Winbond Chairman Chiao Yu-chun said the AI data center investment boom continues, but how long it lasts depends on funding. Over the past six months to a year, markets have begun watching the cash flow strain from cloud providers' aggressive capex. However strong AI demand is, companies must weigh how much capital they can deploy—once funding tightens, investment slows.
Chiao said demand for 2027 and 2028 is relatively clear, with funding and infrastructure largely arranged and most data centers slated to open next year already prepared. But 2029 may be the key inflection point for the next downturn, with funding the biggest constraint. He said no customer has yet outlined a complete 2029 investment plan. 2029 is not destined for a reversal, but a critical checkpoint—as capex runs into funding, land, and power limits, whether AI can clear the next investment bottleneck will shape the semiconductor cycle.
On rising memory prices, Chiao said everything has a limit. He hopes prices have peaked. Sustained increases help memory makers' profits but may hurt demand for phones and PCs. "One family can't eat everything"—AI customers can absorb high prices, but consumers may not. The industry shouldn't chase short-term profit at the expense of the broader electronics sector.
Industry data shows DRAM and NAND prices hit decade highs in early 2026, with some models up over 300%. HBM demand has pulled advanced capacity away from consumer memory, prompting OPPO, vivo, and Honor to raise prices on some models. Q2 2026 conventional DRAM contract prices rose 58–63% sequentially, NAND 70–75%. Five hyperscalers—Microsoft, Google, Amazon, Meta, Oracle—account for most AI server purchases, and their capex decisions directly drive HBM and server DRAM demand.
Winbond began expanding capacity in the second half of last year, with benefits visible next year and capacity near full in the year after. Chiao warned that if the boom persists, makers tend to keep building, only to find oversupply when the market turns—making the downturn sharper.
On reports that TSMC may develop 3D DRAM for Nvidia, Chiao said entering DRAM requires solving technology and patent issues and weighing relationships with existing partners. He believes TSMC would announce only after technology and talent are in place. Neither TSMC nor Nvidia has confirmed the reports. Analysts note 3D DRAM vertically stacks memory cells, unlike HBM's die stacking, and requires long R&D validation across transistors, capacitors, materials, etching, and yield—TSMC's logic and packaging strength does not equal DRAM mass-production capability, making near-term in-house production unlikely.
ICgoodFind Summary: Chiao marks 2029 as the checkpoint for AI data center demand, with funding—not demand—the key variable, and no customer has given a clear 2029 answer. His view on memory pricing matters too: a price peak hurts near-term profits, but once consumer electronics demand is squeezed far enough, the gains from higher prices also hit their limit.
