OpenAI has published the first results for Jalapeño, its own custom ASIC accelerator built for AI inference. The numbers put it ahead of NVIDIA’s Blackwell GB200 and GB300 systems on all three workloads tested. On GPT-OSS 120B it hits 85,448 mixed tokens per second per kilowatt, 1.9 times the comparison system, and 1,459 tokens per second per user at its lowest-latency point. The measurements come from the SemiAnalysis InferenceX benchmark, and they don’t yet prove Jalapeño is faster than Blackwell across every AI workload.
The 20-second version
- Jalapeño delivers 1.5x to 1.9x more performance per watt across the three published models.
- It reaches 1,459 tokens/sec per user on GPT-OSS 120B.
- The comparisons use NVIDIA GB200 and GB300, depending on the model.
- OpenAI also tested DeepSeek R1 670B and Kimi K2.5 1T.
- Internal deployment is expected to start before the end of 2026.
What makes this notable is as much who is behind it as the numbers. OpenAI has been one of the biggest buyers of NVIDIA accelerators for years, and it plans to keep buying them. Jalapeño adds a second track: silicon designed specifically for inference, pulling together accelerator hardware, memory, interconnects, software, and system architecture.
The published results span models with very different profiles: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 at one trillion parameters. SemiAnalysis says Jalapeño beat the NVIDIA, AMD, and Google accelerators it evaluated against several open large models. That widens the comparison, but it should be kept separate from the specific NVIDIA numbers OpenAI put in its own charts.
Up to 1,459 tokens per second per user
One of the standout figures is the top decoding speed per user.
On GPT-OSS 120B, Jalapeño reaches 1,459 tokens/sec per user against 535 for the reference GB200 system, about 2.7 times faster.
The gap grows with DeepSeek R1: 700 versus 169 tokens/sec, roughly 4.1 times. With Kimi K2.5 it manages 694 tokens/sec against 182 from the GB300, about 3.8 times.
| Model | Jalapeño | Compared NVIDIA | Advantage |
|---|---|---|---|
| GPT-OSS 120B | 1,459 tokens/sec/user | GB200: 535 | 2.7x |
| DeepSeek R1 670B | 700 | GB300: 169 | 4.1x |
| Kimi K2.5 1T | 694 | GB300: 182 | 3.8x |
That metric matters most for virtual assistants and AI agents, where the time to produce each response adds up across the steps of a longer task.
OpenAI says one goal for Jalapeño was to avoid the usual trade-off between throughput and latency. A system can handle many requests at once through batching, but that doesn’t guarantee the fastest reply to each user. Jalapeño tries to hold on to both efficiency and low latency in a single architecture.
The end-to-end latency numbers show the same thing: 1.03 seconds versus 1.80 with GPT-OSS, 1.65 versus 5.99 with DeepSeek R1, and 1.56 versus 5.31 with Kimi K2.5.
The real fight: tokens per kilowatt
Raw speed is only part of the story for a data center. Available electricity has become a hard limit on deploying AI infrastructure at scale, so OpenAI leaned on performance per unit of power as one of its main metrics.
Here are the results:
| Model | Jalapeño | NVIDIA | Difference |
|---|---|---|---|
| GPT-OSS 120B | 85,448 tokens/sec/kW | GB200: 44,960 | 1.9x |
| DeepSeek R1 670B | 19,641 tokens/sec/kW | GB300: 11,781 | 1.7x |
| Kimi K2.5 1T | 18,195 tokens/sec/kW | GB300: 11,862 | 1.5x |
The company normalized the comparison using each accelerator’s rated power: 700 W for Jalapeño, 1,200 W for the GB200, and 1,400 W for the GB300. It adds that the chip held steady performance in tests at 550 W or less.
That 85,448 tokens/sec/kW figure needs a caveat. It doesn’t mean one Jalapeño puts out that speed nonstop. The metric normalizes system performance against power so you can compare efficiency.
It also doesn’t prove Jalapeño is 90% more powerful than Blackwell in general. What it shows is that on GPT-OSS 120B, under the specific InferenceX method, Jalapeño gets about 1.9 times the mixed throughput per kilowatt of the GB200 it was measured against.
InferenceX reports results across many configurations and shows how performance, cost, and efficiency shift with model size, precision, sequence length, and how interactive the workload is.
Making sense of the 104x figure
Another eye-catching result is 104.3 times.
Reading that as Jalapeño being 104 times faster than an NVIDIA GB300 would be wrong.
OpenAI ran a different test. It fixed both systems at the top per-user speed NVIDIA’s platform could reach, then measured how much throughput per kilowatt each architecture could hold at that point.
On GPT-OSS at about 535 tokens/sec per user, Jalapeño delivered 22,935 mixed tokens/sec/kW against 427 for the GB200, 53.7 times more.
With DeepSeek R1, at around 169 tokens/sec per user, it was 12,258 versus 118, the 104.3x gap. For Kimi K2.5, at about 182 tokens/sec per user, 6,744 versus 120, or 56.1 times.
The test shows the cost of pushing an architecture toward very low latency. The Blackwell system’s total capacity drops sharply when it has to sustain that per-user speed for everyone at once.
That is where OpenAI wants its ASIC to stand apart.
Language model inference isn’t one repeated operation. During prefill, the accelerator processes the prompt and the compute demand is high. During decode, it generates tokens one after another and memory bandwidth matters far more.
Jalapeño is built to cut data movement between these phases. OpenAI says it can place the model state, including the KV cache, and keep it local while handing out compute, memory, and network resources as each stage needs them.
That is a real difference from architectures that split prefill and decode across separate resources. It doesn’t make that split worse: NVIDIA and modern inference engines already get very competitive results with those techniques, and InferenceX includes GB200 and GB300 setups using disaggregated inference to balance performance and interactivity.
OpenAI wants more control over its own infrastructure
Jalapeño isn’t a break from NVIDIA hardware.
OpenAI says plainly that it will keep deploying plenty of NVIDIA and partner accelerators for both training and inference. For now the ASIC is one more piece of an infrastructure that keeps needing more capacity.
The change is that OpenAI is starting to own part of the silicon its services run on.
SemiAnalysis reports the chip was developed with Broadcom, starting in mid-2024. For a company that buys accelerators by the data center, better tokens per megawatt can carry real economic weight, even without replacing GPUs outright.
There is another twist: AI helped design and program the accelerator itself.
OpenAI says it went from first design to tapeout in nine months, using its models to explore implementations, shorten design and verification cycles, and tune arithmetic circuits. It then used Codex with GPT-Astra to port three open-weight models to Jalapeño in about two months.
In some attention blocks and Mixture of Experts layers of GPT-OSS, AI-generated implementations ran 1.5 to 1.8 times faster than the ones people had written before. OpenAI is careful to say that applies to those blocks, not the whole model.
The next test is bigger than any benchmark. OpenAI plans to start folding Jalapeño into its infrastructure before the end of 2026, while it keeps qualifying the chip for production and building out the software stack. A second-generation chip is already in the works, and a third is being defined.
Until the hardware runs at scale in production, one piece is still missing: how these numbers hold up as availability, total cost of ownership, and sustained performance under real workloads.
Even so, the first results make one thing hard to argue: OpenAI’s first-generation ASIC can go toe to toe with Blackwell on inference and, in the published tests, beat NVIDIA systems on latency and performance per watt. For a company whose token use keeps climbing, that efficiency may matter more than winning a FLOPS race.
Frequently Asked Questions
Is Jalapeño faster than NVIDIA Blackwell?
In the published tests with GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, Jalapeño beats the reference GB200 or GB300 systems on both performance per watt and top decoding speed. That doesn’t automatically mean the same edge holds for every model or workload.
How many tokens per second does OpenAI Jalapeño reach?
Its highest published per-user throughput is 1,459 tokens/sec on GPT-OSS 120B. It also reaches 700 tokens/sec on DeepSeek R1 and 694 tokens/sec on Kimi K2.5.
Will Jalapeño replace NVIDIA GPUs at OpenAI?
Not under the current plan. OpenAI says it will keep deploying NVIDIA and other vendors’ accelerators for training and inference as it gradually brings in its own silicon.
When will OpenAI start using Jalapeño?
It expects to start deploying it in its infrastructure before the end of 2026. Jalapeño is the first of a family, with a second generation in development and a third planned.
