Anthropic has confirmed it is building an in-house silicon team to co-design a custom artificial intelligence chip for its Claude large language models, joining OpenAI in a small but rapidly growing club of frontier AI laboratories designing their own processors. The chip will be co-designed with future Claude models themselves, a workflow that lets Anthropic tune the silicon to the exact memory, compute, and bandwidth profile Claude inference needs rather than accept the mismatches that come with off-the-shelf accelerators. According to a Business Insider report cited by SiliconAngle, the strategic goal is to cut Claude inference costs roughly in half.
Co-design is the technical pattern that makes the move worth doing. Off-the-shelf AI accelerators are designed for a wide range of workloads, which means their memory capacity, bandwidth, and compute layout rarely match what any single model actually needs. An accelerator may have slightly more memory than the model requires, or slightly less, and the inefficiencies compound across every query the system serves. Co-designing the processor with the software it will run eliminates those mismatches and lets the hardware team make tradeoffs that general-purpose silicon cannot.
The Verification Workflow Will Run on Claude Itself
Anthropic is hiring processor designers and verification engineers to lead the chip programme. According to a job listing obtained by Business Insider, the company plans to automate significant parts of the verification flow using its own model. The listing describes building simulations in which Claude learns how to test newly developed chip designs, with particular emphasis on formal verification. Formal verification checks a chip design for flaws by simulating every combination of operating conditions in which it could run, a method that traditional teams carry out by hand over months. Anthropic’s bet is that a sufficiently capable language model can compress that timeline the same way it has compressed timelines elsewhere in the software stack.
“Inference is AI providers’ largest infrastructure line item. Halving it would reshape the unit economics of every Claude-powered product.”
Why Samsung, Not TSMC
In early June, The Information reported that Anthropic may partner with Samsung Electronics to manufacture the chip. Samsung holds a much smaller share of the contract chipmaking market than Taiwan Semiconductor Manufacturing Co, but it has been pushing a packaging technology called zHBM that places memory directly atop logic cores. That arrangement shortens the physical distance data must travel between memory and compute circuits, which lowers power use and improves throughput for memory-bound workloads like transformer inference. If Anthropic adopts zHBM in its first chip, it will be one of the first commercial designs to use the packaging in production silicon.
OpenAI Set the Pace With Jalapeno
Anthropic is following a path OpenAI has already walked. In June, OpenAI debuted a custom inference accelerator called Jalapeno that was developed through a collaboration with Broadcom. The chip took only nine months to design because OpenAI and Broadcom automated substantial parts of the manual workflow with AI tooling. OpenAI’s choice to optimise Jalapeno specifically for inference rather than training is the same strategic logic Anthropic is now applying. Inference, the work of running large language models in production after training is complete, is the largest infrastructure line item for AI providers, and custom silicon is the cleanest lever for reducing it.
The Bigger Pattern Across Frontier Labs
Every major AI laboratory has now moved into chip design in some form. Google designed its own TPUs over a decade ago and continues to ship new generations for both training and inference. OpenAI built Jalapeno with Broadcom. Anthropic is now assembling its own team. Meta is reportedly working with Broadcom on a parallel accelerator programme. The frontier of the AI race is no longer just a race between models. It is a race between silicon design teams, foundry capacity, memory packaging breakthroughs, and the ability to ship the resulting hardware in volume. Anthropic’s confirmation puts the company squarely in that race, with a target of cutting Claude inference cost in half and a verification pipeline that Claude itself will help drive.
Why Inference Cost Now Drives the Roadmap
The shift toward inference-optimised silicon is the most consequential change in AI infrastructure since the move from CPU to GPU training a decade ago. Training is a fixed cost that gets amortised across every query a model eventually serves. Inference is a recurring cost that scales linearly with usage, and at frontier-lab scale that recurring cost now exceeds the training bill. Every percentage point of inference efficiency flows directly to gross margin, which is why the largest AI providers are willing to spend hundreds of millions of dollars on custom silicon programmes that may not produce a working chip for two or three years. Anthropic joining that race with a co-design pipeline that uses Claude to verify the chip itself is the cleanest expression yet of how AI is becoming the tool that builds the next generation of AI infrastructure.