# Frontier Labs Build Their Own Inference Silicon

The largest AI operators keep converting from buyers of merchant accelerators into designers of their own inference chips, permanently removing their highest-volume workloads from the general-purpose GPU market and shifting value toward custom silicon design partners.

- Conviction: 47 / 100 (strengthening)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-08-26T00:00:00.000Z
- Last updated: 2026-08-27T14:00:31.454Z
- Canonical: https://polylog.news/ai/trends/lab-custom-silicon-vertical-integration
- Publisher: Polylog
- Affected regions: United States, China

## Recent score history

- 2026-08-27: 47
- 2026-08-28: 45

## Recent evidence

- [confirms] OpenAI Publishes First Benchmarks for Its Jalapeño Inference Chip Against Nvidia Systems (2026-08-27): OpenAI published its first public benchmarks for the 700-watt Jalapeño inference chip, claiming 1.5 to 1.9 times the throughput per kilowatt of Nvidia's GB300. A frontier lab moving from disclosing a chip program to publishing comparative performance numbers signals the part is close enough to deployment to be defended in public, not a paper design.
- [confirms] OpenAI Publishes First Jalapeño Benchmarks, Claiming Up to 1.9 Times Nvidia's Throughput Per Kilowatt (2026-08-26): OpenAI published its first benchmarks for Jalapeño, the 700-watt inference chip co-developed with Broadcom, claiming up to 1.9x Nvidia's throughput per kilowatt against 1,200W and 1,400W Nvidia rack systems on GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. Moving from announced silicon to published head-to-head numbers is the step that lets a lab justify shifting its highest-volume inference off merchant GPUs, and it hands Broadcom rather than Nvidia the design value.