OpenAI's Jalapeño chip delivers inference 1.9 times faster than Nvidia Blackwell

Summary

OpenAI and Broadcom unveiled the Jalapeño inference ASIC at Hot Chips 2026. The companies claim it delivers up to 1.9x the data throughput of Nvidia Blackwell in SemiAnalysis' InferenceX benchmark. This performance is driven by early HBM4 adoption (15.4 TB/s per package) and a deliberately non-disaggregated prefill/decode architecture. The chip reached tape-out in just 16 months, with AI (including Codex) used for both hardware design and software kernel generation.

Why it matters

This represents the first significant hardware challenge to Nvidia's CUDA-centric inference dominance from a major model lab. It shifts the bottleneck in energy-constrained data centers from raw FLOPS to performance per watt. If the B0 revision (TSMC N3P, 700 W per die) delivers the projected 25% efficiency gain, it could pressure Nvidia's Rubin roadmap (900–1150 W, late 2026). This could also shift bargaining power in the AI infrastructure supply chain.

Highlights

• A 16-month design-to-tape-out cycle leveraging AI-assisted hardware design and software kernels generated with Codex
• HBM4 at 15.4 TB/s per package and a homogeneous chip pool design that avoids costly inter-server KV-cache transfers
• The B0 revision targets 700 W per die, compared to Rubin's 900–1150 W range, with an estimated 25% efficiency improvement

Critical take

The 1.9x figure in the headline relies on a single third-party benchmark (SemiAnalysis InferenceX). The article itself admits the chip is a first-generation ASIC, which is unusual given that first-gen parts are typically not competitive. Additionally, the comparison with Nvidia Rubin is speculative, as Rubin is not expected until late 2026. The source is a French developer portal rather than a primary technical disclosure, meaning the figures remain independently unverified.