Back Research Notes CoreWeave’s Warning: When AI Demand Meets Infrastructure Reality Published on November 13, 2025 By Jordi Visser The Supply-Constrained Revolution It has been a fun year writing about AI in 2025. The AI boom has reshaped global markets over the past three years, caused constant fears of a bubble and has driven record capital expenditures, surging equity valuations, and a historic demand for compute power. Staying bullish on the AI infrastructure trade in 2025 has been the way to make money but it has not been easy. Based on CoreWeave’s latest update this week, I think the next phase of this revolution will be very different for investors: the constraint has moved from capital to concrete and with it, I expect a very different year in 2026. Although I believe 2026 will be another year where the winners will be the ones benefitting from the acceleration in AI, I also believe the sector and name winners will be very different than the prior three years and Coreweave was a warning sign of a transition. Next year for me is about sectors about to benefit from AI adoption and one where the massive buildout means a broadening out PMIs heading higher. With more rising estimate revision competition and less concentration, the bar for AI legacy infrastructure winners will be higher. The artificial intelligence buildout has reached an inflection point where the constraint is no longer demand, capital availability, or technological capability, it’s physical infrastructure delivery. CoreWeave’s earnings call provided the clearest signal yet that the AI industry is running into hard infrastructure bottlenecks that will reshape investment performance across the sector in 2026 and beyond. When do you hear a company say they have “insatiable” demand to fill and the stock falls 16% that day. But here’s what makes this moment so unusual: while AI is advancing, the macro picture feels like we are coming out of a recession. The Fed is cutting rates because of weakness in the jobs market, credit stress is rising in auto loans, housing, and credit cards. Oil prices are down 15% YTD. These are classic signs of economic weakness yet infrastructure bottlenecks are tightening, not loosening. In a normal cycle, Fed rate cuts stimulate demand to fill excess capacity. Right now in the fastest growing part of the economy and stock market, there is no excess capacity. Companies are rationing customers, cutting capex guidance not from lack of demand but from construction delays, and equipment manufacturers are sold out through 2028. This is the K-shaped economy in real time. At the bottom of the K, consumers face credit stress and rising unemployment fears, conditions that would typically trigger recession. Yet the story is circular: AI and other exponential innovations are driving margin expansion, which in turn reshapes labor demand and market concentration. The stock market’s rally increases household net worth but at the same time AI creates labor anxiety that is growing by the day. At the top of the K, The Mag 7 enjoy AI-driven margin gains and are investing over $400 billion annually in infrastructure they can’t yet monetize. The Fed can stimulate consumers, but it can’t accelerate transformer production or data center construction. Perhaps the “AI bubble” critics are asking the wrong question. The issue isn’t return on invested capital. CoreWeave’s 85% backlog growth and Oracle’s $455 billion in contracted revenue prove the returns are real. The bubble is in TIME: the gap between exponential demand growth and linear infrastructure buildout capacity. When equipment lead times stretch to 2-7 years and data center construction faces 12-24 month delays while token demand is growing exponentially, even perfect execution cannot close the gap between what customers want and what the physical economy can deliver. CoreWeave slashed its 2025 capital expenditure guidance by roughly 40% from $20-23 billion down to $12-14 billion not because demand weakened or financing became unavailable, but because a third-party data center developer failed to deliver “power shells” on schedule. This wasn’t a demand problem. The company’s revenue backlog surged 85% quarter-over-quarter to $55.6 billion, driven by contracts with Meta, OpenAI, and Nvidia. Management described customer requests for compute resources as “far exceeding current capacity” and characterized demand as “insatiable.” CoreWeave has the capital, the customers, and the orders but it cannot deploy the infrastructure fast enough. As management stated bluntly, “This is a systemic problem that the industry will have to deal with for the foreseeable future.” This admission marks a regime change in how investors should think about AI infrastructure investments. Oracle Confirms the Pattern: Signed Contracts, Missing Infrastructure CoreWeave’s experience is not isolated. Oracle with its $657 billion market cap and a staggering $455 billion revenue backlog faces identical constraints. CEO Safra Catz acknowledged in June 2025: “We actually currently are still waving off customers or scheduling them out into the future so that we have enough supply to meet demand… We are putting out as much capacity as we possibly can.” Oracle is turning away customers despite having secured massive contracts with OpenAI, Meta, and xAI. The bottleneck isn’t demand, capital, or technology. As CEO Mike Sicilia stated in October 2025: “There is real value in artificial intelligence as demand far outpaces supply.” Sicilia explicitly dismissed “AI bubble” fears, recognizing that compute supply, power availability, and physical facilities cannot keep pace with AI workload growth. Industry analysts covering Oracle’s earnings identified the core constraint: “This is due to supply constraints across land, buildings, energy and GPUs.” Another review was even more direct: “Execution bottleneck: Oracle signed the contracts, but now it has to actually build the data centers… Execution now is about power, GPUs, and the cadence of embedded regions with AWS, Azure, and Google.” Oracle’s situation mirrors CoreWeave’s perfectly both companies have secured enormous customer commitments, both have available capital, and both are constrained by the physical buildout timeline. The pattern is now undeniable: the AI infrastructure industry has transitioned from capital-constrained to delivery-constrained. And the similarities extend to the recent performance for their stocks. GE Vernova: The Equipment Bottleneck Revealed The infrastructure constraint extends beyond data center shells and power availability. It reaches into the physical equipment supply chain itself. GE Vernova (GEV), the dominant supplier of gas turbines, transformers, and grid equipment for AI data centers, has made the bottleneck explicit. CEO Scott Strazik stated in March 2025 at CERAWeek: “I would expect by the end of the summer, we will be largely sold out through the end of ’28 with this equipment” referring to gas turbines, power transformers, and switchgear. The company currently has 50 GW of gas turbines under contract or slot reservation, with lead times stretching into 2028-2030. But Strazik’s most revealing commentary came in a June 2025 LinkedIn post where he laid out the transformer crisis: Distribution transformers now take up to 2 years to deliver, a 4× increase over pre-2022 levels Large power transformers average 120 weeks, with some orders stretching to 210 weeks (4 years) Lead times for gas turbines now range from 3 to 7 years, depending on model His summary was blunt: “No transformer = no project.” GE Vernova purchases nearly $20 billion in materials and components sourced from over 100 countries. The company faces “supply-chain strains” that have caused its Wind division orders to fall 43% year-over-year as “supply chain bottlenecks and project delays persist.” RBC analyst Chris Dendrinos noted that GE Vernova’s recent $41 million investment in generator capacity was necessar