Ambra
Blog / AUGUST 7, 2026

The Modular Data Center Opportunity.

Why the fastest-growing slice of AI infrastructure is being built in factories, not on construction sites.

Every AI infrastructure conversation eventually arrives at the same wall: it isn’t chips, it isn’t capital, it’s power. Finding 2–5 MW in a remote location takes months. Finding 500 MW next to a major metro can take years, if it happens at all. That gap between demand for compute and the availability of electrons is what has turned modular, factory-built data centers from a niche curiosity into one of the fastest-growing categories in the industry.

Why power now decides where AI gets built

As of the end of 2024, roughly 10,300 generation and storage projects were sitting in U.S. grid interconnection queues awaiting a connection date. By the end of 2025 that figure had eased slightly to about 8,200 projects — but the queue itself grew larger in capacity terms, representing over 1,300 GW of generation and nearly 750 GW of storage, according to Lawrence Berkeley National Laboratory’s Queued Up: 2026 Edition. The composition shifted too: renewables retreated from the queue across almost every region in 2025, while gas capacity — the resource AI operators increasingly favor for firm, dispatchable power — grew sharply in the largest markets.

The queue is still the single biggest chokepoint standing between an AI roadmap and an operating data center. Of the roughly 110 large data center projects slated for 2025, about 26% were delayed, and industry trackers expect up to half of the U.S. data center pipeline to slip in 2026 on power constraints, community pushback, and equipment shortages — long-lead transformers and switchgear in particular.

History is not encouraging for anyone waiting in line. Of all generation capacity that applied for grid interconnection between 2000 and 2019, only 13% had become operational by the end of 2024; 77% was withdrawn outright. But that outcome is not evenly distributed by project size — and the size effect is the whole thesis behind modular.

Interconnection queue outcomes by project size
< 20 MW
18%
69%
13%
18% operational · 69% withdrawn · 13% waiting
20–100 MW
14%
76%
10%
14% operational · 76% withdrawn · 10% waiting
100–500 MW
9%
80%
11%
9% operational · 80% withdrawn · 11% waiting
500+ MW
5%
84%
11%
5% operational · 84% withdrawn · 11% waiting
OperationalWithdrawnStill waiting
U.S. interconnection queue outcomes by application size, 2000–2019 cohort, status as of end-2024. Source: Lawrence Berkeley National Laboratory, Queued Up, 2025 Edition.

Sub-20 MW requests clear the queue at roughly triple the completion rate of 500+ MW requests. A hyperscale 500 MW campus competes for the same transformers, transmission upgrades, and utility engineering hours as every other gigawatt-scale project in the interconnection line. A 2–5 MW request behind an existing substation, next to stranded gas, or on a retired industrial site with spare capacity skips most of that queue entirely.

What a modular data center actually is

A modular data center is a prefabricated, self-contained compute facility: power distribution, cooling, fire suppression, and rack space are built and tested as a unit in a factory, then trucked or railed to a site and commissioned in place. Configurations typically run 1–5 MW per module — some vendors ship ISO-container-sized units, others ship larger prefabricated buildings — and modules can be added incrementally as power and demand allow, rather than sized for a final capacity decided years in advance.

DimensionTraditional buildModular deployment
Time to first workload18–36 months8–16 weeks after site prep
Typical power per site50–500+ MW1–5 MW per module
CapEx per site$500M–$2B+$5M–$30M per module
ConstructionCustom, on-siteFactory-manufactured, shipped
ScalabilitySized for final capacity upfrontAdded incrementally, module by module
Location flexibilityNear population centersWherever power is available

The practical effect is compressing deployment from years to months:

Time from decision to first workload
TRADITIONAL~28 months total
Permitting & grid
9mo
Construction
15mo
Commissioning
4mo
MODULAR~6 months total
Site prep
3mo
Module delivery
2mo
Hookup
1mo
06mo12mo18mo24mo30mo
Illustrative deployment timeline: a 50+ MW ground-up campus vs. a 2–5 MW prefabricated module shipped to a site with existing interconnection. Based on industry timelines reported by McKinsey and Data Center Frontier.

A concept the industry already tried — and abandoned — twice

Modular is not a new idea; it has failed to reach mass adoption twice before. Sun Microsystems shipped Project Blackbox, a data center built into a 20-foot shipping container, in October 2006 — it never gained traction because operators of the day had no shortage of real estate or power; their constraint was compute demand, not infrastructure. Microsoft’s 2008 “Generation 4” architecture and Google’s own container-based designs from 2005–2009 followed a similar arc: both companies eventually abandoned the shipping-container form factor for purpose-built modular designs, since sealed metal boxes made poor peace with rack density, maintenance access, and cooling. HP’s POD and Schneider Electric’s prefabricated units saw limited adoption through the 2010s for the same reason traditional construction was never really broken: enterprise IT demand grew predictably enough that an 18-month build cycle was tolerable.

GPU clusters broke that equilibrium. A modern AI training rack draws 40–80 kW, three to eight times a traditional enterprise rack, and demand for that power is now growing faster than utilities can build transmission. Crusoe proved the modular playbook works for AI in 2021, deploying compute directly at oil wellheads on stranded natural gas; IREN did the same starting from a Bitcoin-mining base, now running 810 MW across six sites on renewable power. By 2025, NVIDIA and Bechtel had turned the pattern into a franchise model for full AI-factory construction.

The “edge” theory versus the power reality

For a decade, the industry expected distributed compute to be driven by edge computing — thousands of small facilities near end users, built to shave milliseconds off latency-sensitive applications. That’s not, in practice, what is driving the current wave of small, distributed AI facilities. A 2 MW modular site running inference in rural Texas does not need to sit within 10 milliseconds of its users: inference API calls tolerate 50–200 milliseconds of added network latency without a noticeable effect on the end-user experience. A developer in San Francisco calling an endpoint in West Texas will not notice the extra 30 milliseconds of transit time.

$0.03/kWh power from stranded gas or surplus hydro beats $0.08/kWh grid power in a metro area — even with worse network connectivity.

What is actually driving site selection is energy economics, not latency. Operators go where power is cheap and available, and accept the marginal networking cost.

Who is building modular AI data centers — and how fast the category is growing

JLL projects the modular and micro data center market will reach $48 billion by 2030, growing roughly 35% a year, and McKinsey estimates prefabricated components already account for 40–60% of the average data center’s build, rising to 80–85% at the leading operators. More recent market trackers put the category at roughly $38–48 billion in 2026 alone, with independent analyst consensus converging on a 17–19% CAGR through 2030–2033 — growth pulled by AI training clusters, data-sovereignty requirements, and the operational need to activate capacity in months rather than years.

That growth sits inside a much larger power story. Gartner puts global data center power demand at 104 GW in 2025, rising 27% to 132 GW in 2026 and reaching roughly 290 GW by 2030. The IEA’s April 2026 report estimates global data center electricity consumption at approximately 485 TWh in 2025 — up 17% year over year — with AI-specific facilities growing at 50% annually, and projects total consumption could reach 945 TWh by 2030 and 1,200 TWh by 2035 in its base case.

Global data center power demand
GW
100
200
300
104
132
165
205
290
2025
2026
2027E
2028E
2030E
Global data center power demand, 2025–2030E (gigawatts). Source: Gartner, June 2026 press release; IEA, Key Questions on Energy and AI, April 2026.

Crusoe raised a $1.375 billion Series E at a $10+ billion valuation in October 2025, following a $600 million Series D at $2.8 billion in December 2024. It started by deploying modular facilities at oil wellheads on stranded natural gas and has since secured $11.6 billion for a 1.2 GW campus in Abilene, Texas as part of the Stargate initiative, expanding into grid-connected renewable sites.

IREN (formerly Iris Energy) now operates 810 MW across six North American sites, powered by 100% renewable energy or renewable energy credits — a playbook it established mining Bitcoin: find cheap renewable power in a remote location, deploy modular compute, run it 24/7. Applied Digital signed a $5 billion AI factory lease with a U.S.-based, investment-grade hyperscaler in 2025 and designs and builds its own modular HPC facilities for AI and machine-learning workloads.

Below that top tier sits a fast-growing long tail: private equity firms, family offices, and independent operators deploying sub-$100 million GPU clusters — a few hundred to a few thousand GPUs — either in colocation space with providers like Equinix, QTS, and CyrusOne, or in their own modular facilities near available power. These clusters serve fine-tuning, small-scale training, and inference workloads well; the largest training jobs, north of 10,000 tightly interconnected GPUs at 50+ MW, still rule most modular sites out. NVIDIA is accelerating the smaller-scale tier through a certified-partner ecosystem spanning OEMs (Dell, HPE), colocation providers, managed-services partners, and construction firms like Bechtel; at GTC 2026 it published the Vera Rubin DSX reference design — a full blueprint for AI-factory construction down to power and cooling specs.

How the economics actually work

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. Modular units — power distribution, cooling, and IT space — run $4.5–6.5 million per MW for Tier II–III configurations, though that figure typically excludes land, site prep, and utility interconnection costs that traditional pricing bundles in. Apples to apples, modular per-MW costs land at or slightly below traditional builds at small scale; at larger scale, economies of scale make traditional construction more competitive again.

The real edge isn’t the per-MW build cost — it’s capital risk and speed. A traditional 50 MW facility requires $500 million to $1 billion of upfront capital before it serves a single customer. Modular operators deploy $5–15 million per module, incrementally, and only commit further capital once a module is generating revenue. McKinsey estimates prefabrication cuts delivery timelines by up to 50%, and pre-approved powered shells cut time-to-market by 50–70%.

Power cost compounds the advantage. Sites on stranded gas, surplus hydro, or behind-the-meter generation 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 gap between $0.03/kWh and $0.08/kWh is worth roughly $700,000 a year in power savings alone. Since electricity runs 10–15% of total GPU cluster operating cost, that delta flows straight through to margin.

Cost driverTraditional buildModular deployment
Build cost per MW$9.3M–$15M (2025 avg $11.7M)$4.5M–$6.5M per module, excl. land/site prep
Power cost$0.06–$0.10/kWh (grid, metro)$0.02–$0.04/kWh (stranded/surplus)
Capital at risk per site$500M–$1B+$5M–$15M per module
Time to first workload18–36 months3–6 months (McKinsey: ~50% faster)

The takeaway

Modular doesn’t win by being categorically cheaper to build — at scale, traditional construction still holds a per-MW edge. It wins because it converts a $500 million, three-year bet made before a single customer signs on into a series of $5–15 million bets made against power that’s already available and cheap. As the interconnection queue keeps growing faster than utilities can clear it, and as global data center demand climbs toward 290 GW by 2030, that difference in capital risk and speed — not the container itself — is what makes modular the default answer to “where do we put the next cluster.”

Sources

  • Lawrence Berkeley National Laboratory, “Queued Up: 2025 Edition” (2025) and “Queued Up: 2026 Edition” (June 2026), emp.lbl.gov/queues
  • Sightline Climate / Latitude Media, “Data Center Outlook: Half of 2026 Pipeline May Not Materialize” (2026)
  • Sun Microsystems, Project Blackbox announcement (October 2006)
  • Michael Manos, Daniel Costello, Christian Belady, “Our Vision for Generation 4 Modular Data Centers,” Microsoft (December 2008)
  • Google, “Efficient Computing” data center blog, container-based data center design notes
  • Crusoe Energy Systems, Series D and Series E funding announcements (December 2024, October 2025)
  • IREN Ltd., investor presentations and site data (2025–2026)
  • Applied Digital, $5B AI-factory lease announcement with U.S. investment-grade hyperscaler (2025)
  • NVIDIA Newsroom, “Bechtel to Accelerate AI Data Center Construction with NVIDIA” (October 2025); Vera Rubin DSX AI Factory Reference Design (March 2026)
  • Cushman & Wakefield, “U.S. Data Center Development Cost Guide” (2025)
  • JLL Research, modular and micro data center market outlook (2025); McKinsey & Company, prefabricated data center component analysis
  • Data Center Frontier, “Scalable Modular Data Centers and the Race to ROI” (2025)
  • Gartner, “Data Center Electricity Consumption to Grow 26–27% in 2026” (press release, June 2026)
  • International Energy Agency, “Key Questions on Energy and AI” — Executive Summary (April 2026)
  • Modular data center market sizing: Research and Markets, Precedence Research, MarketsandMarkets, and Grand View Research, 2026 editions (analyst range, not consensus figure)

This note was prepared by Ambra Compute for internal and partner distribution. Figures are drawn from public third-party sources cited above and are believed reliable but not independently audited; where analyst estimates diverge, ranges are shown rather than a single point estimate.

← All posts