No, OpenAI has not expanded into snack foods. The company’s new Jalapeno chip, announced Wednesday morning, is a custom inference chip co-developed with Broadcom and designed to help power the company’s growing AI infrastructure. Jalapeño has not yet been deployed at scale, but it is said to be 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 is not unique. OpenAI joins the likes of Google, Meta, and Amazon, which are launching their own custom chips to give them more control over the infrastructure behind their AI services. Rather, this latest announcement confirms that major providers are disrupting the standard supply of off-the-shelf hardware in favor of systems tailored to their workloads.
“AI as an application is so demanding that it is forcing the industry to shift its strategy towards customization and greater integration,” said Alexander Hallowell, senior principal analyst at Omdia.
In an industry like AI, where supply contracts are worth billions of dollars, this shift is remarkable. But the impact extends beyond the financial statements of AI providers. For enterprise customers, the implications will also be greater as they learn more about the economics and future architecture of AI services than from the technical specifications of a single new chip.
Why now is the time to invest in custom silicon
In the electronics industry, this cycle between standardized commercial products and customized application-specific products is so common that it has earned the name “Makimoto’s Wave.” Within the AI processor space, waves are visible from a distance. Hallowell said Omdia’s analysts are working on the basis that this wave has been experienced since 2022.
OpenAI’s Jalapeño project wasn’t all that surprising.
“Specifically, we’ve been aware of the OpenAI/Broadcom project for some time,” Hallowell said. “It was not only a rumor, it was a given, and Hock Tan blurted it out during the Q3 2025 earnings call.”
The expectation that this will happen is coupled with the clear advantages that custom silicon offers and is further amplified by the current market. Although the initial cost is high, the resulting custom chip has several benefits.
Improved performance where it matters
According to Richard Simon, CTO of T-Systems International, Jalapeno is an application-specific integrated circuit (ASIC) that acts as an “AI accelerator” optimized for AI inference requirements. It is intended to support the day-to-day operation of AI applications. For OpenAI, this includes all prompts sent to ChatGPT.
Simons explained the downstream effects of these proprietary silicons: “Cost efficiency per inference token and performance per watt, reduced latency and faster response for application and API calls, quickly improving and enhancing performance for consumer and enterprise customers.”
Significant cost savings
Perhaps most notable for OpenAI, bringing in-house chips makes a huge difference to the company’s bottom line. This is critical at a time when many providers, including OpenAI, have expensive contracts with their suppliers.
“Every time a user requests an OpenAI model, enterprises incur high computational costs,” said Quentin Reul, director of global AI strategy and solutions at expert.ai. “Based on existing contracts and partnerships, most of the funding from model inference flows directly to infrastructure providers such as Microsoft, OCI, AWS, and NVIDIA.”
By developing its own chips and data centers, OpenAI can avoid third-party margins and reduce these operating costs. This reduces the long-term costs of delivering the model and makes the overall business proposition more sustainable.
Harrowell explained: “NVIDIA’s gross margins are 75% to 78%; your margin. If you replace this with the 30% to 35% margins that ASIC outsourcers like Broadcom typically earn, the loss in profitability is cut in half. ”
Reduce power consumption
One of the biggest challenges currently plaguing the AI field is power consumption. The U.S. government and enterprise technology sector are working together to expand data center capacity, but these projects could take years to materialize, leaving AI providers in the dark. This is where custom chips can have a big impact.
“Customization helps manage power consumption, which is the number one cost driver in data center environments,” said Hallowell.
Customized chips are optimized for specific use cases and therefore require less power to achieve the same results. This allows the company to reduce the accelerator’s thermal design power to 700W to 800W without exceeding a kilowatt, and eliminates liquid cooling entirely, Hallowell explained. This will revolutionize the economics of AI and data centers.
Impact on business customers
Most business customers do not directly interact with jalapeño chips. Organizations leverage AI through applications, platforms, and APIs, but the underlying infrastructure remains largely invisible. However, infrastructure decisions being made today are likely to shape the cost, performance, and availability of enterprise AI services for years to come.
Omdia analysts predict that ASICs will start gaining significant market share in 2027. The price difference is large, so you’ll probably get more quantity than money. Simons is optimistic that this will have a positive knock-on effect on AI pricing for customers.
“IT leaders will benefit from all the economies of scale this will bring,” he said. “Inferential (and therefore token) economies will benefit from lower costs per request at scale.”
Then there are the performance benefits. For each optimized deployment within OpenAI’s products, customers will also benefit from OpenAI’s unique cost savings, likely at a cost equal to or comparable to what they are currently paying.
Finally, Reul observed less obvious benefits for enterprise customers from a data security perspective. “By developing our own chips and building dedicated data centers, OpenAI is now able to reduce the risk of data breaches as data is shared across our cloud infrastructure.”
Of course, it is important to note that the finished jalapeño chip has not yet been released for external testing, so there is no independent confirmation of its effectiveness. However, Harrowell points out that OpenAI uses both the same ASIC shop and the same server OEM (Celestica) as Google, which suggests the chips may be very similar. This shines a positive light on Jalapeno, as Google’s TPU is “definitely capable of competing with Blackwell.”
That said, even if the comparison to Nvidia’s Blackwell chips turns out to be inaccurate, it might not even be relevant. Jalapeno is used for inference rather than model training, so the goalposts are different. As Reul said, “The goal is to develop a chip that is more consistent with that architecture.”
That seems to be the case with jalapeños.
