Anthropic is assembling a custom silicon design team, moving from early discussions to active hiring for engineers who can co-design hardware alongside the company's AI models. Reuters first reported the effort in April 2026 as a preliminary exploration; by August, the company had posted job listings for a "Research Engineer, Chip Design RL (Reinforcement Learning)" role based in San Francisco and New York, categorized under Engineering & Design — Product TechCrunch. Anthropic's careers page also lists a Hardware Security Engineer position, suggesting the scope extends beyond pure architecture into silicon-level security.

The company stated it plans to co-design hardware and models to improve the speed and efficiency of its technology TechCrunch. The Information reported that Anthropic has been scouting Samsung as a potential fabrication partner — the company that would actually manufacture the physical chips — though no partnership has been confirmed TechCrunch.

Anthropic's decision to pursue its own silicon comes as demand for Claude accelerates and AI companies scramble to lock in compute infrastructure. Reuters reported in April that the broader push responds to a chronic shortage of AI computing hardware CNBC. At that time, Anthropic's own-chip plans remained preliminary, with no committed design or assembled team. The August hiring push indicates concrete movement beyond that stage.

Anthropic already operates across a varied hardware stack. The company trains and runs Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs, matching workloads to the appropriate silicon Anthropic. In April 2026, Anthropic expanded its compute partnership with Google and Broadcom. Weeks later, it extended its collaboration with Amazon for up to five years of compute capacity Anthropic. Anthropic also plans to expand its use of Google Cloud technologies to as many as one million TPUs Anthropic. On the AWS side, Anthropic's engineers work closely with Amazon's Annapurna Labs chip design team to extract maximum computational efficiency from Trainium hardware Anthropic. The company has signed deals with AWS, Google, Nvidia, and AMD for AI computing hardware access TechCrunch.

Anthropic is not the first AI lab to move down the silicon stack. Google DeepMind has long relied on Alphabet's TPU chips. Meta has been developing its MTIA accelerators for AI workloads. In June 2026, OpenAI unveiled its Broadcom-built Jalapeño chip, designed specifically for inference — the process of running a trained model to produce outputs TechCrunch. The pattern across these companies is consistent: as model scale and inference volume grow, the cost and availability constraints of merchant silicon — chips sold off-the-shelf by companies like Nvidia — push labs toward custom designs tuned to their own workloads.

What distinguishes Anthropic's situation is the breadth of its existing hardware partnerships. Few AI labs have simultaneous deep engagements with AWS, Google Cloud, Broadcom, Nvidia, and AMD. Custom silicon, in that context, is less about replacing those relationships than about gaining a design lever that existing partners cannot fully provide — whether for inference latency, memory bandwidth, or workload-specific efficiency targets that general-purpose accelerators do not optimize for.

The job listing for a reinforcement learning-focused chip design engineer is worth noting. RL for chip design has produced results at Google, where RL-based floorplanning — the process of arranging components on a chip — surpassed human-designed layouts for TPU blocks. Anthropic hiring specifically for RL applied to silicon suggests the company intends to use its own models in the chip design loop, not merely hand specifications to a third-party design house.

Anthropic's careers page listed 398 open jobs in AI Research & Engineering as of the most recent access, with the custom silicon roles sitting inside that broader hiring surge. The scale of recruiting signals that Anthropic is building out infrastructure capability in parallel with model capability, treating hardware-software co-design as a first-class engineering discipline rather than a procurement problem.

Whether Anthropic reaches a tape-out — the point at which a chip design is finalized and sent to a factory for manufacturing — which foundry it partners with, and whether Samsung materializes as a fab partner all remain open questions. The company has not committed to a specific architecture or product. What is clear is that Anthropic has moved from weighing the possibility of custom chips, as Reuters characterized it in April, to actively staffing a design team four months later. For a company whose entire product line depends on inference economics, that trajectory has a logic to it that requires no speculation to appreciate.