OpenAI’s New Chip Is Built to Make AI Faster and More Efficient

The next noticeable improvement in artificial intelligence may not begin with a new app. It may start inside the data center, where a chip designed for one job can help AI respond faster while using power more efficiently.

OpenAI Has Built Its First Inference Chip

OpenAI published the first measured results for Jalapeño on August 25. The company describes it as its first custom inference chip: hardware designed specifically for the stage when a trained AI model processes a request and produces an answer.

That distinction matters. Training creates a model, but inference is what happens every time someone asks a chatbot a question, gives an agent a task, or calls an AI feature through an application. A faster inference system can make the experience feel more immediate, especially when an agent must complete several steps in sequence.

Jalapeño was developed with Broadcom and other infrastructure partners as part of a multigeneration platform. OpenAI says the chip gives it more control over the speed, efficiency, and economics of serving its models, while continuing to use accelerators from other companies.

The Early Results Focus on Speed and Power

According to OpenAI’s August 25 engineering report, Jalapeño was tested with GPT-OSS 120B, DeepSeek R1, and Kimi K2.5. Across those three models, OpenAI reports that the chip completed 1.5 to 1.9 times more AI work per watt at peak throughput and delivered 1.7 to 3.6 times lower end-to-end latency than the comparison systems. For highly interactive workloads, the company reports performance improvements of 2.1 to 4.1 times.

Those are company-reported benchmark results, not a guarantee that every product or request will improve by the same amount. The tests used the public InferenceX benchmark and compared systems at matched response-speed targets. The useful idea is not one dramatic number; it is the possibility of doing more AI work without increasing power use at the same rate.

OpenAI rates the chip at 700 watts, although it says sustained power remained at or below 550 watts in the workloads it tested. That could become important as the demand for AI data centers continues to put pressure on electrical capacity and operating costs.

Faster Hardware Could Make Agents Feel More Natural

Latency is the delay between a request and a response. A small delay may barely matter for a single answer, but it can compound when an AI agent must search, reason, use a tool, check a result, and then continue.

OpenAI says the practical goal is faster responses, more responsive agents, and more reliable access as demand grows. Its related full-stack overview explains that the company wants to optimize models, serving software, chips, memory, and networking together instead of treating each layer as a separate problem.

If the approach works at production scale, users may notice the result as less waiting rather than as a new button or feature. That is an inference from the reported performance, not a promise about a particular ChatGPT plan or device.

AI Helped Build the Hardware

The chip is also a story about AI being used inside engineering. OpenAI says the design moved from concept to manufacturing tape-out in nine months and that its models assisted parts of the design, programming, and optimization process.

The company says engineers used Codex with an internal model to bring models that were not part of the original chip plan to strong performance within two months. It also reports faster results for selected implementation blocks. That does not mean AI designed the entire chip on its own. It means AI tools were used to accelerate specific engineering work under human direction.

What Happens Next

OpenAI says it plans to begin deploying Jalapeño within its own infrastructure by the end of 2026, while later generations are already in development. The chip is not being presented as a consumer product that people can buy, and the company has not said that every current ChatGPT request is already running on it.

The bigger shift is easy to understand: AI companies are designing more of the machinery beneath their services. If Jalapeño’s reported gains hold up as deployment expands, the most important benefit may be an AI system that can serve more people, respond more quickly, and use expensive data-center power more effectively.

Sources

• OpenAI — Jalapeño’s first measured performance results, August 25, 2026

• OpenAI — The full stack behind abundant intelligence, August 25, 2026

• OpenAI and Broadcom — Jalapeño announcement, June 24, 2026

Advertisement