On August 25, 2026, OpenAI published its first independent benchmark results for Jalapeño, the custom inference chip it built with Broadcom, showing 1.5 to 1.9 times more compute per watt and up to 4.1 times better performance on interactive workloads than Nvidia's Blackwell-based GB200 systems. Less than 24 hours later, Nvidia reported second-quarter revenue of $96.2 billion, up 106% from a year earlier, with data center sales alone hitting $89.0 billion. Two headlines, one week, telling almost opposite stories about the company that has dominated the AI boom. For anyone holding Nvidia stock, an AI-heavy index fund, or simply trying to understand where the next leg of the AI trade is headed, the timing was not a coincidence, and the details matter more than either headline alone.
The short version: OpenAI, Nvidia's single largest customer by some estimates, just showed the industry it can build hardware that rivals Nvidia's own chips for a specific and increasingly important job. Nvidia, meanwhile, just posted numbers so large that Wall Street mostly shrugged off the chip news. Both things are true at once, and understanding why requires separating what OpenAI actually proved from what it merely claimed, and what Nvidia's blowout quarter does and doesn't say about its long-term moat.
What OpenAI's Jalapeño Benchmarks Actually Show
Jalapeño is OpenAI's first application-specific chip designed purely for inference, the work of actually running a trained model to answer a prompt, as opposed to training the model in the first place. OpenAI and Broadcom took the chip from initial design to manufacturing tape-out at TSMC's 3-nanometer node in roughly nine months, a pace semiconductor analysts have called one of the fastest cycles ever for a chip of this complexity. Rather than release marketing slides, OpenAI ran its numbers through SemiAnalysis's public InferenceX benchmark, a third-party test that measures throughput, power draw, and latency together across a full request-processing pipeline, rather than cherry-picking a single metric.
Tested against Nvidia's GB200 systems on the open-source GPT-OSS 120B model, Jalapeño delivered 1.5 to 1.9 times more compute per watt and cut latency by 1.7 to 3.6 times. On the most demanding interactive workloads, the kind where a user is waiting on a live response rather than a batch job running overnight, the gap widened to 2.1 to 4.1 times. OpenAI is also targeting roughly 50% lower cost per inference token than current-generation Nvidia hardware once Jalapeño reaches volume production. Those are real, independently benchmarked numbers, not vendor slideware, which is precisely why the semiconductor world took notice.
The caveats matter just as much as the headline figures. Jalapeño is only entering low-volume production in late 2026, with a fuller ramp expected across 2027 and 2028. It hasn't been tested against Nvidia's next-generation Vera Rubin platform, which is still to come. And crucially, this is a benchmark aimed squarely at inference, the recurring, high-volume, cost-sensitive workload that now makes up a growing share of AI spending. Training frontier models, the far more compute-intensive process of building a model like GPT-5 in the first place, remains overwhelmingly a job for Nvidia's GPUs, and nothing about Jalapeño changes that in the near term.
Nvidia's Earnings Tell a Different Story, For Now
If Jalapeño's benchmarks were meant to rattle Nvidia, the market's answer came fast. Nvidia's fiscal second-quarter results, reported after the close on August 26, showed data center revenue up 117% year over year to $89.0 billion, split between $48.7 billion in hyperscale cloud sales and $40.3 billion in AI clouds, industrial, and enterprise revenue. Both GAAP and non-GAAP gross margins held at 75.0%, meaning three-quarters of every dollar Nvidia books from a chip sale is profit before overhead. CEO Jensen Huang told analysts on the earnings call that demand is "accelerating," not slowing, and the company guided to $108 billion in revenue for the current quarter, a number that would have sounded implausible for the entire AI industry just three years ago.
Perhaps the more telling figure sat further down the release: Nvidia's supply commitments, essentially forward orders and reserved capacity from customers, more than doubled from $119 billion in the prior quarter to $279 billion, driven largely by memory procurement. In plain terms, Nvidia's own customers, quite possibly including OpenAI itself, are still locking in massive volumes of Nvidia hardware well into the future even as some of them build competing chips on the side. That is not the behavior of buyers who believe Nvidia's dominance is about to end.
Key Numbers From Both Announcements
| Metric | OpenAI / Jalapeño (Aug 25 benchmark) | Nvidia Q2 FY2027 (Aug 26 earnings) |
|---|---|---|
| Headline claim | 1.5–1.9x compute per watt vs. GB200; up to 4.1x on interactive workloads | $96.2B total revenue, up 106% year over year |
| Data center / core segment figure | ~50% target reduction in cost per inference token | $89.0B data center revenue, up 117% year over year |
| Timeline | Low-volume production late 2026; full ramp 2027–2028 | Q3 FY2027 guidance of $108B, plus or minus 2% |
| Independent verification | Benchmarked on SemiAnalysis's InferenceX, but not yet MLPerf-certified | Audited quarterly financials filed with the SEC |
| What it signals | Hyperscaler-designed silicon can now credibly compete on inference | Current demand for Nvidia hardware remains effectively unconstrained |
Why Every Major AI Buyer Is Building Its Own Silicon
OpenAI is not acting alone, and that context is what makes Jalapeño worth watching rather than dismissing as a one-off. Meta has committed to deploying 1 gigawatt of custom AI chips built on Broadcom technology as part of a larger multi-gigawatt agreement. Anthropic, OpenAI's closest rival, has pledged more than $100 billion in spending with Amazon Web Services over the next decade that explicitly includes current and future generations of Amazon's own Trainium chips. Google has run its in-house TPU line for years specifically to reduce its reliance on Nvidia. Microsoft is reportedly slated to absorb around 40% of Jalapeño's initial production run for its own Azure infrastructure.
The economic logic is simple once you see Nvidia's margins in black and white. A company posting 75% gross margins on hardware is, by definition, leaving a large amount of money on the table for any customer big enough to design around it. OpenAI, Meta, Amazon, Google, and Microsoft are collectively spending hundreds of billions of dollars a year on AI infrastructure. Even a 30-to-50% reduction in the cost of running inference, the part of the AI pipeline that scales with every single user query rather than every model they train, translates into real savings at that volume. None of this threatens Nvidia's grip on training hardware in the near term, but it chips away, deliberately, at the highest-margin corner of Nvidia's business.
What This Means for Your Portfolio
If you own Nvidia shares directly, through a tech-heavy ETF, or via a target-date retirement fund that has quietly become AI-concentrated over the past three years, none of this is a reason to panic-sell on a single benchmark release. Nvidia's own results, delivered one day after the Jalapeño news, show a company still growing revenue faster than almost any large-cap in market history, with supply commitments that suggest demand visibility well into 2027. The custom silicon trend is a multi-year structural risk to Nvidia's margins, not a multi-week trading catalyst.
That said, it's worth watching two things closely over the next several quarters. First, whether Jalapeño's real-world performance, once independently tested through MLPerf or similar third-party benchmarks, holds up to OpenAI's own numbers; vendor-published results have a way of looking better in a press release than in production. Second, whether Nvidia's gross margins begin compressing as inference workloads shift toward custom chips, even as total revenue keeps climbing. A company can grow its top line and still see its profitability per chip erode if its highest-margin segment faces new competition. That distinction, growth versus margin durability, is the one long-term investors should track rather than reacting to any single week's headlines.
Questions to Ask Before You Trade on This News
- Is this benchmark independently verified, or is it a vendor-published result that hasn't yet been replicated on a standard like MLPerf?
- Does the news affect training demand, inference demand, or both? Nvidia's moat in training remains far more durable than its position in inference.
- What does the company's own forward guidance and supply commitments say about near-term demand, separate from the competitive narrative?
- Am I reacting to a single data point, or to a trend confirmed across multiple quarters and multiple competitors?
- If I'm invested through a fund rather than individual stock, how concentrated is that fund in AI infrastructure names, and does that match my risk tolerance?
What Happens Next
The next real test for Jalapeño isn't a press release, it's production. OpenAI has said prototype deployments begin in its own infrastructure by the end of 2026, with Microsoft absorbing a large share of the initial run. Watch for independent MLPerf benchmark submissions, which would give the industry an apples-to-apples comparison outside OpenAI's own testing. On Nvidia's side, the company's next quarterly report will show whether the $279 billion in supply commitments converts into booked revenue on schedule, and whether gross margins hold at 75% as more of its largest customers bring inference workloads in-house. Neither company's story ends with last week's headlines; both are multi-year bets that will keep showing up in quarterly numbers for years to come.