North America's GPU Capacity Gap: Where the Next 100 GW of Compute Power Will Come From

1. The Supply-Demand Imbalance

The United States is in the middle of the most compressed infrastructure buildout in modern history. Demand for GPU compute has surged faster than the industry's ability to build the facilities and power systems to support it. The result is a structural supply-demand imbalance that will take years to close and billions in capital to bridge.

AI workloads are the primary driver. Token consumption for large language models has increased approximately 50-fold over a few years, and the end of that growth is not visible. Goldman Sachs projects ¹ that U.S. data center power demand will grow 220% from 2023 levels by 2030. The IEA ² projects global data center electricity consumption will nearly double from 415 TWh in 2024 to 945 TWh by 2030. That figure is roughly equivalent to Japan's entire annual electricity consumption materializing as new load within six years.

On the supply side, GPU hardware availability remains constrained. Lead times for data-center-class GPUs, including the H100 and H200 series, run 36 to 52 weeks ³. Chinese technology companies alone have placed orders for more than 2 million H200 chips for 2026 against an estimated 700,000 units in stock. DRAM supply currently supports only approximately 15 GW of AI infrastructure ⁴, a binding physical constraint. Nvidia reported data center revenue exceeding $26 billion in a single quarter in 2025.

Yet chip supply is not the binding constraint on the buildout. Power is.

2. The Numbers: How Big Is the Gap?

The scale of required infrastructure investment is without modern precedent. Consider the following projections from major research institutions and financial institutions:

The four largest hyperscalers, Alphabet, Amazon, Meta, and Microsoft, collectively invested nearly $200 billion in CapEx in 2024. That figure grew over 40% in 2025 to approximately $381 billion, and in 2026, their combined spending is projected to reach $635 to $665 billion. Microsoft alone is running at an annualized CapEx rate of approximately $145 billion; Amazon is expected to deploy $200 billion.

Every 300 MW AI campus, in practice, requires 390 to 450 MW of total generation capacity once cooling, redundancy, and reliability margins are accounted for, meaning announced capacity figures consistently understate real infrastructure need. Gartner projects ¹² that AI-optimized servers will account for 44% of total data center power by 2030, and nearly two-thirds of all incremental demand.

The McKinsey $6.7 trillion figure represents the investment required across the full stack: AI-ready facilities ($5.2T) and traditional IT expansion ($1.5T). In an accelerated scenario, it rises to $7.9 trillion. These numbers are not projections from optimistic startups; they come from institutions running the capital allocations of the world's largest technology companies.

3. Where Existing Capacity Falls Short

Announced capacity and delivered capacity are two fundamentally different things. The gap between them is widening.

Sightline Climate's February 2026 analysis ¹³, tracking 777 large data centers and AI factories above 50 MW announced since 2024, found the following:

Of the 190 GW announced, only 5 GW was under active construction for 2026 delivery. The rest sits in planning, permitting, or interconnection queues. The 30 to 50% delay prediction for 2026 is not pessimism; it reflects an industry pattern documented in 2025 and driven by three compounding bottlenecks.

Equipment shortages. GE Vernova confirmed in December 2025 ¹⁴ that its order backlog had reached a record 80 GW against current annual output of 20 GW, meaning the company is effectively sold out through 2029. Siemens Energy and Mitsubishi Power show the same constraint for heavy-frame gas turbines. Large transformer lead times have extended to 80 to 120 weeks, with some transmission-class units taking 3 to 6 years.

Grid interconnection delays. Approximately 2,300 GW of generation and storage capacity ¹⁵ is currently waiting in U.S. interconnection queues, roughly double the nation's entire existing generating capacity. Only 13% of capacity requesting interconnection from 2000 to 2019 had reached commercial operation by end of 2024. Projects that became operational in 2025 spent an average of 8 years in the queue. Data centers must be operational in 18 to 36 months; the grid operates on a 5 to 8 year timeline.

Community opposition. At least $64 billion in U.S. data center projects ¹⁶ were blocked or delayed between May 2024 and December 2025, with $18 billion fully blocked and $46 billion delayed. At least 142 activist groups across 24 states are organizing opposition. Data center cancellations quadrupled in 2025. Arizona's Tract abandoned a $14 billion data center after the city refused zoning approval.

The math does not resolve itself: the infrastructure needed by 2030 is real, the capital to fund it exists, but the physical supply chain, grid access, and local permitting cannot absorb it at the required pace.

4. The Power Bottleneck

Power has replaced capital as the binding constraint on AI infrastructure expansion. This is a structural shift, not a temporary disruption.

The evidence is most vivid in the extraordinary lengths hyperscalers are going to in order to secure generation:

Microsoft signed a 20-year, 835 MW Power Purchase Agreement with Constellation Energy to restart Three Mile Island Unit 1 ¹⁷, decommissioned in 2019. The restart cost Constellation $1.6 billion; the U.S. DOE contributed an additional $1 billion loan ¹⁸. Microsoft committed to taking 100% of the plant's output through 2043.

Amazon invested $20 billion or more ¹⁹ to convert the Susquehanna nuclear site into an AI-ready campus and backed 5 GW of new small modular reactor projects via X-energy.

Meta signed a 20-year PPA with Constellation Energy for 1.1 GW of nuclear energy ²⁰ from the Clinton Clean Energy Center in Illinois.

Google signed the first-ever corporate SMR fleet deal in the United States, contracting with Kairos Power for 500 MW ²¹ of small modular reactors.

In aggregate, major technology companies signed contracts for more than 10 GW of possible new nuclear capacity in the United States over a single year. This is not a strategic preference for clean energy. This is what desperation for reliable baseload power looks like when the grid interconnection queue stretches beyond the planning horizon.

Even with nuclear deals, the grid cannot scale fast enough. NERC's 2025 Long-Term Reliability Assessment ²² identified five regions, MISO, PJM, ERCOT, WECC Basin, and WECC Northwest, as facing HIGH reliability risk by 2030. Summer peak demand growth projections increased 69% year-over-year between the 2024 and 2025 assessments. NERC Director John Moura stated that the uncertainty and magnitude of load growth, and its impact on planning, carries "significant risk."

The AEP utility, serving parts of Virginia and Texas, has customer commitments for 24 GW of new demand by 2030 ²³, 18 GW from data centers alone. That is five times the utility's current system size. In a single year, PJM's capacity auction revenues increased $7.3 billion (82%) driven primarily by new data center load, with those costs passing through to all ratepayers.

5. Grid vs. Distributed: A Power Source Comparison

The economics of different power sources for data centers diverge significantly across cost, timeline, and reliability:

The stranded gas figure is not a rounding error. At under $10/MWh fully loaded, wellsite gas-to-power generation undercuts every conventional source by a factor of 3 to 10. The gas acquisition cost at wellsite ranges from 10 to 30 cents per MCF, compared to a market price often above $3/MCF, because at the point of flaring, gas has no market and operators pay nothing to burn it.

Off-grid and behind-the-meter approaches share an additional structural advantage beyond economics: they bypass the interconnection queue entirely. A modular power-plus-compute unit can be deployed in 60 to 90 days at a wellsite versus 5 to 8 years for a grid-connected facility. For AI operators who need capacity now, this timeline advantage translates directly into revenue.

Sightline Climate ³¹ found that on-site and hybrid generation approaches represent less than 10% of announced projects by count but nearly 50% of announced capacity by megawatts, driven by a small number of gigascale, grid-independent campuses. The sector is bifurcating: large off-grid campuses (Pacifico Energy's 7.65 GW GW Ranch permit, Chevron's planned 2.5 GW off-grid facility in West Texas) are capturing the high end, while modular distributed deployments are filling the distributed, behind-the-meter tier.

6. The Role of Stranded Energy

While hyperscalers compete for nuclear deals and grid capacity, a largely overlooked resource sits idle across the oil fields of North America: stranded gas that would otherwise burn in open flares.

The global scale of the waste is striking. Global gas flaring reached 151 billion cubic meters in 2024 ³², the highest level since 2007. The economic value of that wasted gas runs between $30 and $50 billion per year ³³. Combusted in open flares, it generates 389 million tonnes of CO2 equivalent emissions ³⁴ annually, including 46 million tonnes from un-combusted methane. To frame the scale another way: the gas flared globally each year contains enough energy to supply roughly two-thirds of Europe's electricity needs.

North Dakota exemplifies the domestic opportunity. The Williston Basin flared approximately 160 to 175 million cubic feet per day ³⁵ throughout 2024 and 2025, totaling roughly 60 to 67 billion cubic feet annually. North Dakota alone accounted for approximately one-fifth of total U.S. gas vented and flared in 2024 ³⁶. Despite a statewide capture rate of 95%, some legacy and remote field capture rates remain as low as 48 to 58%.

The root cause is structural, not operational. The Bakken gas-to-oil ratio has tripled since 2014 ³⁷, from roughly 1 MCF per barrel to approximately 3 MCF per barrel today. With daily oil production averaging about 1.2 million barrels per day, total associated gas production exceeds 3.4 billion cubic feet per day and continues to grow. Pipeline infrastructure, which requires years and hundreds of millions to build, consistently lags behind production. The EDF's 2025 flaring update ³⁸ explicitly attributes the 2024 U.S. flaring increase to the Bakken, where produced gas volumes exceeded takeaway capacity.

The NDIC Director's Cut data ³⁹ indicates that approximately 17% of oil wells in the state flare between 10 and 50% of produced gas even when connected to gathering networks, due to pressure differential issues. Of North Dakota's roughly 19,266 producing wells, that implies more than 3,200 wells with active or intermittent flaring, across terrain where pipeline tie-ins do not pencil out economically.

The distributed power opportunity from current North Dakota flaring alone is estimated at 20 to 30 MW at remote sites ⁴⁰, with total addressable potential including partially flaring connected wells closer to 125 MW. That figure is modest relative to hyperscale ambitions, but it is immediately available, requires no grid connection, generates sub-$10/MWh power, and sits in a regulatory environment that actively incentivizes flare reduction.

North Dakota has progressively tightened its gas capture requirements since 2014, reaching a current mandatory capture target of 91% ⁴¹. The state passed tax incentives for flare mitigation systems meeting specified capture percentages and gives operators credit for on-site beneficial use of captured gas. These policies create a floor of regulatory demand for distributed power solutions that will persist regardless of federal posture. The EPA Waste Emissions Charge was repealed via Congressional Review Act in March 2025 ⁴², reducing federal financial pressure on flaring, but state-level regulations and voluntary ESG incentives remain active.

Beyond the Williston Basin, the same dynamics play out across the Permian, Appalachian, and DJ Basins, as well as internationally across oil-producing regions from Alberta to the Middle East to sub-Saharan Africa. The global flare gas power generation market ⁴³ was valued at $3.1 billion in 2024 and is growing at 10.6% CAGR toward $6.3 billion by 2031.

7. Who Is Building the Bridge?

The distributed stranded-gas-to-compute sector has evolved rapidly over the past three years. Understanding who is operating in it, and who recently left it, is essential context.

Crusoe Energy pioneered the model. Founded in 2018 in Denver, Crusoe developed "Digital Flare Mitigation" technology, deploying fully containerized, mobile data centers at flare sites, converting wellhead gas to electricity at under $0.01/kWh, and using that power for initially Bitcoin mining and later GPU compute. By 2024, the company had converted more than 10.4 billion cubic feet of flare gas ⁴⁴, generating approximately 1.3 TWh of electricity and avoiding 1.3 million metric tonnes of CO2 equivalent.

Then Crusoe pivoted. In March 2025, Crusoe sold its entire Bitcoin mining and Digital Flare Mitigation business to NYDIG ⁴⁵, transferring 425 or more modular data centers across 7 U.S. states and Argentina representing over 270 MW of power generation. Crusoe reoriented entirely toward hyperscale AI, breaking ground on a 1.2 GW facility in Abilene, Texas ⁴⁶ powered by renewables and conventional natural gas, and raising a $1.375 billion Series E at a $10 billion valuation ⁴⁷ in October 2025.

The exit is significant. Crusoe's departure from distributed flare gas was not a verdict that the model does not work; NYDIG acquired it, validating the technology and economics. It was a choice to chase the larger hyperscale market. That choice left the distributed stranded-gas-to-compute space with a visible gap where its most technically credible operator used to be.

Several companies are filling that gap:

Giga Energy, based in Texas and founded by Brent Whitehead and Matt Lohstroh, operates containerized modular data centers and has installed more than 150 MW of containers across Texas, Shanghai, and Argentina. The company claims build timelines of 6 to 8 months ⁴⁸ versus the industry's standard 24 to 36 months, using pod-based construction. Giga established a joint venture with Atlas Power to deploy stranded and flare gas in the Williston Basin ⁴⁹.

MARA (Marathon Digital Holdings) partnered with NGON to deploy a 25 MW micro data center distributed across wellhead sites ⁵⁰ in Texas and the North Dakota Bakken. Sites began energizing in September 2024 and were fully operational by January 2025. Methane mitigation efficiency was reported at up to 99% versus a typical flare's 92%. MPLX, Marathon Petroleum's midstream subsidiary, extended the same model to the Permian Basin through a collaboration on integrated power generation and data center campuses in West Texas.

Applied Digital (Nasdaq: APLD) operates 286 MW in North Dakota grid-connected facilities in Jamestown and Ellendale and is exploring behind-the-meter natural gas generation with Babcock & Wilcox for up to 1 GW ⁵¹ of capacity. Their Polaris Forge 1 facility in Ellendale is fully leased to CoreWeave for an estimated $11 billion in lease revenue over 15 years.

New West Data Corp, based in Alberta, Canada, is deploying generators at uneconomic "orphan" wells to convert stranded gas to on-site compute. The company has reported 300+ bopd of oil production combined with 15 MW of Bitcoin mining ⁵² at 50% lower cost than competing miners, and claims potential for 10 GW of power capacity from orphan well gas without straining the public grid.

NYDIG, which acquired Crusoe's DFM business, now operates the largest single portfolio of distributed flare gas compute units in North America. As a subsidiary of Stone Ridge, which manages a 10 GW natural gas portfolio, NYDIG has financial infrastructure to scale the distributed model further.

The pattern across these operators is consistent: modular, skid-mounted power generation units deployed directly at stranded gas sites, with compute loads running on the produced electricity. The fundamental economics remain unchanged from what Crusoe proved at scale: gas at near-zero cost, power at sub-$10/MWh, and compute revenue that monetizes electrons at 5 to 10 times the value of simple electricity sales.

8. The 100 GW Question

The question of where the next 100 GW of compute power comes from does not have a single answer. The U.S. data center buildout will ultimately draw on every available generation source, and the mix will vary materially by geography, timeline, and workload type.

The realistic supply mix through 2030:

Natural gas: the largest single source. As of 2024, natural gas supplies over 40% of electricity for U.S. data centers ⁵³ and will remain dominant through 2030 due to its 24/7 dispatchable character. Goldman Sachs projects approximately 60% of the 82 GW in new capacity ⁵⁴ required to come from natural gas peaking plants. The bottleneck here is not the gas itself but the turbine supply chain: GE Vernova's backlog alone represents a 4-year production constraint.

Solar: fast to deploy, cannot stand alone. Goldman Sachs projects approximately 27% of required new capacity will come from solar. Solar PPAs are available and deployable within 18 to 36 months, but their 20 to 30% capacity factors mean solar cannot serve the 24/7 power demands of AI compute without substantial storage or firm backup. Solar PPA prices reached $61.67/MWh in Q4 2025 ⁵⁵, up 9% from the prior year.

Nuclear: long-term credibility, near-term constraints. Nuclear provides the firmest possible baseload at large scale and has captured the imagination of every hyperscaler with power problems. But the pace of new nuclear construction is measured in decades, and even nuclear restarts take 3 to 7 years. The Microsoft/Three Mile Island, Amazon/Susquehanna, and Meta/Clinton deals represent the realistic near-term ceiling. Goldman Sachs projects 85 to 90 GW of new nuclear will be needed globally by 2030 to meet data center demand, with less than 10% currently available.

Off-grid gas: the fastest path to firm power. Large off-grid gas campuses, including Pacifico Energy's 7.65 GW GW Ranch project in West Texas ⁵⁶, which received the largest air permit ever granted in the United States, and Chevron's planned 2.5 GW off-grid facility ⁵⁷, bypass interconnection queues entirely. These projects can deliver first power in 24 to 36 months, compared to 5 to 8 years for grid-connected facilities.

Distributed stranded energy: the underestimated tier. Distributed wellsite generation will not close the 100 GW gap alone. North Dakota's 125 MW addressable flare gas opportunity is modest relative to hyperscale demand. But distributed stranded energy serves a distinct function: it delivers the cheapest power available, to locations where grid power is unavailable, on the fastest possible timeline. For operators who need capacity outside the traditional data center markets, including those serving government, defense, or edge compute customers, this tier is uniquely positioned.

The investment required to build the 100 GW is substantial but not incomprehensible. McKinsey's central estimate of $6.7 trillion in global data center investment through 2030 ⁵⁸ works out to roughly $1.1 trillion per year. Hyperscaler CapEx alone will approach that figure in 2026. The capital is available; the constraints are physical and logistical.

The most likely 2030 scenario is not a clean resolution. The NERC 2025 LTRA ⁵⁹ found that summer peak demand could surge by 224 GW, 69% more than the 132 GW projected a year earlier. Those projections carry compound annual growth rates described as the highest since NERC began tracking in 1995. Against that demand trajectory, even the aggressive mix of gas, solar, nuclear, and off-grid generation will leave pockets of scarcity.

The projects that close those pockets will share several characteristics: they will be fast-to-deploy, they will not depend on grid interconnection queues, and they will draw on energy sources that existing infrastructure cannot reach. In the Williston Basin and basins like it, that means the gas that burns in flare stacks today.

The infrastructure gap will not be bridged by a single technology, a single company, or a single regulatory decision. It will be bridged by a parallel buildout across every viable power source, with the fastest and cheapest sources serving as the critical foundation while larger-scale infrastructure catches up. Distributed stranded energy is not the only answer to the 100 GW question, but for operators who cannot wait a decade for the grid, it is often the first answer that works.

Sources cited inline throughout. All data current as of March 2026.