Broadcom’s $230 Billion AI Outlook: Power, Not Chips, Is the Bottleneck

Broadcom CEO Hock Tan’s message at the Goldman Sachs Communacopia + Technology Conference was unusually direct: AI demand is strong enough to support a dramatic multiyear revenue trajectory, but the physical world may determine how quickly that demand turns into deployed systems.

Broadcom expects roughly $58 billion in AI revenue in fiscal 2026, about $115 billion in fiscal 2027, and has said it has line of sight to about $230 billion in fiscal 2028. 17 Those figures imply a near-doubling in each of the next two fiscal years—but they are company projections, not assured results.

The constraint, in Tan’s view, is infrastructure

Tan described the outlook as supply-constrained rather than demand-constrained. The relevant supply issue is not simply whether Broadcom can manufacture semiconductors: it is whether customers can secure powered data-center sites and complete the construction needed to deploy AI systems. Reporting on the conference cited power availability, suitable sites and construction schedules as the key limiting factors. 318

That distinction matters. A customer can have demand for an accelerator and a vendor can have chip supply available, yet revenue can still be delayed if the data center that will run the hardware is not ready. In this framing, electricity and deployment capacity become the gating inputs for the next phase of AI infrastructure growth.

Custom XPUs are central to Broadcom’s AI plan

Broadcom is pursuing the opportunity through custom XPU accelerators developed closely with a concentrated group of frontier-model companies. Tan said the company works with six such developers, including OpenAI, Google and Anthropic. 17

The approach is different from selling a broadly standardized processor to a wide market. Custom silicon can be tailored to a model developer’s workloads, systems and deployment plans, making the commercial relationship deeper—but also making Broadcom’s growth more dependent on the execution and expansion plans of a small number of very large customers.

OpenAI’s Jalapeño illustrates the co-development model

Tan pointed to Broadcom’s work with OpenAI as an example of the pace he believes is possible. He said the companies created the custom XPU called Jalapeño in roughly a year, with a sample delivered within nine months of the initial agreement, according to conference reporting. 317

He also said Broadcom was working with OpenAI on the next two generations after Jalapeño. 19 Earlier company commentary said OpenAI was planned to deploy 1.3 gigawatts of Jalapeño in 2027 and that Broadcom had line of sight to more than 5 gigawatts of Jalapeño and successor XPUs in 2028. 1

Those deployment ambitions reinforce Tan’s core argument: the opportunity is not just a matter of selling a chip. It requires the power, networking, buildings and construction timelines to support systems at gigawatt scale.

Google remains a long-term custom-silicon partner

Tan also highlighted Broadcom’s longstanding work on Google’s TPU program. Reporting from the conference said the companies have an agreement covering TPUs and networking products through 2031. 19

Together with the OpenAI work and the broader six-customer group, the relationship points to Broadcom’s bet that major AI developers will continue to invest in purpose-built compute and networking rather than rely exclusively on general-purpose accelerators.

Where Tan expects AI economics to accrue

Tan’s broader strategic view was that much of the eventual value in AI should accrue to the companies developing frontier models and the applications built on top of them, rather than solely to infrastructure suppliers. Reporting on his remarks described this as a case for value flowing toward frontier models and custom chips. 19

He used a hypothetical example involving one gigawatt of compute, potential annual recurring revenue of $30 billion and operating costs of $10 billion to illustrate the operating leverage he believes successful model platforms could achieve. These figures should be read as Tan’s illustrative framework rather than an independently verified industry forecast.

He also discussed an imbalance between spending on token-generation infrastructure and model-related revenue, as well as the revenue concentration of closed frontier models. The available source material does not independently substantiate every underlying figure, so those remarks are best treated as management’s perspective on AI economics—not settled market measurements.

What the outlook means for Broadcom

The $58 billion, $115 billion and $230 billion figures make Broadcom’s AI thesis clear: custom compute and networking could become far larger businesses if leading model developers keep scaling deployments. 17

But the same thesis contains its main risk. The targets depend on a complex physical buildout beyond Broadcom’s direct control—especially grid power, data-center readiness and construction execution. 318 Tan’s message was therefore both bullish and conditional: demand appears to be there, but the pace of AI revenue growth may be governed by how fast the world can build places to run the machines.