At a critical juncture just prior to the mass production of its first custom silicon, OpenAI has lost a foundational hardware pioneer. Clive Chan, a key engineer on OpenAI’s custom chip team, announced his departure on X, revealing that he has officially joined arch-rival Anthropic.
As the second hardware hire in OpenAI's history, Chan spent the last 30 months co-designing the company's proprietary silicon since joining in January 2024. While praising OpenAI's chip group as having "the best silicon design team in the world," Chan noted he "could not shake the urge to climb a new mountain from the base." He summarized his decision to join Anthropic with three pillars: talent, values, and ambition.
Chan's primary contribution at OpenAI was leading the design of a 10GW custom AI accelerator developed in collaboration with Broadcom, leveraging TSMC's 3nm process node. Chan had previously hinted that the chip would transition to mass production nine months after tape-out. With that milestone now imminent, Chan has elected to transition to his next venture, leaving OpenAI with a fully-realized hardware foundation.
Chan’s pedigree is highly distinguished. Prior to OpenAI, he spent years at Tesla working on the Tesla Dojo supercomputer team, co-designing the training ASIC software framework and datacenter architecture, reporting directly to Elon Musk. In early 2024, he joined Google TPU veteran Richard Ho to build OpenAI's custom silicon initiative from scratch. What began as a highly speculative unit has since matured into a robust 40-person chip-design powerhouse.
This high-profile defection is a massive win for Anthropic. While reports in April suggested Anthropic harbored custom silicon ambitions, their efforts were deemed early-stage and lacked dedicated leadership. By securing Chan, Anthropic is positioned to accelerate its hardware roadmap from conceptual exploration into execution, directly threatening OpenAI's early hardware advantage.
Industry reactions have been mixed. AI critic Gary Marcus noted that talent voting with their feet should make potential OpenAI IPO investors think twice, while others joked that OpenAI has become a pre-onboarding school for Anthropic. As both companies march toward staggering valuations, the war for silicon talent will only intensify.
[AgentUpdate Depth Analysis] The defection of Clive Chan highlights an accelerating trend among frontier AI labs: the transition from generic GPU clusters to highly specialized, custom ASIC hardware architectures. For the next-generation AI Agent ecosystem, compute efficiency and latency are the ultimate bottlenecks. While Nvidia's GPUs remain the gold standard for LLM training, running autonomous agents with continuous multi-step reasoning, real-time environment feedback loops, and multi-modal routing demands localized, low-latency, and cost-efficient custom silicon. By engineering custom accelerators, labs like OpenAI and Anthropic aim to dramatically optimize the unit economics of agent execution. Chan’s arrival at Anthropic will likely catalyze a new wave of hardware-software co-design specifically optimized for agentic workloads, reshaping the competitive landscape of systemic AI deployment.
