Anthropic Builds In-House Chip Team, But Pay Gap Emerges Between AI Trainers and Silicon Engineers

Anthropic has confirmed to Business Insider that it is building an in-house silicon design team to develop its own custom chip, following months of speculation sparked by earlier comments about a partnership with Samsung involving 'logic chips.' The company said it will take a 'multi-chip approach' going forward.

According to Tom's Hardware, Anthropic is hiring engineers to co-design custom ASIC processors specifically for AI inferencing workloads, working with an unspecified manufacturing partner. Job listings reportedly indicate the work is schedule-driven, with hires expected to help get the chip design completed on a set timeline. This move puts Anthropic alongside Google, Meta, Microsoft, Amazon, and OpenAI, all of which have developed their own custom chips as global chip supply pressures push AI companies toward proprietary hardware solutions.

Separately, Wccftech reported on a pay disparity within Anthropic's chip efforts, citing a tweet that compared two job listings. A 'Research Engineer, Chip Design RL' role — focused on building reinforcement learning environments so Anthropic's Claude models can learn to design silicon, including RTL generation, verification, and physical design optimization — is listed with a salary range of $500,000 to $850,000. By contrast, a 'Silicon Engineer' role, tasked with actually designing Anthropic's first ASIC, is listed at $320,000 to $485,000.

Wccftech noted that this pay gap exists despite the two roles requiring largely overlapping skill sets, including expertise in full ASIC/FPGA flow, RTL-to-tape-out processes, UVM/formal verification, physical design, power-performance-area (PPA) optimization, design-for-test (DFT), and EDA tools.

Wccftech also pointed to a separate example of AI-driven chip design in the industry: Moonshot's Kimi K3 model reportedly designed a functional silicon chip autonomously within 48 hours, producing a design with a 4.0 mm² area using a Nangate 45nm library. The effort relied on open-source EDA tools and a custom-built GPU compiler, with the simulated chip achieving over 8,700 tokens per second in decoding throughput.

The two reports differ in focus: Tom's Hardware centers on Anthropic's broader hardware strategy and its place among major tech companies pursuing custom silicon, while Wccftech highlights the internal compensation structure at Anthropic as it builds out this chip design effort, an issue not addressed in Tom's Hardware's coverage.