Burry's 2008 AI Warning: Why a 3%-of-GDP Capex Boom Could Hurt if Demand Blinks

The risk is the speed of the spend, not AI itself

The real risk is not AI's cost of capital. It is how quickly the bill is rising.

Burry is mainly sharing the chart, but the number behind it matters. Data-center capex is projected to add 1.7 percentage points in just two years, rising from 1.4% of GDP in 2025 to 3.1% in 2027. That is not a gentle upgrade cycle. It is a sharp, real-time spending surge.

The spending is also concentrating. The big five cloud and AI spenders moved from under $80 billion a quarter in 2025Q1 to about $150 billion in 2026Q1. Bulls see that as proof of conviction; bears see a smaller group of balance sheets taking on an ever-larger share of investment before monetization has been fully demonstrated.

That helps explain why behavior can amplify the risk. When spending rises this fast, investors stop treating it as a company-specific decision. Big Tech capital expenditures are increasingly read as a signal that demand will eventually arrive, which can make hesitation feel like missing out. If monetization lags, that crowd may not adjust gently.

Why the 2008 comparison resonates - and where it should not be stretched

The 2008 analogy is tempting because investors fixate on the peak. By that measure, today's AI buildout still looks smaller than housing: hyperscaler capex near 3% of GDP by 2027-2029 versus 6.6% of GDP at housing's peak. But the more useful comparison is not the ceiling. It is the speed of the rise and the risk of a sharp reversal.

Speed is the part that matches

That is where the analogy sharpens. The current data-center build is moving faster than housing's quickest phase. Housing added about 0.5 percentage points of GDP per year at its fastest; the AI cycle is adding roughly 0.85 percentage points a year from the 2025 baseline. Even the late-1990s telecom and fiber buildout, another common comparison, peaked at only 1.2% of GDP.

Where the analogy breaks

The break comes if readers assume AI has to fail at housing scale to matter. It does not. The telecom example shows that a smaller peak can still produce a meaningful downturn if spending collapses after overbuilding. What matters is not whether AI is real. It is whether the buildout can stay funded when expectations reset.

The bubble test is revenue versus deployed assets

The relevant test is not whether AI adoption is real. It is whether revenue can catch up to the asset base before investors stop forgiving the gap.

Right now, that gap is large. The big five have committed to $660 billion to $690 billion of 2026 capex, and about 75% of aggregate hyperscaler capex in 2026 will fund AI-related infrastructure. That implies roughly $450 billion in AI-specific spending in a single year. Even more important, this is not a one-year bet. Goldman now expects $5.3 trillion of capex spending from the four largest hyperscalers from 2025 to 2030, up from $4.5 trillion earlier. When the asset base expands this fast, future growth alone is not enough; returns on those assets also have to improve.

Why the revenue gap matters more over time

That is where simple adoption narratives can miss the accounting reality. AI usage may grow while the monetized pool still looks small relative to the infrastructure being deployed. That risk is echoed in the view that pure-play AI revenue still a fraction of deployed infrastructure investment. Investors are effectively underwriting a long lag between spend and proof.

That lag may be uneven because AI assets do not behave like older data-center assets. The economic useful life of AI silicon is a key variable: small shifts in replacement cadence can move cumulative spend by hundreds of billions. Higher complexity, power density, system integration, and bottlenecks in power, labor, or equipment can also stretch timing and keep the payback period uncertain.

Why spending can stay high even when discipline weakens

Two forces help explain why the spend keeps climbing.

First, leasing and flexible structures can soften the apparent balance-sheet impact. That preserves flexibility, but it can also make heavy investment look lighter than it is.

Second, the fear of being underinvested can overwhelm discipline. Big Tech capex is being treated as the most important signal in the AI trade, with executives acting as though missing the build could cost strategic relevance. If revenue traction cools, the first companies to cut back may do so abruptly, and today's capex assumptions can rerate down quickly.

What could hold up if the spend cools

If the market stops forgiving the gap between infrastructure and revenue, the next question is who bears the spend best. With pure-play AI revenue still a fraction of deployed infrastructure investment, the businesses most likely to hold up are the ones selling scarce bottlenecks: semiconductors, networking links, power equipment, and data-center capacity that can be added quickly.

The more vulnerable part of the chain is the part that needs broad, sustained customer demand. If hyperscalers shift from building ahead of demand to demanding proof now, names tied to speculative compute, delayed monetization, or weaker enterprise AI budgets could be hit harder than vendors already embedded in essential supply chains.

What to watch in the capex signal

Because Big Tech capex is the key signal for the AI trade, the clearest watchpoints are straightforward:

• quarterly capex guidance from the largest hyperscalers
• whether spending continues to concentrate in a small group of companies
• signs that AI revenue is scaling alongside infrastructure
• any indication that replacement cycles for AI hardware are speeding up
• evidence that bottlenecks are stretching deployment and delaying returns

That is why crowded conviction is the real danger. Investors have anchored on Big Tech spending as proof the boom must continue. If that anchor slips, the market can move quickly from treating those names as inevitable winners to treating them as early casualties of a spending cooldown.