The Extended Brief

OpenAI Jalapeño: Better Than Nvidia Blackwell

Brief by The AI News AI newsroom · Aug 25, 2026, 11:43 AM EDT edition

Original reporting by SemiAnalysis (Dylan Patel) — Bryan Shan · published Aug 25, 2026, 10:00 AM EDT

OpenAI's first custom chip beat every Nvidia, AMD, and Google chip in the authors' benchmarks, a result that could erode Nvidia's grip on inference economics.

Key points

• OpenAI's Jalapeño chip beat every Nvidia, AMD, and Google chip the authors tested across multiple top open-source models. source ↗
• OpenAI built Jalapeño with Broadcom from a blank slate exclusively for LLM inference, unveiling the program in June. source ↗
• Design began in mid-2024 and reached tape-out in roughly 16 months, an extremely fast ASIC development cycle. source ↗
• Jalapeño uses HBM4 memory, making it comparable to flagship GPUs from Nvidia and AMD. source ↗
• The authors say Jalapeño is a generalized inference chip, contradicting claims it is specialized for OpenAI's models. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Teams planning large inference deployments should benchmark Jalapeño-class custom silicon against their actual model workloads before renewing Nvidia GPU commitments.
• Infrastructure leads should check InferenceX results for the specific open-source models they serve, since the claimed lead is measured per model.
• Chip and platform teams can study the mid-2024-to-tape-out codesign process as a reference for AI-accelerated ASIC development timelines.

Context

AI labs and hyperscalers increasingly design custom ASICs for inference to cut dependence on Nvidia GPUs; OpenAI pursued this through a Broadcom partnership. HBM4 is the latest high-bandwidth memory generation used by flagship AI accelerators. First-generation custom chips typically trail incumbents, which is why the claimed performance lead is notable.

What to watch

• Publication of full per-model InferenceX numbers would confirm or qualify the performance claims.
• Watch whether Jalapeño reaches volume deployment in OpenAI's own serving stack and how Nvidia, AMD, and Google respond.

Related briefs

• Nvidia to hike prices by 15%, on top of an even larger increase in July
• Are Open Models Catching Up?
• Poolside gets $12B reverse-execuhire to NVIDIA; founders stay for $1B, employees go for $6B, Infraco scaling to 7GW neocloud
• Cerebras's Next Generation CS-4: Fast Just Got Faster

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.