AI Power Consumption Drives $725B Infrastructure Spend as Data Centers Enter Unprecedented Power Race

AI's power crunch is sending data-centre builders to the banks

Electricity, not chips, is now the binding constraint on AI, and the scramble to secure it is driving builders to seek billions in financing pledges. The bottleneck in AI is no longer chips. It is power, and the scramble to secure it is sending data-centre builders to the banks for billions in financing commitments, Bloomberg reports.The shift is a big one. For much of the AI boom the scarce resource was Nvidia's GPUs; now it is electricity, along with the contracts, substations, and financing needed to deliver it at scale. The numbers are staggering. The four largest cloud firms are set to spend around $725bn on AI infrastructure this year alone, and single sites now carry price tags that would once have funded a whole company, like Meta's $13bn Texas data centre. Power is what makes those sites so expensive. Server racks that drew about 3kW for ordinary computing now pull up to 150kW for AI, and global data-centre capacity is expected to nearly double to roughly 200GW by 2030. That is where the banks come in. Builders want firm financing pledges before they commit to power deals and construction contracts, and the sums are large enough that no single balance sheet wants to carry them alone. The borrowing has already reshaped credit markets. Lenders have poured into the sector, from a $5.9bn loan for one data-centre operator to a wave of bonds and private-credit deals feeding the build. The financing is getting creative, and opaque. Oracle's $16.3bn data-centre raise leaned on private credit after banks grew cautious, a sign of how far off the traditional path the money now travels. Much of it sits out of plain view. Analysts estimate the sector is carrying around $1.65tn in off-balance-sheet obligations, structured through special vehicles that keep the debt off the tech giants' own accounts. The hunt for power is reshaping energy markets. Builders are reviving gas plants, signing nuclear deals, and jumping utility queues, turning data centres into some of the largest new electricity buyers in a generation. The strain is landing on everyone else. Grids are creaking under the new load, and in some regions households are already paying more as data centres compete for the same electrons. Regulators are watching uneasily. The Bank for International Settlements has flagged 'circular financing' between cloud firms, their suppliers, and construction lessors as one of the bigger risks to financial stability. The building itself is a chokepoint. Fewer than ten firms in the world can deliver a hyperscale project, and one of them, Turner, is sitting on a record backlog of about $44bn with a large share tied to data centres. Banks are not lending blindly, either. As the sums swell and the structures grow more circular, some lenders have pulled back, which is part of why builders are chasing firm pledges rather than assuming the money will be there. Power is being locked up years ahead. Builders are signing long deals for electricity, gas, and even nuclear output, and the bank pledges are what make those commitments credible. Underneath it all is a bet that demand shows up. The spending only pays off if AI usage keeps climbing fast enough to fill the capacity, which is why every quarter of AI revenue is now watched so closely. There is a self-reinforcing quality to it, too. The more money commits to the build, the more the industry needs AI to keep growing to justify it, which is precisely what makes the circularity regulators fear so hard to unwind. The scale is historic. Goldman Sachs sees cumulative spending on this buildout running into the trillions by 2030, a figure that dwarfs past infrastructure booms and leaves little room for error. For now, the money keeps flowing toward the wall socket. The race for compute has become a race for power, and the winners may be decided less by who has the best models than by who can finance the electricity to run them.

The power problem behind the AI boom: Why data centers are racing to upgrade their electrical infrastructure

A GPU cluster can be ordered, shipped, and racked in well under a year. Energizing the building that holds it can take three. That gap sits behind almost every data center headline of the past two years. Capital is not the constraint. Silicon, for the larger buyers, has mostly stopped being one. What actually decides whether a site opens on schedule is electricity, plus everything standing between a substation and a server cabinet: transformers, switchgear, breakers, busway, and an interconnection agreement that has to be signed before any of it means anything. A single rack now pulls what a small office used to For roughly two decades the planning numbers were comfortable. A typical enterprise rack drew five to ten kilowatts, and a hall designed around 100 to 150 watts per square foot could absorb whatever tenants brought in. Facilities were built with headroom because headroom was cheap. AI erased that headroom in about one product cycle. Dense accelerator racks now land somewhere between 40 and 130 kW, and vendor roadmaps discuss cabinets measured in the hundreds of kilowatts as a near-term expectation rather than a thought experiment. Density is not only a cooling problem, though cooling is the part that gets photographed. Each step up in per-cabinet power pushes work back upstream, onto distribution equipment that was specified for a building with a different purpose. The waiting list that starts long before the utility says yes Operators tend to discover the equipment problem in the wrong order. The interconnection study gets attention first, because it has a formal process attached and a queue position that can be quoted to investors. Then procurement comes back with lead times. Large power transformers are the usual shock. Buyers who were quoting under a year before 2020 have spent recent cycles quoting in multiples of that, and the constraint is not simple to unwind: grain-oriented electrical steel, bushings, tap changers, skilled winding labor, and factory floor space that cannot be conjured in a quarter. For a voltage transformer manufacturer, expanding production is far more complex than simply adding another assembly line. Medium- and high-voltage transformers require specialized winding equipment, experienced personnel, lengthy qualification processes, and years of manufacturing know-how. That makes production capacity slow to expand even when demand surges, which is one reason lead times across the industry have stretched so dramatically. Utilities running long-term replacement programs often secure manufacturing capacity years in advance, leaving data center developers competing for available production slots rather than for transformers themselves.. Switchgear, breakers, generators, and even basic busway have stretched in sympathy. So developers started doing something that would have looked reckless in 2019: ordering transformers speculatively, against sites that are not fully designed, and warehousing them until a layout catches up. It is an ugly use of capital. It also works, and it has quietly become normal practice among the larger builders. AI training loads behave less like a data center and more like a steel mill Classic data center load is boring in the best way. Thousands of uncorrelated workloads average out, and the facility draws a fairly flat curve that utilities find easy to plan around. Training clusters can break that assumption. Tens of thousands of accelerators working through synchronized steps may ramp together and, in some large deployments, swing a substantial share of site load within a very short window when a job stalls, checkpoints, or fails outright. How pronounced this gets depends heavily on the scheduler, the framework, and how the operator has configured power management, so the behavior varies more between sites than the headlines suggest. Grid operators in several markets have flagged it regardless, and interconnection agreements increasingly carry ramp-rate limits, ride-through obligations, and power quality conditions that would have been unusual for a data center a decade ago. For the electrical room, this changes what "correctly sized" means. Equipment gets specified against step loads and harmonic content rather than a steady thermal rating. Battery systems installed as a bridge to generator start are being asked to smooth load swings as a routine function, and some operators now shape workloads in software specifically so the electrical side sees a gentler curve. Utilities notice which campuses can promise predictable behavior. It is a far easier interconnection conversation than the alternative. Why copper, not silicon, sets the limit inside the rack Here is the arithmetic that quietly rewrote rack design. Power is voltage multiplied by current, so at a fixed voltage, more kilowatts mean proportionally more amps. A 100 kW cabinet fed at 415 V three-phase needs roughly 140 A coming in, which is manageable. The problem starts after the power supplies convert it down. On a traditional 54 V DC busbar inside the rack, that same 100 kW works out to something in the region of 1,850 A, and every one of those amps has to travel through copper, connectors, and contact surfaces that heat up in proportion to resistance. Copper does not scale gracefully. Doubling the current means roughly quadrupling the resistive losses, so the conductor has to get dramatically thicker to keep the heat under control, which makes it heavier, more expensive, and harder to route through a cabinet already crowded with liquid cooling manifolds. At megawatt-class racks the busbar stops being a component and starts being a structural element. Raising the voltage is the only lever that actually moves. At 800 V, the same 100 kW draws about 125 A instead of 1,850, which is why higher-voltage DC distribution went from a niche argument to a mainstream design direction in the space of two product generations. The physics did not change. The power density finally made the tradeoff obvious. Three routes around the interconnection queue Most large projects combine at least two of the following, and rarely rely on one. What the table cannot show is how often the choice gets made under duress, and the public record is full of examples. xAI's Memphis site is the clearest version of route two. Rather than wait for grid capacity, the company installed gas turbines on the property to run the cluster, and then spent months answering air permit questions from regulators and local groups. Fast, and expensive in ways that do not appear on the electrical drawings. Amazon took route three when it bought a data center campus adjacent to the Susquehanna nuclear plant in Pennsylvania, paying a reported several hundred million dollars for a site whose main asset was proximity to generation that already existed. Microsoft's twenty-year agreement with Constellation to restart the shuttered unit at Three Mile Island belongs to the same logic: buying output from generation nobody has to newly interconnect. The pattern across all three is worth noting. None of these deals were about electricity being expensive. They were about electricity being unavailable on a schedule that matched a hardware order. The retrofit question colocation tenants ask too late Existing halls are where the mismatch gets uncomfortable. A tenant with a modest allocation of GPUs will ask about adding a few high-density racks, and the answer often has nothing to do with the racks. The floor may not carry the weight of liquid-cooled cabinets. The cooling loop may not exist at all. The distribution may be sized for a per-cabinet average the new hardware clears in the first row. And then there is the failure mode nobody expects, which is that the protection equipment becomes wrong in the other direction. Feeding a denser hall usually means a bigger transformer with lower impedance, and lower impedance means more available short-circuit current at the fault. Switchgear that was properly rated when it was installed can end up under-rated for the fault current the new supply can deliver, which is not a performance issue but a safety one. The fix ranges from current-limiting reactors to splitting the bus to replacing the lineup entirely, and none of those are cheap surprises to find halfway through a retrofit. Retrofits that succeed usually take a zone approach, carving out a high-density pod with its own distribution rather than lifting the whole floor. That also matches how the hardware actually arrives, which is to say gradually, and rarely in the order originally planned. Partial upgrades are not elegant. They are frequently the only version that gets built. Where the race is actually being run For most of the last decade, the binding question was who could get the chips. That question has an answer now, and it turned out to be the easy one. The harder one is being settled in transformer factory order books, in interconnection queues, and in permit hearings held in counties most people could not find on a map, and it determines which of the announced gigawatts ever switch on. The next stretch of the AI race will be decided less by who designs the faster chip than by who can get power to it first.

Data centers now face a power race: AI racks need 50 kW or more

Gigawatt campuses, grid delays and liquid cooling are redefining expansion Data center infrastructure has moved into a different phase. This isn't the kind of steady server growth the industry used to plan around. It calls for far more electricity, cooling, land, and supporting infrastructure than traditional expansion ever did. Older central processing unit (CPU)-focused facilities were usually built for 5 kW to 10 kW per rack. Racks loaded with accelerators now often need 50 kW or more, and some designs are already heading toward 100 kW. Power has become the main bottleneck. Data center electricity use has climbed about 12% a year over the past five years, could reach 565 TWh in 2026, and may go past 1,200 TWh by 2030. The server side is shifting even faster: accelerated server platforms alone are expected to jump 84% in 2026 and overtake conventional hardware by 2027. Those power and infrastructure requirements are pushing operators toward gigawatt-scale campuses. To make those sites work, they have to bring together power, water, cooling, fiber, permitting, and community support all at once. Grid interconnection delays and limited capacity are making that harder, so some operators are starting to look at on-site generation, off-grid designs, or hybrid energy. Cooling is changing just as fast. Air systems are running into their limits, and direct-to-chip or immersion cooling is becoming a core option for 100 kW-class racks. NVIDIA's latest infrastructure is reportedly fully liquid cooled. If you work in data center planning, this is a shift to keep close track of, especially with the market projected to grow from $147.3 billion in 2025 to $810.6 billion by 2033, while cost and sustainability pressure keep rising across global hyperscale and AI facilities.