On June 24, 2026, OpenAI unveiled its first custom AI chip, "Jalapeño," developed jointly with Broadcom. The ASIC is purpose-built for large language model inference and is part of a broader push to reduce dependence on Nvidia and optimize costs.
June 24, 2026 · OpenAI × Broadcom × Celestica
OpenAI Unveils "Jalapeño" — Its First In-House Inference Chip
A custom ASIC purpose-built for LLM inference behind ChatGPT, Codex and the API — designed by OpenAI, fabbed with Broadcom, and aimed at gigawatt-scale data center deployment from late 2026.
9 mo
Design-to-tape-out cycle — possibly the fastest high-performance ASIC ever, aided by OpenAI's own models
GW
Gigawatt-scale roadmap across multiple chip generations
Late '26
Targeted start of large-scale data center deployment
Who designs what — three hands, one chip
Full-stack co-design from architecture to rack.
OpenAI
Architecture from scratch
→
Broadcom
Silicon implementation
→
Celestica
Board · rack · system
The custom-silicon race
Hyperscalers building their own AI accelerators to cut cost and Nvidia dependence.
Company
Custom chip
Primary use
Google
TPU
Training / inference
Amazon
Trainium / Inferentia
Training / inference
SpaceX / xAI
Terafab (in-house GPU)
Supply-risk mitigation
The strength
Full-stack co-design: models, kernels, serving + hardware
Early lab perf-per-watt "substantially better" than current state-of-the-art
Lower inference cost could lift margins
The caveats
Still lab-testing — not yet in production
No independent third-party benchmarks
Training likely stays Nvidia-dependent
Detailed report still pending
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…