OpenAI unveils custom chip it designed with Broadcom to boost its AI infrastructure

OpenAI has announced a custom AI accelerator chip designed in partnership with Broadcom, its first significant step toward vertically integrating its own silicon stack. The move follows a broader industry trend of hyperscalers — Google (TPU), Amazon (Trainium/Inferentia), and Microsoft (Maia) — developing proprietary chips to reduce unit economics dependency on NVIDIA's H100/H200 GPUs, which command premium pricing backed by CUDA lock-in.

For Broadcom, the deal is a meaningful revenue validation of its ASIC/custom-silicon strategy. AVGO has been building out its custom AI accelerator business (XPUs) for Google and Meta, and adding OpenAI as a named client confirms that Broadcom's design-and-supply model is scaling. NVIDIA, meanwhile, faces a symbolic and potentially structural headwind: OpenAI is reportedly one of its largest single GPU customers, spending billions annually on H-series chips.

The second-order tension is whether this is a long-term volume threat to NVDA or simply additive demand. Custom chips typically handle inference workloads more efficiently, while NVIDIA retains dominance in frontier model training. OpenAI will still need NVDA GPUs for training next-generation models even as it deploys its own chips for inference at scale.

NVDA's enrichment data shows extraordinary fundamentals — $215.9B revenue at +65.5% YoY growth, 71.1% gross margins, $4.90 diluted EPS — making it expensive to short on fundamentals alone. The real question is whether this headline accelerates a re-rating of NVDA's customer concentration risk. Watch AVGO for a near-term pop and NVDA for whether the market treats this as a structural de-rating event or a one-day headline fade.