Our Take
Vertical integration of chip design cuts the margin tax OpenAI pays to NVIDIA, but the chip is not disclosed to be faster or cheaper per unit—only that OpenAI built it.
Why it matters
Frontier model labs are now competing on infrastructure cost, not just model quality. Custom silicon is becoming table stakes for labs that want to avoid NVIDIA pricing power.
Do this week
Enterprise: audit your AI infrastructure contracts for NVIDIA lock-in clauses and renewal dates so you can plan multi-vendor strategies before 2026.
OpenAI designs its first custom silicon with Broadcom
OpenAI announced a custom chip developed in partnership with Broadcom, a semiconductor design and manufacturing services firm. The chip is intended to support OpenAI's AI infrastructure and reduce operational costs tied to compute hardware.
Neither OpenAI nor Broadcom disclosed technical specifications, performance benchmarks, or deployment timelines in the announcement. The partnership is framed as a response to scaling constraints and cost pressures in training and inference at the scale OpenAI operates.
Custom silicon is now a cost-control lever for frontier labs
OpenAI's move reflects a pattern already established by Meta and Google: building proprietary chips to reduce dependency on NVIDIA GPUs and to optimize silicon for internal workloads. NVIDIA commands 80%+ of the AI accelerator market and has pricing power that erodes margins for labs running inference at scale.
Broadcom is a design partner, not a manufacturer. This suggests OpenAI is outsourcing chip architecture and tape-out while retaining control over specification. Broadcom has experience co-designing custom silicon for hyperscalers (Amazon Trainium, AWS Graviton). The choice signals OpenAI's intent to iterate on chip design alongside model improvements, rather than committing to a fixed hardware roadmap.
For frontier labs, custom silicon is no longer optional. It is now a competitive necessity to defend inference margin and reduce capex lock-in to a single vendor.
What builders should do now
If you are an enterprise AI team, audit your existing NVIDIA commitments for multi-year contracts that carry termination penalties. OpenAI's move will likely accelerate announcements of custom chips from other labs (Anthropic, Meta, potentially others), which will expand vendor optionality over the next 18 months.
If you are running inference workloads on NVIDIA hardware, document your cost per inference token and latency profiles. Custom silicon from OpenAI's infrastructure (once deployed) may be available to API consumers, and comparison pricing will become relevant when you next renegotiate capacity terms.