Inside the $700B Data Center Power Race

Five of the largest data center operators in the world are slated to spend up to $700 billion this year on U.S. data center infrastructure. A significant portion of that capital is targeted at one thing: securing access to power.

New substations. Direct utility offtake agreements. Behind-the-meter natural gas plants. Nuclear partnerships with timelines stretching into the 2030s. The trade press has converged on a single narrative. The bottleneck for AI is power. The race for the next gigawatt is on. Whoever secures the megawatts first wins.

The narrative is half-right. There is a power crisis. Interconnection queues now hold more than 2,060 gigawatts of generation and storage capacity waiting to come online, more than the entire installed capacity of the United States grid. Median wait times have stretched past four years and continue to climb. In some markets, energization timelines now extend to a decade.

But the half of the story almost no one is telling is this: a meaningful portion of the megawatts those operators are racing to build are already inside their facilities. Permitted. Paid for. Energized. And invisible to the metrics the industry has used to manage itself for the past twenty years.

This is stranded power. And the industry's failure to talk about it honestly is becoming one of the largest capital misallocations in modern infrastructure history.

Stranded power, in the data center context, is electrical capacity that a facility has contracted, paid for, and connected to the grid but cannot convert into sustained, useful compute output. The watts arrive at the building. The watts pay their delivery charges. The watts do not produce revenue.

The simplest illustration comes from a recent piece of industry analysis: two data centers operating at the same Power Usage Effectiveness can deliver radically different compute output. Same total facility power. Same overhead ratio. Different productive capacity. The difference is structural, and PUE cannot see it.

What PUE measures is the relationship between total facility consumption and IT load. What PUE does not measure is whether the IT load is actually being converted into work. A rack drawing its full power allocation while throttling under thermal stress is consuming power without producing the compute the operator is being paid to deliver. A cooling system reacting too slowly to a power density spike forces workload shedding upstream. A GPU cluster running below its rated utilization because the facility cannot remove heat fast enough is, by every operational measure that matters, stranded.

This is no longer a theoretical concern. Rack-level power density has increased roughly tenfold in three years, from 10 to 14 kW in conventional configurations to over 100 kW in current AI deployments. NVIDIA's most recent reference architectures push that figure higher still. Most facilities operating today were not designed for that density. They were designed for the world that existed before frontier AI training runs became a normal line item on a hyperscaler's capital plan.

The consequence is that thousands of contracted megawatts across the U.S. industrial base cannot be fully utilized. The power is there. The compute is not.

At Data Center World 2026, executives quantified the gap the industry believes it needs to close. Roughly 200 gigawatts of new AI load demand by the end of the decade, plus another 104 gigawatts of generation retirements happening in the same window. A 300 gigawatt shortfall.

That number is treated as gospel. It drives the entire capital allocation story being pitched to LPs, debt markets, and utility commissions across the country.

But the 300 gigawatt figure assumes that every megawatt of facility power purchased becomes a megawatt of productive AI compute. It does not account for what is being stranded by thermal limits, control fragmentation, and operational drift inside facilities that have already been built. It treats the conversion rate from contracted power to delivered compute as an implicit 1.0, when the empirical evidence from operators willing to measure honestly suggests the ratio is well below that.

What if 10 percent of installed U.S. data center capacity is stranded? At a current sector draw approaching 15 gigawatts and a pipeline that, fully built, could push total demand past 90 gigawatts within five years, the recoverable capacity in already-permitted facilities runs into the gigawatt range. The "gap" the industry is racing to close with new substations, gas plants, and behind-the-meter generation overlaps with capacity that is already paid for.

The capital efficiency implications are stark. New power infrastructure runs $1.5 to $3 million per megawatt for natural gas, multiples of that for nuclear, and timelines of three to ten years from announcement to operational capacity. Recovering stranded power inside an existing facility, by contrast, can bring additional compute online in months. The land already exists. The substation already exists. The permits already exist. The customer relationships, the security perimeter, the operational runbook, all already exist. What does not exist, in most facilities, is the operational discipline to convert what they already have into what they could be delivering.

The metric the industry has used to measure efficiency for almost two decades was built for a different problem. PUE was designed to expose facility overhead and force operators to reduce the power consumed by cooling, lighting, and electrical distribution losses relative to the power consumed by IT equipment. For that purpose, it worked. Global averages fell from north of 2.0 in the mid-2000s to roughly 1.5 by the mid-2010s.

What PUE was never designed to measure is the relationship between IT power consumed and compute actually delivered. A facility that draws its full IT power allocation through chronically throttled GPUs is, by PUE, a successful facility. The number says nothing about whether the work the operator is paid to perform is getting done.

This is why a small but growing chorus of operators and analysts has begun proposing a successor metric. Power Compute Effectiveness, or PCE, reframes the question. Instead of asking how efficiently power flows through the building, it asks how much sustained, usable compute is produced per unit of power consumed under real operating conditions.

PCE makes stranded power visible in a way PUE cannot. The leak does not show up on the dashboard most operators look at. It shows up on the income statement of the compute buyer who is paying for capacity that does not perform.

The industry has not yet adopted PCE at scale. It will. The pressure to do so is coming from AI infrastructure buyers who pay for compute output, regulators who are starting to ask whether public grid capacity is being used efficiently, and investors who are noticing that the relationship between deployed capital and productive output is not what the spreadsheets assumed.

Stranded power persists because the facilities producing it are managed by tools that were never designed to talk to each other. Building management systems handle thermal. DCIM platforms handle assets and inventory. SCADA handles electrical. GPU schedulers handle compute. Each of these systems is local, each is optimized for its own narrow set of metrics, and none of them are responsible for the whole.

The result is sub-optimization at every layer. A BMS that lowers chiller setpoints to protect against an anticipated heat spike does not know that the GPU scheduler is about to shift workload off the affected rack anyway. A power management system that derates a PDU in response to a temperature reading does not know that the cooling system has already responded. A facility runs every component within spec, and the facility as a whole still produces less compute than the contracted power should support.

Operators are not unaware of this. The best of them know exactly where the stranding is happening in their buildings. What they lack is a coordination layer above their existing systems that can make decisions across them in real time. The tools to do this exist in adjacent industries. Process plants, refineries, and modern manufacturing operate with control hierarchies that span every system in the facility and optimize for the outcome the business cares about, not for the local metric each subsystem tracks. The data center industry, for historical reasons, has not yet built that layer.

That gap is now the largest unforced capital efficiency error in the sector.

For the CFO of a data center operator, stranded power is not an engineering problem. It is a balance sheet problem with a P&L expression.

A 30 megawatt facility with a $4 million annual power bill that is stranding 15 percent of its contracted capacity is leaking $600,000 a year in pure overhead. That is the easy line item to see. The harder line item is the revenue that capacity could have produced if it had been converted into compute. At AI infrastructure margins, the foregone revenue can run multiples of the power cost. A facility delivering 85 percent of its contracted capacity to paying workloads is not 85 percent of a fully utilized facility. It is, in operational and financial terms, a facility that is structurally below its own breakeven assumptions.

Now multiply that across an installed base. The U.S. data center sector is approaching 15 gigawatts of total draw with a pipeline that could push past 90 gigawatts within five years. Even a single-digit percentage of stranding represents capital that is contracted, depreciated, and not producing the output the underwriting assumed. The institutional investors who funded the buildout did not model this. The utilities who agreed to provide the capacity did not model this. The boards that approved the CapEx did not model this. And yet the empirical evidence is mounting that it is happening across most of the installed base.

Investors are starting to ask harder questions. The next phase of data center underwriting will price not just power access but power conversion. Operators who can demonstrate that they are converting their contracted megawatts into delivered compute at higher rates than peers will earn capital at better terms. Operators who cannot will increasingly find themselves explaining to LPs and lenders why their effective capacity is less than their nameplate.

This is what stranded power actually means. It is not a thermal problem. It is a capital efficiency problem expressed through a thermal symptom.

The industry's first instinct will be to throw more infrastructure at the problem. More cooling. More liquid. More copper. Faster pumps. Larger CRAH units. Some of that is necessary. Liquid cooling is no longer optional for high-density AI deployment, and operators delaying the transition will pay for the delay in compute they cannot deliver.

But infrastructure alone is not enough. The deeper problem is that even facilities with the right cooling cannot extract their full contracted power if the systems managing thermal, electrical, and compute are not coordinating in real time. The fix is not in any one box. It is in the layer above the boxes, the layer that watches all of them at once and makes decisions across them faster than any human operator could.

That layer is what the industry has historically called operational control. It is what process industries have built into their facilities for decades. It is what data centers, until now, have largely chosen to do without, partly because monitoring felt like enough, and partly because the cost of not having it had not yet shown up clearly in the numbers.

It is showing up now. The operators who close the gap between contracted power and delivered compute will not do it by buying more power. They will do it by becoming the kind of facility that converts the power it already has into the compute it was always supposed to produce.