OpenAI has published its first detailed performance results for Jalapeño, the custom inference chip it built with Broadcom, and the numbers are notable: 1.5 to 1.9 times more throughput per watt and up to 3.6 times lower latency than comparison Nvidia systems, according to results presented at the Hot Chips conference on August 25, 2026.
That’s a meaningful step up from where things stood in June, when OpenAI first unveiled Jalapeño with confident language but no hard data to back it up. This time, the company brought numbers — and, notably, had them checked by an outside benchmarking firm rather than presenting them on its own authority alone.
Here’s what OpenAI actually showed, what’s been independently confirmed, and what’s still an open question.
Table of Contents
What OpenAI Announced
OpenAI’s blog post, published the same day as its Hot Chips presentation, describes Jalapeño delivering “a significant performance advance”: more AI work per unit of power, combined with faster response times, in a single architecture. Historically, chip designers have had to trade one of those for the other. OpenAI says Jalapeño doesn’t.
Across three models — GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T — OpenAI reported 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, compared with the systems it tested against. For highly interactive workloads, like the kind of back-and-forth tasks AI agents perform, the reported advantage climbed to 2.1 to 4.1 times.
“The bottom line is that the results show a very, very significant performance advance over state of the art,” Richard Ho, OpenAI’s head of hardware, said on a press call, according to TechCrunch. “Jalapeño can serve more AI work per unit of power, while also returning responses more quickly.”
These are OpenAI’s own reported figures. They matter, but so does understanding exactly how they were produced.
How the Chip Was Tested
OpenAI ran Jalapeño on InferenceX, a public benchmark suite built and maintained by SemiAnalysis, an independent semiconductor research firm. InferenceX is designed to measure the full process of serving an AI request — not just raw chip throughput, but the combination of latency, power draw, and useful output that determines what a system can actually deliver to users.
That choice matters for credibility. Rather than inventing its own metric, OpenAI used a benchmark that others can, in principle, run themselves.
SemiAnalysis says it verified the benchmark runs in person, in OpenAI’s lab. In its own analysis, the firm wrote that “in general, first-generation chips are not competitive, but OpenAI bucks the trend by being industry-leading… beating every Nvidia, AMD, and Google chip we have been able to test on multiple top open source models.”
That’s a real endorsement — but SemiAnalysis was careful to attach conditions to it. The firm has not yet run its own full InferenceX suite on Jalapeño independently; the numbers so far came from OpenAI, with SemiAnalysis observing rather than executing the tests itself. And Jalapeño hasn’t been tested on AgentX, SemiAnalysis’s newer and more demanding benchmark built around long-context, multi-turn agent workloads that can expose weaknesses simpler tests miss.
Jalapeño vs. Nvidia: What the Numbers Actually Compare
It’s worth being precise about what “beats Nvidia” means here, because the comparison is narrower than the headlines suggest.
OpenAI measured Jalapeño against Nvidia’s GB200 and GB300 systems — commercial Blackwell-generation hardware rated at 1,200 watts and 1,400 watts, respectively. Jalapeño itself is rated at 700 watts, though OpenAI says its measured sustained power stayed at or below 550 watts across the tested workloads.
SemiAnalysis flagged an important caveat: Blackwell is no longer Nvidia’s newest platform. The more relevant rival is Nvidia’s Rubin, which uses next-generation HBM4 memory and is closer to Jalapeño in terms of when it will actually ship at scale. “Vera Rubin systems are starting to ship to customers right now, while it will still be some time before OpenAI has anything beyond engineering samples of Jalapeño,” the firm noted.
When SemiAnalysis compared Jalapeño against Rubin’s own published July 2026 results, Jalapeño still came out ahead on throughput per megawatt. But once total cost of ownership is factored in, the two land close to even on cost per token — a much less dramatic gap than the Blackwell comparison implies.
In short: Jalapeño’s advantage over Nvidia’s currently-shipping commercial hardware is real and independently corroborated in broad strokes. Its advantage over Nvidia’s next-generation hardware is closer to a toss-up.
Jalapeño vs. Nvidia Blackwell (GB200/GB300) — OpenAI’s reported figures
| Metric | Jalapeño advantage | Comparison system |
|---|---|---|
| Throughput per watt (peak) | 1.5–1.9x | Nvidia GB200 / GB300 |
| End-to-end latency | 1.7–3.6x lower | Nvidia GB200 / GB300 |
| Interactive/agentic workloads | 2.1–4.1x | Nvidia GB200 / GB300 |
| Package power rating | 700W (measured ≤550W sustained) | GB200: 1,200W · GB300: 1,400W |
Source: OpenAI, tested on SemiAnalysis’s InferenceX benchmark across GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.
The Technology Behind Jalapeño
Running a large language model isn’t one uniform task. Processing an incoming prompt — a phase called prefill — is compute-intensive. Generating the response, token by token, relies much more on memory bandwidth, a phase called decode. Data also has to move constantly between cores and chips, and that movement can leave parts of a system idle while they wait.
OpenAI says it designed Jalapeño around this reality from the start, rather than optimizing for one phase at the expense of the other. Model state, including the “KV cache” used while generating a response, can be kept local, while the system shifts the right mix of compute, memory, and networking to whichever phase — prefill or decode — needs it at that moment. The chip’s network is built as one large, connected domain, so an entire workload can stay within a single system instead of shuttling data across boundaries.
OpenAI also says its own AI models played a direct role in building the chip. Earlier-generation models helped with initial design and bring-up; more recent models, including Codex, were used to optimize the chip’s arithmetic circuits and to help engineers add support for models that weren’t part of Jalapeño’s original production plan. In one case, AI-generated implementations of certain GPT-OSS attention and mixture-of-experts components ran 1.5 to 1.8 times faster than versions written by human engineers — a notable result, though it applies only to those specific components, not the model as a whole.
Who Built It
Jalapeño wasn’t built by OpenAI alone. Broadcom handled the chip’s silicon engineering and implementation, along with networking hardware built on its Tomahawk Ultra Ethernet technology. According to reporting from outlets including Tom’s Hardware, the chip is fabricated by TSMC on a process described as 3-nanometer-class — the same broad generation used in Apple’s M4 chips and Nvidia’s Blackwell GPUs. That detail comes from industry reporting rather than an official disclosure; neither OpenAI nor Broadcom has confirmed the exact process node in their own materials. Celestica, a systems integrator, is responsible for turning the finished chips into the boards and racks that go into a data center.
OpenAI has described the nine-month span from initial design to manufacturing tapeout as one of the fastest development cycles ever achieved for a chip of this complexity.
Jalapeño’s Development Partners
| Partner | Role |
|---|---|
| Broadcom | Silicon engineering and implementation; networking hardware (Tomahawk Ultra Ethernet) |
| TSMC | Fabrication (reported 3-nanometer-class process; not officially confirmed by OpenAI or Broadcom) |
| Celestica | Systems integration — boards, racks, and data center-ready hardware |
Deployment Timeline and Business Context
Jalapeño Timeline
- June 24, 2026 — Jalapeño publicly unveiled by OpenAI and Broadcom; engineering samples delivered
- August 25, 2026 — First detailed benchmark results published at Hot Chips
- End of 2026 — Small-volume deployment planned within OpenAI’s own infrastructure
- 2027 — Larger-scale deployment ramp planned
Jalapeño remains in the production-qualification stage. OpenAI says it plans to begin deploying the chip within its own compute infrastructure in small volumes by the end of 2026, with a more substantial rollout following through 2027. A second-generation chip is reportedly already deep in development, and early concept work has begun on a third generation.
Crucially, OpenAI isn’t framing this as a break from Nvidia. The company has said explicitly that it will continue to widely deploy Nvidia GPUs, alongside other accelerators, for both training and inference workloads. Jalapeño is additional capacity, not a replacement — at least for now.
The chip fits into a broader strategy OpenAI has described as “full-stack”: designing models, products, serving software, chips, memory, networking, and systems together, so that lessons from real-world workloads can improve every layer at once. It’s a strategy other major AI infrastructure players have already pursued in their own ways — Google with its TPUs, Amazon with Trainium and Inferentia, Microsoft with Maia, and Meta with its MTIA chips. OpenAI is a later entrant to in-house AI silicon, not the first.
The business logic, as OpenAI frames it, is straightforward: getting more useful work out of the same power and hardware can help the company serve growing demand and lower the cost of each response, potentially letting revenue and usage grow faster than the cost of serving them.
Limitations and Open Questions
Several caveats are worth keeping in view.
No lab outside SemiAnalysis’s supervised visit has tested Jalapeño. The chip hasn’t been evaluated on AgentX, the more demanding long-context benchmark SemiAnalysis considers a better proxy for real production agent workloads. And the industry-wide supply of high-bandwidth memory and advanced chip packaging — the kind Jalapeño and its rivals depend on — remains extremely tight; Broadcom’s own CEO has described customer demand as more than the company can currently address.
There are also structural questions common to any application-specific chip: ASICs like Jalapeño are generally less flexible than general-purpose GPUs, and as language model architectures continue to evolve, hardware tuned for today’s workloads may need redesigning for tomorrow’s. OpenAI and SemiAnalysis both push back on the idea that Jalapeño is narrowly built around OpenAI’s own models — SemiAnalysis notes engineers even got the chip running the game Doom using code generated through Codex, as a demonstration of the underlying software’s flexibility — but the tradeoff between specialization and adaptability is one that will only really be tested once the chip is deployed at scale.
Finally, some basic technical details remain undisclosed. OpenAI hasn’t published Jalapeño’s exact memory configuration, and the reported TSMC process node comes from outside reporting rather than an official statement.
What This Means Now
Jalapeño’s first quantified results are a genuine step forward for OpenAI’s hardware ambitions — the company has moved from unverified marketing language to numbers checked, at least in part, by an independent benchmarking firm. That’s more substantive than most first-generation chip announcements get.
But the picture is also more qualified than the “beats Nvidia” framing suggests. The direct comparison was against Nvidia’s outgoing Blackwell generation, not its newer Rubin platform, where the gap narrows considerably on a cost basis. Deployment is still small-scale and months away. And full independent verification — beyond a single supervised lab visit — hasn’t happened yet.
What to watch next: whether SemiAnalysis or another outside party gets to run its own full benchmark suite on Jalapeño, how the chip performs against Nvidia’s Rubin once both are shipping at comparable scale, and whether OpenAI’s end-of-2026 deployment target holds given the tight global supply of memory and advanced packaging capacity that the whole industry is competing for.
Frequently Asked Questions
What is OpenAI’s Jalapeño chip? Jalapeño is OpenAI’s first custom-designed chip, built specifically to run trained AI models (inference) rather than to train them. It was developed with Broadcom and unveiled in June 2026, with detailed performance results following in August 2026.
How does Jalapeño compare to Nvidia’s chips? In OpenAI’s published tests, Jalapeño delivered 1.5 to 1.9 times more throughput per watt and up to 3.6 times lower latency than Nvidia’s GB200 and GB300 Blackwell systems. Against Nvidia’s newer Rubin platform, an independent analysis found Jalapeño still ahead on throughput per megawatt but close to even on cost per token.
Who manufactures the Jalapeño chip? Broadcom handled the silicon design and engineering; the chip is reportedly fabricated by TSMC on a 3-nanometer-class process, though that detail comes from industry reporting rather than an official OpenAI or Broadcom disclosure. Celestica manages the board, rack, and system integration.
When will Jalapeño be deployed? OpenAI plans to begin deploying Jalapeño in small volumes within its own infrastructure by the end of 2026, with a larger rollout through 2027. As of the August 2026 announcement, it remains in production qualification rather than commercial deployment.
Is Jalapeño replacing Nvidia GPUs at OpenAI? No. OpenAI has said it will continue to widely deploy Nvidia GPUs, alongside Jalapeño, for both training and inference workloads.
Were Jalapeño’s benchmark results independently verified? Partially. SemiAnalysis, an independent semiconductor research firm, verified the benchmark runs in person in OpenAI’s lab and offered its own analysis, but it has not yet run its full benchmark suite on Jalapeño independently, nor tested it on its more demanding AgentX benchmark.
What is InferenceX and why does it matter for these results? InferenceX is a public benchmark suite built by SemiAnalysis to measure the full process of serving an AI request — combining latency, throughput, and power efficiency rather than a single narrow metric. OpenAI’s use of a third-party, publicly available benchmark, rather than an internal metric, is part of why these results carry more weight than typical vendor claims.