No, OpenAI hasn't expanded into snack foods. The company's new Jalapeño chip, unveiled Wednesday morning, is a custom inference chip developed with Broadcom and designed to help power its growing AI infrastructure. Although Jalapeño has yet to be deployed at scale, it has been described as comparable to Nvidia's coveted Blackwell chips and Alphabet's tensor processing units — at least, according to Broadcom CEO Hock Tan.
The move to custom silicon isn't unique. OpenAI joins the likes of Google, Meta and Amazon, who have all launched their own custom chips as they seek greater control over the infrastructure behind their AI services. Rather, this latest announcement is confirmation that major providers are disrupting the standard supply of off-the-shelf hardware in favor of systems tailored to their own workloads.
"AI as an application has been so demanding that it's forced the industry to switch strategy to customization and higher levels of integration," said Alexander Harrowell, senior principal analyst at Omdia.
Related:What Apple's AI update reveals about the future of build vs. buy
In an industry like AI, where supply deals are valued in the billions, this pivot is notable. But the ramifications ripple beyond the AI providers' financial statements. For the enterprise customer, there is also impact — less from the technical specifications of a single new chip and more from what it reveals about the economics and future architecture of AI services.
This cycle in the electronics industry between standardized, merchant products and customized, application-specific ones is so common it has a name: Makimoto's Wave. Within the AI processor space, the wave has also been visible from afar; Harrowell said Omdia analysts have been working on the basis that we have been experiencing the wave since 2022.
OpenAI's Jalapeño project was even less of a surprise.
"Specifically, we've been aware of an OpenAI/Broadcom project for some time," Harrowell said. "Not only has it been in the rumor mill, but it was also an obvious thing to happen — and then Hock Tan blurted it out on the 3Q 2025 earnings call."
The expectation that this will happen is tied to the clear advantages custom silicon offers, which are only amplified by the current market. While the initial outlay is significant, the resulting custom chip offers several benefits.
Jalapeno is an application-specific integrated circuit (ASIC), meaning it functions as an "AI accelerator" optimized for AI inference requirements, said Richard Simon, CTO at T-Systems International. It is intended to support the day-to-day operation of AI applications — which in OpenAI's case will include every prompt sent to ChatGPT.
Related:Why bank AI projects stall at approval
Simons described the downstream effects of such proprietary silicon as: "Cost efficiency per inference token and better performance per watt, reduced latency and faster responses for applications and API calls, and rapid improvement and enhanced performance for consumer and enterprise customers."
Perhaps most notably for OpenAI, the introduction of in-house chips makes a big difference to the company's bottom line. This is critically important at a time when many providers — including OpenAI — have expensive contracts with their own suppliers.
"Every time a user prompts an OpenAI model, the company incurs high computational cost," said Quentin Reul, director of global AI strategy and solutions at expert.ai. "Based on its existing agreements and partnerships, most of the money generated from model inference is flowing directly to infrastructure providers such as Microsoft, OCI, AWS and NVIDIA."
By developing its own chips and data centers, OpenAI can reduce these operational costs through bypassing third-party margins. This lowers the long-term cost of serving their models, making the entire business proposition more sustainable.
Related:The agentic shift at the Snowflake Summit: Finding a platform's 'right to win'
As Harrowell explained: "NVIDIA's gross margin is between 75% and 78%, and all of that comes out of your margin. If you replace that with the 30%-35% margin an ASIC outsourcer like Broadcom usually gets, you've halved the drain on your profitability."
One of the biggest challenges currently plaguing the AI sector is power consumption. While the U.S. government and the enterprise technology sector are working in tandem to expand data center capacity, these projects could take years to come to fruition, leaving AI providers in the dark. This is where a custom chip can have an outsized impact.
"Customizing helps manage the power draw, which is the biggest driver of costs in a data center environment," Harrowell said.
Customized chips require less power to achieve the same results, since they are optimized for their specific use case. This enables the company to keep its accelerator's thermal design power to 700W-800W, rather than pushing over the kilowatt, which allows it to skip liquid cooling altogether, Harrowell explained. This substantially changes the economics of AI and the data center.
Most enterprise customers will never interact directly with a Jalapeño chip. Organizations consume AI through applications, platforms and APIs, while the underlying infrastructure remains largely invisible. Yet the infrastructure decisions being made today could shape the cost, performance and availability of enterprise AI services for years to come.
At Omdia, analysts are forecasting that ASICs will start taking substantial market share in 2027, probably much more in volume rather than value as the price gap is large. Simons is optimistic that this will have positive knock-on effects for customer AI pricing.
"IT leaders will benefit from the full spectrum of economies of scale that this will usher in," he said. "Inference (and thus, Token) Economics will benefit from reduced cost-per-request, at scale."
Then there's the performance benefits. For every optimized deployment within OpenAI's products, customers will reap those rewards too, possibly at a similar or equal cost to what they're paying today due to OpenAI's own cost savings.
Finally, Reul observed a less obvious benefit for enterprise customers, in terms of data security: "By developing its own chip and building dedicated data centers, OpenAI can now reduce the risk of data leakage as data is shared across cloud infrastructure," he said.
Of course, it's important to note that the finished Jalapeño chip has not yet been released for external testing, so there has been no independent corroboration of its efficacy. However, Harrowell noted that OpenAI is using both the same ASIC shop and the same server OEM (Celestica) as Google, which suggests that the chip might be quite similar. Since Google's TPUs are "definitely competitive with the Blackwells," this casts the Jalapeño in a favorable light.
That said — even if comparisons to Nvidia's Blackwell chip turn out not to be accurate, they might not even be relevant. Since Jalapeño is not being used for model training but for inference, the goalposts are different. As Reul said, "the goal is to develop chips that are better aligned with its architecture."
With Jalapeño, it seems to have done that.
Senior Editor, InformationWeek
Madeleine Streets is a senior editor at InformationWeek, where she shapes stories and contributes news analysis through a CIO lens.
She comes to InformationWeek from TechTarget’s Learning Content team, in which she authored explainers and features on a range of enterprise IT topics. Before moving to the field of enterprise technology, Madeleine spent several years covering retail, consumer finance, and ecommerce technology for fashion trade publication Footwear News. She has also been published in Women’s Wear Daily, TIME, Associated Press, SELF, and Observer, among others. The thread that ties her coverage together is a commitment to honest, impactful storytelling -- and insatiable curiosity.
Outside of writing, Madeleine can be found studying wine, singing in her local choir, and working her way towards her annual reading goal of 100 books. She is based in New York City, US.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。