Is There a Compute Overcapacity?

The Core Question

Whether there is too much compute does not depend on how much got built. It depends on whether the finished capacity is used, and whether the people using it can pay. That is the substance of Sequoia's question: at the prevailing rate of spend, how much revenue would the AI industry have to produce to earn the money back? Sequoia's own estimate widened from $200B in 2023Q3 to $600B in 2024Q2 — within a year. The gap was not closing.

The Spending Side Is Hard. The Return Side Isn't.

Microsoft, Google, Amazon, Meta and Oracle spent $140.6B on capital expenditure in 2025Q4 combined (Epoch AI, compiled from the companies' 10-Q and 10-K filings). That is five companies, one quarter, and it reconciles line by line.

There is no equally clean public series on the revenue side. That asymmetry is part of the problem rather than a gap in the data: spending is an accounting fact, return is an estimate. When the two are set beside each other, they are not equally reliable.

Three Things That Break Before the Headline Number Does

• Utilization. Built is not the same as busy. The Information reports that roughly 65% of AWS EC2 instances averaged under 20% CPU utilization over 30 days, and that AWS has told engineers to shut idle instances down and upgraded its compute optimizer. Utilization sags before investment does.
• Depreciation. GPUs carry an accounting life of only a few years. If demand grows more slowly than the schedule writes the asset down, the balance sheet devalues before sentiment does — which is why every extension of a depreciation schedule is worth watching on its own. It defers the reported loss, not the physical one.
• Who is funding whom. When a chipmaker finances a customer's purchases and the customer spends the money on chips, that slice of demand is worth less than it looks. The feed files these arrangements separately.

Cash Flow Is the Statement That Talks First

Tencent's 2026Q2 results put this plainly. Capital expenditure for the quarter was RMB 52.8B, up 176% year on year, and free cash flow turned negative at RMB 13.8B against a positive figure a year earlier. Tencent's own announcement notes that excluding AI-related prepayments, free cash flow would have been positive RMB 37.6B — and the RMB 51.4B difference between those two numbers is what one quarter of buying compute ahead of time costs.

An income statement can smooth this through depreciation. A cash flow statement cannot. That is why this topic watches cash flow more closely than the capex headline.

China Takes a Different Route to the Same Problem

China's AI datacenters are pushed jointly by local governments and large platforms, and reports of finished-but-idle capacity arrived earlier and cluster in regional projects. At the platform end, Tencent's stock fell the day it reported — the market's patience for spending-in-place-of-growth has a limit in both countries. Same question, different constraints.

How to Use This Page

The charts below plot only verified points with a public source. The signal list is pulled live from the feed under the capex and chips categories, and moves as collection runs. The site does not hand you a verdict — it puts spending, utilization and cash flow side by side and leaves the judgement to you.