OpenAI's Jalapeño Chip: The Real Benchmarks

• Jalapeño is real and confirmed - OpenAI's first custom AI chip, built with Broadcom, unveiled 24 June 2026, with production benchmarks published 25-26 August 2026.
• It is an inference chip, not a training chip. It serves responses from already-trained models like GPT-5.6 Sol; it does nothing for the far larger training-compute side of OpenAI's business.
• The headline numbers are real but narrow: 1.5-1.9x better performance-per-watt and 1.7-3.6x lower latency than Nvidia's GB200/GB300, on exactly three tested models (GPT-OSS 120B, DeepSeek R1 670B, Kimi K2.5 1T).
• SemiAnalysis, who ran the independent benchmark, flagged real caveats: only one workload profile was tested, the decoding technique used may have been unfair to Nvidia's chips, and Jalapeño arguably should be compared against Nvidia's next-generation Vera Rubin rather than the current Blackwell generation.
• Deployment is small and slow at first - "very small volumes" by the end of 2026, more meaningful scale in 2027, as the first generation of a planned multigenerational chip platform.
• Nvidia is not losing OpenAI as a customer. OpenAI has said explicitly it will keep buying Nvidia chips for training and inference alike, on top of roughly $119 billion in existing supply commitments.
• "PHASEONE" is not a real OpenAI product name - it doesn't appear in any primary source, and looks like YouTube shorthand for "phase one of OpenAI's chip strategy" rather than an actual codename.

Jalapeño is OpenAI's first custom-designed AI accelerator - an application-specific chip (ASIC) built specifically to run inference for large language models, rather than a general-purpose GPU adapted for the job. OpenAI led the architectural design, informed by the systems it runs every day across ChatGPT, Codex, the API and its agentic products, while Broadcom - which has custom-silicon experience through its work on Google's TPUs - handled silicon implementation, networking and manufacturing. Broadcom CEO Hock Tan and President Charlie Kawwas presented an early prototype wafer to OpenAI CEO Sam Altman and President Greg Brockman, a moment captured in the photo used as this article's hero image, with the wafer plaque reading "Jalapeño Intelligence Processor."

The chip is deliberately narrow in purpose. It targets the "decode" phase of autoregressive language-model inference - the token-by-token generation step that dominates the cost of serving billions of daily API requests - and is architected to minimise data movement and communication delays during that phase, particularly around where a model's state and KV cache are physically placed relative to the compute. That is a very different engineering problem to training a frontier model from scratch, which is why Jalapeño doesn't touch OpenAI's training compute needs at all.

The underlying trade-off is a familiar one in chip design: a general-purpose GPU like Nvidia's Blackwell line has to be good at a huge range of workloads, from training to inference to graphics to scientific computing, which means it carries hardware and design complexity that any single workload doesn't fully use. An application-specific chip like Jalapeño gives that flexibility up entirely in exchange for being tuned to exactly one job - serving OpenAI's own models, at OpenAI's own scale, with OpenAI's own request patterns baked into the design from day one. That specialisation is precisely why a first-generation ASIC from a company with no prior chip-design history can plausibly beat a mature, general-purpose product on a narrow efficiency metric, while still losing badly on flexibility, software ecosystem maturity and the sheer breadth of workloads Nvidia's hardware and CUDA software stack already support.

OpenAI and Broadcom first unveiled Jalapeño publicly on 24 June 2026, describing a design process that began roughly in mid-2024 and moved from initial hiring to a finished chip design remarkably quickly for the industry - reported design-to-tape-out timelines range from about nine months for the core chip design cycle once architecture was locked, to around sixteen months counting the full path from early team-building. Either figure is fast by chip-industry standards, where a new accelerator typically takes several years from concept to silicon.

At the June reveal, OpenAI made only a general claim - that early testing showed "performance per watt substantially better than current state-of-the-art" - without publishing specific figures. That changed on 25-26 August 2026, when OpenAI released detailed benchmark results run on SemiAnalysis's independent InferenceX platform, the first time actual numbers, rather than a general claim, were put on the record.

Here is what OpenAI actually published, verified by SemiAnalysis testing in person at OpenAI's facilities. Jalapeño is rated at 700 watts, though its measured sustained power stayed at or below 550 watts across the tested workloads - a meaningful gap between the rated and the real figure that OpenAI used to argue the efficiency comparison is, if anything, conservative. The comparison points were Nvidia's GB200 (rated 1,200W) and GB300 (rated 1,400W) systems, tested across three published models: GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T.

SemiAnalysis's conclusion that Jalapeño sits on the "Pareto frontier" is worth unpacking rather than just quoting. In a chart plotting performance-per-watt against latency, a chip is on the Pareto frontier if no other chip tested beats it on both measures at once - it might lose to a rival on efficiency alone, or on latency alone, but nothing tested was simultaneously better on both. That is a genuinely meaningful claim precisely because it's narrower than "Jalapeño is the best chip": it says there was no tested alternative that was a strict, no-trade-off upgrade over it, not that it dominates every possible metric.

SemiAnalysis's own testing on specific models put real throughput figures on the board too: on DeepSeek R1 at a concurrency of one user, Jalapeño delivered over 700 tokens per second per user; on Kimi K2.5 and GPT-OSS, it reached roughly 1,400 tokens per second per user; and on Kimi K2.5 specifically, SemiAnalysis measured it beating the next best chip tested by more than 9 times on that metric. Richard Ho, OpenAI's head of hardware, summarised the result plainly: "The bottom line is that the results show a very, very significant performance advance over state of the art."

This is the section most of the "beats Nvidia" YouTube coverage skipped, and it matters. SemiAnalysis, the firm that actually ran the independent benchmark, attached several specific caveats to its own numbers rather than simply endorsing OpenAI's framing.

• Decoding technique mismatch. Jalapeño's results used single-token prediction (STP) without speculative decoding, while the competing chips it was measured against used multi-token prediction (MTP) - a more advanced decoding technique. SemiAnalysis called this comparison "somewhat unfair" on its face, since MTP-equipped competitors are working under a harder self-imposed constraint that STP doesn't have to match.
• Narrow benchmark scope. Testing covered only InferenceX's "8k1k" workload profile (roughly 8,000 input tokens to 1,000 output tokens) - not SemiAnalysis's preferred AgentX suite, which is built to stress real, messier production agent workflows rather than a clean, fixed-shape request.
• OpenAI-supplied data, independently checked but not independently generated. SemiAnalysis says "all numbers are provided to us by OpenAI" and were verified in person in OpenAI's own lab, rather than run end-to-end by SemiAnalysis on hardware it fully controlled - a normal constraint for this kind of vendor-hosted benchmark, but a meaningfully different standard of independence than a fully third-party test.
• The generational comparison may be unfair the other way too. SemiAnalysis suggests the more appropriate Nvidia comparison point is the unreleased next-generation Vera Rubin platform, which shares Jalapeño's HBM4 memory generation, rather than the already-shipping Blackwell-generation GB200/GB300 used in OpenAI's published comparison.

None of this means the headline numbers are fake - SemiAnalysis's conclusion was still that Jalapeño sits on the efficiency-latency "Pareto frontier," meaning nothing tested beat it on both metrics simultaneously. It means the fair reading is "a genuinely strong first-generation chip, tested under conditions that flatter it somewhat," not "Nvidia has been definitively beaten."

No primary source - not OpenAI's own announcements, not Broadcom's investor releases, not any of the mainstream outlets that covered the Jalapeño benchmark reveal (Bloomberg, CNBC, TechCrunch, Reuters) - uses the term "PHASEONE" to describe this chip or any related OpenAI product. The confirmed, repeatedly verified name across every primary source is Jalapeño.

The most likely explanation is that "PHASEONE" is YouTube-native shorthand rather than a genuine leak: several creators covering the story this week framed the benchmark reveal as the first concrete proof-point in OpenAI's multi-year, multi-generation custom-silicon plan - which OpenAI itself has described as a "multigenerational platform." Compressing "the first real phase of OpenAI's chip strategy" into a punchy, single-word video title ("PHASEONE") is a reasonable summary of the substance, but it reads, and is being searched for, as if it were an official codename. It isn't one. This is a pattern worth recognising generally on fast-moving AI YouTube: a real, sourced story (the Jalapeño benchmarks) wrapped in an invented-sounding label that didn't come from the company at all.

Two facts sit uncomfortably next to the "beats Nvidia" headlines. First, OpenAI has stated plainly that it "will continue to widely deploy accelerators from NVIDIA and other partners for both training and inference workloads" - not the language of a customer preparing to walk away. Second, Nvidia holds roughly $119 billion in supply-related commitments tied to OpenAI, on top of the 10-gigawatt Nvidia systems partnership the two companies announced separately this year. Nvidia's Data Center revenue reached $75.25 billion in a single quarter (Q1 FY2027), up 92% year-on-year - a growth rate that a single competing chip, tested on three models, does not meaningfully dent.

The structural reason is the inference-only scope covered above: training frontier models remains Nvidia's deepest moat, and Jalapeño doesn't compete there at all. Even within inference, "very small volumes" by the end of 2026 means Jalapeño's near-term impact on OpenAI's overall Nvidia GPU spend is close to zero. What Jalapeño actually threatens is Nvidia's margin on the inference slice of the business over a multi-year horizon, by giving OpenAI a credible in-house alternative and, in turn, negotiating leverage - not Nvidia's revenue this year or next.

OpenAI's own stated timeline is deliberately modest. Richard Ho has said Jalapeño enters production in "very small volumes" by the end of 2026, with meaningfully larger deployment expected across 2027. Both companies have described Jalapeño as the first generation of a multigenerational platform, not a one-off chip - meaning the August 2026 benchmarks are best read as a proof of concept for the architecture and the OpenAI-Broadcom partnership model, ahead of a longer roadmap reportedly targeting around 10 gigawatts of custom silicon capacity by 2029.

OpenAI is late to this specific idea, not early. Google has run its own TPUs for years, Amazon has Trainium and Inferentia, and Meta has its own MTIA chips - all built on the same logic that a company serving its own models at sufficient scale can eventually justify the cost of custom silicon tuned to its exact workloads instead of paying Nvidia's margin on general-purpose GPUs. What is genuinely notable about Jalapeño is the speed: a roughly nine-month-to-sixteen-month design cycle for a first-generation chip is unusually fast, and Broadcom's established role as the custom-ASIC partner behind several of those other companies' chips (including Google's TPUs) is a large part of why OpenAI could move that quickly rather than building a silicon team entirely from scratch.

It also sits alongside a broader theme in this week's AI news of the industry's attention shifting from bigger models to the physical and economic infrastructure underneath them - see our coverage of Anthropic's Model Hardware Standard, published the same week, which is about the physical world rather than chips, but reflects the same underlying shift: 2026's most consequential AI stories are increasingly about plumbing, not benchmarks.

If you're an OpenAI API customer or building on GPT-5.6 Sol at scale, Jalapeño matters over a two-to-three-year horizon: if the efficiency gains hold up in production, OpenAI's inference costs eventually come down, and that has historically shown up in API pricing over time. If you invest in or follow Nvidia, the more useful number here isn't the 1.5-1.9x efficiency claim - it's the reported $119 billion Nvidia supply commitment and the 92% year-on-year Data Center revenue growth, both of which say Nvidia's position is intact for now. If you're simply a ChatGPT or Codex user, nothing changes about the product in front of you today; Jalapeño doesn't reach production volume until deep into 2027 at the earliest.

Jalapeño is a real, independently-tested chip with genuinely strong first-generation numbers - 1.5-1.9x better performance per watt and up to 3.6x lower latency than Nvidia's current flagships is a legitimate result, not vapourware. But it is an inference-only chip shipping in token volumes through 2026, tested under conditions SemiAnalysis itself flagged as somewhat favourable, and arriving alongside an explicit OpenAI commitment to keep buying heavily from Nvidia. "OpenAI built a genuinely competitive first chip and Nvidia's position is still overwhelmingly secure" are both true at the same time - and neither one is called "PHASEONE."