Clive Chan, the second hardware engineer hired for OpenAI's custom chip initiative, has transitioned to its chief rival, Anthropic. This high-profile departure occurs as both entities accelerate toward potential IPOs and Anthropic actively contemplates building its own proprietary silicon.
In a public post, Chan expressed pride in his tenure at OpenAI, praising the extraordinary "density of hardware talent" on the team and predicting that their custom chips would become "one of the most important engines of AGI." Despite these conditions, Chan made the leap to Anthropic. While at OpenAI, he played a critical role in building custom chips from scratch and was deeply involved in the strategic partnership with Broadcom, which reportedly encountered hurdles regarding production costs and credit-worthiness.
It remains to be seen whether Anthropic recruited Chan for complete custom silicon design or to optimize software for existing hardware stacks. Chan's cryptic LinkedIn bio, "perplexity per picojoule," hints at both possibilities. Perplexity measures a language model's predictive performance, while a picojoule is a minuscule unit of energy. The objective is to extract maximum model performance per unit of energy, either through deep software optimization on existing GPUs and TPUs, or via tailored custom silicon.
Anthropic is reportedly exploring custom AI chip designs to follow in the footsteps of OpenAI and Meta. While the plans were in early development stages as of early 2026, Chan's expertise could jumpstart a dedicated hardware unit. Currently, Anthropic runs its Claude models on Google's TPUs and Amazon's chips, and recently solidified a multi-billion-dollar commitment to US computing infrastructure. Developing custom hardware, especially for inference, could dramatically boost profit margins as AI matures from research-driven breakthroughs into a capital-intensive infrastructure race.
Prior to joining OpenAI in early 2024, Chan spent over two years at Tesla's Autopilot division. At Tesla, he focused on custom machine learning training chips, handling software framework integration, datacenter co-design, and energy-efficient numerical formats.
[AgentUpdate Depth Analysis] As the generative AI paradigm shifts from raw pre-training to dynamic AI Agent execution, compute efficiency during inference has become the ultimate competitive moat. AI Agents require multi-step reasoning, real-time environment interaction, and recursive tool execution, which exponentially increase active token generation costs. Chan's philosophy of maximizing 'perplexity per picojoule' addresses this exact operational bottleneck. Compared to competitors like Meta and Google, who already possess in-house silicon (MTIA and TPU), Anthropic's push for custom chip talent is a strategic necessity. By designing dedicated hardware or deeply optimizing software-hardware co-design, Anthropic can drastically lower the cost-per-token for agency-heavy tasks like Computer Use. This move will fundamentally redefine the AI Agent ecosystem by making autonomous workflows financially viable for enterprise-scale deployments, transitioning the market from speculative experimentation to highly optimized production.
