Publication date: 12 October 2026
Broadcom chief executive Hock Tan says capital will become the largest constraint on continued AI expansion, a warning that shifts attention from chip shortages to the financial and physical limits of building computing infrastructure.
Tan made the comment on 12 October as the technology industry gathered around new infrastructure announcements. Reuters reported his assessment that funding, rather than memory availability alone, would be the main hurdle as development continues.
Broadcom’s own outlook illustrates the scale of demand. In its third-quarter earnings call, the company said AI semiconductor revenue had grown 221 percent from a year earlier and projected continued rapid expansion. It is also investing in manufacturing capacity and networking systems needed to connect large clusters of accelerators.
AI capacity is a complete system
The industry’s early bottleneck was access to advanced processors. The constraint then expanded to high-bandwidth memory and sophisticated chip packaging. At data-centre scale, the problem includes land, electricity, cooling, network links, construction time and financing.
A company may have an allocation of chips but still be unable to deploy them if grid connections are delayed or the cost of capital makes the project uneconomic. That is why capital becomes a technical constraint: it determines which designs can leave the planning stage and operate reliably.
Broadcom’s focus on networking is also important. Thousands of processors create value only when data can move between them quickly and efficiently. As clusters grow, networking performance, power use and software coordination can limit usable computing capacity.
The investment race needs better discipline
Technology companies are under pressure to secure infrastructure before competitors do. That can encourage commitments based on optimistic future demand rather than demonstrated customer use. Long construction schedules increase the risk that hardware, models or market needs change before a facility is completed.
Leaders should test AI infrastructure plans against higher interest rates, energy-price volatility and slower-than-expected revenue. Contracts need flexibility, and investment should be released in stages as utilisation becomes visible.
Governments face a related challenge. They want domestic AI capacity, but subsidies and planning approvals can shift risk to taxpayers if projects are poorly chosen. Public support should be tied to energy efficiency, skills, supply-chain resilience and transparent economic benefits.
Tan’s warning does not imply that the AI build-out is ending. It suggests that the industry is moving from a period defined by component scarcity to one defined by capital allocation. The winners will be organisations that deploy computing productively, not simply those that announce the largest capacity.