Subscribe
Share
Share
Embed
Anthropic and OpenAI race to go public
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Tuesday, June second.
In today's briefing we see NVIDIA's Computex keynote delivering Cosmos 3, the first open omnimodel for physical AI and robotics, Anthropic filing a confidential S-1 as it races toward public markets alongside OpenAI, and the AI infrastructure arms race shifting into public market territory while state-level product liability law moves to challenge consumer AI products.
First up - Today in the big model news;
Anthropic filed a confidential draft S-1 with the SEC on June first, formally entering the IPO race alongside OpenAI, which has already filed a public S-1 targeting a Nasdaq listing under SPCX around June twelfth. Anthropic's revenue run-rate hit forty-seven billion dollars in May, up from roughly ten billion a year ago, at a nine hundred sixty-five billion dollar post-money valuation. For enterprise AI buyers, the implication is that quarterly earnings calls will now drive API pricing decisions and margin management in contract negotiations because public-market obligations change the cost accounting for serving large customers.
In other lab news today, NVIDIA's Computex keynote delivered Cosmos 3, the first fully open omnimodel combining language, vision, image generation, video, audio, and action in a Mixture-of-Transformers architecture, purpose-built for robotics and physical AI. Cosmos 3 ships now on Hugging Face and as NIM microservices, with an Edge variant for real-time inference coming soon, and the NVIDIA Cosmos Coalition already has Runway and other partners signed on to build on top of it. For AI PMs building on physical-world agent stacks, the gap between having a language model and having an agent that can reason about what to do next in a physical environment just got narrower because Cosmos 3 provides a fully open, production-available omnimodel rather than requiring teams to assemble separate components.
Alongside Cosmos 3, NVIDIA announced Nemotron 3 Ultra, a five hundred fifty billion parameter open-weights mixture-of-experts model launching June fourth, claiming five-times faster inference and thirty percent lower cost than comparable models at that scale. NVIDIA positions it as the highest-performing open model available, though those numbers need verification once weights ship. The directional move is clear: five hundred fifty billion parameters as mixture-of-experts is becoming self-hostable on Blackwell-class hardware, which meaningfully rewrites the frontier API buy-versus-build calculation for teams that can run their own inference. For AI PMs evaluating cost structures between API consumption and on-premise infrastructure, there is now a meaningful argument to move at least some orchestration onto local hardware because what's runnable on consumer-grade machines has crossed the threshold where it can do real production work.
Elsewhere, MiniMax surfaced with M3, claiming a one-million token context window and fifty-nine percent on SWE-Bench Pro, without releasing weights, which the local LLM community received with skepticism given the pattern of announcements before release. If the benchmark is confirmed at launch, M3 would rank at or near the top of agentic coding benchmarks. For teams building code agents, a million-token context window paired with strong coding benchmarks is significant, but the verification matters because the community has learned to be skeptical of pre-release coding benchmarks.
In the harness, tools and orchestration world; JetBrains released Mellum2, a twelve-billion parameter mixture-of-experts model tuned for low-latency IDE routing and code completion. The move signals that IDE-integrated coding models are now fragmenting by latency and task-specific requirements. For product teams shipping IDE agents, there is now a market tier for latency-optimized coding models competing with general-purpose frontier models because the task-specific coding domain has matured enough to justify dedicated model development.
OpenAI's frontier models including GPT-5.5 and Codex are now available through AWS Bedrock, joining Anthropic's presence on the platform and completing AWS's role as the neutral enterprise distribution layer between competing labs. Both providers flow through a single procurement relationship, unified billing, and AWS's compliance stack. For enterprise procurement teams, expect model access to consolidate through cloud providers because both major labs accepted margin trade-offs for AWS's regulated-industry customer base.
In AI Infra, two major capital signals. OpenAI announced Stargate Michigan, a one-gigawatt data center addition to the broader Stargate buildout - infrastructure at a scale where the natural comparison is power plant output rather than rack count. In parallel, Alphabet announced an eighty-billion-dollar capital raise on June first: thirty billion in immediate public offerings and forty billion via an at-the-market program, plus a ten-billion-dollar private placement directly to Berkshire Hathaway. When Berkshire Hathaway, which has historically avoided tech stocks, acquires ten billion in Google shares specifically for AI infrastructure, it marks a threshold in mainstream capital market acceptance of AI infrastructure as a durable asset class. For enterprise AI teams budgeting capex, the implication is that frontier AI infrastructure spending is now a multi-year, multi-decade public-market commitment because the compute scale required for competitive models exceeds what venture-stage budgets can support.
In AOB: Florida filed the first state-level product liability lawsuit against an AI company on June first, tying OpenAI to an alleged incident where ChatGPT was used to plan an attack, seeking civil penalties, data restrictions, and personal liability from Sam Altman. The legal theory matters most: product liability is a different exposure than privacy or consumer protection frameworks regulators have previously used. If a state court accepts that AI outputs are a defective product rather than a neutral platform service, every consumer-facing AI product team needs to reassess its liability architecture. Multiple state AGs are reportedly monitoring the filing. For product leaders building consumer-facing AI, the shift from platform liability to product liability is material because a successful case establishes precedent that AI outputs are a regulated product category requiring fundamentally different safety architectures and design controls.
That's the briefing. Have a great day.