Who Gets the Megawatt?

How AI Data Centers are Outcompeting American Industry for the Right to Exist

46 States and Still No Power

A site selection consultant recently searched 46 states to find a location for a $4 billion aluminum smelter. She almost came up empty. Not because of land, or labor, or logistics, but because they could not find power.

In a recent article I read on LinkedIn, Didi Caldwell, CEO of Global Location Strategies, who advises major manufacturers on where to build, said in a message to the industry: “We are not short on ambition, but we sure are short on megawatts.”

The electric grid has become the number one constraint in U.S. industrial site selection, ahead of land, labor, and logistics, and yet even where power exists, manufacturers are losing the competition for it.

The competitor is not another manufacturer. It is a demand class that operates on fundamentally different energy economics, one that can not only outbid traditional industry for every available megawatt, but can fund the construction of entirely new megawatts that traditional industry could never afford to build.

The Asymmetry Nobody Has Measured

The Compute Heat Rate™ (CHR)¹ is a metric that quantifies the maximum electricity price that AI data center workloads can profitably sustain, derived from GPU economics. The current blended CHR across major workload types is approximately $6,350/MWh. Individual tiers range from commodity inference at the low end (approximately $250/MWh) to frontier model training and premium inference at the high end ($49,000 to $74,000/MWh+). Check out the full CHR article on Substack.

Now consider the aluminum smelter: aluminum smelting operates on product margins that support electricity costs of roughly $60 to $80/MWh or somewhere in that order of magnitude at least. Above that range, it becomes uneconomic and shuts down. Steelmakers operate similarly, same with chemical plants, paper mills, and other energy-intensive manufacturers.

These are the loads that have historically provided the demand-side brake in wholesale electricity markets. When prices spike, they curtail. Their curtailment reduces demand, which lowers prices, which restores equilibrium. The system self-corrects because the marginal buyer has a clear and relatively low pain threshold.

AI data centers have no comparable curtailment threshold at basically any normal price level. The gap between the two demand classes is 40 to 100 times, depending on the workload tier.

The aluminum smelter curtails at $80/MWh. The data center does not curtail until $250/MWh at the floor, $6,350 blended. In every interconnection queue, every capacity auction, every utility resource plan, the data center outbids the manufacturer. Not because it is trying to. Because the economics are 40 to 100 times different.

410 Gigawatts in Line

The interconnection queue is where the displacement becomes visible.

As of early 2026, ERCOT’s large load interconnection queue reached 410 GW. Texas’s all-time peak demand was 85.5 GW. The queue is nearly five times the capacity of the entire existing grid. Most of those requests are data centers.

Oncor, the transmission utility serving Dallas-Fort Worth and central Texas, has approximately 350 GW of data center interconnection requests against a current system peak of 31 GW.

Even applying aggressive attrition assumptions, the math is daunting. If 80% of the queue never materializes, 70 GW remains. That is still 2.4 times Oncor’s entire current peak, all competing for the same constrained transmission interfaces.

At CERAWeek 2026, Texas PUC Chairman Thomas Gleeson said: “If you look at the wholesale prices in forward markets in ERCOT, they don’t reflect the numbers that are being presented right now. That causes a lot of problems because, obviously, if those numbers are even close to being real, we need a lot more generation in this state, and the prices right now just do not support that.”

The Crowding-Out Spiral

There is a cruel irony in this dynamic.

Data centers require lots of physical materials: steel, aluminum, copper, concrete, specialty glass for fiber optics etc. The facilities that manufacture these materials are energy-intensive industrial operations, exactly the demand class being displaced from the grid by the facilities that need their products.

The risk is that the spiral is self-reinforcing. Data center demand drives up electricity costs in a region. Energy-intensive manufacturers in that region face higher costs and may curtail, relocate, or fail to expand. The physical supply chain for data center construction becomes more constrained. Data center developers pay more for materials and wait longer for delivery. They respond by accelerating their power procurement strategies, further concentrating grid resources toward their needs, further displacing the manufacturers they depend on.

This is already visible here and there. For example, in Louisiana, the Louisiana Energy Users Group has formally opposed elements of the Meta-Entergy deal structure, arguing that allowing tech firms to shoulder only half of infrastructure costs may shift financial risk onto other ratepayers. The area that has defined Louisiana’s industrial economy for decades is now directly competing with hyperscale data centers for the same utility’s generation, transmission, and regulatory attention.

In Illinois, ComEd has petitioned regulators to approve a $15.3 billion grid modernization plan for the Chicago region, where the utility projects data center demand could reach 19 GW by 2030 against a current system load of 23 GW.

In MISO broadly, data center load has grown at 43% annually since 2020, the fastest of any ISO region. MISO was historically a manufacturing-heavy grid: automotive in Michigan, food processing in the upper Midwest, petrochemicals in the Gulf.

What About Reshoring

The timing of this displacement coincides with one of the largest pushes for domestic manufacturing in a generation. Federal policy and bipartisan support for supply chain resilience has created significant incentive for manufacturers to build or expand facilities in the United States.

Many of those facilities are energy-intensive. Unfortunately, these reshoring projects are arriving at the interconnection queue at the same time as data center projects, and they are arriving with fundamentally different economics. The data center hypothetically brings more revenue per megawatt, more capital investment per project, more tax revenue per acre, and more willingness to fund infrastructure improvements.

This is not a market failure. It is a market working exactly as designed, allocating scarce resources to the highest bidder. The question is whether the highest bidder is the same as the highest-value use measured in industrial resilience, job growth, supply chain security, and long-term economic diversification. That is a policy question, not a market question. But it cannot be answered intelligently without first understanding the magnitude of the economic asymmetry. CHR provides that understanding.

What CHR Makes Visible

The crowding-out dynamic described above is happening without anyone having a formal metric for the economic gap driving it. Utilities know data centers pay well. Manufacturers know they are struggling. Policymakers know there is tension. Everyday people, including random friends and family, read a few LinkedIn posts I’ve written, and totally get the vibe even though they’re not at all in the energy industry. But nobody has quantified the structural asymmetry that makes the outcome predictable.

CHR provides this quantification. With it, several things become possible that are currently impossible.

So Who Gets the Megawatt?

The question this article’s title poses is not rhetorical. It is the defining resource allocation question of the American energy economy for the next decade.

Left to market forces alone, the answer is straightforward: the megawatt goes AI compute.

Whether that outcome is desirable is a separate question, one that involves values, priorities, and tradeoffs that markets alone cannot resolve. A nation that allocates all available grid capacity to data centers may lead the world in AI but find itself unable to produce the steel, aluminum, and silicon needed to build the physical infrastructure that AI requires. A nation that reserves grid capacity for traditional industry may preserve its manufacturing base but cede AI leadership to competitors who do not.

These are hard choices. But they are choices that cannot be made intelligently without understanding the economic forces driving them.

Hans Royal is the originator of the Compute Heat Rate™ (CHR) framework. All views are his own and do not represent those of any employer or affiliated organization.

1. Royal, Hans, The Compute Heat Rate: Quantifying AI-Driven Electricity Price Tolerance
and Its Implications for Wholesale Market Repricing (February 28, 2026).
Available at SSRN: http://dx.doi.org/10.2139/ssrn.6322318