The AI industry's power crisis just shifted from theoretical risk to immediate business problem. Two of the world's largest AI infrastructure operators hit real friction on the exact same constraint in the same week: getting enough electricity to a data center on schedule. Crusoe Energy walked away from a $1.25 billion order for jet-derived turbines from Boom Supersonic, while Oracle filed a force majeure notice on its $165 billion Stargate data center campus in New Mexico after a natural gas pipeline meant to power the site was delayed nearly six months. These aren't isolated stumbles; they're signals that the infrastructure buildout underpinning artificial intelligence has hit a wall that money and engineering alone cannot quickly solve.

Why Did Crusoe Cancel a Billion-Dollar Turbine Order?

Nine months ago, the Boom Supersonic deal looked like the workaround the industry had been waiting for. Crusoe, the AI cloud company building Oracle's Stargate campus in Abilene, Texas, signed on as the launch customer for Boom's 42-megawatt Superpower turbines, stationary spinoffs of the jet engines Boom is developing for its Overture supersonic aircraft. The agreement was worth $1.25 billion and was bundled with a $300 million Series B round for Boom when it was announced in December 2025, with first deliveries expected in 2027.

The appeal was straightforward: jet-derived turbines could be trucked in and fired up off-grid in months, bypassing the years-long wait for utility interconnection that typically requires new substations and transmission upgrades. Natural gas, not exotic aviation engineering, was always the actual power source. Boom was simply repurposing an engine it was already building for commercial aviation and selling it as a stationary power plant.

But on September 26, 2026, Boom CEO Blake Scholl announced on X that turbines are no longer part of Crusoe's near-term primary power mix at Abilene and its other sites, so a launch partnership no longer made sense. Crusoe framed the reversal as a shift toward flexibility, pointing to wind, solar, batteries, turbines and ordinary grid power as options it can choose site by site instead of a single, expensive, fixed-infrastructure bet.

The timing, however, tells a different story. Crusoe announced the initial close of a $3.9 billion Series F round on September 17, 2026, at a $30.9 billion valuation, led by Atreides Management, Mubadala Capital and Valor Equity Partners. Earlier this year, the company paused work on a planned large-scale AI campus in Cheyenne, Wyoming. A company that just raised almost $4 billion is not retreating from turbines and a Wyoming campus because it is short on cash. It is retreating because the fixed, capital-intensive infrastructure bets that defined the AI buildout's first phase are looking riskier than they did a year ago, even to the companies flush with new money.

What Is the Real Bottleneck Blocking AI Data Centers?

The conventional wisdom says the United States is running out of electricity. The reality is more specific and more troubling: the country is running out of places to deliver it. National generation keeps growing, and the tighter limit is deliverability, whether transmission lines, transformers, substations and the local network can carry power to one particular site.

AI data centers differ fundamentally from conventional industrial customers in ways that break decades of utility planning assumptions. Consider how a utility plans for a steel mill or chemical plant: the customer picks a site for raw materials, rail access and labor pool; it builds to a known engineering schedule; it stays for decades; and it runs a load profile the utility can estimate from similar plants elsewhere. An AI campus inverts most of that. The developer picks a site largely because power might be available there, so it often files requests with several utilities at once and keeps the one that answers fastest. Phases get energized in stages, and a hyperscaler can cancel a lease or slow a build when chip supply or financing changes.

The scale of this demand is what turns these traits into a planning crisis. Data centers already draw about 5 percent of U.S. electricity, and the Electric Power Research Institute expects that share could triple by 2035. In PJM, the grid operator covering much of the mid-Atlantic and Midwest, data centers make up 94 percent of projected peak-load growth, according to Latitude Intelligence's 2026 analysis. When one customer type explains nearly all of the growth in the country's largest power market, forecasting errors about that customer become forecasting errors about the whole system.

How Do Utilities Evaluate Whether an AI Data Center Can Actually Get Power?

Power delivery involves five separate locks that an AI data center must clear, and most projects stall at the third and fourth. Understanding these constraints explains why announcements of gigawatt-scale campuses often omit critical details like the name of the substation, the delivery date for transformers, and who pays if the building never fills with GPUs.

• Energy Adequacy: Does the region produce enough megawatt-hours over a year to cover the new load? This is the easiest lock to clear and the one most headlines focus on.
• Capacity Adequacy: Is there enough firm supply available at the single worst hour, usually a summer or winter peak? This is also relatively straightforward to assess.
• Deliverability: Can transmission move that supply to the specific node where the campus connects? This is where most projects stall because transmission lines are often full.
• Local Grid Strength: Can the nearby substation, transformers and distribution network handle hundreds of megawatts, including sudden swings, without voltage or stability problems? Large power transformers are custom-built, heavy and slow to procure, with timelines stretching to roughly two to three years.
• Cost Recovery: If the utility builds upgrades, who pays, and what happens if the load never arrives? This is where the politics live, and where utilities are rewriting their rate books to require collateral, minimum payments and curtailment rights from data center operators.

The deliverability gap shows up in concrete numbers. Day-ahead congestion costs in six U.S. markets outside California hit $11.1 billion in 2025, roughly a third higher than the year before. Congestion cost is what the market pays when the cheapest power cannot flow to where it is needed because lines are full. The other side of that coin shows up as wasted renewable output: 23.8 million megawatt-hours of curtailed renewable generation across six ISO/RTO markets through July 2026, up 17.5 percent on the same stretch of 2025. Wind and solar farms are being told to stop producing in one place while buyers elsewhere pay premiums.

What Does the Data Center Infrastructure Market Look Like as Companies Redesign for AI?

While power delivery remains the binding constraint, the broader data center infrastructure market is undergoing a massive transformation. The Data Center Infrastructure Market was valued at $269.86 billion in 2025 and is projected to reach $795.17 billion by 2035, reflecting a 10.3 percent compound annual growth rate from 2026 to 2035.

The market is shifting from component-based procurement toward integrated infrastructure architectures. AI workloads are increasing requirements for rack-scale compute, liquid cooling, high-bandwidth networking, power optimization and infrastructure management. Providers are increasingly positioning complete systems, validated architectures and deployment services rather than isolated products.

Hardware accounted for $171.99 billion in 2025 and is projected to reach $471.02 billion by 2035. Software is projected to increase from $41.16 billion to $152.52 billion, recording the highest projected growth rate at 12.9 percent. Services are forecast to expand from $56.71 billion to $171.62 billion during the same period. The rapid growth in software and services reflects the increasing importance of infrastructure management, orchestration, optimization and operational intelligence alongside physical data center assets.

Major technology companies are reshaping their roles within this market. NVIDIA is extending from accelerated computing into rack-scale AI factories and full-stack data center architecture. In May 2026, NVIDIA announced that its Vera Rubin platform was ramping into full production for next-generation AI factories, combining Rubin GPUs, Vera CPUs, ConnectX-9 SuperNICs, BlueField-4 DPUs and networking technologies into rack-scale systems. The platform is being manufactured through a supply chain spanning more than 350 factories across 30 countries, with the objective of supporting large-scale AI infrastructure deployments by cloud providers and hyperscalers.

Supermicro is integrating servers, racks, liquid cooling, software and deployment services into modular building blocks. In July 2026, the company announced a range of precision-engineered AI racks designed for high-density deployments, with factory pre-assembly, modular construction and support for integrated liquid-cooling components. The company introduced ten rack models across multiple configurations and highlighted load capacities exceeding 5,500 pounds for higher-density computing deployments. The company also introduced additional rear-door heat exchanger options supporting cooling capacities from 10 kilowatts to 120 kilowatts.

Lenovo is expanding liquid-cooled infrastructure and gigawatt-scale AI factory solutions. In January 2026, Lenovo announced an expansion of its Neptune Liquid Cooling technology deployment, with the solution using warm-water cooling at the processor level designed to improve performance and energy efficiency for high-density computing. The company also expanded its AI data center strategy through its Lenovo AI Cloud Gigafactory with NVIDIA program, announced in January 2026, which supports rack-scale AI infrastructure based on NVIDIA Blackwell Ultra and Vera Rubin architectures.

What Does This Mean for the AI Infrastructure Buildout?

The simultaneous cancellation of Crusoe's turbine order and Oracle's force majeure notice on Project Jupiter reveal that the AI power crunch is not a future risk analysts are modeling. It is already forcing well-funded, well-advised companies to cancel billion-dollar contracts and file legal notices on multibillion-dollar campuses, in the same month, over the same bottleneck.

Brookings researcher Stijn Van Nieuwerburgh estimated that the coming AI infrastructure spend will reach $10.3 trillion through 2032, a figure larger relative to the economy than the railroad or highway buildouts that came before it. Power, not chips or capital, is turning out to be the constraint nobody fully priced in.

Crusoe still controls flexible power sources and a war chest nearly four billion dollars deep. What it no longer has is faith that a fleet of jet engines bolted to the ground was the fastest way to keep its data centers lit. That is a narrower bet than the one it made in December 2025, and a more honest one. The question now is whether other hyperscalers will reach the same conclusion, and how quickly utilities and regulators can adapt their planning frameworks to accommodate a customer type that breaks every assumption built into decades of grid infrastructure design.