OpenAI Partnering with Broadcom to Build Custom Inference Silicon
OpenAI is collaborating with Broadcom and TSMC to design its first in-house ASIC for AI inference. The move represents a major strategic shift towards securing independent chip supply and reducing dependency on Nvidia hardware.
This custom silicon strategy means OpenAI will eventually gain a structural cost advantage over competitors running pure Nvidia hardware. For product managers building high-throughput AI features, expect API costs to drop by 40-50% in the 2026-2027 timeframe. When planning multi-year AI architecture, avoid locking into long-term pricing contracts today, as model inference costs are on a steep downward trajectory.
OpenAI is pivoting from a custom foundry model to an ASIC partnership with Broadcom.
Securing the Supply Chain
OpenAI has reportedly put plans for a multi-billion dollar foundry network on hold, pivoting instead to a more focused custom chip design strategy. By partnering with Broadcom for design and TSMC for manufacturing, the organization hopes to secure dedicated inference silicon by 2026. This allows OpenAI to optimize hardware specifically for its proprietary models while shielding itself from Nvidia's pricing premium and supply bottlenecks.
Optimizing for Inference
As AI models move from training to real-world deployment, inference workloads are expanding exponentially. Custom ASICs (Application-Specific Integrated Circuits) can deliver up to 3x performance-per-watt improvements compared to general-purpose GPUs. This specialized silicon will run model weight calculations much faster, directly translating to lower API latency and reduced compute costs for end-users.
Frequently Asked Questions
Production is slated to begin in late 2025, with active deployment in OpenAI's API data centers expected by early to mid-2026.
No. Scale now using existing API models, but design your architecture to easily swap providers or utilize tier-discounts as compute costs fall.
