OpenAI's first Jalapeno numbers buy it leverage, not a procurement input
The 1.7x to 3.6x latency range is set by the baseline systems, not the chip, and the report's own publication date is unsettled. Read it as direction, not evidence.
The Engineer · Build desk
What happened
- OpenAI published its first measured results from working Jalapeno silicon on Tuesday.
- It reported end-to-end latency 1.7 to 3.6 times lower than the Nvidia GB200 and GB300 systems it compared against, plus 1.5 to 1.9 times more work per watt at peak throughput.
- OpenAI ran the tests itself and no one has reproduced them in full; SemiAnalysis says it watched the InferenceX runs without running the whole suite.
Why it matters
- What it limits constraint Because the range endpoints are driven by how the baseline systems performed rather than by Jalapeno's own spread, no single ratio in the set maps onto a buyer's own model mix or serving pattern.
- Who's exposed exposure A page whose publication date is unsettled and unversioned is a weak artefact to cite in a capacity plan, and the cost of that lands on whoever quoted it, not on OpenAI.
- Who has to decide decision Anyone pricing multi-year inference commitments has to decide how much weight to give hardware that will not carry production traffic until the end of 2026.
- What it sets up precedent Publishing on a public harness that only the vendor has executed normalizes treating an open benchmark as an independent one, which makes the next set of first-party numbers easier to wave through.
Next story · Invest · Oct 8
OpenAI's $50 Billion Revenue Figure Reported as Oracle, AMD and TSMC Shares FallConfirmed24 publishers
The range is not a property of the silicon. Work back through the per-model tables: GPT-OSS 120B finished in 1.03 seconds on Jalapeno against 1.80 on a GB200 [23], which is 1.75x [31]. DeepSeek R1 went from 5.99 seconds to 1.65 against a GB300 [15], or 3.63x [26]. Kimi K2.5 went 5.31 to 1.56 [16], 3.40x [27]. Jalapeno's three absolute latencies sit within 1.6x of each other while the comparison systems' span 3.3x [28]. So the top of the headline range is mostly a statement about how a GB300 behaved in OpenAI's DeepSeek R1 configuration at that operating point, and the configuration is precisely what no outside party has run [9].
One methodological choice runs against OpenAI's interest. The per-watt results were normalized on published package ratings, 700 watts for Jalapeno against 1,200 and 1,400 for the Nvidia systems [6], even though measured sustained draw stayed at or below 550 watts [7]. That rating sits 27 percent above the number the chip actually pulled [24]. Nothing in the material says what the Nvidia systems drew, so the discount is not necessarily symmetric, but on its own side of the ledger OpenAI gave away efficiency it says it had.
The 104.3x figure is a different kind of number, and worth separating from the rest [5]. It applies to one DeepSeek R1 operating point, measured as the throughput Jalapeno could hold while matching the fastest time between tokens the comparison system ever reached [14]. On the same model at peak throughput, the advantage was 1.7x [14]. That is a 61-fold gap between the biggest slide number and the ordinary one [29]. The narrow figure is not meaningless, because agent workloads run many steps in sequence and per-token delay compounds across a task [20]. It is simply not a fleet-level efficiency claim.
Then there is provenance. The live engineering report carries an August 25th, 2026 date, the supplied page metadata dates the same page to August 3rd, no public revision history establishes which came first, and OpenAI's X announcement is timestamped August 25th [25]. For a marketing page this is trivia. For a document a buyer wants to cite in a capacity plan, a moving date with no changelog means there is no fixed record of what was claimed when.
What the numbers are actually for is visible in who benefits from them existing. A production deployment gives OpenAI direct control over part of its own inference bill and a source of capacity outside its chip suppliers [10], which is the strategy Greg Brockman described in June as making compute more abundant [18]. Broadcom did silicon implementation and networking, Celestica the boards, racks and system integration [19], and the program reached tape-out nine months from initial design [11]. Richard Ho's hardware team now has to carry laboratory results into production traffic [17], with deployment planned for the end of 2026 [8] and scheduling, utilization and reliability conditions that a controlled comparison cannot settle [22].
SemiAnalysis says it watched the InferenceX runs in person but did not run the full suite itself [9]. Witnessed is not reproduced. Until someone else executes the harness on the same stack, the honest use of 1.7x to 3.6x is as a directional read on where OpenAI thinks its cost curve is going.
What to watch
- A third-party InferenceX run on the same software stack and comparison systems, published in full rather than witnessed.
- Any correction or changelog on the results page that fixes whether it first went up on August 3rd or August 25th.
- Nvidia or another accelerator vendor publishing its own InferenceX numbers on the same three models, which would give the ratios a second reference point.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence52
- Adoption14
- Hype gap+38
- Incentives82
- Confidence68
Perspective Coverage
9 publishers- Builder
- Builder 41%
- Operator
- Operator 31%
- Investor
- Investor 28%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
OpenAI published the first measured results for its Jalapeno custom inference chip on Tuesday.
- [2]
The tests covered GPT-OSS 120B, DeepSeek R1 670B and Moonshot AI's Kimi K2.5 1T using InferenceX, SemiAnalysis' open-source inference benchmark, against commercially available Nvidia systems.
- [3]
OpenAI reported end-to-end latency between 1.7 and 3.6 times lower than the Nvidia comparison systems.
- [4]
OpenAI reported 1.5 to 1.9 times more AI work per watt at peak throughput than Nvidia GB200 and GB300 systems across the three models.
- [5]
OpenAI reported 8.6x to 104.3x higher performance per watt at a prior-best time between tokens.
- [6]
OpenAI normalized the comparisons using each accelerator's published package power rating: 700 watts for Jalapeno, 1,200 watts for the tested GB200 and 1,400 watts for the GB300.
- [7]
OpenAI said Jalapeno's measured sustained power stayed at or below 550 watts during the workloads, though it used the higher 700-watt rating in the normalized results.
- [8]
The results are OpenAI's measurements on engineering hardware, with production deployment planned for the end of 2026.
- [9]
The runs were reported by OpenAI and have not been independently reproduced in full; SemiAnalysis said it observed the InferenceX runs in person but did not run the full benchmark suite.
- [10]
A production Jalapeno deployment would give OpenAI direct control over part of its inference bill and another source of capacity beyond outside chip suppliers.
- [11]
OpenAI said the program went from initial design to manufacturing tape-out in nine months.
- [12]
For selected attention and mixture-of-experts blocks in GPT-OSS, OpenAI said AI-generated implementations ran between 1.5 and 1.8 times faster than versions written by human experts.
- [13]
OpenAI reported 2.1 to 4.1 times higher interactivity across the tested models.
- [14]
The 104.3x performance-per-watt figure applies to a specific DeepSeek R1 operating point, measured as the throughput Jalapeno could deliver while matching the fastest time between tokens previously reached by the comparison system; OpenAI's reported peak-throughput advantage for that model was 1.7x.
- [15]
Against a GB300 system running DeepSeek R1, OpenAI said Jalapeno cut end-to-end latency from 5.99 seconds to 1.65 seconds.
- [16]
On Kimi K2.5, OpenAI said Jalapeno reduced end-to-end latency from 5.31 seconds to 1.56 seconds.
- [17]
Richard Ho, who leads OpenAI's hardware program, now has to carry the laboratory results into production traffic.
- [18]
OpenAI co-founder and President Greg Brockman described Jalapeno in June as part of a long-term full-stack infrastructure strategy to make compute more abundant.
- [19]
Broadcom handled silicon implementation and networking technology, while Celestica worked on boards, racks and system integration.
- [20]
OpenAI said agents need to complete many steps in sequence, so delays can compound across an entire task.
- [21]
On GPT-OSS 120B the Jalapeno system produced 85,448 mixed tokens per second per kilowatt versus 44,960 for the GB200 configuration.
- [22]
An independent reproduction of the configurations, software stack and comparison systems is still needed, and production systems introduce scheduling, networking, utilization and reliability conditions that a controlled comparison cannot settle.
- [23]
On GPT-OSS 120B, OpenAI reported end-to-end latency of 1.03 seconds for Jalapeno against 1.80 seconds for the Nvidia GB200 system.
- [24]
Jalapeno's 700-watt published rating is about 27 percent above its measured sustained draw of 550 watts, so normalizing on the rating understates OpenAI's own per-watt result by that factor.
- [25]
The live engineering report on Jalapeno's first results displays an August 25th, 2026 date while the supplied source metadata dates the same page to August 3rd; no public revision history establishes whether August 25th reflects an update or the original publication, and OpenAI's X announcement is timestamped August 25th.
- [26]
The DeepSeek R1 latency ratio against the GB300 is 3.63x.
- [27]
The Kimi K2.5 latency ratio is 3.40x.
- [28]
Jalapeno's three reported absolute latencies span 1.6x while the comparison systems' span 3.3x, so the width of the 1.7x-3.6x range comes mainly from variation in the baseline systems.
- [29]
For DeepSeek R1, the 104.3x figure is about 61 times the 1.7x peak-throughput advantage reported for the same model.
- [30]
The GPT-OSS 120B throughput-per-kilowatt ratio is 1.90x.
- [31]
The GPT-OSS 120B latency ratio is 1.75x.
Sources
9 independent publishers whose own reporting we read for this story.
- archive.thedeepview.comHow OpenAI's Jalapeno chip surprised the industry
1 article · August 26, 2026
- dev.toOpenAI's first custom chip just benchmarked past NVIDIA. Jalapeño changes the inference equation.
1 article · August 26, 2026
- latent.space[AINews] Hot Chips: OpenAI’s Jalapeño, Cerebras CS-5, Groq 3 LPX, Apple M6
1 article · August 26, 2026
- mezha.netOpenAI розробляє чип Jalapeño для швидкого інференсу ШІ
1 article · August 25, 2026
- runtimewire.comOpenAI says Jalapeno beats Nvidia Blackwell on inference speed per watt
2 articles · August 25, 2026
- servethehome.comOpenAI Jalapeno Custom AI ASIC at Hot Chips 2026 - ServeTheHome
1 article · August 25, 2026
- the-decoder.comOpenAI's first custom chip "Jalapeño" reportedly beats Nvidia's Blackwell and Rubin in inference benchmarks
2 articles · August 25, 2026
- thenewstack.ioOpenAI built a chip in nine months. Then it let AI rewrite the code.
2 articles · August 25, 2026
- tomshardware.comOpenAI’s 700W Jalapeño ASIC outpaces 1,400W Nvidia flagship GPU — claims up to 1.9x throughput per kilowatt and 3.6x lower latency, co-developed with Broadcom
2 articles · August 27, 2026