Broadcom, OpenAI deal hit as infrastructure costs take center stage

The artificial intelligence race is no longer only about building the biggest models or buying the most powerful chips. It is also becoming a cost-control problem.

That is the message behind OpenAI’s Jalapeño, the custom chip it designed with Broadcom (AVGO). The chip gives Broadcom another role in OpenAI’s artificial intelligence infrastructure, but it also raises a broader question now facing the company and its investors: how much the AI buildout will cost, and which companies can profit from making it more efficient.

Jalapeño is built for inference, the process that allows AI products such as ChatGPT to respond to user requests, which makes its cost different from a one-time training expense. As AI products get used more often, inference becomes a recurring cost, adding expense each time a model answers a query, writes code, or completes a task.

Key numbers behind the AI cost problem

Several numbers help explain why the OpenAI-Broadcom chip story is also a cost story:

• $1 trillion: Projected capital-project spending by major technology companies next year, according to Reuters’ Artificial Intelligencer
• $236 billion: AI-related debt issuance in the first five months of 2026, according to the same Reuters report
• End of 2026: OpenAI’s planned initial deployment timeline for Jalapeño
• $16 billion: Broadcom’s expected AI chip revenue in its current third quarter
• $16.36 billion: Analysts’ estimate for Broadcom’s current-quarter AI chip revenue, according to Visible Alpha
• $100 billion: Broadcom’s long-range sales forecast for AI chips
• More than 21%: Broadcom's stock price drop since June 2, resulting from its earnings report falling short of high Wall Street expectations

OpenAI’s chip shows why AI costs matter

OpenAI and Broadcom said in a company statement that Jalapeño is OpenAI’s first Intelligence Processor and the first chip in a multigeneration compute platform that the companies are building together.

The companies said the chip is built around OpenAI’s view of large language model inference. Broadcom is helping industrialize the platform through chip implementation, board and rack system integration, high-performance networking, and scalable production systems.

Related: Broadcom gets major OpenAI boost in AI chip race

The chip is designed to do more than add computing power. It is meant to make that computing power better suited to OpenAI’s own workloads.

For AI companies, that kind of efficiency can become more useful as products scale. They need infrastructure that can run models repeatedly and reliably as usage grows, without letting operating costs climb too quickly.

Broadcom said early testing shows Jalapeño will deliver substantially better performance per watt than the current state of the art. The company also said the architecture reduces data movement and balances compute, memory, and networking resources.

If OpenAI can run more AI workloads with greater efficiency, it may be able to lower the long-term cost of operating products such as ChatGPT. Broadcom’s opportunity is to become one of the suppliers helping AI labs and major technology companies make that shift.

Broadcom sells into AI’s rising cost problem

Large technology companies are spending heavily on data centers, chips, power, cooling, and networking equipment. The spending has supported demand for AI infrastructure suppliers, but it has also made Wall Street more sensitive to whether those investments will generate enough returns.

More AI:

• Goldman Sachs has blunt message for AI stock investors
• Microsoft CEO sends a blunt warning on AI and the tech ecosystem
• The next AI infrastructure race has nothing to do with chips

Broadcom sits in the middle of that debate. The company can benefit when major AI customers decide that standard chips are not enough and that they need custom silicon built around their own workloads.

Reuters reported earlier in June that the boom in inference has made custom chips important to the AI industry, driving more orders for advanced processors and intensifying competition.

But Broadcom also has to prove that the business can scale profitably. Custom chips may help customers control costs, but that does not mean the same margin profile for Broadcom.

Reuters reported that Broadcom’s profit margin on custom AI chips is not as high as some of the company’s other chips, such as networking switches, because AI chips require large amounts of high-bandwidth memory.

JHVEPhoto &sol Getty Images

The OpenAI deal strengthens Broadcom’s custom AI chip story, but it does not settle the margin question.

The company gets another high-profile AI customer working on a custom chip platform. It also gains more exposure to the segment of the AI market where customers are optimizing for performance, power, and cost.

But investors will still want evidence that those projects can produce large revenue streams without putting too much pressure on margins.

Broadcom already has ambitious AI targets. Reuters reported that Broadcom expected AI chip revenue of $16 billion in its current third quarter, slightly below analysts’ estimates of $16.36 billion, according to Visible Alpha. CEO Hock Tan also stuck with the company’s long-range forecast of $100 billion in sales from AI chips.

That forecast shows the scale of the opportunity. Broadcom’s sharp stock drop earlier this month showed the risk of falling even slightly short of Wall Street’s AI expectations.

Can Broadcom turn AI cost pressure into growth?

OpenAI’s Jalapeño chip gives Broadcom another place in the AI infrastructure buildout. 

However, as AI products scale, what matters is not only winning more custom chip work, but also whether companies can help customers run AI systems more efficiently while still protecting their own revenue and margins. And the harder part is proving that custom AI chips can become a business that is both large and profitable.

For Broadcom, the next AI test is whether selling into the AI cost problem can become a durable growth driver.

Related: Nvidia pours cold water on AI fears