AI Bear – Right About the Excess

Download Full Report

This article was authored by: Lance Roberts

AI Bears regard today’s datacenter mania a more dire situation that the great fiber-optic capacity overbuilds that cracked in 2000. The Bank for International Settlements flagged roughly $1.65 trillion in off-balance-sheet obligations held by the largest hyperscalers, exceeding the amounts they carry on their books. Sequoia puts the annual gap between AI infrastructure spending and ecosystem revenue at nearly $600 billion. Allianz measures the capex-to-revenue divergence at about 46%, well past the 32% that marked the 2001 telecom bust. An MIT study suggests most corporate AI pilots have produced no measurable revenue at all. Yes, valuations are stretched and market concentration is worse now than it was in 2000. The circular financing argument is real, too. When Nvidia takes an equity stake in a company that then commits to buying Nvidia chips, part of what gets reported as “demand” is the seller funding its own sales. But leading up to the 2000 overbuild, the financing came from companies that had no business borrowing what they borrowed: WorldCom, Global Crossing, and the upstart carriers that were stringing fiber on debt, and the vendor loans that Lucent and Nortel handed to customers who could not pay them back. When revenue failed to arrive on schedule, those balance sheets could not cover the shortfall, and the structure collapsed into bankruptcy. Today’s buildout is a different animal. Roughly two-thirds of the 2026 capex is funded directly from the operating cash flow and equity of Microsoft (MSFT), Alphabet (GOOGL), Amazon (AMZN), and Meta (META), four of the most profitable enterprises ever assembled. Second, “no revenue” is not the same thing as revenue that simply hasn’t caught up to the spending yet. Microsoft’s AI business is past a $37 billion run rate, Amazon’s AI revenue is growing in the triple digits, and Anthropic went from about $9 billion to a reported $47 billion run rate in a single year. The AI bears predict a glut, yet the binding constraint right now is the opposite of a glut. Microsoft is sitting on something like $80 billion of Azure orders it cannot fill for lack of electricity, with GPUs idle in inventory waiting on power. Today, more than 60% of the data center capacity planned for 2027 is not yet under construction. The question is not whether there is excess, because there plainly is. The real question is what a disciplined investor does with a genuine, extreme, but cash-funded overbuild. Where you take the risk matters as much as how much you take. Not all AI exposure carries the same danger. How do you know when the story is actually wobbling? 1. Hyperscaler capex guidance gets cut. 2. AI revenue growth stalls. 3. AI credit-default swaps widen. 4. Depreciation 2027 – 2029 outruns revenue. 5. A hawkish Fed – rate hikes and quantitative tightening. Bob Farrell’s Rule #9 is always worth repeating here: “When all the experts and forecasts agree, something else usually happens.” With more than half of managers now calling AI a “bubble” and “long the Magnificent 7” ranked the most crowded trade on the Street for nearly two years, the consensus has already tilted bearish. That does not make the AI bear case wrong, but it does suggest the obvious crash may refuse to arrive on the obvious schedule.