Analysis
Anthropic has confirmed it's building an in-house team to co-design custom AI inference chips for Claude, posting listings for a dedicated "custom silicon team" with salaries ranging from $320,000 to $485,000, Tom's Hardware reported, with Samsung Electronics reportedly in talks as a manufacturing partner. The company has also poached an engineer who worked on OpenAI's Broadcom-made chip effort, a hire that signals real technical intent rather than a hedging headline.
Anthropic joins OpenAI and Meta in moving toward custom ASICs, but the strategic framing is narrower than it might sound: the company says it will keep its multi-chip reliance on AWS, Google, Nvidia and AMD rather than replacing any of them, and it already has a long-term agreement with Google and Broadcom for roughly 3.5 gigawatts of custom TPU capacity starting in 2027. Custom silicon here reads as margin engineering on top of an existing multi-vendor stack, not a bet against any single supplier -- Anthropic reported more than $11.5 billion in Q2 2026 revenue and turned adjusted operating income positive for the first time, giving it both the balance sheet and the inference volume to make in-house chip design pay off at scale.
The timing lines up with Anthropic's own IPO preparation -- Pulse has covered the company's $965B Series H and expected late-2026 public listing -- and custom silicon is exactly the kind of gross-margin story that plays well in an S-1: inference cost per token is the single biggest lever on any foundation model's unit economics, and owning more of that stack is a cleaner story for public investors than "we pay Nvidia's list price." The risk is execution timeline: custom ASIC programs routinely run two to three years from design to production silicon, and Anthropic is starting from a standing start against OpenAI and Google, both of which have had chip programs running longer.
