The scarce resource in AI expansion may no longer be chips, but money. At the DBS Asia Future Forward Conference in Singapore on October 12, 2026, Broadcom CEO Hock Tan said that as AI development continues, capital will be the biggest constraint facing the company; Reuters reported the panel the same day. Tan added that the company is getting past wafer and memory shortages, and that frontier labs Anthropic and OpenAI have shown their latest models, deployed for AI personal assistants, can generate significant revenue.
What is missing is not capacity, but capital and proof of returns
On the same Singapore stage, Intel CEO Lip-Bu Tan pointed the same way. His view: the challenge up front is capital and the cost of capital, and in the long run the industry will have to prove the returns justify the spending. That is not the same as weak demand. Tan does not expect momentum to slow significantly over the next two to three years, citing spreading applications and continued demand for computing power. Translated: the queue is still there, but the financing tempo and the pressure to earn have moved to the center.
The same day, Tan also noted that Chinese frontier models such as DeepSeek, Kimi and GLM carry far lower token costs, are spreading very fast and are already good enough, and he stated firm support for open source. That remark exposes the other half of the returns side: as cheaper models keep matching capabilities, the case for premium routes to justify expensive build-outs will be examined ever more strictly.
Why a migrating bottleneck matters
For the past two years, the AI bottleneck list mostly pointed at physical things: whoever secured GPUs first, whoever had power and data-center sites, set the pace. The Singapore remarks move one link of that list toward the balance sheet: as wafer and memory supply constraints ease, the cost of borrowing and returns discipline will decide who can keep expanding.
The spillover reaches three groups. For large platforms, capital expenditure keeps climbing, so investors will ask about actual revenue per added dollar of capacity rather than gigawatt announcements. For startups and smaller labs, financing ability itself becomes a competitive variable, and a closed window can be fatal even with sound technology. For customers, pressure on the returns side usually means faster productization and pricing moves. Broadcom praising frontier labs' revenue ability while naming capital as the constraint is no contradiction: to suppliers, those customers are both the reason to spend and the proof that the spending pays.
Of course, these are still executives' judgments, and an industry whose demand genuinely stays strong for two or three years can carry financing costs far more easily than one whose growth rate slips. The real measure of this forecast will be whether coming earnings seasons start judging capacity spending by revenue instead of plans. If money really becomes the bottleneck, the industry's story changes with it: less about how much capacity can be built, more about how much anyone can afford to build.
