"Coreweave's a Zero" | George Noble on Why He's Short Neoclouds, Why AI is a Shale 2.0 Bubble Headed for Debt/Depreciation Brick Wall, and Why He's "Coming For The Momentum Bros"

George Noble calls AI a bubble, not because he doubts the technology, but because he has not seen a convincing return-on-investment case for infrastructure spending on the scale of hundreds of billions or trillions of dollars. Compute and token prices are falling, Chinese open models can perform the work for about 95% less, and successive models arrive at roughly half the previous cost. He says providers can show rapid revenue growth while selling services below cost, citing about $200 in compute and energy costs for a service sold for $20. Consumers’ willingness to pay around $20 a month, he argues, does not establish a market for $200-a-month services or prove that the economics justify current valuations.

Noble compares AI’s buildout to shale, where companies keep investing because they fear falling behind competitors or China. He calls this a prisoner’s dilemma that “guarantees” a crash, while declining to give a timeline or quantify the likelihood: he expects Silicon Valley’s conviction to keep spending near sell-side expectations. In his view, today’s compute shortage lets some providers earn temporarily elevated returns, but those earnings could fall to about 20% of current levels. A roughly one-year compute payback may be possible now, he says, but only for a limited time. If supply catches up with or exceeds demand in two to three years, semiconductor prices could weaken while interest, depreciation, and fixed costs remain.

CoreWeave is Noble’s biggest short and, in his words, “a WeWork from GPUs”: a debt-heavy business whose chip assets lose value over time. He says he is highly confident it would be worth zero without a bailout and expects it to be “a lot lower” before the cycle ends, but does not know whether it will go bankrupt; Nvidia could potentially buy or take it over. The chips’ useful life—four, six, or ten years—and the effect of newer generations on existing equipment’s value remain open questions. Noble says CoreWeave’s higher cost of capital makes the stock a stronger short, and that if investors doubt its ability to repay bonds, its equity could be marked at zero. He allows for a short-term bounce and admits he could overstay the trade. The interviewer said the stock had fallen more than 40% since Noble recommended shorting it in June and cited a 60% implied default rate on its 2032 notes; Noble also cited CoreWeave bonds yielding 15%.

Noble sees neoclouds such as CoreWeave and IREN as better short opportunities than Nvidia, which he is neither long nor short but could eventually short. He says Nvidia lends customers billions to buy its products, with some chips potentially sitting unused in warehouses. More broadly, he expects high semiconductor returns to attract competition, pointing to advances in China and Japan and reports from Korea. He calls Micron’s reported 85% gross margin abnormal and urges investors to focus on normalized earnings; he expects current semiconductor monopoly profits to erode within a couple of years, though he cannot say exactly when. He warns that data centers are even more capital-intensive than semiconductor factories, making temporary high prices especially risky.

Noble’s depreciation warning rests on a specific assumption: if AI assets and chips reach about $2.5 trillion, hyperscaler depreciation could exceed $500 billion in 2030 and outweigh their combined profits. He stresses that the result depends on the $2.5 trillion estimate. He argues that price-to-earnings ratios can mislead when revenue is booked before associated costs and depreciation arrive; the host countered that continued cloud growth could keep hyperscaler earnings positive in 2030. Noble also says 45% growth may not justify valuations of $1.5 trillion to $2 trillion. He compares the spending boom with the internet buildout, which saw traffic compound at about 43% annually for 25 years and increase 500,000%, yet did not prevent losses at companies including Nortel, Lucent, and Global Crossing.

Noble is also short Tesla and SpaceX, though he says those positions are losing money and expects the short trades to pay over time. He describes SpaceX’s roughly $2.2 trillion valuation and 100–120-times-revenue multiple as driven by narrative and momentum, estimating Starlink at $300–400 billion and the rocket business at tens of billions. He calls a proposed SpaceX-Tesla merger two overpriced companies propping each other up. For Tesla, he cites about $38 billion in cumulative profits over roughly 16–18 years against a valuation near $1.5 trillion, with profits including billions in emissions and tax credits; he says sales are declining and that, with one exception, Tesla has never earned its cost of capital. Noble rejects annualized quarterly revenue and earnings-beat trading as unreliable ways to justify extreme valuations, criticizes market voices for not analyzing financial statements, and says the tide is turning against momentum traders: “I’m coming for you.”