OpenAI has become the fourth major AI player to put hard performance numbers behind a custom chip built to outrun Nvidia at the one job that now eats the industry’s electricity bill: inference. On August 25, 2026, OpenAI published the first benchmark results for Jalapeño, the inference accelerator it built with Broadcom, and the numbers are specific enough to move the conversation past marketing slides. OpenAI says the chip delivers 1.5x to 1.9x more AI work per watt than Nvidia’s GB200 and GB300 systems, with end-to-end latency cut by 1.7x to 3.6x on select workloads.
That puts OpenAI alongside Google, Microsoft and Amazon in the growing group of AI companies fielding their own silicon instead of relying entirely on Nvidia’s GPUs. None of the four has said it plans to walk away from Nvidia hardware. But each new chip narrows the gap between “Nvidia is the only option” and “Nvidia is the default option,” and that shift matters for anyone tracking AI infrastructure costs, GPU pricing, or the shape of the next data center buildout.
What OpenAI Actually Announced on August 25
OpenAI’s announcement was narrower than some headlines suggested. The company did not claim Jalapeño beats Nvidia across the board or that it is abandoning Nvidia hardware. It released the first public benchmark results for a chip it had already unveiled two months earlier, and it did so using InferenceX, a public benchmark maintained by the semiconductor analysis firm SemiAnalysis, comparing Jalapeño against Nvidia’s GB200 and GB300 systems.
According to TechCrunch’s reporting on the announcement, OpenAI’s chip chief Richard Ho said Nvidia’s GB300 was the leading publicly benchmarked system available for comparison, and that Jalapeño is expected to begin deploying “in very small volumes” by the end of 2026, with a more significant production ramp arriving in 2027. That timeline matters: this is not a chip replacing Nvidia GPUs in OpenAI’s data centers next quarter. It is a chip proving it can compete, ahead of a slower rollout.
CNBC reported that OpenAI said the chip beat Nvidia’s Blackwell-generation systems on performance per watt in nearly all tested scenarios, while noting that analysts cautioned Nvidia’s newest platform is the more directly comparable rival going forward. That distinction, an incumbent’s newest chip versus a challenger’s first chip, is the kind of nuance that gets lost fast once a story like this starts trending.
Inside Jalapeño: What OpenAI and Broadcom Actually Built
Jalapeño is OpenAI’s first self-designed AI chip, built in partnership with Broadcom and unveiled on June 24, 2026. OpenAI calls it an “Intelligence Processor” rather than a GPU, a naming choice meant to signal that the chip was architected specifically around large language model inference rather than adapted from a general-purpose graphics design. It is not a training chip. OpenAI still trains its frontier models, including the GPT family, on Nvidia hardware, and the company has been explicit that Jalapeño’s job is serving already-trained models to users at scale, not building them.
The chip is rated at 700W, well under the 1,200W rating of Nvidia’s GB200 and the 1,400W rating of GB300 in OpenAI’s comparison set, and Memeburn reported that in sustained testing Jalapeño ran at 550W or lower. Lower power draw at the chip level compounds across a data center: less power per rack means more racks per substation, and less cooling infrastructure per unit of inference delivered. That is the economic argument OpenAI is making, and it is a different argument than “our chip is faster.” It is “our chip does more work per dollar of electricity,” which is the metric that actually shows up on a cloud infrastructure budget.
The Benchmark Numbers, Model by Model
OpenAI tested Jalapeño against three open-weight models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. On GPT-OSS 120B compared against Nvidia’s GB200, Memeburn reported peak throughput of 85,448 tokens per kilowatt for Jalapeño versus 44,960 for GB200, a 1.9x advantage. Across all three tested models, OpenAI’s own figures put the performance-per-watt gain in the 1.5x to 1.9x range, with end-to-end latency 1.7x to 3.6x lower than the Nvidia systems it was measured against. Memeburn additionally reported that the minimum gap between generated tokens, a metric closely tied to how responsive a chatbot feels to a user, improved by 2.7x to 4.1x depending on the model.
| Metric | Jalapeño | Nvidia GB200 | Nvidia GB300 |
|---|---|---|---|
| Rated power draw | 700W (550W or lower sustained) | 1,200W | 1,400W |
| Peak throughput/kW, GPT-OSS 120B | 85,448 tokens/kW | 44,960 tokens/kW | N/A (tested vs GB200) |
| Performance-per-watt gain (all 3 models) | 1.5x to 1.9x higher | Baseline | Baseline |
| End-to-end latency | 1.7x to 3.6x lower | Baseline | Baseline |
| Min. time between tokens | 2.7x to 4.1x faster | Baseline | Baseline |
| Benchmark used | InferenceX (SemiAnalysis public benchmark) | ||
Those figures come from OpenAI’s own testing and its chosen benchmark, which is a standard caveat for any vendor-published number. SemiAnalysis’s InferenceX is a public, third-party-maintained benchmark rather than an internal OpenAI tool, which gives the comparison more credibility than a self-graded test, but it is still OpenAI picking which models and configurations to publish. No independent lab has yet run a matching test on hardware it controls, so the numbers should be read as a strong opening claim rather than a settled verdict.
Why OpenAI Is Building Its Own Silicon
The motivation is not mysterious: inference is OpenAI’s largest and fastest-growing cost center. Every ChatGPT response, every Codex completion, every API call runs through inference hardware, and that hardware bill scales directly with usage in a way training costs do not. Reuters reported earlier this year that OpenAI was “unsatisfied with some Nvidia chips” and was actively looking at alternatives, even as a company spokesperson told Reuters that Nvidia still powers the vast majority of OpenAI’s inference fleet and delivers the best performance per dollar available today. That combination, dissatisfaction paired with continued dependence, is the most accurate summary of where OpenAI actually stands: hedging, not divorcing.
OpenAI’s own framing, posted alongside the Jalapeño announcement, described the chip as “designed from the ground up by OpenAI and brought to production with Broadcom,” built to be “purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products,” according to the company’s announcement. That is a company describing a supply chain hedge, not a company declaring independence from its largest hardware supplier, which remains Nvidia under a partnership that has included a $100 billion investment commitment and a planned 10 gigawatts of Nvidia systems starting with the Vera Rubin platform.
Joining a Crowded Field: Google, Microsoft, Amazon Got There First
OpenAI is the newest entrant to a strategy that Google, Amazon, and Microsoft have been running for years. Google has shipped multiple generations of its Tensor Processing Units, most recently deploying TPU 8t and 8i variants built with Broadcom and MediaTek. Amazon has pushed its in-house Trainium line through Annapurna Labs, with Trainium3 already running production workloads for customers including Uber under commercial terms reported to undercut Nvidia pricing. Microsoft has developed its own Maia accelerators for internal Azure workloads. Each of the three has made the same basic bet OpenAI is now making: that owning silicon tuned to your own model architecture beats buying general-purpose GPUs at Nvidia’s margins, at least for the specific, repetitive job of inference.
What makes OpenAI’s move notable is not novelty, it is scale and timing. OpenAI is the company whose usage growth has been most directly cited as the reason Nvidia’s data center revenue keeps climbing quarter after quarter. If the company that buys the most Nvidia inference capacity starts routing a meaningful share of that demand to its own chips, even slowly and even while continuing to buy Nvidia hardware for training and overflow capacity, the long-term demand curve for Nvidia’s inference-focused products shifts. That is the “problem” framing outlets like Memeburn and scanx.trade have used, and it is grounded in a real structural shift rather than pure speculation.
How the Four Custom Chip Programs Compare
| Company | Chip / Program | Silicon Partner | Primary Workload | Reported Status |
|---|---|---|---|---|
| OpenAI | Jalapeño | Broadcom | Inference | Small-volume deployment planned late 2026, larger ramp in 2027 |
| TPU 8t / 8i | Broadcom, MediaTek | Training and inference | In production for Google and cloud customers | |
| Amazon | Trainium3 | Annapurna Labs (in-house) | Training and inference | Deployed at scale, including third-party customers such as Uber |
| Microsoft | Maia | In-house design | Training and inference (Azure internal) | In active development and internal deployment |
| Nvidia | Blackwell (GB200/GB300), Vera Rubin platform | Merchant silicon, sold broadly | Training and inference | Current market leader; incumbent in all four benchmark comparisons above |
Broadcom’s role across three of the four programs is the pattern worth watching closely. The company that used to be known mainly for networking chips and enterprise software is now the common design partner behind Google’s TPUs and OpenAI’s Jalapeño, and it has separately been reported in talks around Anthropic’s own custom chip ambitions. When Broadcom CEO Hock Tan told Reuters that the chip OpenAI’s team built is “as good as the Blackwell chips made by Nvidia or the tensor processing units designed by Alphabet’s Google,” he was describing a company positioning itself as the arms dealer for every major AI lab that wants to reduce its Nvidia dependence without building a chip design team from scratch.
Nvidia’s Position: Still the Incumbent, Still Growing
None of this means Nvidia is losing ground today. OpenAI’s own spokesperson told Reuters that Nvidia delivers the best performance per dollar for inference currently available, and that Nvidia hardware still runs the vast majority of OpenAI’s inference fleet. The $100 billion partnership between the two companies, announced in 2025, remains in place, with Nvidia’s Vera Rubin platform slated to anchor 10 gigawatts of planned deployment. Reports have also pointed to Nvidia in discussions to guarantee as much as $250 billion tied to a major U.S. data center project connected to OpenAI’s infrastructure buildout. Those are not the actions of a company being displaced. They are the actions of a company whose biggest customer is diversifying its supply chain while continuing to place enormous orders.
The Register’s analysis of the announcement captured the skepticism directly: comparing Jalapeño to GB300, rather than to whatever Nvidia ships next, understates how fast Nvidia’s own roadmap moves. Nvidia has shipped a new architecture generation roughly every one to two years for the better part of a decade, and a custom chip that beats today’s Blackwell parts on paper may be facing a different Nvidia product by the time it reaches meaningful production volume in 2027.
Market Impact: What Changes for AI Infrastructure Spending
The practical effect of Jalapeño’s benchmark release is not a sudden drop in Nvidia orders. It is a shift in negotiating leverage. Every hyperscaler and AI lab now has, or is building, a credible alternative supplier for at least part of its inference workload, and that alternative does not need to fully replace Nvidia to matter. It only needs to be good enough to serve as a bargaining chip in pricing discussions and a hedge against Nvidia’s own supply constraints, which have periodically limited how quickly AI companies could scale up GPU-based capacity.
For enterprise buyers and developers, the more immediate consequence is a widening menu of inference options with different cost and performance tradeoffs. Companies serving high-volume, latency-sensitive AI products, chat assistants, coding tools, real-time agents, now have a reason to watch custom silicon announcements the way they once watched GPU generation launches. Inference cost per token has become one of the defining unit economics of the AI industry, and a chip claiming a 1.5x to 1.9x efficiency gain on that specific metric is directly relevant to anyone’s cloud bill, not just to OpenAI’s.
Historical Context: From GPU Monopoly to Multi-Vendor Inference
Five years ago, the AI infrastructure conversation had one axis: how many Nvidia GPUs can you get, and how fast. Google’s TPU program was the outlier, running for internal workloads with limited outside visibility. That single-vendor picture started breaking down as training costs plateaued relative to inference costs, since a model trained once gets queried billions of times, and each of those queries costs real money on someone’s infrastructure bill. Amazon’s Trainium push, Microsoft’s Maia development, and now OpenAI’s Jalapeño all follow the same logic: training is a one-time capital-intensive project that benefits from Nvidia’s mature software stack and raw flexibility, while inference is a repeatable, predictable workload that rewards purpose-built silicon once volume gets high enough to justify the design cost.
What is different about 2026 compared to the earlier custom-silicon efforts is that the company building the chip, OpenAI, is also the company whose product usage is most directly responsible for inference demand growth industry-wide. Google and Amazon build custom chips partly to serve their own cloud customers at better margins. OpenAI is building a chip almost entirely to serve its own workload, which makes Jalapeño less a cloud-platform play and more a direct read on how OpenAI itself expects its own infrastructure costs to evolve over the next several years.
What Industry Voices Are Saying
OpenAI described the chip in its own announcement: “We’ve designed and built our first AI chip: Jalapeño. Designed from the ground up by OpenAI and brought to production with Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products,” the company said, according to its official announcement.
Broadcom CEO Hock Tan offered the most direct competitive comparison available. “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,” Tan said in an interview reported by Reuters.
CNBC’s coverage of the August 25 benchmark release noted that OpenAI said testing results of its new custom inference chip “show a significant performance advance,” a characterization the company has stood behind even as it continues placing large Nvidia orders.
OpenAI has also been careful to temper expectations about a full Nvidia exit. A company spokesperson, in a statement reported by Reuters, said OpenAI relies on Nvidia to power the vast majority of its inference fleet and that Nvidia delivers the best performance per dollar for inference currently on the market, a striking admission to publish in the same window as a chip announcement designed to show Nvidia can be beaten on specific benchmarks.
OpenAI’s own product description of the chip called Jalapeño “the first AI accelerator in a multi-generation compute platform the companies are building together to make advanced AI faster, more reliable, and more accessible to more people,” language that frames this as the start of a roadmap rather than a single product launch.
Competitive Comparison: Jalapeño’s Real Position in the Market
Stacked against its three hyperscaler rivals, Jalapeño’s advantage is specificity. Google’s TPUs and Amazon’s Trainium chips are built to serve broad customer bases across cloud platforms running many different model architectures. Jalapeño is tuned narrowly to the inference patterns of OpenAI’s own model family, a tighter design target that plausibly explains its power-efficiency numbers but also limits its usefulness to anyone outside OpenAI’s own infrastructure, at least for now. Microsoft’s Maia sits in a similar internal-use category, built primarily to serve Azure’s own AI workloads rather than to be sold or licensed externally.
Nvidia’s structural advantage remains its software ecosystem. CUDA, the toolchain most AI researchers already know, runs on Nvidia hardware, and switching a production inference pipeline to a new chip architecture carries real engineering cost even when the underlying silicon is faster or cheaper to run. That switching cost is a big part of why OpenAI’s own spokesperson can simultaneously praise Jalapeño’s benchmark results and confirm that Nvidia still handles the vast majority of the company’s inference traffic. Benchmark wins do not automatically translate into workload migration.
What Comes Next: Five Predictions
- Jalapeño’s 2026 rollout stays small. OpenAI’s own chip chief has said deployment begins in very small volumes by year-end, so any near-term impact on Nvidia’s order book will be limited through early 2027.
- Broadcom becomes the default custom-silicon partner. With Google’s TPU program and OpenAI’s Jalapeño both running through Broadcom, expect more AI labs, potentially including Anthropic, to explore Broadcom-designed accelerators rather than building chip teams from scratch.
- Nvidia keeps growing even as diversification spreads. Total AI compute demand is expanding faster than any single custom chip program can absorb, so Nvidia’s overall revenue trajectory is unlikely to reverse even as its share of inference workloads gradually narrows.
- Independent benchmarking pressure increases. As more vendor-published numbers like Jalapeño’s circulate, expect louder calls for neutral, reproducible testing on hardware outside any single company’s control, following the model InferenceX has started to establish.
- Inference cost-per-token becomes a headline metric. Just as GPU generation launches once dominated AI hardware coverage, expect more public benchmarking around inference efficiency, since it now directly determines the unit economics of consumer AI products.
The Bigger Picture for Nvidia Investors and AI Buyers
Nothing in OpenAI’s announcement changes Nvidia’s near-term financial picture. The company remains the dominant supplier for AI training and the largest single supplier for AI inference, and its partnership commitments with OpenAI, including the previously announced 10-gigawatt Vera Rubin deployment, remain intact. What has changed is the competitive narrative: for the first time, all four of the companies buying the most Nvidia inference capacity, Google, Amazon, Microsoft, and now OpenAI, have publicly benchmarked custom silicon against Nvidia’s current-generation hardware and published results claiming a real advantage on efficiency. That is a different story than any one of those announcements happening in isolation, and it is the reason outlets across financial and technology media framed the OpenAI news as adding to Nvidia’s competitive pressure rather than as an isolated product launch.
Whether that pressure shows up in Nvidia’s next earnings report depends on how fast Jalapeño’s 2027 production ramp actually moves and how much inference volume OpenAI is willing and able to shift onto hardware still working through its first year of real-world deployment. For now, the most accurate read is that OpenAI has proven it can build a competitive chip, not that it has replaced the one it has been buying by the billions.
Frequently Asked Questions
What is OpenAI’s Jalapeño chip?
Jalapeño is OpenAI’s first custom AI chip, an inference-focused accelerator co-designed with Broadcom and unveiled on June 24, 2026. It is built specifically to serve, rather than train, large language models like those behind ChatGPT and Codex.
Does Jalapeño actually beat Nvidia GPUs?
On the specific benchmarks OpenAI published on August 25, 2026, using the InferenceX benchmark against Nvidia’s GB200 and GB300 systems, Jalapeño showed 1.5x to 1.9x better performance per watt and 1.7x to 3.6x lower latency. These are OpenAI’s own published results on a third-party benchmark, not independently verified figures from a neutral lab.
Is OpenAI dropping Nvidia hardware?
No. OpenAI has said Nvidia still powers the vast majority of its inference fleet and delivers the best performance per dollar currently available. Jalapeño is planned for small-volume deployment by the end of 2026, with a larger production ramp in 2027, alongside continued Nvidia purchases.
How does Jalapeño compare to Google TPU, Microsoft Maia, and Amazon Trainium?
All four are custom accelerators built by major AI companies to reduce reliance on Nvidia GPUs for training and inference. Google’s TPUs and Amazon’s Trainium chips serve broad cloud customer bases, while Jalapeño and Microsoft’s Maia are built primarily for each company’s own internal workloads.
Who manufactures the Jalapeño chip?
Broadcom is OpenAI’s design and production partner for Jalapeño. Broadcom also partners with Google on its TPU program, making it a common thread across two of the largest custom AI silicon efforts in the industry.
When will Jalapeño actually be used at scale?
OpenAI’s chip chief Richard Ho has said deployment starts in very small volumes by the end of 2026, with a more significant production ramp expected in 2027, according to TechCrunch’s reporting on the announcement.
What benchmark did OpenAI use to test Jalapeño?
OpenAI used InferenceX, a public benchmark maintained by the semiconductor research firm SemiAnalysis, testing three open-weight models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, against Nvidia’s GB200 and GB300 systems.
Why does this matter for Nvidia’s stock and business?
It does not change Nvidia’s near-term revenue, since Jalapeño’s rollout is small through 2026. But it adds OpenAI, historically one of Nvidia’s largest customers, to the list of companies with credible custom silicon, alongside Google, Amazon, and Microsoft, which analysts see as a long-term structural pressure on Nvidia’s inference market share even as overall AI compute demand keeps growing.