OpenAI's Jalapeño chip challenges Nvidia's AI lead
August 28, 2026
OpenAI reports substantially more performance per watt and lower latency from its first custom inference chip. The measurements matter, but they are not yet an independent production test.
What this is about
OpenAI published the first measured results for its custom Jalapeño inference chip on August 25, 2026. Developed with Broadcom, the chip is designed to run language models after training with high speed and energy efficiency. OpenAI plans to begin deploying Jalapeño in its own infrastructure by the end of 2026.
The published numbers address Nvidia’s strongest market position directly: serving large models in data centers. Depending on the model, OpenAI reports 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems. The tests covered GPT-OSS 120B, DeepSeek R1, and Kimi K2.5.
What Jalapeño actually does
Jalapeño is not a general-purpose processor or a training chip for every task. It is optimized for inference, when an already trained model processes requests and generates responses. Inference alternates between compute-heavy and memory-heavy phases. OpenAI says the chip, memory, network, and software were co-designed to reduce data movement and waiting time.
The chip is rated at 700 watts; OpenAI says measured sustained power remained at or below 550 watts in the tested workloads. Testing used InferenceX, a public benchmark from SemiAnalysis. In the GPT-OSS 120B test, OpenAI reports about 1.9 times higher peak throughput per kilowatt than the Nvidia GB200 comparison system. The reported advantage was about 1.7 times for DeepSeek R1 and 1.5 times for Kimi K2.5.
Why it matters
The operating cost of large AI services depends heavily on electricity, cooling, utilization, and response time. A chip that completes more requests per kilowatt can reduce the cost of each useful response. Lower latency matters especially for agents because delays accumulate across many sequential model calls.
Strategically, this is bigger than one benchmark. OpenAI is becoming a chip designer and can tune models, software, and hardware together. That does not immediately displace Nvidia: OpenAI explicitly says it will continue widely deploying accelerators from Nvidia and other partners. It does show that the largest model providers want to reduce their dependence on standard hardware over time. CNBC therefore described Jalapeño as a potential threat to the high margins of AI accelerators.
In plain language
Imagine a large restaurant kitchen. A general-purpose cook can prepare many dishes but loses time moving between tools and workstations. Jalapeño is like a kitchen laid out specifically for one frequently ordered meal: ingredients, stove, and pickup counter are close together, so it produces more portions with less energy. That makes the kitchen efficient but does not prove it wins for every menu.
A practical example
An AI service processes 100,000 requests during a peak hour. Each request triggers several model steps, and every extra second lengthens the overall task. If a specialized system handles more requests per kilowatt at comparable output quality, the operator can run the same load on less active hardware or serve more users.
The calculation is not simply “1.9 times faster means half the cost.” Purchase price, manufacturing yield, memory, networking, cooling, software maintenance, and real utilization determine total cost. Only sustained operation in production data centers will show whether the advantages persist beyond selected benchmarks.
Scope and limits
- The central figures come from the vendor: OpenAI published the results and selected the models, operating points, and comparison systems. InferenceX is public, but independent reproductions of the full Jalapeño tests are not yet available.
- A benchmark is not production: Reliability, manufacturing yield, software defects, maintenance, and changing request patterns are only partly visible in a performance chart.
- It is not a complete Nvidia replacement: Jalapeño targets inference. Training and many other workloads still need other accelerators, and OpenAI says it will continue broad Nvidia deployments.
- Cost savings remain unproven: Better performance per watt can reduce costs, but OpenAI has not disclosed unit pricing or complete operating costs.
SEO & GEO keywords
OpenAI Jalapeño, AI chip, inference, Nvidia GB200, Nvidia GB300, Broadcom, InferenceX, performance per watt, data center, AI hardware, latency, custom silicon
💡 In plain English
OpenAI built a custom chip intended to run trained language models faster and with less power. Early benchmarks look strong, but they largely come from OpenAI and still need validation in sustained production use.
Key Takeaways
- →OpenAI plans to deploy Jalapeño in its own infrastructure by the end of 2026.
- →The vendor reports 1.5 to 1.9 times more peak throughput per watt, depending on the model.
- →Jalapeño is optimized for inference and is not a general replacement for training accelerators.
- →The chip was compared with Nvidia systems using a public SemiAnalysis benchmark.
- →Independent reproductions and complete cost data are not yet available.
FAQ
What is an inference chip?
It runs an already trained AI model and generates responses. Training the model is a different and usually much more compute-intensive task.
Is Jalapeño faster than every Nvidia GPU?
That has not been established. OpenAI reports advantages for selected models and operating points; independent full tests are still missing.
Will OpenAI stop using Nvidia chips?
No. OpenAI explicitly says it will continue widely deploying accelerators from Nvidia and other partners.