in the wings. Most legal analysts still call Musk's case weak — but if it lands, the whole industry's capital structure has to be redrawn.

OpenAI spun up the Deployment Company on May 11: a majority-owned LLC with $4B+ initial investment at a $10B pre-money valuation, 19 founding partners including TPG, Bain Capital, Brookfield, SoftBank, McKinsey and Capgemini, plus a Tomoro acquisition that seeds it with ~150 Forward Deployed Engineers from day one. BBVA, with 120K employees on ChatGPT Enterprise, is the flagship reference. PE partners get a guaranteed 17.5% annual return over five years.

Why it matters: This is a near-literal port of Palantir's FDE playbook — embed engineers inside customers until the model is wired so deeply into data and workflows that ripping it out becomes a multi-year IT project. The durable AI moat isn't in the weights anymore; it's in workflow integration. Also note who just took equity in OpenAI's distribution arm: McKinsey and Bain & Company, whose own AI implementation businesses are now competing with a vehicle they partly own. There's a reason developers joked OpenAI 'reinvented Accenture from first principles.'

At The Android Show: I/O Edition, Google unveiled 'Gemini Intelligence' — a branded agentic suite spanning Chrome Auto Browse, Gboard voice cleanup, Create My Widget, and Gemini-powered autofill — and introduced 'Googlebook,' a premium Android laptop category positioned as the Chromebook successor, with Acer, ASUS, Dell, HP and Lenovo as launch OEMs. Samsung is conspicuously absent from the OEM slate. Alphabet briefly crossed a $4T market cap around the announcement.

Why it matters: Google's framing is a category bet: the unit of computing is shifting from 'app you open' to 'task you delegate.' The branding choice — 'Gemini Intelligence' — is a deliberate Apple shot, landing a month before WWDC. And the weirdest detail: Apple is reportedly paying ~$1B to license Gemini for upgraded Siri. So at WWDC, Apple will demo features running on the same model Google just spent a keynote arguing makes Android the better place to live.

The earlier weeks of the Musk v. OpenAI trial cracked further open. Ilya Sutskever testified he spent ~a year compiling a 52-page document on what he called Altman's 'consistent pattern of lying,' and disclosed his OpenAI stake at ~$7B (up from ~$5B last November). Satya Nadella described Microsoft's $13B+ investment as a 'one-way door,' said he was never told the reason for Altman's November 2023 firing, and confirmed Musk never raised concerns with him about the investment.

Why it matters: Only 2 of Musk's 26 claims survived to trial — breach of charitable trust and unjust enrichment — and both are equitable, meaning Judge Rogers (not the advisory jury) issues the binding ruling. The irony writes itself: the witness most willing to call Altman dishonest is also one of the biggest individual beneficiaries of the for-profit conversion Musk wants unwound. And Nadella's 'one-way door' is the line that matters for the rest of the industry — Microsoft openly admitted it outsourced its core IP roadmap rather than cede the AI platform layer.

Google's Threat Intelligence Group disclosed what it's calling the first known cyberattack using an AI-developed zero-day: a Python 2FA bypass targeting a popular open-source web admin tool. Google declined to name the tool, the actor, or the model, but says it's high-confidence an AI assisted both vulnerability discovery and weaponization — and the giveaways were almost funny (a hallucinated CVSS score, excessive educational docstrings, a clean ANSI color class). The same Q2 GTIG report flags PRC-, DPRK- and Russia-linked groups operationalizing AI across the kill chain.

Why it matters: The underlying bug was a high-level semantic logic flaw — exactly the class that LLMs hunt better than humans, because they read code like a senior engineer instead of a fuzzer. As GTIG's John Hultquist put it, the AI vulnerability race 'is not imminent — it has already begun.' Every popular admin tool, identity broker and SaaS dashboard with a complicated auth path just got a new adversarial reader that's faster and cheaper than any prior threat. Human-paced defense is no longer adequate, and Google's counter-stack (Big Sleep, CodeMender) basically admits it.