OpenAI Unveils Custom Inference Chip, Sending Nvidia Shares Tumbling

SAN FRANCISCO — OpenAI on Saturday took the wraps off its first in-house inference accelerator, a chip the company says delivers a 4x cost-per-token advantage over current Hopper-class GPUs and reduces tail latency on long-context workloads by more than half. The reveal sent Nvidia shares down 6% in after-hours trading and triggered an immediate re-rating of the entire AI silicon supply chain.

The Chip, Briefly

Designed in partnership with Broadcom and fabricated on TSMC's N3P process, the accelerator is purpose-built for transformer inference rather than training. OpenAI claims 1.4 PFLOPS of FP8 throughput per package, 384 GB of stacked HBM4, and a custom interconnect that scales linearly to 256-chip pods. First deployments are already running production GPT-class traffic in OpenAI's Texas and Norway data centers.

Why Nvidia Slid

Inference is the larger and faster-growing half of the AI compute pie, and OpenAI is Nvidia's single largest customer. A credible alternative — even if it ships only into OpenAI's own fleet — caps the upside on Hopper and Blackwell volume forecasts that had baked in OpenAI demand through 2027. Bernstein cut its FY27 Nvidia data-center revenue estimate by $14 billion within hours of the keynote.

The Bigger Pattern

OpenAI joins Google (TPU), Amazon (Trainium/Inferentia), Meta (MTIA), and now Anthropic-via-Akamai in building or buying around Nvidia. The pattern is no longer a hedge — it is a structural redistribution of AI margin from a single supplier toward an oligopoly of hyperscaler-designed silicon.

What to Watch

Three signals matter next: Broadcom's next earnings call for booked AI ASIC revenue, TSMC's N3P allocation disclosures, and whether Microsoft — OpenAI's largest backer and a competing chip designer with Maia — accelerates its own roadmap. The era of one chip ruling them all is officially over.