What this shift means for makers and artists
When Clive Chan leaves OpenAI for Anthropic, the ripples extend far beyond corporate strategy. For the creators and builders relying on these systems, the move signals a fierce intensification in the race for efficiency. The industry is no longer just about raw compute power; it is about squeezing maximum intelligence out of every unit of energy. Whether you are training models or running inference for your own applications, the underlying hardware architecture is set to change, potentially lowering costs and increasing the accessibility of powerful AI tools for independent studios and smaller labs.
The hardware talent exodus
Clive Chan, who identifies himself as the second hardware hire in OpenAI’s custom silicon programme, has announced his departure to join Anthropic. This transition occurs as both rivals prepare for their initial public offerings, with reports suggesting Anthropic is actively considering the development of its own proprietary chips.
In a public statement, Chan expressed pride in his tenure, describing the concentration of hardware expertise on the team as extraordinary. He noted that he did not believe a better chip design group existed anywhere, anticipating that the resulting silicon would become a primary driver of artificial general intelligence.
Despite praising the environment at OpenAI, where he worked on building custom chips from the ground up and participated in the strategic alliance with Broadcom, he is now heading to OpenAI’s primary competitor. That partnership at the former company reportedly encountered significant hurdles regarding production expenses and the firm’s credit standing.
The precise objective of his move remains ambiguous. His LinkedIn profile lists his role as “perplexity per picojoule,” a phrase that could refer to either designing custom silicon or optimising software for off-the-shelf hardware. Perplexity measures how accurately language models predict text, while a picojoule represents a minute unit of energy. The implied goal is to maximise model performance relative to energy consumption, achievable through better software on existing GPUs and TPUs or through bespoke silicon tailored to specific models.
Custom silicon could widen profit margins
Reporting from Reuters indicates that Anthropic is evaluating the design of its own AI chips, mirroring the paths already taken by OpenAI and Meta. As of April 2026, these plans were still in their infancy, with no dedicated team established. Chan’s arrival could facilitate this effort, bringing his extensive knowledge of chip architecture from his time at OpenAI.
Currently, Anthropic operates Claude on Google’s TPUs and Amazon chips. The company has recently secured a long-term agreement with Google and Broadcom, committing to invest $50 billion in US computing infrastructure.
Proprietary chips would provide Anthropic with a distinct financial advantage. For inference specifically, purpose-built silicon could generate significantly larger margins over time. This factor becomes increasingly critical as AI shifts from a narrative of research breakthroughs to a fundamental infrastructure play.
A lineage of custom silicon
According to his LinkedIn profile, Chan spent approximately two and a half years at Tesla’s Autopilot division prior to joining OpenAI in January 2024. During that period, he worked on a custom chip for machine learning training, managing software framework integration, datacentre co-design, and energy-efficient numerical formats.
Key takeaways
- Clive Chan is moving from OpenAI to Anthropic, bringing rare expertise in custom chip design and energy-efficient hardware optimisation.
- Chan’s new role likely focuses on improving the “perplexity per picojoule” ratio, ensuring models run efficiently regardless of whether they use custom silicon or existing hardware.
- Anthropic is exploring its own chip development to secure long-term financial margins, following the infrastructure-led strategies of Meta and OpenAI.
- This talent shift underscores a broader industry trend where hardware efficiency is becoming as critical as raw processing power for AI scalability.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
