OpenAI is bringing a custom chip into the fight for AI inference, and its first results put Nvidia’s Blackwell systems directly in the crosshairs. Called Jalapeño, the chip is expected to begin deployment inside OpenAI’s computing infrastructure by the end of the year, creating a new path for faster responses, responsive agents, and more reliable access.
That shift matters because OpenAI has been one of Nvidia’s largest customers, buying huge volumes of GPUs to train and run large AI models. Now, the company is developing its own hardware with Broadcom and says Jalapeño has already delivered “industry-leading speed and efficiency.”
Jalapeño Targets AI Inference
OpenAI designed Jalapeño for inference, the stage where trained AI models generate answers and carry out requests. The company announced the chip in June, describing it as “built from the ground up for current and future LLMs across the industry.” On Tuesday, OpenAI unveiled the first benchmarking results for the semiconductor.
OpenAI said, “Jalapeño beat Nvidia’s Blackwell systems on performance per watt in nearly all tested scenarios.” Performance per watt measures how much computing work a chip delivers for the power it consumes, making it a central metric for large AI systems that require vast amounts of electricity, cooling, and supporting infrastructure.
Research firm SemiAnalysis reached the same broad result, finding that Jalapeño beat Blackwell on performance per watt in nearly all tested scenarios. The firm also described the comparison as “somewhat incomplete and unfair” because Jalapeño uses newer HBM4 memory.
Nvidia’s Rubin platform offers a closer comparison because it also uses HBM4. Vera Rubin systems are starting to ship to customers right now, while OpenAI still needs time before it has anything beyond engineering samples of Jalapeño. That leaves an important gap between an impressive benchmark and a large working deployment.
Even with that limitation, the result sends a clear signal. Adrien Sanchez told CNBC that a “hyperscaler-designed chip can now match or beat Nvidia’s Blackwell-class GPUs on inference efficiency.”
Efficiency Could Reshape AI Infrastructure
OpenAI said its first benchmarking results could help users get “faster responses, more responsive agents and more reliable access.” Those gains would matter across the systems that serve large language models, where every improvement in chip efficiency can affect how much computing infrastructure an AI company must build and operate.
Alexander Harrowell called Jalapeño an “impressive achievement, most of all in terms of efficiency.” He added, “In a large-scale deployment, this would save power, cooling, and power distribution infrastructure, and contribute a lot to their unit economics.”
OpenAI is not stopping with one chip. The company said it is already working on Jalapeño’s second and third generations, showing that the project reaches beyond a single experiment or limited hardware test.
The timing connects Jalapeño to a wider move among major AI companies. In April, a group of deals highlighted the growing push toward custom AI chips, also known as application-specific integrated circuits, or ASICs.
- Google unveiled new chips for AI training and inference called tensor processing units, or TPUs.
- Meta agreed to deploy 1 gigawatt of custom AI chips using Broadcom technology as part of a multi-GW deal.
- Anthropic committed to spending more than $100 billion on AWS tech over the next 10 years, including current and future generations of Amazon’s custom AI chips, Trainium.
Omdia expects custom ASIC chips like Jalapeño to exceed GPUs in volume by 2028. About half of all capital expenditure on AI infrastructure comes from hyperscale cloud providers that either have a custom chip program or could reasonably have one.
Nvidia’s Biggest Customer Builds Its Own Route
The rise of custom silicon does not erase Nvidia’s position. Nvidia’s Blackwell systems remain part of the comparison, and Vera Rubin systems are already starting to ship to customers while Jalapeño remains on the road from engineering samples to deployment.
But OpenAI’s move changes the relationship. The AI lab has relied on Nvidia GPUs to train and run its large AI models, and having its own chip could affect that connection. Sanchez said OpenAI had been “one of the largest single consumers of Nvidia GPUs,” adding that Jalapeño “raises the stakes for Nvidia’s largest customer relationship specifically.”
That pressure reaches beyond one company. Google, Meta, Amazon through AWS, Anthropic, and OpenAI are all tied to the expanding contest over who designs the hardware powering AI training and inference. Broadcom’s role in both Jalapeño and Meta’s custom chip deployment also shows how semiconductor partnerships are becoming part of that contest.
Jalapeño now has a demanding test ahead: moving from engineering samples into OpenAI’s infrastructure by the end of the year. If its efficiency results hold at scale, OpenAI will have more control over the systems serving its models, while Nvidia will face a customer building an alternative from the ground up.
The next generation of AI competition may not be decided by models alone. It may be decided by which chips deliver the most useful intelligence for every watt, every cooling system, and every part of the infrastructure underneath.
Based on
