On June 25, the frontier AI labs made moves on three layers of their stack at the same time. OpenAI and Broadcom unveiled Jalapeño, a custom chip built to run, not train, large language models at scale.
On Tuesday, June 25, the frontier AI labs stopped pretending they were just model companies. In a single news cycle, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom inference chip, built to run, not train, large language models at the kind of scale the company says it will need by the end of the decade. Anthropic went public with an accusation that Alibaba had mounted a large-scale distillation campaign against its Claude models. And the TLDR AI daily, summarizing a Bloomberg wire report, said two Google DeepMind researchers, Jonas Adler and Alexander Pritzel, had joined Anthropic.
Each of these is, on its own, a tidy news item. Together, they describe a single shift: the companies competing at the frontier of AI are no longer content to compete on the model. They are now competing on the silicon beneath the model, the legal perimeter around the model, and the small population of researchers who can build the next one.
The silicon layer. An inference chip is not a general-purpose GPU. It is a specialized processor tuned to run an already-trained model as cheaply and quickly as possible, and it sits at the heart of every AI lab's cost structure. The joint announcement from OpenAI and Broadcom frames Jalapeño as the first member of a planned family of LLM-optimized accelerators, engineered for performance per watt and built for what OpenAI calls gigawatt-scale data center deployments. The company says the chip was designed in nine months with AI-assisted development.
That framing comes from OpenAI and Broadcom, not from independent benchmarks. Any direct comparison to NVIDIA's H200 or B200 silicon on throughput, cost, or availability should be hedged until third parties weigh in. TechCrunch's coverage of the launch makes the same point, and The Verge notes that the production timeline is a roadmap rather than a shipped product. Even so, the move matters. Building a custom inference accelerator is a long, expensive bet that a company's model roadmap is stable enough to commit silicon to, and it pulls the most expensive layer of the AI cost stack inside the lab's own four walls.
The IP layer. Distillation is the practice of training a smaller, cheaper model to mimic the outputs of a larger one. In legitimate research it is openly used; in commercial settings it can be a way to absorb a competitor's capability without paying for the underlying training run. Anthropic's blog post alleges that Alibaba ran a coordinated campaign against Claude, using the words "brazenly" and "illicitly." CNBC and Business Insider both carried Anthropic's framing.
The accusation is one-sided by design. Alibaba has not, in the visible reporting, publicly responded to the specifics. Anthropic's post reads as a public legal posture, not as an adjudicated finding. The interesting strategic move is the choice to publish rather than to litigate quietly: Anthropic is using the open-web blog format to set the perimeter around its models, in the same way OpenAI uses a chip launch to set the perimeter around its compute.
The talent layer. The defection of Adler and Pritzel, as the TLDR AI daily summarized it, is on its own a personnel note. The pattern it sits inside is not. The same period has seen Noam Shazeer return to the frontier and John Jumper, a DeepMind director, leave for Anthropic. The cluster is small enough that a few dozen researchers can meaningfully shift the trajectory of a frontier lab, and the labs themselves behave as if they know it.
The three layers connect. Custom silicon is a long-horizon bet on a specific model trajectory. Public IP accusations are a long-horizon bet on a specific legal perimeter. Talent movement is a long-horizon bet on a specific research culture. Each bet becomes more valuable when the other two are in place. That is the through-line the day's three items share: frontier AI is organizing itself as a vertically integrated stack, not as a model market.
What to watch next is whether the same consolidation shows up in the places the day's news did not reach: in the cloud providers that lose volume when labs bring silicon in-house, in the Chinese model labs that Alibaba's distillation campaign implies exist as competitors, and in the regulatory posture of jurisdictions that have so far treated AI as a software market rather than an infrastructure one.