AI data-centers risk excess capacity
- SMIC co-CEO Zhao Haijun warned in February that companies are trying to build a decade of AI data-center capacity in one or two years. - China is already trying to route surplus compute through a state-backed national network after a three-year local-government building boom left capacity stranded. - That matters because AI infrastructure spending is still accelerating, even as weak utilization hints this cycle could destroy returns before demand catches up.
AI data centers are the factories of the current tech cycle. They soak up chips, power gear, steel, cooling equipment, and huge piles of capital. The problem is that a factory boom only works if somebody actually needs the output. In China, that gap is starting to show. New reporting over the past year and fresh warnings from SMIC in February 2026 point to the same issue — too much AI capacity may be getting built too fast, with real demand lagging behind. ### What exactly is being overbuilt? Not “the internet” in some vague sense. It’s AI-oriented compute capacity — racks of GPU servers, networking gear, power systems, and specialized data halls designed to run model training and inference. China spent the last few years pushing these builds hard through local governments, telecom operators, cloud firms, and state-backed investors, with hundreds of projects announced across provinces. But a meaningful share of that capacity appears underused or mismatched to what customers actually want. ### Why did China build so much? Because AI became both an industrial policy target and a growth story. Local officials wanted headline projects. Investors wanted exposure to the AI boom. Big platforms wanted to secure compute before export controls and supply bottlenecks got worse. Alibaba alone said in February 2025 that it would invest at least RMB 380 billion over three years in cloud and AI infrastructure. Goldman Sachs has also projected a sharp jump in China cloud capex and data-center power capacity. (technologyreview.com) ### So what changed? Two things. First, real customer demand has not filled all the new supply. Second, the market shifted toward efficiency. MIT Technology Review described many new Chinese AI data centers sitting unused as speculative projects ran into weak demand and changing AI economics. DeepSeek’s rise sharpened that point — if competitive models can be built and served more efficiently, the case for brute-force infrastructure expansion gets weaker at the margin. (alibabagroup.com) ### What did SMIC actually warn about? SMIC’s co-CEO Zhao Haijun put it bluntly in February 2026: companies want to build “10 years’ worth” of data-center capacity in one or two years. That’s the cleanest version of the risk. Pull future demand forward too aggressively, and today’s boom becomes tomorrow’s idle inventory. It’s the same problem manufacturers hit when they mistake a shortage for permanent demand. (technologyreview.com) ### Why doesn’t surplus compute just get used somewhere else? Because compute is not perfectly fungible. Location matters. Latency matters. Power costs matter. Chip mix matters. Software stacks matter too — especially when hardware ranges from Nvidia systems to domestic alternatives. Reuters reported last year that Beijing was exploring a national, state-run platform with MIIT and the three big telecom carriers to connect and sell surplus computing power. You only build a resale network like that if stranded capacity is already a real problem. (businesstimes.com.sg) ### Is this only a China story? No — but China is the clearest stress case. In the US, hyperscalers still have stronger monetization paths, deeper software ecosystems, and bigger paying customers. Even there, though, capex is exploding. Bloomberg reported Meta was working on roughly $13 billion in financing for an El Paso data-center project this month. So the broader question is global: how much of this buildout is durable demand, and how much is everyone racing not to be left behind? (tech.yahoo.com) ### Why should anyone outside tech care? Because data-center booms ripple outward. They drive demand for semiconductors, transformers, cooling systems, construction, grid upgrades, and electricity. If utilization disappoints, the pain doesn’t stay inside AI. Returns on servers fall first, but suppliers, utilities, industrial firms, and commodity demand can all feel it next. UBS has even flagged an AI bubble burst as a downside risk to China’s outlook. (bloomberg.com) ### Bottom line? The AI buildout is real. But real booms still overshoot. China now looks like the place where that overshoot is easiest to see — not because AI demand is fake, but because infrastructure got built ahead of proven usage. If that pattern spreads, this cycle stops looking like a clean growth story and starts looking like a classic capacity glut with GPUs instead of steel. (ubs.com)
