#数据中心或在2028年面临过剩 The second half of AI infrastructure: what exactly is the market watching?

It’s not just chips or power—it’s about where the bottleneck shifts next. The focus will move from power back to chips, but the key metric that will really determine what goes up or down will be “utilization.”

The past two years: everyone was scrambling for chips. Back then, GPUs were impossible to get, CoWoS packaging was booked out a year in advance, and HBM prices were soaring. The whole story was simply “a chip shortage.”

Now, in 2027: the bottleneck is power, not chips. Here’s the interesting part—chips are actually easier to buy now. Even SemiAnalysis says “buying GPUs is easier than getting power.” What’s really hard now is getting a grid connection, which takes four to seven years; transformer deliveries are measured in years. So the giants are generating their own power, and companies in North America have already reserved nearly 75GW of behind-the-meter power plants. At this stage, the market is watching the “powering up” segments: power equipment, gas turbines, nuclear power, and liquid cooling.

From 2028 onward: the tide is set to turn again. SemiAnalysis’s model shows data center construction doubling in a surge in 2027. As a result, by the second quarter of 2028, data center capacity may shift from “not enough” to “oversupply,” with nearly 10GW of excess capacity by year-end—in other words, “the buildings are ready, but there aren’t enough chips to fill them.” At that point, the bottleneck will shift back to chips, but not packaging—it’ll be HBM and advanced process nodes. So their message to the market is simple: stop worrying about a chip glut; worry about a surplus of computing services. Whether the chip supply chain can expand fast enough is the metric most worth watching next.

More importantly, the market is starting to do the math. The actual utilization rate of H100s worldwide is only around 15% to 18%, but making money generally requires utilization above 60%. On one hand, spending keeps rising—with major tech companies’ capital expenditures heading toward more than $1 trillion next year, a third of it financed through bond issuance. On the other, returns haven’t kept pace. That’s why the sharp drop in US semiconductor stocks over the past few days, and Meta’s increased spending turning its free cash flow negative, have sparked so much concern about ROI. Put simply, the market has already started measuring the industry by “utilization and payback.”


This round of AI infrastructure: after chips, power, and then back to chips and the numbers—power is no longer the bottleneck; utilization is the star of the second half#WCTCS9