OpenAI’s Jalapeño Puts Nvidia’s AI Chip Dominance Under Pressure

Nvidia’s near-monopoly over advanced artificial intelligence chips is coming under increasing pressure as OpenAI and other major technology companies develop custom semiconductors designed to reduce their reliance on the chip giant, analysts told CNBC.

OpenAI unveiled its first AI chip, Jalapeño, on Tuesday and said initial testing showed “industry-leading speed and efficiency.” Developed in partnership with Broadcom, the chip is designed primarily for AI inference, the process of running trained models to generate responses and perform tasks for users.

The development has become of interest because inference is becoming one of the fastest-growing sources of AI computing demand. As companies deploy AI agents and models to millions of users, the amount of computing required to operate those systems after training can become enormous.

Nvidia remains the dominant supplier of AI accelerators, with its GPUs powering much of the training and inference infrastructure used by leading AI companies and cloud providers. Its CUDA software ecosystem has also created a significant barrier to switching because developers have built years of tools and applications around Nvidia’s architecture.

But the emergence of custom chips threatens to weaken that advantage, particularly among the largest technology companies that have the scale and engineering resources to design their own silicon.

Adrien Sanchez, a technology analyst at Yole Group, said Jalapeño demonstrated that a chip designed by a major cloud or AI company can now compete with Nvidia’s Blackwell-class GPUs on inference efficiency.

“Nvidia still owns the vast majority” of AI compute and retains strong software ecosystem lock-in through CUDA, Sanchez said. But OpenAI’s chip represents a “threat to Nvidia’s inference margins, which is the field growing the most at the moment.”

Nvidia’s biggest vulnerability may not be an immediate loss of its overall AI chip leadership, but pressure on the economics of individual workloads.

OpenAI is one of Nvidia’s largest customers, buying huge volumes of GPUs to train and operate its AI models. If the company can shift a meaningful portion of inference workloads to its own chips, it could reduce its dependence on Nvidia and gain greater control over computing costs.

OpenAI said Jalapeño would enable faster responses, more responsive AI agents and more reliable access as demand increases. The chip is expected to be deployed in OpenAI’s infrastructure by the end of the year, and the company said it is already developing second- and third-generation versions.

That creates the possibility of a broader strategic shift. Instead of relying almost entirely on merchant GPUs, OpenAI could eventually operate a mixed infrastructure in which Nvidia hardware is used for workloads where its flexibility and performance provide the greatest advantage, while custom accelerators handle predictable, high-volume inference tasks.

Alexander Harrowell, senior principal analyst at Omdia, described Jalapeño as an “impressive achievement,” particularly in efficiency.

“In a large-scale deployment, this would save power, cooling, and power distribution infrastructure, and contribute a lot to their unit economics,” he said.

The savings could be substantial at the scale of OpenAI’s infrastructure. AI data centers consume enormous amounts of electricity, and the cost of cooling and distributing that power adds significantly to the cost of operating large clusters. A more efficient accelerator can therefore improve economics even if its headline computing performance is not dramatically higher.

Still, Jalapeño does not mean Nvidia’s dominance is about to disappear.

Fion Chiu, an analyst at TrendForce, said OpenAI’s custom chip could reduce its reliance on Nvidia for inference over time, but Nvidia GPUs are likely to remain important for computationally intensive workloads such as large-scale model training and frontier AI.

Nvidia retains advantages in programmability, performance, its software ecosystem, and the ability to support a broad range of workloads, Chiu said.

Benchmark comparisons also require caution.

Research firm SemiAnalysis tested Jalapeño after visiting OpenAI’s facilities and found that it delivered better performance per watt than Nvidia’s Blackwell architecture in nearly all of the scenarios it tested.

But SemiAnalysis said the comparison was “somewhat incomplete and unfair” because Jalapeño uses newer HBM4 memory. Nvidia’s forthcoming Rubin platform also uses HBM4, making Rubin a more appropriate comparison.

“Jalapeño is really competing against chips like Rubin that also use HBM4,” SemiAnalysis analysts said.

Nvidia’s Rubin systems are already beginning to ship to customers, while OpenAI still has engineering samples of Jalapeño, highlighting another challenge for OpenAI: bringing its custom architecture into large-scale production and deployment.

OpenAI is also far from alone in pursuing custom silicon.

Google has developed its own tensor processing units for AI training and inference, while Meta has committed to deploying custom AI chips using Broadcom technology. Amazon has developed its Trainium family of AI accelerators, with Anthropic committing to more than $100 billion of spending on AWS technology over the next decade, including current and future generations of Trainium.

The shift is becoming notable because hyperscalers account for a substantial share of global AI infrastructure investment.

Omdia expects custom application-specific integrated circuits, or ASICs, to surpass GPUs in unit volume by 2028, although GPUs are likely to remain ahead in revenue because they are substantially more expensive.

“This is the biggest competitive threat to NVIDIA,” Harrowell said, arguing that about half of AI infrastructure capital expenditure comes from hyperscale cloud providers that already have custom-chip programmes or have the resources to develop them.

The economics explain why the hyperscalers are willing to invest heavily in semiconductor design. Building an AI chip requires enormous upfront engineering costs, but at sufficient scale the savings from owning the architecture can outweigh the cost of development.

Custom silicon can also be optimized for a company’s particular models and workloads rather than being designed to serve the broadest possible customer base. That could gradually erode one of Nvidia’s traditional advantages. Nvidia sells general-purpose accelerated computing platforms that can support a wide range of AI workloads. Hyperscalers, by contrast, can design chips around their own software stacks, models and data-center architectures.

The competitive field is widening beyond the major technology companies as well. Startups including Cerebras, SambaNova, D-Matrix, Etched and Fractile are developing specialized AI processors targeting different parts of the AI computing market.

For Nvidia, however, the biggest issue may be the changing relationship with its largest customers.

OpenAI has been one of the biggest single consumers of Nvidia GPUs. If it succeeds in deploying Jalapeño at scale, the company could gain bargaining power over future Nvidia purchases while simultaneously lowering its dependence on the supplier.

Sanchez said Jalapeño “raises the stakes for Nvidia’s largest customer relationship specifically.” That does not necessarily mean OpenAI will stop buying Nvidia GPUs. More likely, the AI industry is moving toward a heterogeneous computing model in which Nvidia GPUs, custom ASICs and other accelerators coexist.

Nvidia’s challenge will be maintaining its technological lead and software advantage while its largest customers increasingly have an economic incentive to develop alternatives. The bigger change is that AI chip competition is shifting from a market dominated by a single merchant-chip supplier toward one in which the largest AI companies increasingly control part of their own semiconductor stack.