OpenAI’s Jalapeño Wins the First Power Test. Production Is the Harder Benchmark.
The first 8K/1K results are strong. Agent workloads, production silicon, and system economics will decide whether OpenAI's custom ASIC becomes a platform.
Crafting seamless user experiences with a passion for headless CMS, Vercel deployments, and Cloudflare optimization. I'm a Full Stack Developer with expertise in building modern web applications that are blazing fast, secure, and scalable. Let's connect and discuss how I can help you elevate your next project!
OpenAI has published the first engineering-sample results for Jalapeño, its custom inference ASIC. The fixed 8K-input, 1K-output InferenceX benchmark covered three open-weight models. Jalapeño delivered 1.5 to 1.9 times more AI work per watt and 1.7 to 3.6 times lower end-to-end latency than the tested commercial systems. That is a meaningful first result. It is not proof that Jalapeño universally beats every NVIDIA, AMD, or Google accelerator.
The distinction matters because these are engineering samples running selected models under a controlled single-turn workload. OpenAI has not published AgentX results for long-running, multi-turn agent workloads, and production silicon is still ahead. The larger story is not one benchmark win. OpenAI is building a feedback loop that connects models, kernels, chips, and data-center capacity.
The Decision in One Table
| Evaluation area | What is verified | What remains open |
|---|---|---|
| Performance per watt | Peak throughput efficiency improved 1.5–1.9x across three tested models | Long-context, continuous multi-turn, and mixed-tenant workloads |
| Interactive latency | End-to-end latency fell 1.7–3.6x; minimum token interval improved 2.7–4.1x | Time to first token, tail latency, and complete agent-task duration |
| Power | The package is rated at 700W TDP; sustained power stayed at or below 550W in the tested workloads | Production clocks, cooling, yield, and rack-level efficiency |
| Software | Three models outside the original specification reached high performance in two months | Broader model coverage, tooling maturity, and operational cost |
| Schedule | OpenAI plans initial internal deployment by the end of 2026 | The reported 2027 production ramp, 100MW scale, and supply-chain execution |
What Jalapeño Is and What OpenAI Actually Announced
OpenAI introduced Jalapeño on June 24, 2026 as its first Intelligence Processor. Co-developed with Broadcom and Celestica, it is a blank-sheet LLM inference accelerator designed around OpenAI’s current and future models rather than a general-purpose GPU with incremental specialization.
The August 25 update did not introduce a second chip. It reported the first measured results from engineering samples. OpenAI used SemiAnalysis’s InferenceX methodology to test GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T with a fixed 8,000-token input and 1,000-token output. The comparison set consisted of leading commercially available systems selected for the benchmark. It did not include every chip generation, every software stack, or every production configuration in the market.
That boundary changes the headline. The public evidence supports a narrower and more useful claim: Jalapeño outperformed the tested systems for these three models, under this benchmark, on throughput efficiency and latency.
Reading the Three Results Correctly
OpenAI reports three related measurements: peak AI work per watt, end-to-end latency, and the fastest sustainable token interval in an interactive operating point.
| Model | Peak AI work per watt | End-to-end latency | Minimum token interval | Approximate generation rate per user |
|---|---|---|---|---|
| GPT-OSS 120B | 1.9x | 1.7x lower | 2.7x faster | 1,459 vs. 535 tok/s |
| DeepSeek R1 670B | 1.7x | 3.6x lower | 4.1x faster | 700 vs. 169 tok/s |
| Kimi K2.5 1T | 1.5x | 3.4x lower | 3.8x faster | 694 vs. 182 tok/s |
The results suggest that Jalapeño was not optimized only for aggregate data-center throughput. It also targets the cadence of a single interactive request. That matters for coding agents, voice interfaces, and interactive research systems, where average tokens per second can hide a poor user experience. Time to first token, streaming consistency, and tail latency under concurrency all matter.
The minimum token interval is still a best operating point in a controlled benchmark. It is not a service-level guarantee. Model routing, KV-cache hit rate, network topology, batching, speculative decoding, and request length can all move the observed result.
Why an 8K/1K Test Does Not Represent an Agent Workload
InferenceX uses a fixed sequence length to make systems comparable. It is useful for asking how much work a platform can complete, at what latency and power, for a known input and output shape. It is not designed to reproduce a long-running agent.
A real agent may call tools repeatedly, ingest new documents, accumulate conversation state, compact memory, and alternate between prefill and decode for tens of minutes. Those behaviors expose cache policy, routing, scheduling, and long-context management. SemiAnalysis created AgentX to cover more of this multi-turn, long-context behavior, but no Jalapeño AgentX result has been published.
The defensible conclusion today is that Jalapeño has earned a place in the next test round. It has not shown that it leads every agent workload. Capacity planners should not translate the 8K/1K ratios directly into complete-task cost.
Reducing Data Movement Is the Architectural Bet
OpenAI describes Jalapeño as a system that minimizes data movement while balancing compute, memory, and networking. Large-model inference is frequently constrained not by arithmetic alone but by moving weights, activations, and KV-cache data through the memory hierarchy. Every longer or more frequent transfer adds latency and energy cost.
Jalapeño keeps KV-cache data local to processing resources and coordinates the system through a large-scale interconnect. It also avoids a permanently fixed split between prefill and decode hardware pools. Operators can adapt the allocation to model and traffic characteristics. For OpenAI, whose models and serving patterns change quickly, that flexibility may be more valuable than optimizing one static model to an isolated peak.
SemiAnalysis’s engineering-sample analysis adds details such as HBM4, 15.4 TB/s of memory bandwidth, and a scale-up domain of 2,048 accelerators. These specifications help explain the system direction, but they come from independent analysis and should not be presented as final production commitments from OpenAI.
What AI-Assisted Chip Development Has Actually Proven
Jalapeño is also a test of AI-assisted hardware and software co-design. OpenAI says Codex and GPT-Astra helped bring three open-weight models that were not in the original chip specification to high performance in two months. That is evidence of faster model bring-up, kernel work, and performance tuning. It is not evidence that an AI autonomously designed the entire ASIC.
OpenAI also reports that AI-generated implementations of selected GPT-OSS attention and mixture-of-experts blocks ran 1.5 to 1.8 times faster than existing human-expert implementations. The scope is selected compute blocks, not the complete GPT-OSS model.
If this approach generalizes, it changes two development loops. Software teams can map models released after tapeout onto the fixed hardware more quickly. Model researchers can see hardware constraints earlier and adjust operators, data formats, and memory behavior. That model-to-silicon feedback loop is one of the hardest parts of OpenAI’s vertical integration for competitors to copy.
Nine Months and Sixteen Months Use Different Starting Lines
OpenAI says the chip moved from initial design to manufacturing tapeout in nine months. SemiAnalysis counts approximately 16 months from the start of core-team hiring in mid-2024 to tapeout. The figures are not contradictory; they measure different phases.
Saying only that the chip took 16 months obscures team formation and architectural exploration. Saying only nine months understates the preparation that preceded the formal design clock. The decision-relevant point is that OpenAI has assembled model research, hardware design, and AI-assisted software engineering into a shorter iteration loop.
This Is Not a Declaration of Independence From NVIDIA
The immediate value of Jalapeño is an additional supply path optimized for OpenAI’s own traffic and models. It may lower the cost of selected inference workloads, strengthen negotiating leverage, and move stable high-volume demand away from general-purpose accelerators.
OpenAI also states explicitly that it will continue broad deployment of NVIDIA and other partner accelerators. Successful engineering samples do not replace CUDA maturity, manufacturing volume, networking, rack-level systems, or operational tooling. SemiAnalysis’s separate modeling of Vera Rubin suggests that total cost of ownership may be close under some assumptions. The systems also use different speculative-decoding settings, so combining their headline numbers into one direct race would be misleading.
The nearer-term outcome is a heterogeneous data center. GPUs will handle workloads that require broad compatibility and rapid change. Custom ASICs will absorb large, stable traffic patterns where hardware and models can be co-designed. Compiler quality, scheduling, networking, and model-update speed may matter more than the peak number on one package.
Four Gates Before Production
1. A0 results must survive B0 and production silicon
The published measurements come from A0 engineering samples. SemiAnalysis reports that B0 is in fabrication and is expected to improve performance per watt by roughly 25 percent. That is forward-looking information, not a validated result. Frequency, yield, cooling, and packaging supply can all change the final outcome.
2. AgentX and mixed production loads
OpenAI needs evidence from long context, multi-turn tool use, concurrent models, and tail-latency conditions. Its internal deployment data will also need to separate gains from the chip, the software stack, and unique traffic patterns.
3. Supply and deployment cadence
OpenAI plans initial deployment in its own infrastructure by the end of 2026. The 2027 production ramp and 100MW target are reported by SemiAnalysis, not formal OpenAI commitments. Moving from samples to sustained deployment requires packaging, racks, networking, cooling, and field operations.
4. Economics must be calculated at system level
Throughput per watt is important, but it is not a bill. Full cost includes accelerator acquisition, servers, power conversion, cooling, networking, redundancy, utilization, software engineering, and model-migration work. TDP and token rate alone can produce the wrong data-center decision.
A Practical Evaluation Framework
Start by segmenting inference traffic. Record model, input and output length, prefill-to-decode ratio, concurrency, KV-cache hit rate, service-level objective, and measured power. Without that baseline, no vendor multiplier can be converted into an internal capacity plan.
Next, separate comparable from non-comparable benchmarks. Direct comparisons require similar models, precision, sequence lengths, batches, speculative-decoding methods, power boundaries, and measurement procedures. When the settings differ, treat the result as directional evidence.
Finally, track software portability. Jalapeño becomes a platform only if new models, custom operators, quantization formats, and debugging workflows can be supported on a predictable schedule. The two-month Codex and GPT-Astra result is a useful early indicator, but it needs repetition across more models and production incidents.
Frequently Asked Questions
Has Jalapeño universally beaten NVIDIA Blackwell?
No. It led the tested commercial systems for three models in the fixed InferenceX 8K/1K benchmark. That does not cover every model, deployment setting, or NVIDIA product.
Is 700W the measured power draw?
Seven hundred watts is the rated package TDP. OpenAI says sustained power stayed at or below 550W for all three tested workloads. A production deployment must still measure server- and data-center-level power.
Did AI design the Jalapeño chip?
Public evidence shows AI assisting model bring-up, kernels, and selected compute-block optimization. It does not support the claim that AI autonomously designed the complete chip.
Author Insight
Jalapeño’s first results move OpenAI’s custom ASIC beyond the “can it run?” stage. The next question is whether it can become a system advantage. Harder evidence will come from AgentX, production yield, supply cadence, and the speed of mapping new models onto fixed silicon.
For infrastructure teams, the message is not to replace the GPU column in a procurement sheet. It is to prepare for model-aware capacity planning across heterogeneous compute, where software portability and complete-task economics determine the winner.
