Anthropic’s decision to hire Clive Chan, OpenAI’s second hardware engineer, is a small personnel move that lands in a very large strategic window. Both labs are preparing for IPOs, and that timing is sharpening a race that is increasingly being decided not just by model quality or product velocity, but by how efficiently each company can turn power into usable inference.
That matters because custom silicon is no longer a side project in frontier AI. It is becoming part of the narrative investors will use to judge whether a lab can scale economically after listing. In that framing, Chan’s move reads less like a routine talent transfer and more like a sign that Anthropic wants more control over the hardware stack, whether that means building its own chips outright or tightening the software-hardware loop around existing accelerators.
Chan said publicly that OpenAI had exceptional hardware talent and that the team’s density of chip-design expertise was unusually strong. He also said he believed the chips under development there would become one of the most important engines of AGI. That is a telling comment coming from someone who helped build the program from an early stage: it suggests that at OpenAI, hardware is being treated as a foundational capability, not merely a procurement function.
At OpenAI, Chan worked on custom chips from scratch and was involved in the company’s partnership with Broadcom, a relationship that illustrates the practical limits of this strategy. Building custom silicon is not just an engineering challenge; it is also a cost, credit, and manufacturing problem. Reported snags around production costs and OpenAI’s creditworthiness underline how much execution risk sits beneath the headline ambition. Even with a strong internal team, the path from chip concept to deployable accelerator is constrained by supplier economics and the realities of volume production.
That is why energy efficiency sits at the center of this story. For AI labs running massive inference workloads, the important metric is not merely raw throughput. It is performance per watt, and ultimately cost per unit of useful output. A better accelerator is one that can deliver lower energy consumption at the same or higher density, because that changes deployment economics, capacity planning, and the cost structure that will matter when markets start asking IPO-era questions about margins and scale.
Anthropic’s hire suggests it may be leaning into that same logic. The company has reportedly considered building its own chips alongside its IPO plans, which would place it on a more hardware-centric path. But there is a second possibility: Chan could be deployed to improve Anthropic’s software-hardware alignment, helping the company squeeze more performance out of existing systems through better system design, compiler work, or workload tuning. The ambiguity is important. It reflects a broader strategic fork between designing around bespoke accelerators and extracting more efficiency from available infrastructure.
OpenAI and Anthropic now look like rivals in a hardware domain that used to sit behind the software race. OpenAI’s approach, at least as revealed by its Broadcom work, suggests a willingness to absorb the complexity of custom chip development in exchange for control over cost and performance. Anthropic’s move indicates that it does not want to concede that advantage. If Chan’s expertise translates into a tighter accelerator strategy, Anthropic could narrow the gap on deployment efficiency and reduce its dependence on off-the-shelf hardware.
The IPO context makes that competition sharper. Public-market scrutiny tends to reward clear stories about scale, defensibility, and operating leverage. For AI labs, that story increasingly includes chips. A company that can show it has a credible path to more energy-efficient inference has a stronger answer to questions about unit economics than one that relies entirely on rented compute. That is especially true if model demand keeps rising and compute costs remain a central constraint.
Still, the risks are substantial. Custom chip efforts can become expensive, slow, and fragmented. Partnerships can stumble if production costs rise or supplier confidence weakens. And if the two firms pursue different hardware strategies, the competitive landscape could split between one lab optimizing around external partnerships and another trying to internalize more of the stack. Neither path is guaranteed to win.
What Chan’s move does do is make the race legible. In the run-up to IPOs, hardware talent is being treated as strategic capital, and energy-efficient accelerators are moving from an engineering preference to a market-facing requirement. The companies that can best align model ambition, chip design, and deployment economics are the ones most likely to control the tempo after they go public.
