Anthropic has publicly confirmed for the first time that it is building an in-house silicon team to design custom chips for Claude. Business Insider reported on this plan on August 5, based on interviews with an Anthropic spokesperson. The spokesperson explained to Business Insider that the company intends to co-design hardware and models to run Claude faster and more efficiently at the scale customers need.
This shift is not about the completion of a chip itself. Reuters reported in April that Anthropic was exploring designing its own chips. At that time, it was still at the exploratory stage, whereas this time, the explanation that a dedicated in-house team is being built has been made public for the first time. Moreover, Anthropic's official job posting not only calls for recruiting semiconductor designers but also outlines work that brings Claude into the design process itself.
Custom silicon can be designed to match the computation, memory, and communication that a model requires. It could become an option for narrowing the fit gap with general-purpose GPUs or other companies' accelerators, and for tuning inference speed, power, and cost. On the other hand, Anthropic has not disclosed whether this chip is intended for training, inference, or both. Building an in-house team is a stage separate from announcing product specifications or mass production.
What the Job Posting Reveals About Chip Design Work
The position Anthropic is recruiting for is "Research Engineer, Chip Design RL." According to the official job posting, the Code RL team develops reinforcement learning environments and evaluations for agentic RTL generation, formal verification, and physical design optimization. RTL is the design stage that describes circuit behavior, and formal verification is the process of mathematically confirming that a design satisfies its intended conditions.
The requirements include ASIC or FPGA design and UVM. Formal methods, synthesis and place-and-route, and timing closure are also covered. It further addresses PPA optimization, where PPA refers to performance, power, and area. If the goal were simply to have a model write circuit descriptions, there would be little need to list these downstream processes as hiring requirements. The job posting can be read as anticipating a workflow that handles the process from design generation through verification and physical design using reinforcement learning, with an eye toward moving to actual silicon.
However, the fact that tape-out experience is required does not mean Anthropic has already completed a tape-out. What the job posting shows is a hiring plan for design and research, not a mass-production contract or a productization decision.
Even With In-House Design Underway, the Multi-Chip Strategy Continues
Anthropic has already secured substantial external computing resources. In an official announcement from October 2025, the company said it would expand its use of Google Cloud, with plans including up to 1 million TPUs and over 1GW of capacity in 2026. The company has explicitly stated a three-platform strategy using Google TPUs, Amazon Trainium, and NVIDIA GPUs, describing AWS as its primary training cloud partner.
Further, in April 2026, the company announced securing up to 5GW of new capacity with AWS and committing over $100 billion to AWS technology over ten years. It also indicated plans to use approximately 1GW combined across Trainium2 and Trainium3 by the end of 2026. According to its explanation to Business Insider, the company will continue a "multi-chip approach" centered on hardware from AWS and Google, as well as NVIDIA and AMD.
Therefore, the building of an in-house silicon team cannot be read as a declaration that will replace existing suppliers. In parallel with building the in-house team, Anthropic will continue its policy of using hardware from multiple companies. At the very least, Anthropic's official information and its explanation to Business Insider do not indicate a policy of consolidating onto a single chip.
OpenAI's Nine Months, and What Anthropic Has Not Yet Shown
OpenAI officially announced that, together with Broadcom, it developed "Jalapeño," a chip for LLM inference, taking it from design through the manufacturing readiness stage in nine months, and plans initial deployment starting from the end of 2026. This is a case of a model development company advancing a custom chip all the way to manufacturing readiness.
Anthropic's current disclosure is at a different stage. What the company has made public is the building of an in-house team and the design-support research position that appeared in the job posting; it has not disclosed the target use case, manufacturing readiness, or the timing of initial deployment. There is also no basis for simply applying OpenAI's nine-month development period to Anthropic's own timeline.
When comparing the two companies, what matters is not whether they have a custom chip or not, but how concretely the design has progressed. OpenAI has indicated an inference use case and a planned initial deployment. Anthropic has explained a direction of co-designing models and hardware and is recruiting talent for that purpose. That is the extent of what has been disclosed so far.
The Numbers Needed to Judge Productization Have Not Yet Emerged
Anthropic has not disclosed the chip's manufacturing partner, process node, or memory configuration. The same applies to packaging, performance, and investment amount. The timing of the start of mass production is also unknown. The value of custom silicon cannot be judged from design-stage intent alone; it can only be assessed once the model's requirements and the manufacturing and implementation conditions are both in place.
What needs to be confirmed next is which of Claude's processing tasks will be handled by in-house design—training, inference, or both. Once the use case (training or inference), the manufacturing partner, and the mass-production timeline are disclosed, it will become possible to judge how the workload will be divided between the existing fleet of TPUs, Trainium chips, and GPUs and the in-house design.