Broadcom's Hock Tan says capital may be a tighter limit on AI buildout than chips, memory or power

Put Hock Tan's dates side by side. He suggested wafer shortages could ease by 2028 or 2029 ^[4]. He also said demand could outrun supply through 2029 and potentially into 2030 ^[6]. On his own numbers, the wafer constraint lifts between zero and two years before the imbalance ends ^[11]. In the years when chips are no longer scarce and demand is still ahead of supply, the remaining limit on spending is whether someone will fund it. That is our inference from his timeline. He did not put it in those terms.

His sentence was hedged. "[What] I think might be the biggest constraint is capital," Hock Tan said at the DBS Asia Future Forward Conference in Singapore on Monday ^[2]. Lip-Bu Tan split the problem in two. "If there's any challenge in the front, it's capital and the cost of capital, and then the other part is in the longer run, people want to see the ROI," he said ^[5].

The first half can be read off a bond yield today. The second cannot be read until the projects run. CNBC noted that higher yields could raise the cost of financing chip purchases and the building of data centers ^[12] for projects that may take years to generate returns ^[13].

The counter-thesis sits in the same remarks. Hock Tan cautioned that demand for computing infrastructure could slow if adoption of advanced AI models disappoints or improvements in their capabilities plateau ^[8]. That outcome would also cut data center starts, but the cause would be weak demand, and cheaper financing would not fix it. A third case is cheaper money: if yields fall, the limit goes back to wafers, memory and power, on the schedule he gave.

We think the financing reading is the likelier of the three over the horizon Lip-Bu Tan gave. He said the industry was unlikely to see a significant slowdown over the next two to three years ^[7], and both executives stayed bullish on near-term demand ^[10]. CNBC's report does not include a yield level, a financing total or a capex figure, and everything here rests on one conference appearance.

We would be wrong if data center projects keep getting funded at higher yields through 2029 while wafers are still short, because supply would then still be the binding limit. For operators, Hock Tan's line that "Memory shortage today might not be that way a year from now" ^[3] argues for budgeting memory as a one-year problem and spending the planning effort on the funding line.