Nvidia just posted one of the biggest quarters in corporate history, and Wall Street barely had time to celebrate before the conversation turned to a different number: four. That is how many of the world’s largest AI buyers now design their own chips instead of relying solely on Nvidia silicon. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft has Maia. And as of late August 2026, OpenAI has Jalapeño, its first in-house inference accelerator built with Broadcom. The addition of OpenAI to that list is the story reshaping how investors, developers and rival chipmakers think about Nvidia’s next decade, even as the company’s own numbers remain enormous.
Nvidia reported $96.2 billion in total revenue and $89.0 billion in data center revenue for its second quarter of fiscal 2027, released August 26, 2026, up 117% year over year, according to CNBC’s coverage of the earnings cycle. Shares traded near $227.98 by market close and market capitalization jumped by roughly $442 billion in a single session, per figures reported by multiple outlets. Those are not the numbers of a company in retreat. But they arrived within days of OpenAI publishing its first Jalapeño benchmark results, and the timing put a spotlight on a trend that has been building quietly since 2023: every major Nvidia customer with enough capital is now also a competitor.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
OpenAI becomes the fourth hyperscaler-grade chip designer
OpenAI unveiled Jalapeño, its first custom silicon project, in partnership with Broadcom in June 2026, and published benchmark results on August 25, 2026. The chip is an inference-only accelerator, meaning it is built to run trained models rather than train them, and OpenAI says it is designed specifically to serve ChatGPT and Codex-style workloads. According to the CNBC report on the announcement, testing results showed what OpenAI called a significant performance advance, with the company stating that Jalapeño “can process more AI workloads per unit of power while also delivering faster responses.” That phrasing matters because performance-per-watt and latency are usually a trade-off in chip design; OpenAI is claiming it improved both at once.
Coverage from scanx.trade on August 26 put numbers to that claim: Jalapeño showed 1.5x to 1.9x higher performance per watt than Nvidia’s GB200 and GB300 systems, alongside 1.7x to 3.6x lower end-to-end latency on the workloads OpenAI tested. A full breakdown of those numbers is available in our earlier report on the Jalapeño benchmark results. Separate reporting from 24/7 Wall St. framed the comparison in raw power terms, noting that a 700-watt Jalapeño part outperformed Nvidia accelerators rated between 1,200 and 1,400 watts on the efficiency metrics OpenAI published. OpenAI has already said it is working on second and third generations of the chip, and plans to begin deploying it inside its own infrastructure by the end of 2026, expanding from there. That roadmap is tied to a broader plan, reported by The Motley Fool, to deploy 10 gigawatts of OpenAI-designed accelerator capacity between the second half of 2026 and the end of 2029.
None of this makes OpenAI a chip company overnight. Crucially, OpenAI itself has been explicit that it is not walking away from Nvidia. As the company put it in a statement reported by Yahoo Finance, “We will continue to widely deploy accelerators from Nvidia and other partners for both training and inference workloads.” Jalapeño covers a narrow slice of OpenAI’s total compute footprint. But it completes a set: the four companies that buy the most AI compute on the planet now all design at least part of their own silicon, and Nvidia has to sell into that reality every quarter.
Where Jalapeño Deployment Stands Now
OpenAI said in June 2026 that first chips were planned for data-center use by the end of 2026, with broader deployment expected in 2027. Jalapeño is not reported as a generally available product: OpenAI plans internal deployment first, with production targeted for the first half of 2027. The current plan is internal use for OpenAI’s compute requirements, so teams should not expect to rent or order the chip soon.
The benchmark disclosures reported performance across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T, which indicates support for both OpenAI and external models. That is a wider workload mix than a chip tuned only to ChatGPT would suggest, though it does not change the internal-first deployment plan. Broadcom is identified as the development partner for Jalapeño, and no acquisition of the project or chip business by Broadcom is confirmed. For readers tracking the competitive picture, the practical signal is whether the rollout ever widens beyond OpenAI’s own fleet, since that is the step that would turn a cost-control project into a product rivals must price against.
Amazon offers a useful yardstick. Its official account says it acquired chip-design company Annapurna Labs in 2015 and has since built four custom-silicon families, and Trainium was already being used by customers by October 2026. Jalapeño sits at the start of a path that Amazon began years earlier.
The four-chip roster now facing off with Nvidia
Each of the four custom chips takes a different approach, and none of them is a drop-in Nvidia replacement. Google’s newest TPU generation, known as Ironwood or TPU7x, is the most mature and the most broadly deployed, built for both training and inference and available to Google Cloud customers. Microsoft’s Maia 200, announced January 26, 2026 and shipped in 2026 with peak compute exceeding 10 PFLOPS FP4, is aimed primarily at inference workloads for Azure and Copilot-scale services. Amazon’s Trainium2, generally available since December 2024, handles both training and inference and is the backbone of Anthropic’s Project Rainier cluster, while AWS’s own Trainium pages now put Trainium3 UltraServers front and center. OpenAI’s Jalapeño is the newest and narrowest of the four, focused purely on inference for now.
| Chip | Company | Role | Memory per chip | Peak compute (low precision) | Power (TDP) |
|---|---|---|---|---|---|
| Ironwood (TPU7x) | Training + inference | 192 GB HBM3e | ~4,614 TFLOPS FP8 | Not disclosed per chip | |
| Maia 200 | Microsoft | Primarily inference | 216 GB HBM3e | >10 PFLOPS FP4 / >5 PFLOPS FP8 | ~750 W |
| Trainium3 | Amazon (AWS) | Training + inference | Not listed | Not listed | Not listed |
| Jalapeño | OpenAI + Broadcom | Inference only | Not publicly disclosed | 1.5x-1.9x perf/watt vs Nvidia GB300 | ~700 W |
The specs alone do not tell the full story. Google’s Ironwood can scale to 9,216 chips in a single liquid-cooled pod, delivering an aggregate 42.5 exaflops of FP8 compute and 1.77 petabytes of shared HBM memory, according to Google’s own technical materials. Microsoft’s Maia 200 clusters can reach up to 6,144 accelerators connected over Azure’s Ethernet fabric. Jalapeño’s cluster-scale figures have not been published, which is itself notable: OpenAI is comfortable publishing chip-level efficiency claims but has not yet detailed how it plans to network thousands of these chips together, the kind of systems-level engineering Nvidia has spent a decade perfecting with technologies like NVLink and InfiniBand.
Why four of Nvidia’s biggest customers all became competitors
The economics are the simplest explanation. Broadcom CEO Hock Tan has said Jalapeño delivers roughly 50% cost savings compared with typical AI GPUs, a figure reported by multiple outlets covering the OpenAI-Broadcom partnership. At the scale these four companies operate, even modest per-chip savings compound into billions of dollars a year once you are running hundreds of thousands or millions of accelerators. Amazon’s Anthropic deal illustrates the scale involved: Anthropic is reportedly running more than 1 million Trainium2 chips as part of Project Rainier, one of the largest AI compute clusters built to date, with an expanded agreement securing up to 5 gigawatts of capacity for training and deploying Claude models.
Supply is the second driver. Nvidia’s GPUs have been supply-constrained for most of the current AI buildout, and companies that can design their own silicon reduce their exposure to that bottleneck, along with Nvidia’s pricing power. A third factor is workload specificity: a chip built only to run a company’s own inference stack can strip out general-purpose flexibility that Nvidia GPUs carry for a broad customer base, trading versatility for efficiency on a narrower set of tasks. None of the four companies has said it is trying to fully replace Nvidia. Each has instead described its custom silicon as a complement, aimed at the workloads where cost and control matter more than raw flexibility, while continuing to buy Nvidia GPUs for training and for workloads that benefit from Nvidia’s mature software ecosystem.
Broadcom’s unusual position: arms dealer to almost everyone
The pressure on Nvidia is not limited to custom silicon. The company has also faced consumer and AI chip pricing scrutiny in 2026 and moved to diversify its own AI software stack through its reported acquisition of Hugging Face. One company sits at the center of nearly all of this activity without being a household name: Broadcom. It co-designed Jalapeño with OpenAI and has worked on custom silicon programs for Google’s TPUs in the past. Broadcom’s business model depends on hyperscalers wanting to design chips rather than buy off-the-shelf GPUs, which puts it in direct tension with Nvidia’s merchant-silicon approach even as both companies rely on the same manufacturing partners, chiefly TSMC. The scale of that business shows in Broadcom’s own results, which reported $16.7B in AI-semiconductor revenue in Q3 fiscal 2026, up 221%. Hock Tan’s public comparisons of Jalapeño to both Nvidia’s Blackwell generation and Google’s TPUs, reported by Reuters, underline how directly Broadcom is positioning itself as the enabler of Nvidia alternatives industry-wide, not just for OpenAI.
Nvidia’s response: confident, but no longer dismissive
Nvidia’s public posture toward custom silicon has shifted from years of framing it as a non-issue to a more measured acknowledgment that the trend is real. In comments made after the Jalapeño benchmark release and reported by financial media covering Nvidia’s earnings cycle, CEO Jensen Huang said he was not taking OpenAI’s chip personally, noting that many custom silicon projects get announced and relatively few reach large-scale production. That framing, that in-house chip programs face a steep path from announcement to deployment at Nvidia’s scale, has been Huang’s consistent line for several years, and Nvidia’s Q2 FY2027 numbers give it real weight: 117% year-over-year growth in data center revenue is not the trajectory of a company losing customers to internal silicon.
Financial commentators covering the reaction have been split. Some framed the Jalapeño benchmark as an uncomfortable moment for Nvidia given the specific efficiency numbers involved, while others, including sentiment reported around CNBC’s market coverage, noted that analysts still treat Nvidia’s newest GPU platforms as the more relevant apples-to-apples comparison rather than the GB300 generation OpenAI benchmarked against. The split reflects a genuine ambiguity: Jalapeño’s numbers are real and independently reported, but they describe one company’s narrow inference workload, not a broad market shift away from Nvidia hardware.
Nvidia by the numbers: still the market’s center of gravity
| Metric | Figure | Period / Source |
|---|---|---|
| Total revenue | $96.2 billion | Q2 FY2027, reported Aug 26, 2026 |
| Data center revenue | $89.0 billion | Q2 FY2027, up 117% YoY |
| Adjusted EPS | $2.22 | Q2 FY2027 |
| Share price (after-hours) | ~$217.88 | Aug 28, 2026 |
| Single-session market cap gain | ~$371.7 billion | Post-earnings, Aug 2026 |
| OpenAI Jalapeño power draw | ~700 W | vs Nvidia GB300 at 1,200-1,400 W |
The scale gap between Nvidia and any single custom chip program remains wide. Nvidia’s $89 billion in quarterly data center revenue dwarfs the disclosed scale of any one hyperscaler’s in-house chip deployment, and every one of the four companies building custom silicon, OpenAI included, has publicly committed to continuing large Nvidia purchases. What has changed is the shape of the negotiation. Four years ago, Nvidia’s biggest customers had limited leverage over pricing and allocation because there was no credible in-house alternative for frontier-scale AI workloads. In 2026, all four of the largest AI compute buyers can point to a working alternative for at least part of their stack, even if none of them can replace Nvidia GPUs entirely yet.
Historical context: a decade-long build-out, not a sudden pivot
Nvidia is not the only chipmaker facing new accelerator competition; startups like Cerebras have also pitched purpose-built AI inference hardware against traditional GPU architectures. Custom AI silicon is not a 2026 invention. Google has been building TPUs since 2015, originally for internal search and translation workloads, long before the current generative AI boom. Amazon introduced its first Inferentia chip in 2019 and its first Trainium chip in 2020, well before ChatGPT existed. Microsoft’s Maia program became public in 2023. What is new in 2026 is that OpenAI, a company that just three years ago was almost entirely a Nvidia customer with no hardware ambitions of its own, has joined that list, and that all four programs have matured enough to post competitive, independently reported benchmark numbers against Nvidia’s current-generation hardware rather than just roadmap slides.
That maturation curve is the real story. Google’s Ironwood, unveiled in April 2025 and rolling out through 2026, represents roughly a decade of TPU iteration. Amazon’s Trainium2 delivers about 4x the performance of the original 2020-era Trainium chip. Microsoft went from no custom AI silicon to Maia 200, its second generation, in about three years. OpenAI compressed a similar arc into roughly 14 months between forming its Broadcom partnership and publishing first-generation benchmark results. Each company’s path was different, but the destination, a credible in-house alternative to at least some Nvidia GPU workloads, is now shared across all four.
What this means for enterprise AI buyers
For companies buying AI compute rather than building it, the practical effect of this shift shows up in pricing and platform choice rather than in any single benchmark. As hyperscalers move more of their own internal inference workloads onto in-house silicon, freed-up Nvidia GPU capacity can, in theory, become more available or better priced for third-party cloud customers, though none of the four companies has published data confirming that effect yet. Enterprises evaluating AI infrastructure now also have a genuine choice question that did not meaningfully exist three years ago: whether to build on Google Cloud TPUs, AWS Trainium instances, Azure’s Maia-backed services, or Nvidia GPU instances across any provider, each with different price-performance characteristics depending on the workload.
The trade-off is portability. Code and models optimized for Nvidia’s CUDA software stack generally do not run unmodified on TPUs, Trainium, Maia, or Jalapeño, each of which has its own software toolchain, though Amazon’s TorchNeuron, introduced on October 19, 2026, now targets that gap for PyTorch code on Trainium. That is one reason analysts continue to expect Nvidia to retain a large share of external, third-party AI workloads even as internal hyperscaler workloads migrate to custom chips: switching an entire software stack off CUDA carries real engineering cost that a raw performance-per-watt number does not capture.
The software moat question
Every custom chip in this comparison faces the same structural challenge: Nvidia’s advantage was never just silicon. CUDA, Nvidia’s software and developer ecosystem, has more than a decade of tooling, libraries, and trained engineers built around it. Google, Amazon, and Microsoft have each built their own competing software layers, JAX and XLA for TPUs, the Neuron SDK for Trainium, and Microsoft’s internal stack for Maia, but none has achieved anything close to CUDA’s third-party developer adoption outside of their own walled-garden cloud platforms. OpenAI’s Jalapeño inherits this same problem in a more acute form, since it does not yet have a public software stack or external customers at all; it is currently a single-tenant chip built to run OpenAI’s own models.
Competitive comparison: custom silicon versus Nvidia’s platform play
Framed side by side, the contrast is less about raw chip specs and more about business models. Nvidia sells a general-purpose, merchant-silicon GPU platform to virtually every AI company on the planet, backed by CUDA and a multi-generation product cadence that now includes the Blackwell and upcoming Rubin architectures. Google, Amazon, Microsoft, and OpenAI are each building narrower, workload-specific silicon primarily for internal use, with Google’s TPU and Amazon’s Trainium going furthest toward external cloud availability. Jalapeño sits at the earliest stage of that spectrum: internal-only, inference-only, single-generation. The gap between “OpenAI has a competitive chip” and “OpenAI is a chip company that competes with Nvidia for external customers” remains wide, and closing it would take years, not one benchmark cycle.
Market impact: pricing power, not market share, is the near-term battleground
Reporting around the Jalapeño benchmark release consistently frames the near-term impact in terms of Nvidia’s pricing leverage rather than an immediate loss of customers. With four of its largest customers now each holding at least a partial in-house alternative, Nvidia has less room to raise prices unilaterally on the workloads those chips can handle, even while overall AI compute demand keeps growing fast enough that Nvidia’s revenue continues climbing. That is consistent with Nvidia’s own Q2 FY2027 results: data center revenue up 117% year over year even as all four major custom-silicon programs were active and, in OpenAI’s case, newly public. Demand growth across the entire AI compute market has so far outpaced the share any one custom chip program can capture, although TrendForce forecasts custom-ASIC shipment growth of 44.6% for 2026, versus 16.1% for merchant GPUs.
Predictions: where the custom silicon race goes from here
Based on the trajectory each company has publicly described, a few outcomes look likely over the next 12 to 24 months. First, expect OpenAI to formally announce a second-generation Jalapeño chip, likely widening its scope beyond pure inference, given the company has already said follow-on generations are in development. Second, expect Nvidia to keep leaning on its Rubin architecture roadmap and CUDA ecosystem lock-in as its primary counter to custom silicon, rather than competing purely on raw performance-per-watt numbers against narrower, single-purpose chips. Third, expect at least one of the four hyperscaler-adjacent custom chip programs to sign a named external customer beyond its own parent company, following the pattern Amazon set with Anthropic’s Trainium2 usage, which would mark a meaningful shift from internal cost-saving tool to genuine commercial rival.
Fourth, expect Nvidia’s data center revenue growth rate to gradually moderate from the current 117% pace as custom silicon absorbs a larger share of inference workloads at the four companies involved, even if absolute revenue keeps rising due to overall market growth. Fifth, expect continued ambiguity in year-over-year comparisons between custom chips and Nvidia GPUs, since companies publishing their own benchmarks, as OpenAI did with Jalapeño, tend to choose comparison points and workloads that favor their own hardware, a pattern likely to continue as Google, Amazon, and Microsoft publish their own competitive figures in response.
What developers and infrastructure teams should watch next
For a broader look at where AI accelerator design is headed across the industry, see our AI chips 2026 hub. For teams making infrastructure decisions this year, the practical signal to track is not any single benchmark but whether Google, Amazon, Microsoft, or OpenAI start offering meaningfully cheaper external access to their custom silicon compared with equivalent Nvidia GPU instances on the same cloud. AWS has already published Trainium2 price-performance claims of 30-40% better economics versus comparable Nvidia-based EC2 instances, which is the clearest existing example of that dynamic playing out commercially rather than just internally. If Jalapeño-class chips or Maia 200 follow with similar external pricing moves, that would mark the point where the OpenAI-Google-Microsoft-Amazon custom silicon trend stops being primarily a cost-control story for those four companies and starts becoming a genuine pricing pressure story for the broader AI infrastructure market Nvidia currently dominates.
Frequently asked questions
Is OpenAI’s Jalapeño chip actually faster than Nvidia GPUs?
On specific efficiency metrics that OpenAI chose to publish, performance per watt and end-to-end latency, Jalapeño outperformed Nvidia’s GB200 and GB300 systems by 1.5x to 1.9x and 1.7x to 3.6x respectively, according to scanx.trade’s coverage of the August 25, 2026 benchmark release. That is not the same as being faster across every workload; the comparison covers inference tasks OpenAI selected.
Is OpenAI going to stop buying Nvidia GPUs?
No. OpenAI has stated it will continue to widely deploy accelerators from Nvidia and other partners for both training and inference workloads, according to its own statement reported by Yahoo Finance. Jalapeño currently covers inference only, a fraction of OpenAI’s total compute needs.
Which companies now have their own custom AI chips instead of relying only on Nvidia?
Google (TPU, currently Ironwood/TPU7x), Amazon (Trainium, currently Trainium3 in its Trn3 UltraServers), Microsoft (Maia), and now OpenAI (Jalapeño, built with Broadcom) all operate custom AI silicon programs as of October 2026.
Did Nvidia’s stock fall after the Jalapeño benchmark news?
No public reporting shows a decline tied specifically to the Jalapeño benchmark. Nvidia’s shares actually rose after its Q2 FY2027 earnings release in the same window, adding an estimated $371.7 billion in market capitalization in a single session, per figures reported by scanx.trade.
Who makes OpenAI’s Jalapeño chip?
Jalapeño was co-designed by OpenAI and Broadcom. Broadcom CEO Hock Tan has publicly compared the chip’s performance to Nvidia’s Blackwell generation and to Google’s TPUs, according to Reuters.
Can I buy or rent access to Jalapeño, Maia, Trainium, or TPU chips today?
Google’s TPUs and Amazon’s Trainium chips are available to external customers through Google Cloud and AWS respectively. Microsoft’s Maia 200 currently powers Azure and Microsoft’s own AI services rather than being sold as a standalone external product. OpenAI’s Jalapeño is not currently offered to external customers; OpenAI plans internal deployment first, with production targeted for the first half of 2027.
Does custom silicon threaten Nvidia’s business in the near term?
Nvidia’s own Q2 FY2027 results, $89.0 billion in data center revenue, up 117% year over year, show no near-term revenue decline. Analysts and Nvidia’s own leadership frame custom silicon as a long-term pricing and margin pressure rather than an immediate threat to current sales.
What is the difference between Jalapeño and Nvidia’s GB300?
GB300 is a general-purpose Nvidia GPU platform sold to a broad customer base for both training and inference. Jalapeño is a narrower, inference-only chip built specifically for OpenAI’s own workloads, currently deployed only internally rather than sold commercially.
