Nvidia just posted the best quarter in its history, and Wall Street barely noticed the chip that might matter more for the next five years. On August 25, 2026, OpenAI said its first custom silicon, code-named Jalapeño, beat Nvidia’s current-generation systems on two of the metrics that decide who wins the inference era: work done per watt and speed of response. A day later, Nvidia reported second-quarter fiscal 2027 revenue of $89 billion and a 75.0% gross margin, according to CNBC. Both stories are true at once, and that tension is the real story: Nvidia’s near-term numbers have never looked stronger, while the long-term threat to its pricing power has never looked more concrete.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What OpenAI’s Jalapeño Chip Actually Is
Jalapeño is not a GPU, and OpenAI does not want anyone to call it one. The company describes it as its first “Intelligence Processor,” an accelerator built specifically to run large language model inference rather than to train models from scratch. OpenAI and Broadcom (NASDAQ: AVGO) unveiled the chip on June 24, 2026, saying it was the first product in what they called a multi-generation compute platform, according to CNBC’s coverage of the announcement.
Broadcom’s role goes beyond a logo on a press release. The chipmaker handled silicon implementation and contributed its Tomahawk networking technology, while systems integrator Celestica built out the physical racks, according to Yahoo Finance. OpenAI said the design went from concept to working samples in about nine months, a pace that would have been unthinkable for a custom chip program five years ago, when Google’s TPU took years to mature and Amazon’s Inferentia line needed multiple generations to find real customers.
OpenAI president Greg Brockman framed the project as a bet on vertical control rather than a break from Nvidia. “By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access,” Brockman said in the June announcement. That is corporate language for a blunt reality: running ChatGPT and its API business at OpenAI’s current scale means inference costs now rival or exceed training costs, and every point of efficiency gained on inference chips drops straight to the bottom line, or lets the company drop prices without losing money.
The August 25 Benchmarks That Rattled the Chip Market
Two months after the unveiling, OpenAI published results from real testing rather than projections. Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput than the comparison systems, and cut end-to-end response latency by 1.7 to 3.6 times, according to figures OpenAI released and Bloomberg reported on August 25. OpenAI’s head of hardware, Richard Ho, did not hedge when describing the results. “The bottom line is that the results show a very, very significant performance advance over state of the art,” Ho told reporters, according to TechCrunch’s coverage of the press call.
The comparison system in those tests was Nvidia’s GB300, the Blackwell Ultra platform that anchors most large-scale inference deployments running in 2026. Broadcom CEO Hock Tan went further, telling Reuters that Jalapeño’s economics could cut the cost of running an AI model by roughly half compared with a typical GPU-based inference setup, a figure Yahoo Finance’s headline distilled to “50% Cost Savings” when the chip first launched in June. OpenAI itself had signaled the direction back at the unveiling, saying in a statement that “early testing shows that Jalapeño will deliver performance per watt substantially better than current state-of-the-art,” a claim CNN reported at the time.
Sam Altman, never one to undersell a launch, called the outcome better than expected. “Jalapeño has surpassed our expectations and delivered both deep savings and excellent performance,” Altman said in OpenAI’s official announcement. The company has not published a full technical whitepaper with die size, transistor count, or process node, which means outside chip analysts are still working from OpenAI’s own benchmark disclosures rather than independent lab testing. That gap matters. Every number in this story so far comes from the company selling the chip or the company that benefits from its success, not from a neutral third party running MLPerf-style suites.
Jalapeño vs. Nvidia Blackwell: The Numbers Side by Side
| Metric | Jalapeño (OpenAI/Broadcom) | Nvidia GB300 Blackwell Ultra | Source |
|---|---|---|---|
| Performance per watt (peak throughput) | 1.5x–1.9x higher | Baseline | Bloomberg, Aug. 25, 2026 |
| End-to-end response latency | 1.7x–3.6x lower | Baseline | Bloomberg, Aug. 25, 2026 |
| Estimated inference cost per token | ~50% lower (Broadcom claim) | Baseline | Yahoo Finance, June 2026 |
| Chip category | Custom ASIC, inference-only | General-purpose GPU (training + inference) | CNBC |
| Time from design start to samples | ~9 months | Multi-year design cycle | Yahoo Finance |
| Manufacturing/systems partners | Broadcom (design, Tomahawk networking), Celestica (systems) | TSMC-fabbed, Nvidia-designed, Nvidia reference systems | Yahoo Finance |
| Deployment status, August 2026 | Prototype scale, gigawatt deployment targeted for late 2026 | Already in high-volume production | CNBC, Yahoo Finance |
Read the table with one caveat in mind. OpenAI chose the workloads it benchmarked, and for its largest public model tested, the reported peak performance-per-watt advantage sat at the lower end of that 1.5x–1.9x range rather than the top, based on OpenAI’s own published methodology notes. Vendor benchmarks always flatter the vendor. That does not make the numbers false, but it does mean the real-world gap will only be settled once Jalapeño runs production traffic at scale, something that will not happen in earnest until the gigawatt-scale rollout OpenAI has targeted for late 2026.
Nvidia’s Response: A Record Quarter, and a Muted Stock Move
If Jalapeño was meant to spook Nvidia’s shareholders, it did not work on the day it mattered most. Nvidia’s Q2 FY2027 earnings, released August 26, 2026, showed data center revenue of $89.0 billion, up 117% year over year and 18% sequentially, beating Wall Street’s $85.7 billion estimate, per CNBC’s earnings coverage. Gross margin came in at 75.0%, up from 72.4% a year earlier. Nvidia shares actually rose 7.56% that day, a gain of $15.85 per share, while Broadcom climbed 3.37%, or $11.98, according to the same CNBC report. Both stocks went up together, which tells you the market read this less as a zero-sum fight and more as confirmation that AI infrastructure spending keeps expanding fast enough to feed everyone.
Analyst sentiment held up too. Needham raised its Nvidia price target to $300 from $270 on August 27 and kept a buy rating, according to MarketBeat’s summary of the note. CNBC’s Jim Cramer, watching the same headlines, argued he sees no real competitor to Nvidia yet, even with Altman touting Jalapeño’s numbers in the same week. Futurum Group CEO Daniel Newman struck a similar note, crediting Broadcom’s engineering for Jalapeño’s results while cautioning against calling it an “Nvidia killer.” The split reaction, cheer the chip, keep buying the stock, is the clearest signal yet that Wall Street sees this as a multi-year erosion story rather than a single event that changes 2026 numbers.
Why Nvidia’s 75% Margin Is the Actual Target
Strip away the chip specs and the fight is about one number: Nvidia’s gross margin, which has sat between 72% and 76% for the past six quarters. That margin funds Nvidia’s R&D pace, its software moat around CUDA, and its ability to outspend rivals on next-generation packaging and memory supply. The Motley Fool put it plainly back in June, writing that a credible custom-silicon threat “could pressure a critical element of the bull case: Nvidia’s pricing power, and the gross margin of about 75% it protects,” in its analysis of the OpenAI-Broadcom deal.
The math behind that concern is straightforward. Every dollar an OpenAI, Google, or Amazon spends on its own inference silicon is a dollar that does not flow through Nvidia’s data center segment at a 75% margin. Nvidia’s own revenue keeps growing because the total AI compute market is expanding faster than any single company can capture, but the company’s share of that market has become the subject of genuine analyst disagreement rather than consensus. Estimates for Nvidia’s AI accelerator market share in 2026 now range from roughly 70% up to 88% depending on methodology, according to a spread of research cited by Tom’s Hardware, which tracks custom ASIC programs across the industry. That is a wide enough range to matter: a five-to-ten point share shift on a market this size is tens of billions of dollars in annual revenue changing hands.
Tom’s Hardware’s reporting also flags a trend line that predates Jalapeño: custom AI chip shipments from cloud providers are set to grow at roughly triple the rate of merchant GPU shipments in 2026, a forecast the outlet attributes to TrendForce research and comments from Alchip Technologies chairman Johnny Shen. Jalapeño did not start that shift. It is the highest-profile confirmation yet that it is real, coming from the company that buys more Nvidia GPUs than almost anyone on the planet.
The Custom Silicon Club: Jalapeño Joins a Crowded Field
OpenAI is late to this game, not early. Google’s seventh-generation TPU, code-named Ironwood, reached general availability in 2026 with 4,614 FP8 TFLOPS per chip, 192GB of HBM3e memory, and pods that scale to 9,216 liquid-cooled chips for roughly 42.5 exaFLOPS of combined compute. Anthropic became the anchor tenant for that hardware, committing to access up to one million TPU v7 chips and more than a gigawatt of capacity coming online through 2026, according to industry tracking from SemiAnalysis. That is a direct hit on Nvidia from a company, Anthropic, that competes head-on with OpenAI for the same enterprise customers.
Amazon’s answer is Trainium3, its first 3-nanometer AI chip, built on TSMC’s process and shipping since early 2026. AWS says Trainium3 delivers 30% to 40% better price-performance than its Trainium2 predecessor, which itself claimed roughly 30% better price-performance than comparable GPUs. The chip is reportedly nearly sold out. OpenAI is a customer here too: the company expanded its AWS compute agreement by $100 billion over eight years in an announcement AWS made in early March 2026, committing to roughly 2 gigawatts of Trainium capacity spanning both Trainium3 and the upcoming Trainium4. Uber separately expanded its own AWS contract to run workloads on Trainium3 and Graviton4 processors. Microsoft’s Maia 200 and 300 chips round out the field, running internal Azure AI workloads and giving Microsoft the same insulation against Nvidia pricing that Google and Amazon are building for themselves.
What makes Jalapeño different is who is building it. Google, Amazon, and Microsoft are hyperscalers with decades of chip-design experience and captive cloud businesses to absorb custom silicon internally. OpenAI has no cloud business, no chip design legacy, and only nine months of development time behind its first product. That it managed a shippable design at all, in a market Google needed years to enter credibly, says as much about Broadcom’s ASIC design expertise as it does about OpenAI’s ambitions. For deeper context on how Nvidia’s next-generation Rubin platform stacks up against these same rivals, see our comparison of Nvidia Rubin, AMD Helios, and Microsoft Maia 300.
Table: The Hyperscaler Custom Silicon Landscape in 2026
| Company | Chip | 2026 Status | Key Spec | Primary Use |
|---|---|---|---|---|
| OpenAI / Broadcom | Jalapeño | Samples tested; gigawatt-scale deployment targeted late 2026 | 1.5x–1.9x perf/watt vs. GB300 | Inference only |
| TPU v7 “Ironwood” | Generally available | 4,614 FP8 TFLOPS/chip, 192GB HBM3e, pods to 9,216 chips | Training + inference (Anthropic: ~1M chips committed) | |
| Amazon | Trainium3 | Shipping since early 2026, nearly sold out | 3nm (TSMC), 128–144GB HBM3e | Training (OpenAI: ~2GW committed) |
| Microsoft | Maia 200 / 300 | Running internal Azure workloads | Specs not fully public | Internal Azure AI inference and training |
| Nvidia (reference) | Blackwell (GB300), Rubin upcoming | Volume production; ~70%–88% share estimates | Full-stack GPU + CUDA software ecosystem | Merchant training + inference, all customers |
Historical Context: Nvidia Has Faced This Kind of Threat Before
Custom silicon challenging a dominant incumbent is not a new plot. Apple spent years quietly designing its own Arm-based chips before cutting Intel out of the Mac lineup entirely in 2020, a move that looked risky until Apple Silicon out-performed Intel’s laptop chips on both speed and battery life within two product cycles. Google built its first TPU back in 2015 specifically because running its own search and translation models on GPUs was becoming too expensive at Google’s scale, a decision that took nearly a decade to mature into the Ironwood generation now anchoring Anthropic’s compute. Amazon’s Inferentia and Trainium lines followed the same arc: years of unglamorous iteration before AWS could credibly tell customers to skip Nvidia for specific workloads.
The pattern in every case is the same. The incumbent’s near-term numbers stay strong for years after the first credible alternative ships, because switching an entire AI infrastructure stack away from CUDA, Nvidia’s software layer, carries real engineering cost. Nvidia’s moat was never just the silicon. It was the tooling built on top of it over fifteen years. Jalapeño runs OpenAI’s own models on OpenAI’s own infrastructure, which sidesteps the CUDA lock-in problem entirely, the same trick Google and Amazon used for their internal workloads. That is precisely why this threat is more structural than a single benchmark headline suggests, even while Nvidia’s quarterly results keep beating estimates.
Market Impact: What Wall Street Is Actually Pricing In
Broadcom is the clearest near-term winner regardless of how the Nvidia-versus-custom-silicon fight ultimately resolves. The company gets paid for designing Jalapeño and every future generation of it, plus its Tomahawk networking gear ships into the same racks. Broadcom stock has traded below its June all-time high near $495 even after the Jalapeño news, but Wall Street’s consensus remains a strong buy with an average target above $500, and RBC raised its own target to $400, reflecting confidence in the custom-ASIC business regardless of which hyperscaler wins the underlying AI race.
For Nvidia, the picture splits by time horizon. Short term, demand for Blackwell and the upcoming Rubin platform remains high enough that Nvidia is reportedly sold out of capacity well into 2027, and its own Q2 FY2027 numbers back that up. Longer term, an AI Margin Pressure Score compiled by independent research (cited across several analyst notes) rates Nvidia’s structural risk from custom silicon as significant over a three-to-five-year horizon, even while near-term cycle strength masks it. Translation: the stock is not being punished today, but the assumption baked into current valuations, that Nvidia holds roughly 74% to 75% gross margins for years to come, is exactly the assumption Jalapeño, Trainium3, and TPU v7 are all built to attack.
Competitive Comparison: How the Chip War Lines Up Today
Line the four programs up and a pattern emerges. Google has the most mature technology and the deepest bench, proven by the sheer scale of Anthropic’s TPU commitment. Amazon has the most aggressive sales motion, converting AWS’s existing cloud relationships (Uber, OpenAI itself) into Trainium customers even while competing with those same customers elsewhere. Microsoft has kept its cards closest to its chest, using Maia almost entirely for internal Azure workloads rather than chasing external validation. OpenAI has the least experience but the most direct incentive: no other company on this list burns as much money on inference relative to revenue, which is exactly why a nine-month design cycle made sense for a company that could not afford to wait years for a mature chip program.
Nvidia’s answer to all four threats is the same: keep shipping a faster chip every year, keep CUDA’s software advantage wide enough that switching costs stay painful, and keep pricing high enough to fund both. That strategy worked against AMD and Intel for a decade because neither ever matched Nvidia’s software ecosystem. It is a different fight against four companies that do not need to sell chips to third parties. They only need their own workloads to run cheaper than renting Nvidia GPUs, a much lower bar to clear. Our recent look at Nvidia Rubin’s chip lineup against Google’s TPU roadmap covers how Nvidia is trying to out-engineer that math with its next architecture.
What Analysts and Insiders Are Actually Saying
The public commentary this week ranges from bullish on OpenAI to dismissive of the whole premise. Richard Ho’s characterization of the benchmarks as “a very, very significant performance advance over state of the art” is about as strong a claim as a company can make about its own first-generation chip, and it came directly from a press call reported by TechCrunch. Broadcom CEO Hock Tan was equally direct when he told Reuters that the chip “is as good as the Blackwell chips made by Nvidia or the tensor processing units designed by Alphabet’s Google,” putting Jalapeño in the same sentence as the two most established alternatives on day one.
Sam Altman’s own framing, that “Jalapeño has surpassed our expectations and delivered both deep savings and excellent performance,” leans into the cost story more than the raw-speed story, which tracks with what OpenAI actually needs: cheaper inference at massive scale, not necessarily the fastest chip on a synthetic benchmark. OpenAI president Greg Brockman’s comment that the goal is to “serve more intelligence with greater efficiency” reinforces that the company sees this as an efficiency play first, a competitive threat to Nvidia second. Whether Wall Street agrees with that framing, or reads it as the opening move in a much bigger break from Nvidia, will not be settled by a press release. It will be settled by how many gigawatts of Jalapeño actually go live in 2027 and what OpenAI’s own infrastructure costs look like a year from now.
Predictions: Where This Goes Through 2027
- Nvidia’s data center gross margin drifts lower, not off a cliff. Expect the 75% figure to compress toward the high 60s or low 70s by 2027 as inference workloads shift to custom ASICs, while training work stays largely on Nvidia GPUs for another cycle.
- Broadcom keeps winning no matter who “wins.” As the design partner behind Jalapeño and a long-time partner on Google’s TPU program, Broadcom’s custom-ASIC revenue grows regardless of which hyperscaler’s chip gains the most share.
- Jalapeño’s second generation adds training capability. OpenAI and Broadcom already describe Jalapeño as the first entry in a multi-generation platform. Expect the next iteration to target at least partial training workloads, not just inference.
- Nvidia responds with pricing flexibility, not just new silicon. Look for more aggressive Blackwell and Rubin pricing for the largest AI labs specifically, a defensive move Nvidia has quietly made before when a major customer threatened to build its own chip.
- The overall AI compute market keeps growing faster than any single vendor’s share shifts. Nvidia’s absolute revenue likely keeps climbing through 2027 even as its percentage share erodes, because total demand for AI infrastructure is still expanding faster than custom silicon can absorb it.
What This Means for Developers and IT Buyers
None of this changes what most engineering teams should do this quarter. Jalapeño runs OpenAI’s own models on OpenAI’s own infrastructure. It is not a chip you can rent, buy, or provision through a cloud console the way you can spin up an Nvidia H100 or H200 instance today. Teams evaluating GPU rental costs for training or fine-tuning should keep comparing providers on today’s hardware. Our RunPod vs. Lambda vs. Vast.ai pricing comparison is still the more relevant read if you are actually provisioning compute this month. What Jalapeño does change is the pricing conversation over the next 12 to 18 months. If OpenAI’s inference costs drop by anything close to the 50% figure Hock Tan cited, expect that savings to show up in ChatGPT Enterprise and API pricing before it shows up anywhere else, simply because OpenAI has the strongest incentive of any company on this list to pass efficiency gains straight to customers and grow usage.
The Bottom Line
Nvidia is not in trouble in 2026. Its data center revenue tripled in two years, its margins are near record highs, and its stock rose on the same day OpenAI’s custom chip made headlines. What changed this week is the credibility of the alternative. Jalapeño is the fourth major hyperscaler-adjacent custom chip program to reach real silicon, real benchmarks, and a real deployment date, joining Google’s TPU, Amazon’s Trainium, and Microsoft’s Maia. Four companies building their own chips does not end Nvidia’s dominance overnight. It does mean the 75% gross margin that has defined Nvidia’s stock story for three years is now a number every one of Nvidia’s biggest customers is actively working to make smaller.
Frequently Asked Questions
What is OpenAI’s Jalapeño chip?
Jalapeño is OpenAI’s first custom AI accelerator, co-designed with Broadcom and unveiled on June 24, 2026. OpenAI calls it an “Intelligence Processor” built specifically to run large language model inference rather than to train models.
Who manufactures the Jalapeño chip?
Broadcom handled the chip’s silicon implementation and contributed its Tomahawk networking technology, while systems integrator Celestica built the surrounding rack infrastructure, according to Yahoo Finance’s reporting on the June 2026 unveiling.
Does Jalapeño actually beat Nvidia’s GPUs?
In OpenAI’s own August 2026 benchmarks against Nvidia’s GB300 Blackwell Ultra systems, Jalapeño delivered 1.5 to 1.9 times more AI work per watt and 1.7 to 3.6 times lower response latency. These figures come from OpenAI and Broadcom, not an independent lab, so real-world performance at scale is still unproven.
When will Jalapeño actually be deployed?
OpenAI and Broadcom have targeted small-scale prototype deployment in late 2026, followed by gigawatt-scale rollout in OpenAI’s data centers by the end of the year, according to CNBC and Yahoo Finance’s coverage of the program’s timeline.
Will Jalapeño replace Nvidia GPUs at OpenAI entirely?
No. Jalapeño is designed for inference workloads only, not training. OpenAI is expected to keep using Nvidia GPUs, along with AWS Trainium3 capacity it has separately committed to, for training its frontier models for the foreseeable future.
How did Nvidia’s stock react to the Jalapeño benchmarks?
Nvidia shares rose 7.56% on August 26, 2026, the day after the Jalapeño benchmark news, driven largely by Nvidia’s own record Q2 FY2027 earnings report released the same day. Broadcom shares rose 3.37% on the same day.
What is Nvidia’s current share of the AI chip market?
Estimates vary by research firm and methodology, ranging from roughly 70% to 88% of AI accelerator revenue in 2026. Most analysts agree the share has started to decline gradually from a 2024 peak near 87%, as custom silicon from Google, Amazon, Microsoft, and now OpenAI takes on a growing share of inference workloads.
How does Jalapeño compare to Google’s TPU and Amazon’s Trainium?
Jalapeño is newer and less proven than either. Google’s TPU v7 “Ironwood” is already generally available with Anthropic committed to up to one million chips, and Amazon’s Trainium3 has been shipping since early 2026 and is reportedly nearly sold out. Jalapeño is still moving from tested samples to its first prototype deployments.
