Clive Chan, who says he was the second hardware employee in OpenAI's custom chip program, is moving to Anthropic. The departure comes as both companies prepare for IPOs and Anthropic is reportedly considering building its own chips.
In a public post, Chan wrote that he's proud to have been part of the program. The "density of hardware talent" on the team was extraordinary. "I don't think there's a better chip design team anywhere," he said, expecting that the chips developed there would become "one of the most important engines of AGI."
Despite those apparently stellar conditions, he's now joining Anthropic, OpenAI's biggest rival, which, like Chan's former employer, is on the verge of going public. At OpenAI, Chan worked on building custom chips from scratch and was involved in the strategic partnership between OpenAI and Broadcom, although that partnership reportedly hit snags over production costs and OpenAI's creditworthiness.
Whether Anthropic hired Chan to design custom chips or to optimize software for existing hardware isn't clear. His LinkedIn role description, "perplexity per picojoule," could mean either. Perplexity is a common metric for how well language models predict text, and a picojoule is a tiny unit of energy.
So the goal is to squeeze as much model performance as possible out of every unit of energy. That can happen through better software on existing GPUs and TPUs or through custom silicon tailored to Anthropic's models.
According to Reuters, Anthropic is weighing the idea of designing its own AI chips, following the path OpenAI and Meta have already taken. The plans were still in early stages as of April 2026, with no dedicated team yet. Chan could help build one, and he's bringing all his OpenAI chip knowledge with him.
Anthropic currently runs Claude on Google's TPUs and Amazon chips. The company recently signed a long-term deal with Google and Broadcom as part of a commitment to invest $50 billion in US computing infrastructure.
Custom chips would give Anthropic a real financial edge. For inference especially, purpose-built silicon could deliver bigger margins over time. That matters even more as AI increasingly becomes an infrastructure play rather than a research-breakthrough story.
Before joining OpenAI in January 2024, Chan spent about two and a half years at Tesla's Autopilot division, according to his LinkedIn. There he worked on a custom chip for ML training, handling software framework bring-up, datacenter co-design, and energy-efficient number formats.
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。