Big Tech will spend $650 billion on AI infrastructure in 2026—data centers, GPUs, power grids—chasing a $3 trillion buildout over five years. On February 10, SMIC co-CEO Zhao Haijun stood up during his company’s earnings call and said the quiet part out loud: we have no idea if anyone will actually use this stuff.
China’s largest chipmaker just warned that the country’s rushed AI data center expansion—cramming 10 years of capacity into two—risks leaving billions in infrastructure sitting idle. And the problem isn’t just China’s. It’s a preview of what happens when speculative infrastructure meets unproven demand at global scale.
This matters now because the AI infrastructure boom has become a self-fulfilling prophecy—until utilization data suggests otherwise.
China built 10 years of AI capacity in 2 years. Most of it sits empty.
China’s western data centers—built under the government’s “Eastern Data, Western Computing” initiative to leverage cheap power in remote regions—run at 20-30% utilization. Not the full capacity most people assume when they hear “AI boom.” Ghost towns of GPUs.
The physics don’t work. Western data centers are too far from eastern users for real-time AI applications—voice assistants, video processing, autonomous systems. Latency kills the model. You can’t run a chatbot from Inner Mongolia when the user is in Shanghai and expects sub-100ms response times. The infrastructure exists, but no one stress-tested whether “build it and they will come” survives contact with the speed of light.
And this isn’t a China problem disguised as a tech problem. Western hyperscalers—Alphabet, AWS, Meta, Microsoft—are making the same bet at 10x scale. Zhao warned on SMIC’s February 10-11 earnings call that “utilizing ballooning capacity has not been fully thought through.” He’s talking about China. He might as well be talking about everyone.
The chipmaker making the chips is warning about the chips
Here’s the irony that makes this story stick: SMIC’s utilization hit 93.5% in 2025, up 8 percentage points from 2024. They shipped 9.7 million wafers, up 21% year-over-year. They’re maxed out making chips for data centers that sit idle.
And they’re not slowing down. SMIC is adding 40,000 12-inch equivalent wafers per month by the end of 2026—on top of the 50,000 they added in 2025. They spent $8.1 billion on capex in 2025, up 10.5% from 2024. But 2026 capex? Flat. That’s the tell.
SMIC is caught between two realities. Short-term demand is real—US sanctions forcing Chinese firms toward homegrown alternatives created a chip rush. But long-term? Those chips go into data centers with no proven workloads. The company is doubling down while simultaneously hedging. That’s cognitive dissonance at industrial scale.
Meanwhile, chip shortages starving non-AI sectors like smartphones reveal the opportunity cost of this bet. Zhao himself warned that AI-driven memory demand is squeezing supplies for everything else. We’re building exascale infrastructure for apps that don’t exist yet while low-end device orders collapse.
The honest problem: we don’t know what AI infrastructure is actually for
AI hype promises infinite demand—AGI, autonomous everything, infinite context windows. But 2026 reality? Most AI apps are chatbots, image generators, and code assistants. None of those need exascale data centers in remote deserts.
The buildout is speculative. It’s betting on applications that don’t exist yet, running on infrastructure that can’t pool capacity due to latency constraints and hardware fragmentation. AI apps remain unproven at the scale this infrastructure assumes—most fail at real-world tasks that require reliability over novelty.
Here’s what I don’t know: which specific facilities failed, which investors lost how much money, or what idle capacity costs per month. SMIC’s warning is the only named source calling this out. The story is structurally sound—utilization rates, capex trends, latency physics all point the same direction. But the specifics are thin because no one wants to admit the emperor has no clothes.
SMIC is still adding 40,000 wafers a month. Big Tech is still spending $650 billion this year. And somewhere in western China, rows of GPUs sit idle, waiting for apps that may never justify their existence. The bet is already made. We just don’t know who’s holding the bag when the music stops.
