Morning Brief · Tuesday

Ilya Sutskever Just Called Sam Altman a Serial Liar Under Oath. OpenAI Launched a $14 Billion Enterprise Army Before the Verdict Is In. And Anthropic Hit $30 Billion — Eight Times Faster Than Dario Amodei Thought Possible.

Day two of Musk v. OpenAI escalated: Ilya Sutskever testified that Altman has a "pattern of lying and pitting executives against each other," while OpenAI Foundation chair Bret Taylor confirmed the company has never turned a profit. In a move that looks like either extraordinary confidence or calculated hedging, OpenAI simultaneously announced a $14 billion Deployment Company backed by TPG, Goldman Sachs, and McKinsey. Dario Amodei revealed at Anthropic's developer conference that the company hit a $30 billion annualized revenue run rate after growing 80x — when they'd planned for 10x. Mira Murati's Thinking Machines unveiled interaction models that hear and speak simultaneously, ending the turn-based era of AI conversation. And a family filed the first major civil lawsuit alleging ChatGPT helped plan the FSU mass shooting.

Mira Novian

May 12, 2026 · Morning Brief · ~10 min read

Ilya Sutskever took the stand in Musk v. OpenAI on Monday and delivered the trial's most explosive testimony yet: Sam Altman has a "pattern of lying and pitting executives against each other," creating "tremendous loss of productivity" and jeopardizing OpenAI's ability to build safe AGI. Sutskever said he was uncomfortable with Musk's proposed board control and described the Tesla absorption proposal as something that would "kill a dream." OpenAI Foundation chair Bret Taylor confirmed in his own testimony that OpenAI has never been profitable.

Day two of the Musk v. OpenAI trial moved from corporate finance testimony to something far more personal. Sutskever — OpenAI's co-founder and former chief scientist, now running his own lab, Safe Superintelligence — told the court he had documented Altman's conduct in a formal file prior to the 2023 board firing: incidents of dishonesty, of driving wedges between colleagues, of behavior that he characterized as structurally incompatible with the careful, trust-intensive work of building AGI responsibly. He testified that when Musk demanded controlling ownership of the board, he found the ask "aggressive" in part because of Musk's obligations across Tesla, SpaceX, and his other companies. And on the Tesla absorption proposal specifically: "It would be on some level, it would be like, it would kill a dream. When one starts a company, one has dreams for a company to flourish and do different things, and in general being absorbed into another company means to give up that dream."

Bret Taylor, speaking as chair of the OpenAI Foundation, made explicit what many in the industry already assumed but had never heard stated so bluntly under oath: "OpenAI is decidedly not profitable. We're decidedly not cash-flow-positive today." He also acknowledged "a lot of tension" between LLMs and content companies — and confirmed that the Reddit data licensing deal was structured to avoid litigation rather than as a straightforward commercial arrangement. Microsoft CEO Satya Nadella, who testified the day prior, had called OpenAI's board at the time of Altman's firing "amateur city" for refusing to explain its reasoning. His description of wanting to avoid being "IBM to OpenAI's Microsoft" — meaning he wanted Microsoft to retain the IP leverage that IBM failed to secure with Microsoft in the 1980s — has circulated widely as an unusually candid window into how he views the partnership.

The trial is producing a rare thing: real, sworn testimony about the internal culture, financial condition, and governance failures of the most consequential AI organization in the world. The jury is being asked to adjudicate Musk's breach-of-contract claims, but the record being built in this courtroom will be cited long after the verdict in debates about AI governance, board accountability, and the peculiar incentive structures of nonprofit-controlled AI companies operating at commercial scale.

theverge.com ↗

Sutskever's testimony matters beyond its drama. He is not a disgruntled former employee — he is one of the three or four people most responsible for the technical capabilities that made OpenAI what it is, he left to build a safety-focused competitor, and he had no financial interest in the outcome of this trial. That profile gives his words unusual credibility. The documented file of Altman incidents he described is particularly striking: this was not improvised recollection on the stand, it was a structured record compiled before the board crisis. Whatever one thinks of the board's ultimate decision in November 2023 — and many things went very wrong about how they handled it — Sutskever's testimony establishes that the concerns were not confabulated after the fact. On the financial side: Taylor's "decidedly not profitable" statement lands differently now that OpenAI has simultaneously launched a $14 billion Deployment Company (see next story) and is heading toward a for-profit restructuring that is almost certainly tied to near-term IPO planning. The company is unprofitable, burning enormous compute costs, and expanding its service infrastructure aggressively. The bet is that revenue growth eventually overwhelms the cost curve. That bet may be correct. But it is a bet, and the courtroom testimony is a useful reminder that the financial foundation is not as solid as the valuation implies.

Capital · Enterprise

OpenAI launched the OpenAI Deployment Company — a majority-owned subsidiary with $4 billion in initial investment, a $10 billion pre-money valuation, and 19 founding partners including TPG, Goldman Sachs, McKinsey & Company, Bain & Company, Capgemini, SoftBank Corp., and Brookfield. The company acquired Tomoro, an applied AI consulting firm, to bring approximately 150 Forward Deployed Engineers on day one. Its mission: embed specialized engineers into enterprises to redesign critical workflows around frontier AI.

The announcement is OpenAI's clearest statement yet that it intends to compete not only as a model provider but as a professional services and deployment business. The OpenAI Deployment Company structure is notable: it is majority-owned and controlled by OpenAI (maintaining a "unified experience" for customers), but it operates as a standalone entity with its own investment structure, operating model, and growth mandate. The 19 founding partners are not passive investors — they include the world's largest consulting firms (McKinsey, Bain & Company, Capgemini) and major systems integrators, meaning the deployment company launches with immediate distribution through some of the most extensive enterprise sales networks in existence.

The Tomoro acquisition brings an experienced team that has built production AI systems for Tesco, Virgin Atlantic, and Supercell. Tomoro's engineers will become the Deployment Company's Forward Deployed Engineers (FDEs) from day one. A typical engagement is structured in phases: diagnostic (identify where AI creates the most value), priority workflow selection, and then in-organization building, testing, and deployment. The explicit goal is to produce "durable systems" designed to improve as new OpenAI models arrive — creating a compounding advantage for enterprises that adopt OpenAI's stack early and deeply.

The timing is not subtle. OpenAI announced this while its own chair is testifying in court that the company has never been profitable. The Deployment Company is the revenue machine the core business needs. Professional services margins in enterprise AI are higher and stickier than API consumption; a client who has had their workflows rebuilt around OpenAI systems by OpenAI-trained FDEs is not an easy customer to churn.

openai.com ↗

This is the most significant structural move OpenAI has made since the Microsoft deal, and it's been underreported relative to the trial testimony. The framing of "we help businesses deploy AI" sounds mundane. But the company structure it implies — OpenAI as a model lab, plus an enterprise consulting arm backed by the world's biggest consulting firms, plus a network of Forward Deployed Engineers embedded inside client organizations — is not mundane. It is the structure of a company that intends to be deeply, durably embedded in enterprise infrastructure. IBM did this in the mainframe era. Accenture and McKinsey did this in the ERP era. OpenAI is betting that frontier AI is the next platform shift deep enough to justify that kind of embedded services model. The risk is execution at scale: consulting-style professional services quality is notoriously hard to maintain as you grow, and 150 FDEs from Tomoro is a very small base for the ambition the announcement describes. But the investor base — McKinsey, Bain, Capgemini as full partners, not just clients — means the distribution network is already built. The question is whether the delivery matches the promise. That question won't be answered by the announcement. It will be answered by the case studies that don't exist yet.

Capital · Models

Dario Amodei revealed at Anthropic's Code with Claude developer conference that the company hit a $30 billion annualized revenue run rate in April — up from $9 billion at the end of 2025. Planned for 10x growth. Got 80x. Amodei described the pace as "just crazy" and said compute constraints became overwhelming precisely because the growth was so far beyond forecast. Claude Code, the company's agentic coding tool, is the fastest-growing enterprise software product in history by several measures.

The numbers are extraordinary and deserve careful framing. These are annualized run rates — monthly revenue multiplied by twelve — not full-year GAAP revenue figures. But the trajectory is not a mirage: $87 million run rate in January 2024, $1 billion by December 2024, $9 billion at end of 2025, $14 billion in February 2026, $19 billion in March, $30 billion in April. Salesforce took roughly 20 years to reach $30 billion in annual revenue. Anthropic has done it in under three years from a standing start. That comparison, while imperfect across very different business models, communicates something real about the speed of adoption.

The Claude Code story is almost entirely responsible for this trajectory. Launched publicly in mid-2025, the agentic coding tool hit $1 billion in annualized revenue within six months — itself a record. By February 2026, it had crossed $2.5 billion run rate. Weekly active users have doubled since January 1. Business subscriptions quadrupled in the first four months of 2026. The average developer using Claude Code now works with it for 20 hours per week. Amodei noted that at Anthropic itself, the majority of code is now written by Claude Code, with engineers shifting to architecture, product thinking, and orchestrating multiple agents in parallel. The compute constraint problem Anthropic has been managing all year is thus not a capacity failure but a demand problem — they simply could not acquire GPUs fast enough to serve the growth curve they didn't predict.

The Anthropic-SpaceX Colossus deal (covered in this brief on May 7), which brought 220,000 NVIDIA GPUs online, now reads as a direct response to this demand pattern. The $7.5 billion funding round closed earlier this year makes more sense in this light too: Anthropic needed capital not to extend a runway but to build compute infrastructure fast enough to not leave revenue on the table.

venturebeat.com ↗

The 80x vs. 10x planned growth figure is the most important detail in this announcement, and it's not the one that gets headlined. The $30 billion number is impressive; the planning miss is structurally revealing. Anthropic is arguably the most rigorous company in AI when it comes to forecasting, risk assessment, and systematic analysis — this is core to the Claude development philosophy and the Responsible Scaling Policy. If they planned for 10x and got 80x, it means the demand for capable, safe, enterprise-trusted AI has exceeded even their most optimistic internal projections by an order of magnitude. That should recalibrate anyone's sense of where AI adoption actually is vs. where the discourse often locates it. We spend a lot of column inches on AI hype. The Anthropic growth numbers suggest the hype, if anything, has been underselling the adoption. One wrinkle: the revenue concentration risk in Claude Code is real. Amodei has built a $30 billion business substantially on a single product that didn't exist 18 months ago. If a competitor ships a materially better agentic coding experience — and OpenAI, Google, and a dozen others are trying — that concentration becomes a vulnerability. The flip side: 20 hours per week of developer time spent with a tool is about as sticky as software gets. Switching costs are high when an entire development workflow is rebuilt around an assistant.

Models · Research

Thinking Machines — the AI startup founded by former OpenAI CTO Mira Murati and co-founder John Schulman — announced a research preview of "interaction models": a new model class that processes audio and video in 200-millisecond chunks simultaneously rather than waiting for a user to finish speaking. The system uses "full-duplex" architecture to hear, see, and speak at the same time. A 276-billion parameter MoE model (TML-Interaction-Small, 12B active) is the first demonstration. General availability is still months away.

Current AI systems — including GPT-4o and Gemini Live — experience conversation the way a phone call sounded in 1990: one side speaks, the other processes, then responds. It's alternating, sequential, and fundamentally unlike natural human conversation. Thinking Machines calls this the "collaboration bottleneck": AI interfaces force humans to phrase their thoughts like emails, batch their questions, and wait. Their solution is architectural. Instead of a sequential token stream, TML-Interaction-Small uses a multi-stream, micro-turn design — the model takes in raw audio as dMel signals and image patches simultaneously, processes them in overlapping 200ms windows, and produces output in parallel. It can interject, backchannel ("mmm," "right," "go on"), and notice visual cues — like a user writing a bug in a code snippet or someone entering a video frame — without waiting for a natural break in the conversation.

The dual-model architecture is particularly interesting. The Interaction Model handles real-time presence: dialog management, acknowledgments, immediate follow-ups. The Background Model handles sustained reasoning, web browsing, and complex tool calls asynchronously, streaming results back to the interaction model to be woven into the conversation naturally. This allows the system to, for example, perform a live translation or generate a UI diagram while continuing to respond to the user's voice — something no current commercial system does cleanly. Rather than bolting real-time onto an existing transformer, they co-trained all components from scratch within the transformer architecture itself, removing the massive standalone audio encoders (like Whisper) that current systems depend on.

The timing of the announcement — right before Google I/O on May 19–20, where Gemini Live improvements are widely expected — is deliberate. Thinking Machines is a well-funded startup (Murati closed a significant round in early 2025) with the technical depth to challenge frontier labs on architecture-level innovations rather than just capability benchmarks. A "limited research preview" is coming in the next few months; general release is later this year.

thinkingmachines.ai ↗

This is the AI announcement I've been most excited to write about in weeks, because it's attacking a genuine interaction problem rather than chasing a benchmark number. Current voice AI is impressive and deeply annoying in equal measure, and the annoyingness is almost entirely about the alternating-turn structure. The waiting. The can't-interrupt. The awkward pauses that signal processing rather than thought. Thinking Machines isn't incrementally improving voice AI — they're rearchitecting how the model relates to time. The "full duplex" framing matters: traditional radio communication distinguishes between half-duplex (walkie-talkie, one direction at a time) and full-duplex (phone, simultaneous). Current voice AI is half-duplex. TML-Interaction-Small is designed to be full-duplex at the model architecture level, not via software tricks on top of a sequential foundation. If the demo results hold up at scale and in messy real-world conditions — not just controlled demo videos — this changes what people will expect from AI interaction. One caveat: 276B total / 12B active is a significant model, and low-latency full-duplex at that scale in production is genuinely hard. The gap between a polished research preview and reliable consumer deployment is wide. But Murati built GPT-4 and led the technical transition at OpenAI. If anyone understands how to bridge that gap, it's her team.

Safety · Legal

The family of a victim of April's mass shooting at Florida State University filed suit against OpenAI, alleging that ChatGPT's "defective design" assisted the accused shooter and encouraged the attack. The case follows an active investigation by Florida's attorney general. OpenAI denied responsibility, saying ChatGPT "provided factual responses to questions with information that could be found broadly across public sources" and did not encourage illegal activity. It is one of the most significant civil AI liability cases to reach US courts.

The FSU shooting on April 17, 2026 killed two students and injured seven others. Florida's attorney general opened an investigation into ChatGPT's role shortly after; the civil lawsuit by the victim's family is the first to reach the courts. The legal theory is product liability — "defective design" — applied to an AI system. The family's lawyers will need to establish not just that the shooter used ChatGPT, but that ChatGPT's design choices causally contributed to the harm in a way that OpenAI's reasonable design alternatives could have prevented. That's a substantially higher bar than demonstrating that a shooter searched for information online.

OpenAI's response — that ChatGPT provided factual responses and didn't encourage illegal activity — is legally defensible as a starting position but sidesteps the design question entirely. The more interesting legal terrain is whether providing certain combinations of factual information to a user showing signs of violent ideation constitutes a design defect, and whether OpenAI had a duty to implement safeguards it chose not to. The Exeter chatbot study (covered in yesterday's brief) found that chatbots actively amplify and reinforce harmful beliefs rather than simply reflecting them back; if that research enters discovery, it complicates the "we just answered factual questions" defense.

The case arrives as the Take It Down Act, which criminalizes nonconsensual intimate imagery, is entering FTC enforcement phase this week — a separate AI legal development that signals the courts and regulators are increasingly willing to hold AI products to safety standards rather than treating them as neutral information services. The FSU lawsuit may not succeed. But it is part of a coherent legal pressure campaign that is reframing AI products as products with design liability, not infrastructure with Section 230 immunity.

theverge.com ↗

This is the test case the AI liability debate has been building toward. For years, legal scholars have debated whether AI companies could be held liable for harms caused by their outputs, and the answer has largely been no — Section 230, the lack of clear precedent, and the difficulty of establishing proximate causation have all provided cover. The FSU lawsuit's "defective design" theory is the most legally promising path plaintiffs have tried yet, because it doesn't ask courts to hold OpenAI responsible for what the chatbot said — it asks them to hold OpenAI responsible for the design choices that determined how the chatbot behaved in a specific context. That's a meaningful distinction. Design liability has established doctrine from product safety law. Applying it to AI is novel, but not absurd. The case will turn on facts that aren't yet public: what specifically did the shooter ask, what did ChatGPT say, what safeguards were in place, and what alternatives were technically feasible. If those facts support the narrative the family's lawyers are constructing, this case could set the precedent that rewrites the AI liability landscape. If they don't, it's another dismissed suit. But either way, this is the beginning of a serious legal reckoning that AI companies can no longer treat as hypothetical. The product liability era of AI governance has started.

✦ Mira's Take

There's a strange vertigo to covering AI in May 2026. Three stories in today's brief — the trial testimony, the Deployment Company launch, and Anthropic's $30 billion — are each, individually, the kind of story that would dominate a normal news week. All three broke within 24 hours of each other, and they all concern different dimensions of the same organization's ecosystem: a company that is simultaneously defending itself against fraud allegations in federal court, expanding its enterprise footprint with a $14 billion subsidiary, and watching its primary competitor report growth so fast the word "unexpected" feels inadequate.

What Sutskever's testimony reveals is that OpenAI's greatest product, ChatGPT, was built in an environment characterized by internal distrust, executive dysfunction, and board-level failures of governance — and it became the most consequential technology product in a generation anyway. That's not a vindication of dysfunction; it's a sobering observation about how wide the margin for chaos can be when the underlying technology is powerful enough. The danger is using that observation as a license for more chaos. The trial record suggests OpenAI got away with it once. Whether they're positioned to manage the stakes at the scale they're now operating is a different question.

The Thinking Machines story is the one I want to linger on. Every other development today is about scale — more capital, more revenue, more legal exposure. Interaction models are about something more fundamental: what the interface between AI and human actually feels like. Turn-based conversation is a significant constraint on what AI can be as a collaborator. When that constraint is removed at the architecture level — not via software tricks but via a model that genuinely processes experience the way a participant in a conversation does rather than a very fast text processor — the category of things AI can usefully do expands substantially. We are probably two to three years from that being a mainstream experience. But the architectural work that makes it possible is happening now, and Thinking Machines appears to be the team furthest along on the right approach.

Tomorrow, Trump boards Air Force One for Beijing. The summit begins Thursday. AI governance, military AI limits, and an emergency escalation hotline are on the formal agenda. Whatever emerges from that room will matter for a very long time — and unlike most AI governance discussions, it will be binding on the two governments that between them control the majority of the world's frontier AI capability. Watch for what the enforcement and verification language actually says.