The Modular Data Center Opportunity

Power is the bottleneck for AI infrastructure. Finding 2-5 MW in a remote location takes months. Finding 500 MW near a major metro takes years, if it happens at all. Of all generation capacity that applied for grid interconnection in the U.S. from 2000 to 2019, only 13% became operational by end of 2024. 77% was withdrawn. ^{[1]Lawrence Berkeley National Laboratory, "Queued Up: 2025 Edition" (2025)https://emp.lbl.gov/queues} Modular data centers, prefabricated GPU facilities deployed to wherever power is available, are how a growing number of operators are sidestepping the interconnection queue.

Why power decides where AI gets built

As of end 2024, roughly 10,300 projects were waiting to connect to the U.S. power grid. ^{[1]Lawrence Berkeley National Laboratory, "Queued Up: 2025 Edition" (2025)https://emp.lbl.gov/queues} The interconnection queue, the formal process for connecting a new facility to the electrical grid, has become the single biggest chokepoint in AI infrastructure. Of 110 data center projects slated for 2025, 26% were delayed. ^{[2]Sightline Climate, "Data Center Outlook: Half of 2026 Pipeline May Not Materialize" (2026); Latitude Media coveragehttps://www.latitudemedia.com/news/up-to-half-of-the-worlds-data-centers-may-be-delayed-this-year/} Up to half of planned data centers may face delays in 2026, driven by power constraints, community opposition, and grid equipment shortages.

Smaller power requests, in locations where generation already exceeds local demand, move faster through the queue. Remote locations with stranded power assets (hydroelectric dams, natural gas wellheads, wind farms, retired industrial sites with grid connections) offer capacity that no one else wants. For AI infrastructure operators, that overlooked capacity is an opportunity.

The sub-$100M GPU cluster operators are the natural fit for modular deployment. These operators deploy clusters of a few hundred to a few thousand GPUs, total project costs under $100 million, serving both training and inference workloads. Enterprises with large training jobs (10,000+ GPUs) typically need tightly interconnected clusters at 50+ MW, which rules out most modular sites. Smaller clusters can handle fine-tuning, small-scale training, and inference, all workloads where 2-5 MW is enough.

What a modular data center actually is

A modular data center is a prefabricated, self-contained compute facility built in a factory and shipped to site. Power distribution, cooling, fire suppression, and rack space are assembled offsite, tested as a unit, then transported by truck or rail to wherever power is available.

Typical configurations range from 1 to 5 MW per module. Some vendors ship ISO container-sized units (20 or 40 feet). Others build larger systems or prefabricated buildings that bolt together on site.

Deployment timelines run months from site preparation to first workload, compared to years for a traditional ground-up data center build. The modules are also stackable: an operator starts with one 2 MW module, proves the economics, and adds modules as demand and power availability allow.

There are trade-offs. A modular facility at a remote site won't match the network connectivity, staffing depth, or physical security of a Tier III colocation facility in Ashburn, Virginia. Large-scale AI training across thousands of nodes also favors larger purpose-built facilities. For AI inference, fine-tuning, and training runs under 1,000 GPUs, modular sites work.

Who tried this before AI

The modular data center concept has been attempted three times in the last two decades, each time for traditional web and enterprise workloads.

Sun Microsystems launched Project Blackbox in October 2006, a standard 20-foot shipping container packed with servers, storage, and networking. ^{[3]Sun Microsystems, Project Blackbox announcement (October 2006)https://en.wikipedia.org/wiki/Sun_Modular_Datacenter} Sun marketed it as a portable data center you could drop anywhere with power and a network connection. The product shipped but never gained traction. Data center operators at the time had no shortage of real estate or power. The constraint was compute demand. Sun was acquired by Oracle in 2010, and Project Blackbox was discontinued.

Microsoft took the idea further. In December 2008, Microsoft engineers Michael Manos, Daniel Costello, and Christian Belady described a "Generation 4" modular data center architecture that used pre-assembled containers as building blocks. ^{[4]Microsoft, "Our Vision for Generation 4 Modular Data Centers" blog post by Michael Manos, Daniel Costello, and Christian Belady (December 2008)https://loosebolts.wordpress.com/2008/12/02/our-vision-for-generation-4-modular-data-centers-one-way-of-getting-it-just-right/} Microsoft deployed container-based systems at its Chicago data center (opened 2009) and later facilities. Google followed a similar path, using container-based designs for some of its early data centers between 2005 and 2009. ^{[5]Google, "Efficient Computing" data center blog: container-based data center designs (2005-2009)https://www.google.com/about/datacenters/}

Both Microsoft and Google eventually moved away from containers toward purpose-built modular designs. The shipping container form factor proved awkward: limited rack density, difficult maintenance access, and cooling challenges in a sealed metal box. The modular concept survived but the container form factor died.

HP offered the POD (Performance Optimized Datacenter), Schneider Electric sold prefabricated modular data centers, and companies like Elliptical Mobile Solutions and SGI (now part of HPE) sold containerized units through the 2010s. None reached mass adoption. Traditional data center construction, while slow, was adequate for the pace of demand. Enterprise IT workloads grew predictably. Operators could plan 18-month builds without losing customers.

AI changed it up. GPU clusters now draw 40-80 kW per rack, 3-8x traditional IT. Demand is growing faster than the grid is ready for. Only small deployments where there are pockets of power available can be built quickly.

The "edge" theory vs. the power reality

For years, the data center industry predicted that "edge computing" would drive a network of small facilities close to end users. The thesis: latency-sensitive applications (autonomous vehicles, AR/VR, real-time gaming) would require compute within 5-10 milliseconds of the user, forcing operators to build thousands of small data centers at the edges of population centers. Companies like Vapor IO, EdgeConneX, and Lumen invested heavily in edge infrastructure.

Operators are now building distributed networks of small facilities, but less because of latency. The location decision is driven by "where can we source energy?" and "what can we afford?"

A 2 MW modular data center in rural Texas running inference on B200 GPUs does not need to be within 10 milliseconds of its users. Inference API calls tolerate 50-200 milliseconds of network latency without noticeable impact on user experience. A developer in San Francisco sending a prompt to an API endpoint in West Texas will not notice the extra 30 milliseconds of network transit.

A site with $0.03/kWh electricity from stranded natural gas or surplus hydro beats a site with $0.08/kWh grid power in a metro area, even with better network connectivity. The location decision comes down to two questions: where can I get 2-5 MW without a 3-year interconnection queue, and what does that power cost?

Who is building modular AI data centers

For a broader look at the compute landscape, see who is building compute.

JLL projects the modular data center and micro DC market will reach $48 billion by 2030, growing at 35% annually. ^{[6]JLL, "Modular Systems and Micro DCs Could Fetch $48 Billion by 2030" (2025)https://w.media/special-feature-modular-systems-and-micro-dcs-could-fetch-us-48-billion-by-2030-jll/} McKinsey estimates that prefabricated and modular solutions already make up 40-60% of data center components on average, with leading operators at 80-85%. ^{[7]McKinsey, "Scaling Bigger, Faster, Cheaper Data Centers with Smarter Designs" (2025)https://www.mckinsey.com/industries/private-capital/our-insights/scaling-bigger-faster-cheaper-data-centers-with-smarter-designs} The market has two tiers. The top tier is building at massive scale, hundreds of megawatts across multiple sites. Below them, a long tail of smaller operators is deploying 2-10 MW modules at individual sites.

Crusoe raised $1.375 billion in a Series E at a $10+ billion valuation in October 2025, after a $600 million Series D at $2.8 billion in December 2024. ^{[8]Crusoe, Series E: $1.375 billion at $10+ billion valuation (October 2025); Series D: $600 million at $2.8 billion valuation (December 2024)https://crusoe.ai/} Crusoe started by deploying modular data centers at oil wellheads, using stranded natural gas (gas that would otherwise be flared or vented because no pipeline exists to move it) to power GPU compute. The company secured $11.6 billion for a much larger, more traditional 1.2 GW campus in Abilene, Texas as part of the Stargate initiative, and has since expanded to grid-connected renewable sites.

IREN (formerly Iris Energy) operates 810 MW across six sites in North America, all powered by 100% renewable energy or renewable energy credits. ^{[9]IREN, investor presentations and site data (2025-2026)https://iren.com/} IREN started with Bitcoin mining, which established the operational playbook for compute: find cheap renewable power in remote locations, deploy modular compute, and run 24/7.

Applied Digital signed a $5 billion AI factory lease with a U.S.-based investment-grade hyperscaler in 2025. ^{[10]Applied Digital, $5 billion AI factory lease with U.S.-based investment-grade hyperscaler (2025)https://www.applieddigital.com/} The company designs and constructs modular HPC facilities for AI and machine learning workloads.

Below these larger players, this market is growing. Private equity firms, family offices, and entrepreneurial operators are deploying sub-$100M GPU clusters in colocation space or in modular facilities near available power. These operators don't build the data centers themselves. They partner with modular facility providers or rent space from colocation providers like Equinix, QTS, and CyrusOne.

NVIDIA is accelerating this with a franchise-style ecosystem. The company certifies partners at every layer, OEMs like Dell and HPE for servers, colocation providers for power and cooling, managed services partners for operations, and construction firms like Bechtel for physical builds. ^{[11]Bechtel, "Bechtel to Accelerate AI Data Center Construction with NVIDIA" (October 2025)https://www.bechtel.com/press-releases/bechtel-to-accelerate-ai-data-center-construction-with-nvidia/} At GTC 2026, NVIDIA published the Vera Rubin DSX reference design, a full blueprint for AI factory construction down to the power and cooling specs. ^{[12]NVIDIA Newsroom, Vera Rubin DSX AI Factory Reference Design (March 2026)https://nvidianews.nvidia.com/news/nvidia-releases-vera-rubin-dsx-ai-factory-reference-design-and-omniverse-dsx-digital-twin-blueprint-with-broad-industry-support} For a modular operator deploying 2-5 MW in a remote location, the hardware configuration, cooling requirements, and software stack come pre-validated.

How the economics work

The per-MW economics of modular depend on what you include in the comparison. Cushman & Wakefield's 2025 Data Center Development Cost Guide puts traditional builds at an average of $11.7 million per MW across 19 U.S. markets, ranging from $9.3 million in San Antonio to $15 million in Reno. ^{[13]Cushman & Wakefield, "U.S. Data Center Development Cost Guide" (2025)https://www.cushmanwakefield.com/en/united-states/insights/data-center-development-cost-guide} A modular unit itself (power distribution, cooling, IT space) runs $4.5-6.5 million per MW for Tier II-III configurations. ^{[14]Data Center Frontier, "Scalable Modular Data Centers and the Race to ROI" (2025)https://www.datacenterfrontier.com/special-reports/article/11427784/scalable-modular-data-centers-and-the-race-to-roi} But that figure often excludes land, site preparation, and utility interconnection, costs the traditional number includes. Apples to apples, modular per-MW costs are comparable to or slightly below traditional builds at small scale. At larger scale, economies of scale make modular less competitive versus traditional builds.

The real advantage is capital risk and speed. A traditional 50 MW facility requires $500 million to $1 billion upfront before it serves a single customer. A modular operator deploys a $5-15 million module, fills it, and adds the next one. McKinsey estimates prefabricated solutions cut delivery timelines by up to 50%, with pre-approved powered shells reducing time to market by 50-70%. ^{[7]McKinsey, "Scaling Bigger, Faster, Cheaper Data Centers with Smarter Designs" (2025)https://www.mckinsey.com/industries/private-capital/our-insights/scaling-bigger-faster-cheaper-data-centers-with-smarter-designs}

Power cost is where modular operators gain the most. Operators placing facilities at stranded gas sites, surplus hydro locations, or behind-the-meter with dedicated generation can achieve $0.02-0.04/kWh, roughly half the $0.06-0.10/kWh that grid-connected metro data centers pay. On a 2 MW facility running at 80% utilization, the difference between $0.03/kWh and $0.08/kWh is roughly $700,000 per year in power savings. For a GPU cluster where electricity is 10-15% of total operating cost , that delta flows directly to margin.

References

1. Lawrence Berkeley National Laboratory, "Queued Up: 2025 Edition" (2025)
2. Sightline Climate, "Data Center Outlook: Half of 2026 Pipeline May Not Materialize" (2026); Latitude Media coverage
3. Sun Microsystems, Project Blackbox announcement (October 2006)
4. Microsoft, "Our Vision for Generation 4 Modular Data Centers" blog post by Michael Manos, Daniel Costello, and Christian Belady (December 2008)
5. Google, "Efficient Computing" data center blog: container-based data center designs (2005-2009)
6. JLL, "Modular Systems and Micro DCs Could Fetch $48 Billion by 2030" (2025)
7. McKinsey, "Scaling Bigger, Faster, Cheaper Data Centers with Smarter Designs" (2025)
8. Crusoe, Series E: $1.375 billion at $10+ billion valuation (October 2025); Series D: $600 million at $2.8 billion valuation (December 2024)
9. IREN, investor presentations and site data (2025-2026)
10. Applied Digital, $5 billion AI factory lease with U.S.-based investment-grade hyperscaler (2025)
11. Bechtel, "Bechtel to Accelerate AI Data Center Construction with NVIDIA" (October 2025)
12. NVIDIA Newsroom, Vera Rubin DSX AI Factory Reference Design (March 2026)
13. Cushman & Wakefield, "U.S. Data Center Development Cost Guide" (2025)
14. Data Center Frontier, "Scalable Modular Data Centers and the Race to ROI" (2025)