# Jalapeño (OpenAI chip)

> Source: https://aiwiki.ai/wiki/openai_broadcom_jalapeno
> Updated: 2026-08-27
> Categories: AI Hardware, AI Infrastructure, OpenAI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Jalapeño (OpenAI chip)." aiwiki.ai, 27 Aug 2026. https://aiwiki.ai/wiki/openai_broadcom_jalapeno
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Jalapeño** is a custom [AI accelerator](https://aiwiki.ai/wiki/ai_chip) designed by [OpenAI](https://aiwiki.ai/wiki/openai) with silicon implementation and networking technology from [Broadcom](https://aiwiki.ai/wiki/broadcom). The companies unveiled it on June 24, 2026 as OpenAI's first "Intelligence Processor" and the first chip in a planned multi-generation compute platform. It is designed specifically for [large language model](https://aiwiki.ai/wiki/large_language_model) inference rather than as a general-purpose training accelerator.[1][2]

At launch, Jalapeño was not deployed at scale. Engineering samples were running machine-learning workloads in OpenAI's laboratory at the intended production frequency and power. OpenAI published its first measured results on August 25, 2026, while production qualification and software work were still under way. It continued to target initial deployment within its compute infrastructure by the end of 2026. No independently reproduced benchmark or general commercial release had been announced.[1][5][9]

| Field | Disclosed information |
| --- | --- |
| Official name | Jalapeño |
| Product description | OpenAI's first Intelligence Processor; custom LLM inference accelerator |
| Announcement date | June 24, 2026 |
| Architecture owner | OpenAI |
| Silicon and networking partner | Broadcom |
| Board, rack, and system partner | Celestica |
| Reported foundry | TSMC, according to Reuters |
| Status on August 25, 2026 | Engineering samples in testing; first measured results published; production qualification continuing[9] |
| Named sample workload | GPT-5.3-Codex-Spark |
| Initial deployment target | By the end of 2026 |
| Published package power rating | 700 W[9] |
| Sustained power in published tests | At or below 550 W, according to OpenAI[9] |
| First-result workloads | GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T[9] |
| General commercial availability | None announced |

## Identity and launch status

OpenAI's published name for the chip is simply Jalapeño. "Intelligence Processor" is the company's product category for it, not a disclosed architectural standard. The launch followed the October 2025 announcement of the broader [OpenAI-Broadcom accelerator program](https://aiwiki.ai/wiki/openai_broadcom_accelerators), which covers multiple chip generations, network systems, and a long-term deployment plan.[1][3]

The June announcement showed that the first design had advanced beyond simulation. OpenAI said engineering samples were running at target frequency and target power, including a GPT-5.3-Codex-Spark workload. Target settings do not establish production readiness by themselves: the companies did not disclose the number of samples, manufacturing yield, clock frequency, power draw, duration of testing, or reliability results.[1][2]

OpenAI and Broadcom say Jalapeño moved from initial design to manufacturing tape-out in nine months. OpenAI models assisted parts of the design and optimization process. The companies described this as possibly the fastest development cycle for a high-performance application-specific chip, but did not publish a comparison set or independent evidence for that superlative. Tape-out means the design package was completed for manufacture; it is not the same milestone as volume deployment.[1][2]

In its August results report, OpenAI said its models helped explore implementations, shorten design, measurement, and verification loops, and optimize arithmetic circuits. It also said engineers using Codex with GPT-Astra brought the three open-weight benchmark models to high performance on Jalapeño within two months, while noting that each model family still required new kernels and model-specific optimization.[9]

## Inference workload

Jalapeño is a blank-slate design for LLM inference. Inference is the phase in which a trained model processes a request and produces an output. OpenAI says it shaped the hardware around the kernels, memory movement, networking, scheduling, and serving patterns used by its products. Named workloads include [ChatGPT](https://aiwiki.ai/wiki/chatgpt), [OpenAI Codex](https://aiwiki.ai/wiki/openai_codex), the OpenAI API, and future agentic products.[1]

The design goal is to combine the power and throughput of leading AI accelerators with latency closer to specialized inference systems. OpenAI also says the architecture reduces data movement and balances compute, memory, and networking resources so that useful performance can remain closer to theoretical peak. These statements described the intended system behavior at launch. The August 2026 report added fixed-sequence benchmark measurements, discussed below, but not production traces or independently reproduced results.[1][8][9][11]

Although OpenAI designed the chip around its own operating experience, it says Jalapeño is flexible enough to run current and future LLMs across the industry. That compatibility claim does not mean the chip is a merchant product. Reuters reported that Jalapeño chips and their associated server systems would be used only by OpenAI.[5]

## Design and production roles

OpenAI owns the accelerator architecture. It says the design reflects its model roadmap, inference kernels, serving software, and product needs. Broadcom contributed silicon implementation and the networking and connectivity technologies required to turn that architecture into a production system. Celestica supplies board, rack, and system expertise, and Reuters reported that Celestica would build the server systems.[1][5]

Broadcom specifically names Tomahawk networking silicon as part of the surrounding platform. The launch material does not identify a Tomahawk generation, network topology, link rate, or per-rack configuration. The earlier program announcement says the planned racks use Broadcom Ethernet for both scale-up and scale-out, together with Broadcom PCIe and optical connectivity. That program-level description does not provide the detailed fabric specification for the first Jalapeño system.[1][3]

Reuters reported that OpenAI sent the completed design to TSMC for manufacturing. Neither OpenAI nor Broadcom named the semiconductor process node in the product announcement. Broadcom's Form 10-Q for the quarter ended May 3, 2026 says that TSMC produced about 95% of the wafers made by Broadcom's contract manufacturers across the company. This company-wide figure does not identify Jalapeño's node, allocated wafer volume, packaging line, or yield.[5][7]

## First published performance results

On August 25, 2026, OpenAI published its first quantitative Jalapeño results from tests using SemiAnalysis's InferenceX benchmark. Across GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, OpenAI reported 1.5-1.9 times more work per watt at peak throughput and 1.7-3.6 times lower end-to-end latency than the listed comparison systems.[9][10]

All three published cases used nominal 8k input/1k output single-turn workloads in single-token prediction (STP) mode. The following values are OpenAI's reported results, not independently reproduced measurements.[9][11]

| Workload | Comparator | Peak mixed TPS/kW, Jalapeño vs. comparator | Reported work-per-watt ratio | End-to-end latency, Jalapeño vs. comparator | Reported latency ratio |
| --- | --- | --- | --- | --- | --- |
| GPT-OSS 120B | GB200 | 85,448 vs. 44,960 | About 1.9x | 1.03 s vs. 1.80 s | About 1.7x lower |
| DeepSeek R1 670B MXFP4 | GB300 | 19,641 vs. 11,781 | About 1.7x | 1.65 s vs. 5.99 s | About 3.6x lower |
| Kimi K2.5 1T MXFP4 | GB300 | 18,195 vs. 11,862 | About 1.5x | 1.56 s vs. 5.31 s | About 3.4x lower |

For the peak-work-per-watt calculation, OpenAI normalized throughput with published package thermal design power (TDP), not sustained measured power: 700 W for Jalapeño, 1,200 W for GB200, and 1,400 W for GB300. OpenAI separately said Jalapeño's measured sustained power stayed at or below 550 W on these workloads. The 550 W figure was not the denominator for the headline ratios.[9][12]

OpenAI's appendix labels the three headline comparisons as STP, and Tom's Hardware reported that the principal GB200 and GB300 comparisons used STP on both sides. SemiAnalysis separately compared Jalapeño STP with other systems' best MTP configurations, but those cross-mode comparisons are not the source of the headline ranges above.[9][11][12]

SemiAnalysis said it verified the InferenceX runs in person with OpenAI engineers, but OpenAI provided the numbers and SemiAnalysis did not run the full suite or see AgentX results. The published cases were fixed 8k/1k, single-turn tests rather than long-context, multi-turn production workloads. They establish performance only for the disclosed configurations and workloads, not a universal production advantage.[11]

## Technical specifications not disclosed

At the June 24 launch, OpenAI and Broadcom did not publish a conventional datasheet. The launch materials left the following undisclosed:

- Process node and transistor count.
- Die area and chiplet organization.
- Package technology.
- Memory generation, capacity, stack count, and bandwidth.
- Supported numerical formats and sparsity modes.
- Peak compute throughput.
- Clock frequency and a per-workload power breakdown.
- Chip-to-chip and network link rates.
- Accelerators per board or rack.
- Sample count, yield, production volume, and unit cost.

The August results supplied the 700 W package rating and the sustained-at-or-below-550 W observation for the tested workloads, but not a conventional datasheet.[9]

Launch photographs prompted third-party estimates of package details. Tom's Hardware explicitly treated its visual analysis as speculative and said the photographs could not establish the internal compute datapath. Other outlets produced different memory-stack counts from the same images. These estimates are not disclosed Jalapeño specifications and should not be used as such.[8]

Reuters reported that Broadcom buys high-bandwidth memory from SK Hynix and Samsung for its custom-chip business. The report did not assign a particular memory generation, supplier, capacity, or stack count to Jalapeño, so those company-level supply relationships do not establish this chip's memory configuration.[5]

## Relationship to the 10 GW program

OpenAI and Broadcom announced a multi-year plan in October 2025 to deploy 10 gigawatts of OpenAI-designed custom accelerators and associated network systems. The plan calls for racks to begin deployment in the second half of 2026 and for the program to reach its target by the end of 2029. The systems are intended for OpenAI facilities and partner data centers and use Broadcom Ethernet and other connectivity products.[3][4]

Jalapeño is the first accelerator in that multi-generation platform. The 10 GW figure applies to the full program, not necessarily to first-generation Jalapeño alone. Later accelerator generations can account for an unspecified part of the capacity. The companies did not disclose the number of chips represented by 10 GW, and electrical capacity cannot be converted into a device count without rack power and configuration data.[1][3]

The June 2026 launch set a narrower target of initial Jalapeño deployment by the end of 2026, followed by expansion in later years. Broadcom said gigawatt-scale data centers with [Microsoft](https://aiwiki.ai/wiki/microsoft) and other partners would begin deployment in 2026. That wording identifies data center partners, not buyers of generally available Jalapeño hardware.[1][2]

## Deployment and commercialization

Axios reported that the first chips were expected to enter commercial use by the end of 2026, with the main volume arriving in 2027. At the announcement, however, only laboratory engineering samples were confirmed. On August 25, OpenAI reiterated that it planned to begin deploying Jalapeño within its own compute infrastructure by the end of 2026. It said production qualification, software maturation, preparation for scaled operation, and testing on additional models were still in progress. The late-2026 date remained a target rather than a completed deployment.[1][6][9]

Reuters said the chips and Celestica-built server systems would be used only by OpenAI. No price, board, server product, licensing program, developer kit, or public-cloud instance was announced. There was also no announcement that Microsoft or another hosting partner would offer the accelerator to its own customers. Jalapeño is therefore best described as internal OpenAI infrastructure that can be installed in partner data centers, not as a generally purchasable Broadcom accelerator.[5]

The financial terms of the OpenAI-Broadcom program were not disclosed. The October 2025 announcement referred to existing co-development and supply agreements and a term sheet for rack deployment, but it did not publish purchase commitments, per-chip prices, or financing terms.[3][4]

## Limitations and rollout risks

Broadcom's investor release classifies its large-scale production and deployment statements as forward-looking. It identifies risks involving demand timing, contract manufacturing, limited suppliers, manufacturing capacity and quality, regulation, trade restrictions, and customer financing. Those cautions apply to the rollout plan; they are not evidence that Jalapeño has suffered a particular defect or delay.[2]

Broadcom's May 2026 Form 10-Q likewise describes company-wide dependence on a small number of contract manufacturers and materials suppliers. It warns that complex semiconductor products can face delays, defects, yield problems, and capacity constraints. The filing does not name Jalapeño or disclose its contract, node, order volume, or production economics.[7]

The distinction between a running benchmark sample and a proven production platform remains important. The August report supplied quantitative results for three public models on fixed 8k/1k STP workloads, but not production-scale reliability, yield, rack deployment, or long-context multi-turn AgentX data. Initial deployment and later gigawatt expansion remain future plans. Broader cost and performance claims will require results from production systems under disclosed workloads and power boundaries.[9][11]

## References

1. OpenAI, "OpenAI and Broadcom Unveil LLM-Optimized Inference Chip," June 24, 2026. https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
2. Broadcom, "OpenAI and Broadcom Unveil LLM-Optimized Intelligence Processor," June 24, 2026. https://www.globenewswire.com/news-release/2026/06/24/3316887/19933/en/openai-and-broadcom-unveil-llm-optimized-intelligence-processor.html
3. OpenAI, "OpenAI and Broadcom Announce Strategic Collaboration to Deploy 10 Gigawatts of OpenAI-Designed AI Accelerators," October 13, 2025. https://openai.com/index/openai-and-broadcom-announce-strategic-collaboration/
4. Broadcom, "OpenAI and Broadcom Announce Strategic Collaboration to Deploy 10 Gigawatts of OpenAI-Designed AI Accelerators," October 13, 2025. https://www.globenewswire.com/news-release/2025/10/13/3165560/0/en/OpenAI-and-Broadcom-announce-strategic-collaboration-to-deploy-10-gigawatts-of-OpenAI-designed-AI-accelerators.html
5. Max A. Cherney, "OpenAI Unveils Custom Chip It Designed with Broadcom to Boost Its AI Infrastructure," Reuters, June 24, 2026. https://m.investing.com/news/stock-market-news/openai-unveils-custom-chip-it-designed-with-broadcom-to-boost-its-ai-infrastructure-4758233?ampMode=1
6. Ina Fried, "OpenAI Moves Beyond Nvidia," Axios, June 24, 2026. https://www.axios.com/2026/06/24/openai-jalapeno-ai-chip-broadcom-nvidia
7. Broadcom, Form 10-Q for the Quarter Ended May 3, 2026, filed June 9, 2026. https://www.sec.gov/Archives/edgar/data/1730168/000173016826000054/avgo-20260503.htm
8. Anton Shilov, "Broadcom and OpenAI Unveil Custom-Built Jalapeño Inference Processor," Tom's Hardware, June 24, 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/broadcom-and-openai-unveil-custom-built-jalapeno-inference-processor-openais-first-chip-is-a-massive-reticle-sized-asic-built-in-an-ultra-fast-nine-month-development-cycle
9. OpenAI, "Jalapeño's First Results Show Industry-Leading Speed and Efficiency in AI Inference," August 25, 2026. https://openai.com/index/jalapeno-first-results/
10. Sarah Friar, "The Full Stack Behind Abundant Intelligence," OpenAI, August 25, 2026. https://openai.com/index/the-full-stack-behind-abundant-intelligence/
11. Bryan Shan, Myron Xie, Jordan Nanos, et al., "OpenAI Jalapeño: Better Than Nvidia Blackwell," SemiAnalysis, August 25, 2026. https://newsletter.semianalysis.com/p/openai-jalapeno-better-than-nvidia
12. Luke James, "OpenAI's 700W Jalapeño ASIC Outpaces 1,400W Nvidia Flagship GPU," Tom's Hardware, August 25, 2026. https://www.tomshardware.com/tech-industry/semiconductors/openai-says-its-jalapeno-chip-beats-nvidias-gb300-in-first-published-benchmarks