Broadcom (NASDAQ:AVGO) CEO Hock Tan said enterprise adoption of generative artificial intelligence remains in an early stage, with companies still working to identify use cases that offer a meaningful return on investment. Speaking at the Goldman Sachs Communacopia + Technology Conference, Tan said Broadcom has involved roughly 15,000 of its 30,000 engineers in evaluating AI tools for hardware and software development.
Tan said the company has found AI’s clearest near-term value in improving engineering quality and accelerating the creation of new technology rather than broadly replacing workers or delivering straightforward productivity gains. In semiconductor development, avoiding a chip redesign can be especially valuable, he said, because a respin of leading-edge silicon can cost about $30 million and add six months to a project timeline.
Value Expected to Favor Leading Models
Tan said he expects the greatest economic value in AI to accrue to developers of the most capable frontier models and the applications built around them. He argued that the best technology tends to attract demand, comparing the dynamic with leading semiconductor products.
Using an illustrative example, Tan said a strong frontier model supported by 1 gigawatt of computing capacity could generate $30 billion in annual recurring revenue, while the annual cost of operating that capacity could be about $10 billion. In that scenario, he said, the model developer and associated applications would retain the larger portion of the value, while cloud providers, power providers, chipmakers and memory suppliers would compete over the remaining spending.
Tan also expressed a preference for closed frontier models over open-weight models, while acknowledging that both approaches could coexist. He said open-weight offerings may appeal to enterprises seeking lower token costs, but argued that long-term value would be concentrated in models that continue to improve intelligence and performance.
He cited his assessment of inference economics, saying global token-generation infrastructure costs are about $200 billion annually, while model-related revenue is about $150 billion. According to Tan, closed frontier models account for at least 75% of that revenue despite representing roughly half of token generation.
Custom Silicon and Data-Center Constraints
Tan said Broadcom’s custom AI accelerators, or XPUs, are designed in close collaboration with customers developing frontier models. Customer teams define the algorithms and workloads their models require, while Broadcom works with them on chip architecture, microarchitecture and physical design.
He pointed to Broadcom’s work with OpenAI, saying the companies produced an XPU known as Jalapeño within about a year and that it performs as well as, or better than, leading general-purpose accelerators. Tan said Broadcom is now working with OpenAI on the next two generations following Jalapeño.
On the company’s longer-term AI outlook, Tan said infrastructure availability is a key limiting factor. While Broadcom can plan around supplies of leading-edge wafers, memory and substrates, he said the less predictable constraint is the availability of power-ready data-center sites.
Building AI data centers requires land, power, buildings, equipment and construction capacity, as well as permitting, Tan said. Sites intended to be ready in 2028 need construction to begin now, he added. The long lead times provide Broadcom with visibility into expected demand, though they also create uncertainty around precise timing.
Google Partnership and Semiconductor Technology
Tan said Broadcom’s relationship with Google spans more than a decade and has included every generation of Google’s tensor processing units since the first version. He said the companies’ long-term agreement for TPUs and networking products through 2031 expands and strengthens the technical and strategic relationship.
As traditional process-node advances provide more limited performance improvements, Tan said Broadcom and its customers are pursuing other methods to increase accelerator performance. These include combining multiple dies into a single chip. He said the company has progressed from one die to two dies and is working on a generation using four dies, with further expansion possible.
Tan said Broadcom’s competitive advantages include intellectual property in high-speed interconnects, SerDes technology, multiplier density, multi-die communication and advanced packaging. He said these engineering capabilities form barriers for competitors, including customers that may seek to bring more semiconductor work in-house.
Financing and Capital Allocation
Tan also discussed Broadcom’s financing platform involving Apollo and Blackstone, which is intended to support compute capacity for AI labs. He said Broadcom currently has six customers developing frontier models and views supporting those customers as strategically important, particularly because it is difficult to determine which models will ultimately lead the market.
For customers without sufficient cash flow to fund needed computing capacity, Tan said the financing structure brings in financial partners and banks that assume part of the financing risk. Broadcom’s role includes providing residual guarantees tied to equipment used as collateral, he said. Tan stressed that the arrangement is not circular financing because, in his view, it supports existing demand rather than creating demand.
Looking ahead, Tan said Broadcom expects to end fiscal 2026 with a record cash balance as revenue grows, driven in part by AI. He said management and the board will review capital allocation at their December meeting. Potential uses of cash include dividend increases and stock repurchases, while debt reduction appears less compelling because much of the company’s approximately $56 billion in debt was issued at relatively low interest rates.
About Broadcom (NASDAQ:AVGO)
Broadcom Inc (NASDAQ: AVGO) is a global technology company that designs, develops and supplies semiconductor and infrastructure software solutions for a broad range of markets. The company’s semiconductor business provides components and systems for wired and wireless communications, enterprise and cloud storage, networking and broadband access, serving original equipment manufacturers, cloud service providers, telecommunications carriers and industrial customers worldwide. Broadcom is headquartered in Irvine, California, and operates globally with research, development and sales organizations across North America, Europe and Asia.
On the semiconductor side, Broadcom’s portfolio includes system-on-chip (SoC) and application-specific integrated circuit (ASIC) solutions, radio-frequency and connectivity components, Ethernet switching and PHY devices, storage adapters and controllers, optical transceivers and other networking silicon.
