OpenAI's Jalapeño chip claims better efficiency than Nvidia

Key Takeaways

• OpenAI's Jalapeño is an inference-only chip, meaning Nvidia hardware likely remains essential for training
• OpenAI used its own AI models to help design the chip, a recursive approach that could accelerate future iterations
• The move signals OpenAI's intent to control every layer of the AI stack, from models to silicon

OpenAI has unveiled Jalapeño, its first custom-built inference processor, developed with Broadcom. The company claims early testing shows the chip delivers significantly better performance-per-watt than current alternatives. That's a direct shot at Nvidia, whose GPUs have powered nearly every major AI breakthrough of the past decade.

The chip is purpose-built for inference, the process of running trained AI models in response to user queries. Training, which requires far more compute, will likely continue on Nvidia hardware. But inference is where OpenAI racks up ongoing costs every time someone uses ChatGPT, Codex, or any of its agentic products. Even modest efficiency gains there compound fast.

Why did OpenAI build its own chip?

The short answer: cost and control. OpenAI reportedly spends billions annually on compute, most of it flowing to Nvidia. Custom silicon lets the company optimize specifically for its own workloads rather than relying on general-purpose GPUs.

"We have a deep understanding of the workload," OpenAI president Greg Brockman said on the company's podcast. "We've really been looking for specific workloads that are underserved, [and asking] how can we build something that will be able to accelerate what's possible?"

Google and Amazon have followed similar logic. Google's TPUs power its internal AI systems and cloud offerings. Amazon's Trainium and Inferentia chips serve AWS customers. Both companies saw vertical integration as a way to reduce Nvidia dependency and capture more value in-house.

What makes Jalapeño different from Nvidia GPUs?

Nvidia's GPUs are general-purpose workhorses. They handle training, inference, gaming, scientific simulation, and more. That flexibility comes with tradeoffs. A chip designed for one specific task can strip away everything else and optimize ruthlessly.

Jalapeño targets inference on OpenAI's coding models, where the company emphasized low operating costs. For a product like Codex that processes millions of code completions daily, even a 20% efficiency gain translates to real money.

OpenAI also disclosed that its own AI models assisted in designing the chip. That's a recursive loop worth watching. If AI-assisted chip design produces better AI chips, which then train better AI models, the cycle could accelerate quickly.

Does this threaten Nvidia's dominance?

Not immediately. Inference-only chips don't replace training infrastructure, and Nvidia's H100 and B200 GPUs remain the gold standard for large-scale model training. OpenAI still needs Nvidia for that side of the business.

But the broader trend is clear. Every major AI company is looking to reduce Nvidia exposure. Meta, Google, Amazon, Microsoft, and now OpenAI all have custom chip efforts at various stages. Nvidia's roughly 80% market share in AI accelerators faces pressure from multiple directions.

For Nvidia, the risk isn't one competitor winning. It's the aggregate effect of every large customer chipping away at the parts of the workload they can handle themselves.

OpenAI's full-stack ambition

The company framed Jalapeño as part of a broader strategy to control the entire AI stack. "OpenAI is not only developing frontier models or building products on top of them; it is designing the infrastructure underneath them: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience," the company wrote in its announcement.

That's the Apple playbook applied to AI. Own the silicon, the software, and the services. Optimize each layer around a unified goal. It's ambitious, expensive, and potentially powerful if executed well.

The Broadcom partnership makes execution more feasible. Broadcom has decades of experience in custom chip design and manufacturing relationships with foundries like TSMC. OpenAI brings the AI expertise and the specific workload knowledge. Neither could do this alone.

Frequently Asked Questions

What is OpenAI's Jalapeño chip used for?

Jalapeño is designed specifically for AI inference, the process of running pre-trained models to respond to user queries. It's not intended for training new models.

Does Jalapeño replace Nvidia GPUs for OpenAI?

No. OpenAI will likely continue using Nvidia GPUs for training, which requires different hardware characteristics. Jalapeño targets inference workloads only.

Who manufactured OpenAI's Jalapeño chip?

Broadcom designed and manufactured the chip in partnership with OpenAI. The partnership was announced in October 2025.

How does Jalapeño compare to Google's TPU?

Both are custom AI accelerators. Google's TPUs handle both training and inference at scale. Jalapeño appears focused solely on inference, optimized for OpenAI's specific model architectures.

When will Jalapeño be available?

OpenAI says the chip is still being tested. No public availability date has been announced. It's intended for internal use powering OpenAI's products.

Source: TechCrunch / Russell Brandom