Can Sam Altman’s Jalapeño chip beat Nvidia?
Could OpenAI's own AI chip challenge Nvidia? Jalapeño delivers lower latency and better performance per watt in selected tests.
OpenAI has shared its first performance results for Jalapeño, its custom AI inference chip, and the numbers are big enough to get the AI hardware world’s attention.
The company says Jalapeño can handle more AI work per watt while responding faster than leading commercial systems in its tests. That sounds like a direct challenge to Nvidia.
But the reality is a little more interesting. Jalapeño may have an edge in specific inference workloads, while Nvidia still has a huge advantage when it comes to the wider AI hardware ecosystem. Here's what you need to know!
What Jalapeño is actually built for
Jalapeño is not trying to do everything. It is built for inference, the stage where an AI model answers questions, writes code, searches data or takes actions as an agent. Training is where models learn. Inference is where they are used repeatedly, often by millions of people.
That makes inference a high-cost and performance problem for OpenAI. As AI agents take on more tasks, they may need to complete several steps in sequence. Even small delays can add up. A custom chip that reduces latency, or waiting time, could make those experiences feel much faster.
The numbers are hard to ignore
OpenAI says Jalapeño was faster than Nvidia’s systems in all three tests. For GPT OSS 120B, it returned a response in 1.03 seconds, compared with 1.80 seconds for Nvidia’s GB200.
Whereas DeepSeek R1 took 1.65 seconds, Nvidia’s GB300 took 5.99 seconds. And for Kimi K2.5, it took 1.56 seconds compared with 5.31 seconds.
The chip also used less power. Jalapeño is rated at 700 W, compared with 1,200 W for Nvidia’s GB200 and 1,400 W for the GB300. OpenAI says its chip’s actual power use stayed at or below 550 W during the tests.
Overall, OpenAI says Jalapeño delivered 1.5 to 1.9 times more AI work for each watt of power. In simple terms, it could do up to 90% more work without using more electricity.
That is a big deal when electricity, cooling and data centre capacity are becoming increasingly expensive parts of running AI.
Nvidia still has a very big head start
Here is where things get interesting. Nvidia is not just selling chips. It has built a huge hardware and software ecosystem that developers, cloud providers, researchers and businesses already know how to use. Jalapeño is different. It is first-party silicon designed primarily around OpenAI’s own workloads.
OpenAI says it plans to start deploying Jalapeño within its compute infrastructure by the end of the year, while continuing to widely deploy Nvidia and other accelerators for training and inference.
OpenAI wants more control
The more interesting part of Jalapeño may be what it says about OpenAI’s strategy. The firm says AI helped speed up the chip’s development, including taking it from initial design to tapeout in nine months.
It also says AI-generated implementations for selected GPT-OSS blocks ran 1.5 to 1.8 times faster than human-written existing versions. Those gains apply only to selected blocks, not the full model.
Still, they point to an interesting possibility: OpenAI could increasingly design its models, software and chips together. That could give the company more control over performance and costs as its AI workloads grow.
So, can Jalapeño beat Nvidia?
On specific inference workloads, OpenAI’s early results suggest it can outperform Nvidia systems on speed and power efficiency. But beating Nvidia across the wider AI hardware market is a completely different challenge. Nvidia’s ecosystem, software support and scale are difficult to replicate.
For now, the more realistic outcome is coexistence. Nvidia remains important to OpenAI, while Jalapeño could give the company a more efficient way to run its own products.