OpenAI's In-House Chip Jalapeño Outperforms NVIDIA GB300 in Inference
Nashnova编辑部
OpenAI published its first benchmark data for Jalapeño, its custom inference chip, showing it beats Nvidia's GB300 on both performance and efficiency. The catch: Jalapeño handles inference only — OpenAI still depends on Nvidia for training.
What did Jalapeño actually score?
Benchmarks covered three models: DeepSeek R1, Kimi K2.5 with 1 trillion parameters, and OpenAI's latest open-source model.
Across all three, Jalapeño beat Nvidia's GB300 and other rival chips on both performance and energy efficiency.
This means → OpenAI now has hard data showing its own silicon can outperform the market leader — at least for inference workloads.
If it beat Nvidia, why can't OpenAI quit Nvidia?
Jalapeño's key limitation: it only handles inference (running trained models), not training new ones.
In plain terms = inference is "taking the exam"; training is "studying from scratch" — Jalapeño can take the exam but cannot study.
On the training side, OpenAI remains heavily dependent on Nvidia and other external suppliers. The custom chip cannot replace Nvidia's core role.
Where do production and commercialization stand?
A limited number of Jalapeño systems ship this year; capacity scales up next year. The second-gen chip is in deep development; third-gen planning has begun.
VP of Hardware Richard Ho: "Our own needs are so large, we can't imagine when we'd be able to sell to anyone else."
This means → Jalapeño is a purely internal tool for now — it will not become a chip-selling business.
What does this signal for the industry?
In-house chip design is now a trend across AI-model companies: Google has TPU, Amazon and Microsoft each have custom AI chips, and Anthropic confirmed this month it is developing its own.
OpenAI deliberately benchmarked against open-source models so outsiders can compare apples to apples — yet it has no plans to supply chips externally.
This reflects a core logic: large-model companies build chips not to sell silicon, but to reduce dependence on Nvidia and control their own compute costs.
What metric should investors watch next?
The share of OpenAI's total compute that Jalapeño and its successors ultimately account for is the key gauge of its compute-independence progress.
In plain terms = the higher the in-house chip share, the stronger OpenAI's bargaining power against Nvidia — and the more controllable its operating costs.
Richard Ho framed it clearly: "Jalapeño is one part, Nvidia is one part, AMD is one part" — a multi-supplier strategy, not a single-vendor bet.
Content is for reference only, not financial advice.