Quick answer: Standard hyperscale shell-and-core construction now averages roughly $11.3 million per megawatt globally, per JLL's 2026 outlook. Once GPU-ready electrical, liquid-cooling, and tenant fit-out costs are layered in, fully built AI facilities routinely land between $20 million and $37 million per megawatt. A single 1-gigawatt AI campus now carries an estimated $38 billion upfront price tag, and the four largest hyperscalers alone are on pace to deploy roughly $725 billion in AI capital spending this year.
Key Takeaways
- Global shell-and-core construction costs have climbed to an average of $11.3 million per megawatt (MW) in 2026, according to JLL's Global Data Center Market Outlook — up from $10.7 million in 2025 and $7.7 million in 2020.
- Layering in AI-specific power density and tenant technology fit-out pushes many fully built AI facilities to $20 million–$37 million per MW, depending on cooling architecture and redundancy, per JLL, Turner & Townsend, and Archdesk benchmarking.
- Epoch AI's total-cost-of-ownership model puts a single 1-gigawatt AI campus at roughly $38 billion in upfront capital, with annual operating costs exceeding $900 million.
- McKinsey projects cumulative AI-related data center investment could reach $5.2 trillion by 2030 in its base case, and as much as $7.9 trillion under a high-growth scenario.
- The four largest hyperscalers — Amazon, Alphabet, Microsoft, and Meta — are collectively guiding toward roughly $725 billion in 2026 capital expenditure, up about 77% from 2025.
- Power delivery, not concrete, has become the primary constraint on how fast new AI capacity can actually be energized.
In this briefing
AI Data Center Cost Dashboard — 2026
$11.3M
Global standard shell-and-core cost per MW
JLL, 2026 Outlook$20M–$37M
Fully built AI-optimized cost per MW
Archdesk / Axis Intelligence$38B
Total cost of one 1 GW AI campus
Epoch AI$725B
Big Four 2026 AI capex, up 77% YoY
Company guidance, FT/CreditSightsA New Cost Reality for Developers
The AI infrastructure boom has rewritten the economics of building a data center. A few years ago, the job of a developer was fairly linear: acquire land, secure a utility interconnection, and put up a hyperscale shell built for CPU-driven cloud workloads. In 2026, that playbook barely applies.
Training and serving large AI models means packing tens of thousands of GPUs into a single facility, running them near-continuously at extreme utilization, and keeping them cool enough to avoid throttling. These facilities consume far more power, reject far more heat, and demand tighter redundancy than the cloud data halls built even five years ago.
That shift has turned AI data center development cost into one of the most closely tracked figures in infrastructure finance — watched by hyperscalers, sovereign wealth funds, utilities, GPU suppliers, and the real estate developers racing to keep pace with demand. The debate has moved past whether this capacity gets built. It is now about how much it costs, how fast it can be energized, and whether returns can keep up with capital outlay of this size.
The Global Benchmark: $11.3 Million Per MW, and Climbing
The most widely cited reference point in the industry remains JLL's 2026 Global Data Center Market Outlook. JLL's research team places the global average shell-and-core construction cost at roughly $11.3 million per MW this year, up from $10.7 million in 2025 and $7.7 million in 2020 — a compound growth rate near 7% a year since the start of the decade.
JLL's analysis also separates out what AI actually adds on top of that shell. Tenant-side AI technology fit-out — the electrical distribution, rack infrastructure, and cooling loop needed to actually run GPU clusters — can add up to another $25 million per MW, according to the same report. Turner & Townsend's 2025–2026 Data Centre Cost Index adds a further data point: AI-optimized shell-and-core construction itself runs 7–10% above a standard build, before any fit-out is even counted.
Put together, a fully built AI-ready facility no longer resembles a marginally upgraded cloud data center. It is closer to a different asset class, with a cost structure dominated by compute hardware and power delivery rather than concrete and steel.
Global Shell-and-Core Construction Cost Growth
| Year | Cost per MW |
|---|---|
| 2020 | $7.7M |
| 2025 | $10.7M |
| 2026 | $11.3M |
Source: JLL, 2026 Global Data Center Market Outlook
Why AI Facilities Cost Two to Three Times More
Independent benchmarking firms are converging on a similar range for what a truly AI-optimized facility costs once every layer is included. Archdesk's April 2026 research puts an AI-optimized build running 40–80 kW rack densities at roughly $20 million per MW, excluding IT equipment and land entirely. Axis Intelligence Research goes a step further, combining the AI shell-and-core premium with JLL's tenant fit-out figure to arrive at an all-in AI cost of about $37.3 million per operational MW — a multiple of roughly 3.3x over the standard shell-and-core baseline, before land, permitting, interconnection deposits, or GPU hardware are even added.
Goldman Sachs Research reached a comparable conclusion in its May 2026 note on AI build-out assumptions, estimating traditional hyperscale facilities near $10 million per MW against AI-optimized builds in the $15–20 million range, with costs climbing further depending on power density and redundancy choices. GetCostIdea's 2026 construction guide and several buyer-side benchmarking services place similar figures at $15–20 million-plus per MW for AI-heavy builds, versus $7–12 million for standard capacity.
Standard vs. AI-Optimized Facility Economics
| Metric | Standard | AI-Optimized |
|---|---|---|
| Cost per MW | $7M–$12M | $20M–$37M |
| Cooling | Air (~$1.8M/MW) | Liquid (~$4.5M–$5.2M/MW) |
| Rack density | 10–20 kW | 40–100+ kW |
| Power demand | Moderate | Extreme |
| Capital intensity | High | Very high |
Sources: JLL 2026 Outlook; Archdesk, April 2026; Goldman Sachs Research, May 2026
The Three Forces Driving the Cost Gap
GPU and Server Infrastructure
Compute hardware remains the single largest line item. High-end AI accelerators frequently cost more than $25,000 per unit, and a large training cluster can hold tens of thousands of them. Epoch AI's cost-of-ownership modeling found that servers alone account for roughly 60% of total ownership cost on a large AI campus — in many projects, the hardware bill now exceeds the cost of the building that houses it.
Power Distribution
AI clusters draw electricity at a scale that traditional utility planning wasn't built for. Developers increasingly need dedicated substations, high-capacity transformers, redundant feeds, and on-site backup generation just to get a facility energized on schedule. In several major markets, securing firm power capacity has become harder — and slower — than arranging construction financing itself, with transformer and switchgear lead times now stretching multiple years.
Liquid Cooling
Air cooling can no longer keep pace with the heat output of dense GPU racks. Direct-to-chip loops, rear-door heat exchangers, and immersion systems are becoming standard rather than optional. Archdesk's cost breakdown shows liquid-cooled infrastructure running $4.5–$5.2 million per MW, versus roughly $1.8 million per MW for conventional air cooling — a gap that alone can decide whether a facility clears the $20 million-per-MW threshold.
Where the Money Goes: AI Facility Cost Breakdown
| Component | Share of Total Cost |
|---|---|
| GPU / AI servers | 60% |
| Facility construction | 20% |
| Power infrastructure | 12% |
| Cooling systems | 5% |
| Networking & other | 3% |
Source: Epoch AI, "Total Cost of Ownership of a One-Gigawatt AI Data Center," May 2026
Inside Epoch AI's $38 Billion Gigawatt Model
One of the most granular cost studies available comes from Epoch AI researchers Amelia Michael and Ben Cottier, who published Total Cost of Ownership of a One-Gigawatt AI Data Center in May 2026. Their model puts a 1 GW AI campus at approximately $38 billion in upfront capital expenditure, with annual operating costs exceeding $900 million once running. Servers account for roughly 60% of lifetime ownership cost, and annualized ownership works out to about $8.5 million per MW per year.
The takeaway for investors is that hardware, not real estate, now drives the return profile of these assets. That reframes how a gigawatt-scale project should be evaluated — less like a building lease and more like a fleet of rapidly depreciating compute assets that happen to sit inside a building.
Where Does $38 Billion Go?
| Category | Estimated Share | Value |
|---|---|---|
| GPU infrastructure | 60% | $22.8B |
| Buildings | 20% | $7.6B |
| Power systems | 12% | $4.6B |
| Cooling | 5% | $1.9B |
| Network & other | 3% | $1.1B |
Real-World Project Benchmarks
Alpha-Matica's November 2025 research on a 100 MW hyperscale facility offers a mid-scale reference point: construction cost of $900 million to $1.5 billion, plus IT equipment running $2.5 billion to $4 billion or more, for a total project cost between $3.4 billion and $5.5 billion-plus. Construct Elements' February 2026 field data corroborates that pattern, finding that AI-optimized projects above 50 MW now routinely surpass $1 billion in investment before equipment purchases are even complete.
➡️ Read Also: Investing in Data Center Commercial Real Estate in 2026: AI, Power Demand, and the Next CRE Boom
Hyperscaler Capex Is Accelerating Faster Than the Cost Curve
The cost-per-MW story only tells half of the picture — the other half is how much capital hyperscalers are actually committing. Amazon, Alphabet, Microsoft, and Meta have each raised full-year 2026 capex guidance during recent earnings cycles, with combined spending now tracked at roughly $725 billion, up about 77% from an already-record $410 billion in 2025. Amazon alone is guiding toward roughly $200 billion, most of it directed at AWS data centers and custom silicon; Alphabet is in the $175–185 billion range on the back of a Google Cloud backlog that has swelled past $460 billion; Meta raised its guidance range to as high as $145 billion, citing rising memory and component costs; and Microsoft's spending has been tracked anywhere from $120 billion to roughly $190 billion depending on the reporting window, with company leadership pointing to an AI business now running above a $37 billion annual revenue rate.
Goldman Sachs' infrastructure model now projects combined capex for the same four companies reaching $5.3 trillion between fiscal 2025 and fiscal 2030, with a broader aggregate figure — spanning compute, data centers, and power — approaching $7.6 trillion between 2026 and 2031. Rising memory and component pricing is itself becoming a capex driver: memory now accounts for close to 30% of hyperscaler data center spending, roughly four times its share in 2023.
Nvidia sits at the center of that spending wave. The company's data center revenue has continued compounding on the back of hyperscaler orders, reinforcing why GPU allocation — not just capital availability — has become a gating factor for how fast new AI capacity can actually be deployed.
Regional Costs Still Vary by 40% or More
Location continues to be one of the biggest swing factors in project economics. Turner & Townsend and Cushman & Wakefield both continue to find construction cost spreads exceeding 40% between markets, driven by electricity availability, labor costs, land pricing, incentive structures, and how congested the regional supply chain is.
Regional benchmarking for 2025–2026 shows the U.S. national average sitting near $11.3 million per MW, with Northern Virginia — still the largest data center market in the world by inventory — near the lower end of the range, and Silicon Valley running significantly higher. Frankfurt and London are tracking closer to $14 million per MW because of land scarcity and higher grid connection costs, while Tokyo sits near $12 million per MW due to land constraints and regulatory complexity. Canadian markets remain comparatively affordable at $7–12 million per MW, though Toronto carries a premium tied to urban land pricing and hydro capacity.
Cushman & Wakefield's 2026 Asia Pacific Data Centre Construction Cost Guide, led by Pritesh Swamy, found regional construction costs rising roughly 10% year-over-year across many APAC markets, driven by sustained demand growth and continuing supply-chain pressure. Separately, the firm's broader cost tracking shows per-square-foot pricing for U.S. facilities pushing toward roughly $1,000 — up about 50% year-over-year — as project specifications shift further toward AI-ready density.
➡️ Read Also: How Canada's Renewable Energy Advantage Is Attracting AI Data Center Investment
McKinsey's Multi-Trillion-Dollar Buildout Forecast
McKinsey's The Cost of Compute: A $7 Trillion Race to Scale Data Centers remains one of the most-cited long-range forecasts in the sector. The firm's base case puts cumulative AI data center investment at $5.2 trillion by 2030, with total compute infrastructure spending — including chips, networking, and power — reaching $6.7 trillion, and a high-growth scenario climbing to $7.9 trillion. McKinsey estimates roughly 125 gigawatts of additional AI-equipped capacity could be added globally between 2025 and 2030, a scale of build-out that helps explain why infrastructure investors, pension funds, and private equity are competing aggressively for access to shovel-ready sites with secured power.
Global AI Infrastructure Spending Forecasts
| Source | Forecast |
|---|---|
| McKinsey | $5.2T base case by 2030 ($7.9T high-growth) |
| Goldman Sachs | $5.3T, Big Four, FY2025–FY2030 |
| JLL | ~$3T real estate + IT, 5-year window |
| Big Four hyperscalers | $725B in 2026 alone (+77% YoY) |
What Investors Are Watching Most Closely
| Factor | Importance |
|---|---|
| Power availability | ★★★★★ |
| Cost per MW | ★★★★★ |
| GPU supply / allocation | ★★★★★ |
| Time to market | ★★★★☆ |
| Land cost | ★★★☆☆ |
| Tax incentives | ★★★☆☆ |
➡️ Also Read: AI Data Centers and the Global Electricity Surge: Why Power Is Becoming the New Bottleneck of Digital Infrastructure | AI Data Center Power Crisis 2026: How It Is Impacting U.S. Real Estate and Infrastructure Investment
Advisory: What Each Stakeholder Should Do Next
The numbers above point to one conclusion: cost-per-MW is no longer a single figure anyone can plan around — it is a range shaped by power access, cooling architecture, and regional grid readiness. Here is how that translates into action for the groups actively deploying capital into this sector.
For Developers & Builders
- Underwrite projects on power-secured timelines, not construction timelines — a firm interconnection date is now the binding constraint, not permitting or steel.
- Model liquid cooling into base-case budgets rather than as an add-on; at 40–80 kW+ rack densities it is no longer optional.
- Separate shell-and-core, tenant fit-out, and active IT hardware into distinct line items when pricing a bid — blended benchmarks routinely understate true AI-ready cost.
For Investors & Capital Allocators
- Evaluate gigawatt-scale assets on total cost of ownership and GPU utilization, not construction cost alone — hardware depreciation now drives returns more than the real estate does.
- Stress-test deals against a capex-to-revenue gap: hyperscaler spending is growing faster than disclosed AI revenue, and any deceleration in cloud growth could compress valuations quickly.
- Treat regional cost spreads (40%+) as a genuine diligence item, not a rounding error, when comparing competing site options.
For Utilities & Grid Planners
- Prioritize transformer and switchgear procurement well ahead of committed capacity, given multi-year lead times already reported across major markets.
- Engage hyperscalers early on flexible load agreements and on-site generation partnerships to avoid becoming the critical-path bottleneck on projects that are otherwise fully financed.
For Policymakers & Regulators
- Coordinate interconnection queue reform with data center permitting so that power availability and construction approval move on comparable timelines.
- Track cumulative regional power and water commitments across projects, since site-by-site review can understate aggregate grid and resource strain.
Frequently Asked Questions
How much does it cost to build an AI data center per megawatt in 2026?
Standard shell-and-core construction averages about $11.3 million per MW globally. A fully built AI-optimized facility, including tenant fit-out and liquid cooling, typically runs $20 million to $37 million per MW depending on rack density and region.
Why are AI data centers so much more expensive than traditional ones?
GPU servers, high-capacity power distribution, and liquid cooling infrastructure account for the bulk of the premium. Servers alone can represent roughly 60% of total ownership cost on large AI campuses.
What does a 1-gigawatt AI data center cost in total?
Epoch AI estimates roughly $38 billion in upfront capital expenditure for a 1 GW facility, plus annual operating costs above $900 million once running.
How much are hyperscalers spending on AI infrastructure in 2026?
Amazon, Alphabet, Microsoft, and Meta combined are tracking toward roughly $725 billion in 2026 capital expenditure, up about 77% from 2025, according to company guidance compiled by industry analysts.
The Spending Race Is Accelerating, Not Slowing
Research from JLL, Goldman Sachs, McKinsey, Epoch AI, Turner & Townsend, Archdesk, Cushman & Wakefield, and hyperscaler earnings calls all point in the same direction: AI data center development costs keep rising because the underlying infrastructure requirements keep changing. Facilities that once cost around $10 million per MW now frequently require $20 million or more, and full project costs climb further once GPU clusters, networking, and power infrastructure are counted.
For developers, the challenge is balancing construction cost, power availability, and deployment speed against a backdrop where demand still outpaces buildable capacity. For investors, the opportunity — and the risk — sits in the same place: a market where capital is being deployed faster than the assets it buys can be fully utilized. The projects that separate winners from overextended balance sheets in this cycle will be the ones that treated power and GPU utilization, not square footage, as the real constraint from day one.
Core Insights Review contributors publish research-based analysis and editorial insights on commercial real estate, PropTech, smart infrastructure, sustainable construction, industrial real estate, and emerging technologies shaping the future of the built environment.
Check for more information: Core Insights Review
