OpenAI put numbers behind its chip ambitions on August 25, 2026, publishing the first independent benchmark results for Jalapeño, the inference accelerator it co-designed with Broadcom. Presented at the Hot Chips conference in Silicon Valley, the figures claim Jalapeño beats Nvidia’s Blackwell-generation GB200 NVL72 and GB300 NVL72 systems on throughput per watt, end-to-end latency, and cost per token. Nvidia still ships the vast majority of AI training and inference silicon on the planet, but the release marks the moment OpenAI’s chip program stopped being a slide deck and started being a benchmark table.
The claims are OpenAI’s own, run on its own hardware, against public open-weight models. That matters for how much weight to put on them. It also doesn’t make them irrelevant: Jalapeño is real silicon, taped out in roughly nine months, and OpenAI says it plans to put it into production inference capacity before the end of 2026. Here’s what the numbers actually say, how they hold up against outside scrutiny, and what it means for Nvidia, Google, Amazon, and Microsoft’s own competing chip bets.
What OpenAI’s Jalapeño Chip Actually Is
Jalapeño is OpenAI’s first custom silicon, developed with Broadcom on the design and integration side, manufactured by TSMC, and assembled into racks by Celestica. OpenAI and Broadcom first unveiled the chip on June 24, 2026, describing it as a reticle-sized ASIC built specifically for inference, not training. That distinction matters. Training chips need to move enormous batches of data across thousands of GPUs at once. Inference chips need to serve one user’s prompt back to them as fast as possible, over and over, at a cost that doesn’t bankrupt the product.
Tom’s Hardware, which got early access to technical details, described Jalapeño as carrying six HBM memory modules on a single package rated at roughly 700 watts, well below the 1,200 watts Nvidia specs for a GB200 package or the 1,400 watts of a GB300. OpenAI’s argument, laid out across its August 25 benchmark release, is that Jalapeño does more useful work per watt of that lower power draw, not just less work at lower power.
The Benchmark Numbers OpenAI Is Publishing
OpenAI ran Jalapeño against Nvidia’s GB200 and GB300 NVL72 racks on three public models: GPT-OSS 120B, DeepSeek R1, and Kimi K2.5. Across those workloads, OpenAI reports Jalapeño delivering 1.5x to 1.9x more throughput per watt at peak load, 1.7x to 3.6x lower end-to-end latency, and 2.1x to 4.1x faster response times in interactive, chat-style sessions where per-user speed matters more than raw batch throughput.
On the GPT-OSS 120B model specifically, OpenAI’s figures put Jalapeño at around 59,800 mixed tokens per second per package, against roughly 53,950 for a GB200 package running the same workload, a case where Jalapeño wins on raw throughput and not just efficiency. On tokens generated per kilowatt, Jalapeño’s reported peak of about 85,448 TPS/kW compares to roughly 44,960 TPS/kW for a GB200 baseline, close to a 1.9x gap.
Latency, Model By Model
The latency gap widens on larger, more complex models. OpenAI’s release shows GPT-OSS 120B completing a request in about 1.03 seconds on Jalapeño versus 1.80 seconds on GB200. DeepSeek R1, a much larger 670-billion-parameter model, drops from roughly 5.99 seconds on GB200 to about 1.65 seconds on Jalapeño. Kimi K2.5, a trillion-parameter model, goes from around 5.31 seconds to 1.56 seconds. That’s a 3.4x to 3.6x cut in wait time on the two biggest models tested, the kind of gap a user would actually notice mid-conversation.
| Metric | OpenAI Jalapeño | Nvidia GB200 NVL72 | Nvidia GB300 NVL72 |
|---|---|---|---|
| Package power rating | ~700W | ~1,200W | ~1,400W |
| Peak perf/watt vs Jalapeño | Baseline | 1.5x–1.9x lower | 1.5x–1.9x lower |
| GPT-OSS 120B latency | ~1.03s | ~1.80s | N/A (not tested) |
| DeepSeek R1 latency | ~1.65s | ~5.99s | N/A (not tested) |
| Kimi K2.5 latency | ~1.56s | ~5.31s | N/A (not tested) |
| Claimed cost-per-token savings | Baseline | ~50% higher (per Broadcom) | ~50% higher (per Broadcom) |
Figures above come from OpenAI’s own August 25, 2026 benchmark release, as reported by CNBC. Independent third-party verification of these specific numbers had not been published as of this writing.
The 50% Cost Claim, and What It’s Based On
The number getting repeated most in coverage of Jalapeño isn’t a throughput multiplier, it’s cost. Broadcom CEO Hock Tan has said the chip offers roughly a 50% cost advantage per token compared with current-generation GPU-based inference systems. That figure comes from Broadcom’s own cost modeling, not a neutral third party, so it deserves the same caveat as the throughput numbers: it’s the seller’s math.
“The chip made by the team is as good as the Blackwell chips made by Nvidia or the tensor processing units designed by Alphabet’s Google,” Hock Tan said, according to CNBC’s coverage of the original June unveiling. That’s a notably direct comparison for a CEO to make in public, and it set the bar OpenAI’s August benchmarks were then measured against.
OpenAI itself was more careful with its language in the August release, framing the result narrowly: “Jalapeño beat Nvidia’s Blackwell systems on performance per watt in nearly all tested scenarios,” the company said, per CNBC. Nearly all, not all, and per watt, not universally faster.
Nine Months From Design to Tape-Out
What made Jalapeño possible on this timeline is as much a story about Broadcom’s design process as it is about OpenAI’s chip architecture. Coverage from Tom’s Hardware describes the roughly nine-month span from initial design to manufacturing tape-out as one of the fastest development cycles ever achieved for a chip of this complexity. Custom ASIC programs at this scale typically run two to three years before first silicon; Jalapeño compressed that into three quarters.
OpenAI designed the chip’s architecture around specific bottlenecks it had already identified in its own inference stack: data movement between compute and memory, the balance of compute-to-memory bandwidth for transformer-style workloads, and network efficiency across racks. Broadcom handled the physical design and integration work. TSMC manufactured the silicon on an advanced process node. Celestica built the racks and system-level integration that got the chip from a wafer to a deployable server.
Why OpenAI Wants Its Own Silicon
OpenAI’s compute bill is the single biggest line item standing between the company and profitability, and nearly all of that spend has historically gone to Nvidia. Building an in-house inference chip attacks that dependency from two directions: it gives OpenAI a hardware platform tuned specifically to how its own models actually run, and it gives the company leverage in future GPU price negotiations with Nvidia, regardless of how big a discount OpenAI actually gets today.
Inference, not training, is where this bet is aimed. Training runs happen periodically and can tolerate custom hardware quirks because a small internal team manages the whole pipeline. Inference runs constantly, at massive scale, serving live traffic from hundreds of millions of users, and even small efficiency gains compound fast across a bill that size. That’s why Jalapeño was built as an inference-only chip rather than a general-purpose Nvidia replacement.
Jalapeño in the Custom Silicon Field: TPU, Trainium, Maia
OpenAI isn’t the first AI lab to build its own chip, and analysts have been quick to frame Jalapeño as the latest entry in a race that Google, Amazon, and Microsoft already started. Google has run its own TPU line internally for years and increasingly rents TPU capacity externally. Amazon has pushed its Trainium chips as a lower-cost option inside AWS specifically to reduce its own Nvidia dependency. Microsoft has developed its Maia accelerators for internal Azure workloads. None of the three has published head-to-head benchmarks against Jalapeño, and OpenAI’s own release didn’t include direct TPU or Trainium comparisons either, so any ranking between them right now is analyst inference, not measured fact.
| Company | Custom Chip Program | Primary Use | External Availability |
|---|---|---|---|
| OpenAI | Jalapeño (with Broadcom, TSMC, Celestica) | Inference only | Internal, OpenAI’s own infrastructure |
| In-house TPU line | Training and inference | Internal plus Google Cloud rental | |
| Amazon | Trainium | Training and inference | Available via AWS |
| Microsoft | Maia | Internal Azure inference | Internal only, not externally sold |
| Nvidia | Blackwell (GB200/GB300 NVL72) | Training and inference, merchant silicon | Sold broadly across the industry |
What sets Jalapeño apart from the TPU and Trainium programs isn’t the ambition, it’s the speed and the direct benchmark challenge. Google and Amazon built their custom silicon programs over the better part of a decade, iterating quietly through multiple generations before making bold public comparisons to Nvidia’s flagship parts. OpenAI went from unveiling to a public Blackwell-beating benchmark claim in about two months.
How Analysts Are Reading the Numbers
Analyst reaction has landed closer to “credible incremental threat” than “Nvidia crisis.” Adrien Sanchez, a technology analyst at Yole Group, told CNBC that the results show a “hyperscaler-designed chip can now match or beat Nvidia’s Blackwell-class GPUs on inference efficiency,” a notable statement given how dominant Nvidia’s inference share has been, but one that stops well short of predicting a broad shift in spending.
The comparison itself has a built-in limitation that several analysts flagged: Jalapeño is being measured against GB200 and GB300, Nvidia racks that launched in 2024 and 2025, not against Nvidia’s upcoming Vera Rubin platform. By the time Jalapeño reaches meaningful production volume, Nvidia will likely have shipped its next generation, which resets the comparison OpenAI is currently winning.
Nvidia’s Market Reaction
Nvidia stock dipped modestly following the August 26 coverage of Jalapeño’s benchmarks, coming just ahead of Nvidia’s own upcoming earnings release, but shares stayed close to record levels. That muted reaction lines up with how markets treated Google’s TPU expansion and Amazon’s Trainium push: real competitive pressure, not an existential threat to a company whose GPUs still underpin the large majority of AI training and inference workloads industry-wide.
Nvidia has not issued a formal rebuttal of OpenAI’s specific benchmark numbers as of this writing. The company has historically responded to custom-silicon competition by emphasizing the breadth of its software ecosystem (CUDA, and the surrounding developer tooling) rather than disputing individual chip-versus-chip comparisons, since raw silicon specs are only part of what determines real-world deployment cost.
Historical Context: A Decade of Chip Independence Attempts
The push to escape Nvidia’s pricing power isn’t new, it’s just accelerating. Google’s TPU program dates back to 2015, born out of the same cost pressure OpenAI is responding to now: serving inference at Google’s scale on merchant GPUs was getting too expensive. Amazon followed with its own Inferentia and later Trainium chips through the 2020s, explicitly pitched to AWS customers as a cheaper alternative to Nvidia instances. Microsoft’s Maia effort, revealed in 2023, took longer to reach meaningful internal deployment.
What’s different about Jalapeño is the speed of the public challenge. Google and Amazon each spent years quietly proving out their chips internally before making direct performance claims against Nvidia’s top parts. OpenAI, by contrast, went public with a head-to-head Blackwell comparison less than a year after Jalapeño’s design work reportedly began, and roughly two months after the chip’s public unveiling. That’s either confidence in the results or pressure to show investors and Microsoft (its largest backer) that its compute costs have a credible ceiling.
What This Means for AI Infrastructure Costs
If OpenAI’s cost figures hold up in production, and that’s a real if given they’re self-reported, the practical effect would show up first in ChatGPT’s operating margins rather than in prices consumers pay directly. Inference cost is the primary variable cost behind serving every chat completion, every API call, every agentic workflow OpenAI runs at scale. A real-world 30-50% reduction in that cost changes the math on how aggressively OpenAI can price its API, subsidize free-tier usage, or fund the next generation of model training.
It also changes OpenAI’s negotiating position with Nvidia on its remaining GPU purchases, and with Microsoft on Azure capacity commitments. Even a chip that only handles a fraction of OpenAI’s total inference load gives the company a credible walk-away number in those conversations, something it didn’t have a year ago.
Deployment Timeline: What Happens Next
OpenAI has said Jalapeño is expected to begin deployment inside its own computing infrastructure before the end of 2026, with broader rollout continuing over the following several quarters. That’s a notably fast turnaround from a June unveiling to production hardware, though “deployment” at this stage likely means a limited slice of inference traffic rather than a wholesale replacement of OpenAI’s existing Nvidia fleet.
Nvidia GPUs aren’t going anywhere from OpenAI’s infrastructure in the near term. OpenAI still relies heavily on Nvidia hardware for model training, and Jalapeño so far only targets inference. Expect a hybrid infrastructure for years: Nvidia for training and a meaningful chunk of inference, Jalapeño for the inference workloads where its efficiency profile pays off fastest.
Predictions: Where This Goes From Here
- Nvidia will likely respond to the Jalapeño benchmarks by accelerating public messaging around its Vera Rubin platform rather than disputing OpenAI’s GB200/GB300 comparisons directly.
- Expect Microsoft to push its own Maia roadmap further into the open over the next two quarters, given its financial stake in OpenAI’s compute economics and its own Azure cost pressures.
- Independent, third-party benchmarks of Jalapeño (from firms like SemiAnalysis or MLPerf-style consortiums) will likely surface within the next two to three months and probably narrow the gap OpenAI is currently claiming.
- OpenAI’s initial Jalapeño deployment will likely stay concentrated on high-volume, latency-sensitive consumer ChatGPT traffic first, with enterprise API workloads migrating later once the chip’s reliability track record builds up.
- Other AI labs without hyperscaler backing (Anthropic, for instance) will face growing pressure to secure their own custom silicon partnerships or long-term Nvidia supply deals to avoid a widening cost gap against OpenAI, Google, and Amazon.
Competitive Comparison: Jalapeño’s Real Rival Isn’t Blackwell
The framing of “Jalapeño vs. Blackwell” makes for a clean headline, but it slightly misstates the actual competitive dynamic. Nvidia sells merchant silicon to everyone: OpenAI, Google, Amazon, Microsoft, Meta, and thousands of smaller AI companies all buy the same GB200 and GB300 hardware. Jalapeño isn’t a merchant product; OpenAI isn’t selling it to anyone else. Its real competitive set is the other hyperscaler-only chips built for internal use: Google’s TPU line and Amazon’s Trainium chips, both of which have had years longer to mature.
Where Jalapeño genuinely stands apart is in how it was benchmarked. OpenAI didn’t just claim efficiency gains in the abstract, it published specific per-model latency and throughput figures against named Nvidia products, which is a more aggressive public disclosure than Google or Amazon have typically made about their own custom chips relative to Nvidia.
What to Watch Before Believing the Benchmarks
Every figure in OpenAI’s August 25 release came from OpenAI’s own testing on OpenAI’s own infrastructure, using models OpenAI chose. That doesn’t make the numbers false, but it’s a meaningfully different standard than a neutral benchmark like MLPerf, where multiple vendors submit results under shared rules and third parties can audit the methodology. Until an outside group replicates these comparisons, or Nvidia publishes its own counter-benchmarks on the same three models, treat the specific multipliers (1.5x-1.9x, 1.7x-3.6x, and so on) as OpenAI’s best case rather than an industry-verified baseline.
The 50% cost figure carries the same caveat, doubled. It comes from Broadcom, the company that gets paid to design and build Jalapeño, describing the economics of a chip it has a direct financial interest in promoting.
Frequently Asked Questions
What is OpenAI’s Jalapeño chip?
Jalapeño is OpenAI’s first custom AI chip, designed for inference workloads and built in partnership with Broadcom, manufactured by TSMC, and assembled into server racks by Celestica. It was first unveiled June 24, 2026.
Does Jalapeño actually beat Nvidia Blackwell?
By OpenAI’s own August 25, 2026 benchmarks, Jalapeño outperforms Nvidia’s GB200 and GB300 NVL72 systems on performance per watt and latency across three tested models. These are self-reported figures, not independently audited results, so they should be treated as OpenAI’s best-case claims rather than confirmed industry benchmarks.
When will Jalapeño be deployed in production?
OpenAI has said it expects to begin deploying Jalapeño within its own infrastructure before the end of 2026, with broader rollout over subsequent quarters.
Is Jalapeño for sale to other companies?
No. Unlike Nvidia’s GPUs, which are sold broadly across the industry, Jalapeño is built for OpenAI’s own internal infrastructure, similar to how Google’s TPUs and Amazon’s Trainium chips are primarily used internally or rented through their respective cloud platforms.
How does Jalapeño compare to Google TPU and Amazon Trainium?
No direct, apples-to-apples public benchmark exists between Jalapeño, Google’s TPU line, and Amazon’s Trainium chips as of this writing. All three are custom inference-focused chips built by AI labs and hyperscalers looking to reduce reliance on Nvidia GPUs, but each was built independently and tested under different conditions.
Will Jalapeño replace Nvidia GPUs at OpenAI?
Not in the near term. Jalapeño is built for inference only, not training, and OpenAI is expected to continue relying heavily on Nvidia GPUs for model training and a substantial share of inference for years to come.
Why did Nvidia’s stock only fall modestly on this news?
Analysts have generally characterized Jalapeño as a credible but incremental competitive pressure rather than an existential threat, similar to how markets reacted to Google’s TPU and Amazon’s Trainium expansions. Nvidia still supplies the large majority of AI compute industry-wide.
What does the Jalapeño benchmark mean for AI pricing?
If OpenAI’s claimed cost savings hold up at production scale, the most likely near-term effect is improved margins on OpenAI’s existing products rather than an immediate cut in prices for consumers or API customers.
