OpenAI's first custom chip tape-out with Broadcom is a bet it can out-engineer Nvidia's margins

OpenAI's confirmation that its custom AI accelerator, co-designed with Broadcom under the multi-year partnership the two companies announced previously, has reached the tape-out stage - the point at which a chip design is finalised and sent for manufacturing - marks a significant milestone in OpenAI's stated ambition to reduce its dependence on Nvidia's GPUs for at least a portion of its enormous and growing inference and training workload, following a path that Google, Amazon and Meta have each pursued with varying degrees of success through their own custom-silicon programmes. The economic logic behind OpenAI's custom-chip investment is straightforward even if the engineering execution is difficult: Nvidia's high gross margins on its GPUs represent, from OpenAI's perspective as the buyer, a cost that a chip designed specifically and only for OpenAI's own workload patterns - without needing to serve the broad, general-purpose market that forces Nvidia's chip designs to make compromises - could potentially reduce meaningfully, even after accounting for the substantial upfront design and manufacturing costs and the risk that a custom chip underperforms a general-purpose GPU on workloads that do not match its specific design assumptions. Broadcom's role as OpenAI's design and manufacturing partner reflects the company's established position as the leading provider of custom-silicon design services to hyperscalers, a business Broadcom has built over years serving Google's TPU programme among others, giving OpenAI access to manufacturing relationships and process expertise it could not efficiently build from scratch on its own timeline. The risk profile of the custom-silicon bet is asymmetric in a way that mirrors the broader industry's approach to AI infrastructure diversification: if the custom chip underperforms expectations, OpenAI can simply continue relying primarily on Nvidia GPUs and AMD's chips at limited additional cost beyond the sunk design investment, but if it succeeds even partially, the cost savings compound across the massive scale of OpenAI's ongoing inference workload in a way that could meaningfully improve the company's overall unit economics. What to watch: when OpenAI's custom chip actually enters production deployment following the tape-out milestone, how its real-world performance compares with Nvidia and AMD alternatives on OpenAI's actual workload mix, and whether other frontier labs accelerate their own custom-silicon programmes in response to OpenAI's progress.

Original source: Reuters