The AI Stack Is Being Repriced, Layer by Layer

CoreWeave's revenue is growing from $1.9 billion to $5.1 billion, and its stock has fallen 40% from its 52-week high. Meta's core ad revenue is up 33% year over year, and its stock has dropped 12%. Nvidia's revenue is up 85%, and its stock has gained less than 15% over the past year. ^[1][2][3][4] Three companies at three different layers of the AI stack. The same paradox: record growth, stagnant or falling valuations. Something in how the market prices AI spending has changed. The change began at the most speculative layer, where the gap between revenue and profit is widest. CoreWeave, the neocloud provider that rents out GPU capacity, burned $8 billion in free cash flow over twelve months. In one quarter it spent $1.9 billion on capital equipment against $1.4 billion in revenue. ^[5] Its GPUs require costly upgrades every time Nvidia releases a new product, which means the hardware depreciates faster than it can generate returns. Even a $99.4 billion revenue backlog has not convinced investors the model works. ^[1] CoreWeave's valuation has dropped 56% from its peak as credit markets demanded higher loan yields and CDS spreads widened; investors, as one analysis put it, "shifted their focus from headline growth to the pace of free cash flow accretion and margin recovery." ^[6] This is cost discovery in its rawest form. The market once asked how fast a company could grow its top line. Now it asks whether the growth will ever produce free cash flow. The same question has reached the financing structures that underwrite the buildout. Nvidia has invested $100 billion in OpenAI, $6.3 billion in CoreWeave, and $2 billion in xAI. That capital then flows back to Nvidia when those customers buy its chips. ^[7] Short seller Jim Chanos captured the circularity with a single question.

[Don't] you think it's a bit odd that when the narrative is 'demand for compute is infinite', the sellers keep subsidizing the buyers? — Jim Chanos

Jeff Bezos called the AI investment surge an "industrial bubble" driven by novelty and FOMO. ^[7] OpenAI's partners have accumulated roughly $100 billion in debt to fund infrastructure, with Oracle's interest payments now consuming 20% of its operating income. ^[8] The financing is real, the debt is real, and the market has begun to price both. The inflection point arrived in February 2026, when Amazon announced $200 billion in capital spending for the year, well above the $147 billion analysts had expected. The stock plunged 11.5%. Free cash flow dropped from $38.2 billion to $11.2 billion year over year, even as AWS grew 24%. ^[9] The market punished the capex announcement itself, not the results. Jefferies warned the same month that "the trend of investors starting to question the returns from AI capex has only just started" and that Big Tech's share of the S&P 500 had fallen from 27.4% to 24.7% in three months. ^[10] What began as skepticism toward the most speculative layer had reached the hyperscalers. The numbers from the proven layer now tell the same story in a different register. Alphabet reported its first-ever negative free cash flow of $5.9 billion and raised its 2026 capex guidance to $205 billion. Meta's free cash flow plunged 91% to $784 million. ^[11] Microsoft fell nearly 20% even as Azure grew 40% and Copilot generated $37 billion in annual revenue. Meta's stock dropped 12% while its core ad business grew 33%. ^[2] These are the most profitable enterprises on earth, and the market is now discounting their AI spending as a drag on value rather than a claim on the future. Alphabet's CFO Anat Ashkenazi said AI spending "will rise significantly further in 2027." ^[12] The hyperscalers are escalating their commitments precisely as investors begin demanding proof of return. That is the central tension of the current market: management conviction and investor skepticism are now on a collision course. Nvidia sits at the top of the stack, the one company that sells the physical infrastructure every other layer depends on. It became the first $5 trillion company on $81.6 billion in quarterly revenue, growing 85% year over year. ^[4] Demand for its chips runs at a 12-to-1 ratio over supply, and yet its stock has gained less than 15% over the past year. Even the picks-and-shovels winner, the company Morgan Stanley identifies as the consistent beneficiary across every AI scenario, is experiencing valuation stagnation. ^[3][13] None of this means the demand for AI infrastructure is fake. U.S. data center vacancy rates have fallen to record lows: 0.3% in Northern Virginia, 6.7% globally, even as inventory grew 33% year over year. CBRE's Pat Lynch says the supply-demand imbalance will persist "at least three years." ^[14] Amazon Web Services grew 28% to $37.6 billion. Google Cloud grew 63% to $20 billion, with a $460 billion backlog. ^[12] The physical buildout is real and accelerating. The repricing is not about whether AI demand exists. It is about whether the spending will ever produce returns commensurate with its scale. Bain & Company quantified the gap: current AI infrastructure commitments run at roughly $500 billion per year, while consumers spend only $12 billion annually on AI services. ^[15] Closing that gap would require $2 trillion in new revenue. The market has stopped treating that gap as a problem the future will solve and started treating it as a question each layer must answer on its own evidence. A year ago, AI capex was priced as one bet: spend big, win big. Now it is priced as a stack of separate bets, each judged on whether its own economics work. The layers furthest from revenue were repriced first and hardest. The layers closest to revenue are being repriced now. The market's question has changed from "how much are you spending?" to "show me the return." And it is asking that question of every layer, in order, until someone answers.