Nvidia built the most valuable chip company on Earth by selling GPUs to the same handful of companies that are now racing to stop buying them. OpenAI, Google, Amazon, Microsoft and Meta collectively account for the bulk of Nvidia’s data-center revenue, and each one is now shipping, testing, or scaling a custom AI chip designed to do the same job a Nvidia GPU does, often for less money. Reports from Stocktwits, CNBC, BigGo Finance, Firstpost, Digitimes and Ad-hoc-news.de this week frame the story the same way: Nvidia’s biggest customers are becoming its biggest rivals, and the “custom silicon race” is no longer a side project. It’s now a structural threat to how Nvidia prices its most important product.
The trigger for the latest round of coverage, first detailed by CNBC’s technology desk, is JalapeÃąo, OpenAI’s first custom AI chip, built in partnership with Broadcom. According to CNBC, OpenAI said testing results for the new inference chip “show a significant performance advance,” and internal benchmarks reported in late August 2026 put JalapeÃąo at 1.5 to 1.9 times more work per watt than Nvidia’s GB200 and GB300 systems, with end-to-end latency running 1.7 to 3.6 times lower. The company is already working on generations two and three of the semiconductor while the first version has not even shipped in production yet. That single data point, more than any other, is why this story is moving markets and headlines simultaneously: the company that arguably needs Nvidia GPUs the most is also the company moving fastest to need them less.
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 Happened: OpenAI’s JalapeÃąo Chip Challenges Nvidia on Inference
OpenAI and Broadcom first revealed JalapeÃąo in June 2026, describing it as OpenAI’s entry into custom artificial intelligence silicon. CNBC’s follow-up reporting on August 25 and August 26 says the chip is aimed squarely at inference, the process of running a trained model to answer prompts and serve outputs to ChatGPT and other OpenAI products, rather than training new models from scratch. Broadcom is manufacturing the chip, and CNBC reports OpenAI plans to deploy it inside its own compute infrastructure by the end of 2026, with expansion in following years as later generations arrive.
Broadcom CEO Hock Tan told CNBC that JalapeÃąo would go through “small prototype development” in late 2026 before scaling up, a cautious timeline that matches how hyperscalers have typically rolled out first-generation custom silicon: quietly, in limited production, with the loud claims saved for generation two or three once the architecture is proven at scale. Stocktwits framed the news bluntly, reporting that “OpenAI’s first custom AI chip is showing an edge over Nvidia Corp.’s Blackwell systems on key inference benchmarks,” and noting that “the bigger threat comes from custom chips from hyperscalers, who are also Nvidia’s biggest customers.” That assessment was echoed by independent semiconductor analysts at Yole Group, who reported in August 2026 that JalapeÃąo had matched or beaten Nvidia’s Blackwell architecture on inference efficiency, lending third-party confirmation to OpenAI’s own benchmark claims.
That framing matters because OpenAI is not a peripheral Nvidia customer. It is one of the largest buyers of Nvidia GPU capacity in the world, tied into multi-year, multi-billion-dollar compute commitments. A chip built in-house, even one limited to inference workloads at first, changes the negotiating dynamic between the two companies regardless of how many Blackwell and future Nvidia GPUs OpenAI continues to buy for training.
The Full Custom Silicon Lineup: Who’s Building What
JalapeÃąo is the newest entrant, but it’s joining a field that has been building for years. Google, Amazon, Microsoft and Meta all run mature custom-silicon programs, and each has shipped or announced a new generation within the last several months. BigGo Finance’s analysis, published this week, groups all of them under one umbrella: hyperscalers “defecting to custom ASICs” to pry open Nvidia’s pricing power amid what it calls an “increasingly severe inference cost crisis.”
Google’s latest silicon is the TPU 8t and TPU 8i pairing, which launched in April 2026 and covers both training and inference workloads. Amazon’s newest accelerator, Trainium3, launched in December 2025 and likewise spans training and inference, building on the Inferentia and Graviton lines AWS has run for years. Meta’s MTIA 300, launched in March 2026, is a training accelerator tuned specifically for the ranking and recommendation models that power its ad and feed products. Microsoft continues to build out its Maia AI accelerator alongside its Arm-based Cobalt CPU for Azure data centers. ByteDance, according to BigGo Finance, is also expanding its own custom ASIC efforts, though specific chip names were not detailed in that report.
Broadcom’s Pivotal Role as the Arms Dealer
Nearly every hyperscaler custom chip in this story shares one supplier in common: Broadcom. The company is not just building OpenAI’s JalapeÃąo, it is also deeply embedded in Google’s TPU program and works with Meta and Anthropic on custom silicon design. Reports describe Broadcom’s custom AI chip business generating $8.4 billion in a single quarter, up 106% year over year, with a backlog of roughly $73 billion and six custom-silicon customers, including Google, Meta, OpenAI (tied to a reported 10-gigawatt compute deal) and Anthropic.
That backlog figure is itself a proxy for how fast this shift is happening. Broadcom doesn’t build speculative chips for hyperscalers on spec, it builds against firm, funded commitments. A $73 billion backlog means these are not experimental side projects; they are production roadmaps with real capital behind them. CNBC has also reported that Broadcom’s position isn’t unchallenged either: Google, its flagship custom-chip customer, has separately struck a deal with Marvell Technology, raising questions about how much of Google’s future custom silicon work stays with Broadcom.
The Numbers: Custom ASICs Are Outgrowing GPUs for the First Time
The most consequential number in this week’s coverage comes from TrendForce data cited by BigGo Finance: custom ASIC shipment growth is projected to hit 44.6% in 2026, compared to 16.1% growth for commercial GPUs. BigGo describes this as the first year custom chip growth has clearly outpaced general-purpose GPU growth, a reversal of the pattern that has held since the generative AI boom began in 2023. Market researcher Omdia projects the trend compounds further out: by 2028, Omdia expects custom ASIC shipment volume to exceed commercial GPU shipment volume outright, not just outgrow it in percentage terms, which would mark a structural shift in how the industry builds AI compute rather than a one-year anomaly.
That doesn’t mean Nvidia is shrinking. Reports describe Nvidia’s market capitalization sitting around the $5 trillion mark in mid-2026, and its GPUs remain the default choice for training frontier models, where raw flexibility and software maturity through CUDA still matter more than cost per inference query. But the growth curve is the story: if custom ASICs keep expanding nearly three times faster than GPUs, Nvidia’s share of the total AI compute market shrinks even while its absolute GPU sales keep climbing.
BigGo Finance describes the likely near-term outcome as a “GPU plus ASIC dual-track” model: Nvidia GPUs continue to dominate training, where general-purpose flexibility matters most, while custom ASICs take an increasing share of inference, where a chip narrowly optimized for one company’s specific models can be cheaper to run at scale. Inference is also the workload growing fastest as AI products move from research demos to daily-use consumer and enterprise tools, which is exactly why hyperscalers are targeting it first with in-house silicon.
Custom AI Chips at a Glance
| Company | Chip | Launch | Primary Workload | Manufacturing Partner |
|---|---|---|---|---|
| OpenAI | JalapeÃąo | Announced June 2026; deployment by end of 2026 | Inference | Broadcom |
| Google (Alphabet) | TPU 8t / TPU 8i | April 2026 | Training + Inference | Broadcom (with Marvell reportedly gaining ground) |
| Amazon (AWS) | Trainium3 | December 2025 | Training + Inference | In-house design |
| Meta | MTIA 300 | March 2026 | Training (ranking/recommendation) | In-house design |
| Microsoft | Maia AI accelerator | Ongoing generational updates | Training + Inference | In-house design |
Why Hyperscalers Are Building Their Own Silicon Now
None of this is happening in a vacuum. Amazon, Alphabet, Microsoft and Meta are on track to spend roughly $725 billion combined on capital expenditures in 2026, up about 77% from the prior year, according to reporting cited by CNBC. A large share of that spending goes directly into data center buildout and, increasingly, into custom chip design and manufacturing capacity. When a company is already spending that kind of money on AI infrastructure, the economics of designing a chip tuned to its own workloads start to look a lot more attractive than paying Nvidia’s margins indefinitely.
Inference cost is the specific pain point driving the shift. Training a model happens once (or periodically, for fine-tuning); inference happens every time a user sends a prompt, which for a company running hundreds of millions of daily AI queries adds up to a massive recurring bill. BigGo Finance describes this as an “increasingly severe inference cost crisis” for the biggest AI platforms, and a custom ASIC narrowly built to run a company’s own model architecture efficiently can cut the per-query cost meaningfully compared to a general-purpose GPU that was designed to handle any workload thrown at it.
There’s also a supply and control argument. Nvidia GPU allocation has been a bottleneck for AI companies for the better part of three years, and building custom silicon gives a hyperscaler a second supply chain that isn’t subject to Nvidia’s production priorities or pricing decisions. Even if a company never fully replaces Nvidia GPUs, having a credible in-house alternative changes the leverage in every future purchasing negotiation.
Historical Context: How We Got Here
Custom AI silicon isn’t new. Google’s TPU line dates back more than a decade, originally built for internal search and ad workloads long before the current generative AI boom. Amazon’s Inferentia and Graviton chips have existed for years as cost-optimization tools inside AWS. What’s changed is scale and intent. In the early TPU and Inferentia era, these chips were mostly used to trim internal cloud costs at the margins. Today, they are explicitly positioned as alternatives capable of running frontier-scale AI workloads that would otherwise require Nvidia GPUs.
An earlier analysis from May 2026, cited alongside this week’s coverage, argued that evidence had been “piling up” that Nvidia’s AI chip dominance may be coming under sustained pressure as its largest customers design chips for their own data centers. That prediction reads differently now that OpenAI, historically viewed as one of Nvidia’s most loyal and dependent customers because of the sheer scale of its compute needs, has joined Google, Amazon, Microsoft and Meta on the list of companies building their own silicon.
Nvidia’s traditional chip rivals, AMD, Broadcom (in its non-custom lines) and Intel, have never posed the kind of threat hyperscaler custom silicon now represents, largely because none of them controlled both the demand and the chip design simultaneously the way Google, Amazon, Microsoft, Meta and now OpenAI do. A traditional competitor has to win Nvidia’s customers over one deal at a time. A hyperscaler building its own chip simply routes its own workloads away from Nvidia by decision, not by sales pitch.
Market Impact: What This Means for Nvidia’s Business
Nvidia’s near-term financial position remains strong by any conventional measure, with reports putting its market capitalization around $5 trillion in mid-2026. Training workloads, where Nvidia’s CUDA software ecosystem and general-purpose flexibility remain hardest to replace, are still overwhelmingly running on Nvidia GPUs across the industry. The company’s Blackwell architecture is still the benchmark against which OpenAI is measuring JalapeÃąo’s early inference results, not the other way around.
The pressure point is inference, and specifically inference at hyperscaler-plus-frontier-lab scale. If TrendForce’s 44.6% versus 16.1% growth split holds through 2026 and into 2027, an increasing share of new AI compute deployment goes to custom ASICs rather than commercial GPUs, even as total spending on AI infrastructure keeps rising. That’s a share-of-growth problem more than an absolute-revenue problem for Nvidia in the short term, but share-of-growth problems compound. Every inference workload that moves to an in-house chip is a workload that doesn’t need a new Nvidia GPU purchase next refresh cycle.
There’s also a pricing power dimension. Even hyperscalers that continue buying large volumes of Nvidia GPUs gain negotiating leverage simply by having a credible custom alternative in production. BigGo Finance’s framing, that hyperscalers are moving to “defect” from Nvidia specifically to pry open its pricing power, suggests the near-term financial impact may show up less in unit sales and more in gross margin pressure on future GPU generations, as Nvidia’s biggest buyers use their own silicon roadmaps as leverage in contract talks.
Competitive Comparison: Custom Silicon vs. Nvidia GPUs
| Factor | Nvidia GPUs (Blackwell and successors) | Hyperscaler Custom ASICs |
|---|---|---|
| Primary strength | General-purpose flexibility, mature CUDA software stack | Cost efficiency for a single company’s specific model architecture |
| Best-fit workload | Training frontier models, mixed/varied workloads | High-volume inference at massive, repetitive scale |
| Supply chain | Nvidia-controlled allocation, third-party priced | Owned or co-designed, insulated from Nvidia allocation decisions |
| 2026 shipment growth (TrendForce, cited by BigGo Finance) | 16.1% | 44.6% |
| Market maturity | Dominant incumbent since the generative AI boom began | Rapidly scaling, several programs now in second or third generation |
What Nvidia Can Still Do About It
Nvidia is not standing still while its customers build competing chips. The company continues to push its own roadmap forward, with its next-generation Rubin architecture positioned to compete directly against the wave of custom silicon detailed above, including Google’s TPU line and Microsoft’s Maia accelerators. Nvidia’s argument to hyperscalers has generally been that even a highly optimized in-house ASIC can’t match the pace of Nvidia’s annual architecture refresh cycle or the breadth of its software ecosystem, which spans everything from CUDA libraries to networking (via its Mellanox-derived interconnects) to full-rack systems. Nvidia has also moved to co-opt the custom-silicon wave rather than simply out-compete it: in August 2026, the company opened its NVLink Fusion interconnect to third-party custom XPUs, including MediaTek-designed accelerators, letting hyperscalers plug their own chips into Nvidia’s NVLink-C2C links, HBM memory, chiplets and MGX rack architecture. That move effectively offers customers building in-house silicon a way to stay inside Nvidia’s rack ecosystem even as they reduce their GPU purchases.
That argument still resonates for training workloads, which explains why every hyperscaler building custom inference silicon is simultaneously placing enormous GPU orders for training capacity. None of the companies named in this story, OpenAI included, are walking away from Nvidia. They are diversifying away from total dependence on Nvidia for one specific, high-volume, high-cost category of workload: inference. Nvidia’s own data center business page continues to promote its full-stack approach as the reason customers stay, even as those same customers build parallel supply chains.
Predictions: Where the Custom Silicon Race Goes Next
- Inference workloads keep migrating first. Expect the split between Nvidia-dominated training and increasingly ASIC-served inference to sharpen through 2027, following the pattern BigGo Finance calls the “GPU plus ASIC dual-track.”
- OpenAI’s later JalapeÃąo generations matter more than the first. CNBC reports OpenAI is already working on generations two and three; the real test of whether JalapeÃąo becomes a meaningful Nvidia alternative, rather than a cost-saving side project, will show up in those later versions once they scale past prototype volumes.
- Broadcom’s custom-silicon dominance faces its own competition. With Google reportedly working with Marvell alongside Broadcom, expect other hyperscalers to hedge their custom-chip manufacturing partnerships the same way they’re hedging their Nvidia dependence.
- Nvidia’s pricing, not its unit sales, absorbs the first wave of pressure. Expect Nvidia’s biggest customers to use their in-house chip roadmaps primarily as negotiating leverage in the near term, rather than as outright GPU replacements, given how dominant Nvidia remains in training.
- Custom ASIC shipment growth likely stays ahead of GPU growth into 2027. If the capex trend holds, with hyperscalers spending roughly $725 billion in 2026 alone, expect the 44.6% versus 16.1% growth gap TrendForce identified to persist or widen rather than close.
What This Means for Developers and Enterprise Buyers
For engineers building on top of these platforms, the practical impact shows up gradually rather than overnight. If an application calls OpenAI’s API today, the underlying inference could increasingly run on JalapeÃąo-based infrastructure rather than Nvidia GPUs once deployment scales past 2026, with the transition largely invisible at the API layer. The same is already true for many AWS Bedrock and Google Vertex AI workloads, which can route to Trainium or TPU-backed instances depending on configuration and pricing tier.
Enterprise buyers evaluating cloud AI contracts should expect more granular, chip-specific pricing tiers going forward, since running the same model on custom silicon versus Nvidia GPU instances now carries meaningfully different cost structures for the providers themselves. That cost difference is likely to filter down into customer-facing pricing over time, particularly for high-volume inference use cases like customer service automation or code generation, where query volume rather than model sophistication drives most of the compute bill.
Related Coverage
For more on the companies and chips shaping this story, see our coverage of the Nvidia Rubin vs. AMD Helios vs. Microsoft Maia 300 AI chip race, our breakdown of Nvidia Rubin’s chips against Google’s TPU 8i and 8t, and our report on Nvidia’s reported $12.9B Hugging Face acquisition. Also read about the Nemotron 3.5 model and Nvidia’s $89B GPU demand bet, and our look at Cerebras’ CS-4 accelerator, which claims 30x faster AI inference than GPUs.
Frequently Asked Questions
What is OpenAI’s JalapeÃąo chip?
JalapeÃąo is OpenAI’s first custom AI chip, built in partnership with Broadcom and announced in June 2026. It’s designed for inference, meaning it runs trained models to serve responses in products like ChatGPT, rather than training new models. CNBC reports OpenAI plans to deploy it within its own infrastructure by the end of 2026.
Is OpenAI dropping Nvidia GPUs entirely?
No. Reporting indicates OpenAI continues to rely heavily on Nvidia GPUs for training, and JalapeÃąo is initially targeted at inference workloads only. The chip is described as reducing OpenAI’s reliance on Nvidia for some workloads, not replacing Nvidia hardware across the board.
Which companies have custom AI chips besides OpenAI?
Google (TPU 8t/8i), Amazon (Trainium3, Inferentia, Graviton), Meta (MTIA 300) and Microsoft (Maia, Cobalt) all run active custom silicon programs, alongside reported efforts from ByteDance.
How fast are custom AI chips growing compared to Nvidia GPUs?
TrendForce data cited by BigGo Finance projects custom ASIC shipment growth of 44.6% in 2026, compared to 16.1% growth for commercial GPUs, marking the first year custom chip growth has outpaced general-purpose GPU growth.
What role does Broadcom play in the custom silicon race?
Broadcom manufactures custom AI chips for several hyperscalers, including OpenAI’s JalapeÃąo and work tied to Google’s TPU program. Reports describe Broadcom’s custom AI chip business generating $8.4 billion in a single quarter, up 106% year over year, with roughly $73 billion in backlog across six custom-silicon customers.
Does this threaten Nvidia’s stock price?
Reports do not cite a specific Nvidia share price move tied directly to the JalapeÃąo news. Nvidia’s market capitalization is reported around $5 trillion in mid-2026, and the company remains dominant in training workloads. The competitive pressure from custom silicon is concentrated in inference, not training, at least for now.
Why are hyperscalers building their own chips instead of just buying more Nvidia GPUs?
Cost and supply control are the two main drivers. Custom ASICs tuned to a company’s specific model architecture can lower per-query inference costs at scale, and owning chip design gives hyperscalers a second supply chain independent of Nvidia’s production allocation and pricing decisions.
What comes next in this story?
Watch for JalapeÃąo’s second and third generations, expanded deployment volumes at OpenAI, and further movement in Broadcom’s hyperscaler customer relationships, including Google’s reported work with Marvell as an additional custom-chip manufacturing partner.
