Aimed primarily at powering increasingly complex models, OpenAI’s custom silicon comes amid widespread industry efforts to optimise performance and cut costs. The company anticipates this internal deployment will enhance efficiency without immediately entering external markets. Broadcom, for its part, has flagged over $10 billion in AI infrastructure orders tied to an unnamed client, widely believed to be OpenAI.
The collaboration with Broadcom builds on a design phase led by a growing team within OpenAI, overseen by Richard Ho, formerly of Google. That team has nearly finalised the chip’s architecture, which will be sent to Taiwan Semiconductor Manufacturing Co. for fabrication at a 3-nanometre process node. Production and shipments are expected to commence in 2026, positioning OpenAI alongside tech giants such as Google, Amazon, and Meta—all of which have developed tailored AI accelerators.
The drive for bespoke hardware stems from constraints in supply and escalating costs of commercially available Graphics Processing Units. OpenAI has, in tandem, expanded its use of AMD and Google Cloud’s Tensor Processing Units, supplementing but not replacing its reliance on Nvidia. Developing its own chips affords the company flexibility to tailor compute capabilities and manage supply chain challenges more strategically.
Broadcom stands to gain significantly if the partnership bears fruit. The firm has already adjusted its fiscal 2026 AI revenue forecast upwards, citing firm orders from a new client that correspond closely with the OpenAI timeline. Its backlog, buoyed by AI commitments, continues to grow.
Industry watchers note, however, that producing AI silicon comes with notable risks—design delays, performance bottlenecks and manufacturing complexities. OpenAI must also navigate geopolitical headwinds tied to supply chains and fabrication in Taiwan. Moreover, Nvidia retains a strong software advantage through its CUDA ecosystem, which continues to underpin much of AI development.
Topics
Technology