OpenAI’s New AI Chip Just Took Aim at Nvidia, And the Numbers Are Interesting

OpenAI’s New AI Chip Just Took Aim at Nvidia, And the Numbers Are Interesting

For years, the AI boom has had a familiar supporting character: Nvidia.

Every time AI models get smarter, companies need more computing power to run them. And for much of that journey, Nvidia’s powerful GPUs have been the hardware behind the scenes.

Now, OpenAI wants to change that.

On August 25, OpenAI revealed the first public benchmark results for Jalapeño, its custom AI inference chip developed with Broadcom. The results are turning heads because, in several tests, OpenAI says its new chip delivered more AI work per watt and lower response times than Nvidia’s high-end Blackwell systems.

And yes, the name really is Jalapeño.

So, How Good Is It?

OpenAI tested Jalapeño using SemiAnalysis’ public InferenceX benchmark across three large models: GPT-OSS 120B, DeepSeek R1, and the 1-trillion-parameter Kimi K2.5.

According to OpenAI, Jalapeño delivered between 1.5x and 1.9x more AI work per watt at peak throughput, while end-to-end response latency was between 1.7x and 3.6x lower than the Nvidia systems used for comparison.

In simpler terms: OpenAI is saying its chip can potentially do more AI work while using less power and return responses faster.

That matters because running AI at massive scale is incredibly expensive. Every extra watt, server, and second of processing adds to the bill.

But There Is a Catch

The headlines make the result sound like OpenAI has just defeated Nvidia.

That would be an oversimplification.

First, Jalapeño is designed for inference, meaning it is built to run already-trained AI models. It is not a replacement for the hardware used to train those models. OpenAI will therefore continue relying on Nvidia and other computing platforms for major workloads.

Second, the benchmark results came from OpenAI's testing. SemiAnalysis verified the InferenceX runs in person with OpenAI engineers but said it did not independently run the entire benchmark suite itself. That makes the results more credible than a simple company claim, but they are still not the same as a completely independent benchmark. 

And perhaps most importantly, Nvidia's next-generation Vera Rubin platform was not part of the direct comparison. Jalapeño was mainly compared with Nvidia's current Blackwell-generation systems.

There is another wrinkle: some comparisons use single-token prediction, while Nvidia systems can use multi-token prediction techniques that can improve efficiency in production. SemiAnalysis notes that when the comparison is adjusted for multi-token prediction, Jalapeño's advantage becomes narrower.

Why OpenAI Is Building Its Own Chips

This is probably the most important part of the story.

OpenAI is not simply trying to make a faster chip.

It is trying to control more of the AI stack.

The company already builds AI models and products such as ChatGPT and Codex. Now it is increasingly designing the hardware and software needed to run those products.

Jalapeño was reportedly developed from design to tape-out in roughly nine months, with OpenAI's own AI models helping engineers accelerate parts of the development process.

OpenAI says the chip is designed to reduce data movement and keep important model information closer to where it is being processed. That approach matters especially when serving huge numbers of AI requests simultaneously.

And this is not supposed to be a one-off experiment.

OpenAI says Jalapeño is the beginning of a multi-generation custom computing platform, with more chips already in development.

What Happens Next?

Do not expect Jalapeño-powered ChatGPT to appear everywhere tomorrow suddenly.

OpenAI says deployment will begin in very small volumes toward the end of 2026, with a much larger rollout expected in 2027. 

So, this is not the end of Nvidia's dominance.

But it could be the beginning of something much more interesting.

If OpenAI can repeatedly design chips that run its AI systems faster and more efficiently, it could reduce some of its dependence on expensive third-party hardware while potentially lowering the cost of delivering AI to millions or even billions of users.

The real story is not that OpenAI beat Nvidia in one benchmark.

It is that one of the world's biggest AI companies is now building the machines its AI needs to survive.

And if Jalapeño is only the first generation, the next chip could be where things get really spicy.