AI Data Center Cost per MW 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: June 27, 2026 | Next scheduled update: Q4 2026 | License: CC BY 4.
Quick Answer
Building an AI-optimized data center in 2026 costs $15M–$25M per MW for shell and infrastructure — roughly double what a standard cloud facility runs. Add GPU fit-out and the all-in figure climbs to $30M–$45M per MW. According to Axis Intelligence Research’s cross-source analysis, annualized total cost of ownership across a 1 GW facility runs $8.5M per MW per year, with servers absorbing 60 cents of every dollar spent.
Key Findings
- According to Axis Intelligence Research’s synthesis of JLL and Turner & Townsend 2025–2026 benchmarks, the global average construction cost for a standard AI-capable data center shell reached $11.3M per MW in 2026 — up from $7.7M in 2020, a 47% increase in six years.
- Axis Intelligence Research calculates that AI-optimized facilities incur a liquid cooling premium of $2.7M–$3.7M per MW over standard air-cooled construction, derived from Turner & Townsend’s disclosed cost structure shift (mechanical rising from 22% to 33% of total facility cost) applied to the 2026 benchmark.
- According to Epoch AI’s May 2026 model, a 1 GW AI data center carries $38B in upfront CapEx and $0.9B in annual OpEx, yielding an annualized total cost of ownership of $8.5B per year — of which $5B (60%) is servers and GPUs, not facility infrastructure.
- Axis Intelligence Research’s analysis of four hyperscaler earnings reports finds the Big Four (Amazon, Alphabet, Microsoft, Meta) will spend approximately $725B in combined capital expenditure in 2026 — a 77% increase over 2025’s $410B — with roughly 75% directed at AI data centers, GPUs, and power.
- According to Axis Intelligence Research’s cross-analysis of IEA and LBNL data, AI-focused data centers drove a 17% surge in global data center electricity consumption in 2025, against 3% growth in global electricity demand overall. The gap between those two numbers is the whole story of AI infrastructure in 2026.
What Does It Cost to Build an AI Data Center per Megawatt?
The honest answer is that “cost per megawatt” is three different numbers depending on which layer you’re buying — and the industry loves to quote the cheapest one. Shell and core, all-in facility, full stack with GPUs: these are not the same thing, and treating them interchangeably will cost you hundreds of millions of dollars on a large build.
Shell and core is what gets quoted in construction indices. It covers the building envelope, structural work, foundations, raised floors, and the power and cooling distribution to the rack row. It does not include the GPUs, servers, networking, storage, or the interconnection scope that can easily add $5M–$25M to a single project. JLL’s 2026 global data center outlook puts the shell-and-core global average at $11.3M per MW, sourced from Turner & Townsend’s Data Centre Construction Cost Index 2025–2026 — and that number has climbed 6% year-over-year from $10.7M in 2025.
For AI-optimized facilities, the shell number is the floor, not the ceiling. Turner & Townsend’s 2025–2026 index found that 60% of industry professionals expect construction cost inflation of 5–15% in 2026, with another 21% expecting it to exceed 15%. An AI hall running 40–80 kW per rack needs liquid cooling rather than air, and liquid cooling changes the cost structure materially: mechanical systems jump from 22% of facility cost to 33%, according to Turner & Townsend’s cost structure data, while electrical drops from 54% to 48%. That shift alone adds $2.7M–$3.7M per MW to a build, a figure Axis Intelligence Research derives by applying Turner & Townsend’s disclosed cost structure percentages to the 2026 shell benchmark.
The most expensive markets in Turner & Townsend’s index are Tokyo ($15.2/W), Singapore ($14.5/W), and Zurich ($14.2/W). On the cheaper end, Dublin and Madrid sit at $10.0/W. Northern Virginia — the world’s largest data center market at roughly 4,000 MW of inventory — runs lower than the US average because of market scale; Silicon Valley runs materially higher.
The table below summarizes the 2026 construction cost tiers as synthesized by Axis Intelligence Research:
| Facility Type | CapEx per MW (2026) | Cooling | Source |
|---|---|---|---|
| Standard shell & core (global average) | $10.7M–$11.3M | Air-cooled | JLL Data Center Outlook 2026; Turner & Townsend |
| AI-optimized shell + liquid cooling | $12.0M–$15.0M | Liquid | Turner & Townsend; Archdesk |
| AI facility + tenant IT fit-out (excl. GPUs) | $15M–$25M | Liquid | JLL (up to $25M/MW for tenant fit-out) |
| Full stack with GPU (GB200 NVL72, 1 GW) | $30M–$45M | Liquid | Bernstein; Epoch AI |
| Gigawatt campus (infrastructure only, hyperscale) | $15B–$20B total | Liquid | Data Center POST; Archdesk |
Source: Axis Intelligence Research synthesis, June 2026. Shell-and-core figures cover construction only; IT hardware is a separate line.
AI Data Center Construction Cost per MW by Region
Geography still decides a significant fraction of the budget — sometimes 25–40% of total cost, according to construction benchmarks. The drivers are land scarcity, local labor markets, grid interconnection timelines, and the density of the existing contractor pool.
The most striking divergence right now is between major US markets and constrained European ones. A new 50 MW facility in Amsterdam faces a 10-year grid connection wait, according to JLL Research cited by Archdesk (April 2026). London is 8 years. That queue time doesn’t show up in the $/MW benchmark — it shows up in your pro forma when you model when the facility actually starts generating revenue.
| Region / Market | Benchmark Cost per MW (2026 shell & core) | Grid Wait (new 50 MW) |
|---|---|---|
| Tokyo | $15.2M ($15.2/W) | Significant |
| Singapore | $14.5M ($14.5/W) | Constrained |
| Zurich | $14.2M ($14.2/W) | Moderate |
| Frankfurt / London | ~$14.0M | 8+ years (London) |
| Asia-Pacific average | ~$12.0M | Varies |
| United States (average) | ~$11.3M | 18–36 months (grid-constrained markets) |
| Northern Virginia | Below US average | Moderate (4,039 MW existing inventory) |
| Silicon Valley | Above US average | Constrained |
| Canada | $7M–$12M (province-dependent) | Moderate |
| Dublin / Madrid | $10.0M ($10.0/W) | 8+ years (Dublin) |
Sources: Turner & Townsend Data Centre Construction Cost Index 2025–2026; JLL via Archdesk (April 2026). Grid wait estimates: JLL Research via SymTerra, 2026. Axis Intelligence Research synthesis.
This is not to say developers avoid constrained markets. They can’t always. London and Dublin exist in the data center map because demand from financial services, hyperscaler edge nodes, and enterprise cloud is real and sticky. But the pipeline math gets brutal fast. Sightline Climate’s tracking of the 2026 pipeline estimates 30–50% of projects will slip into 2027 or later because of permitting, zoning, and power procurement delays.
BloombergNEF tracked 23.1 GW of IT capacity under active construction globally at end-September 2025, spread across 831 sites. The US alone held 15.9 GW of that total — about 69% of global activity concentrated in one country. EMEA and Asia-Pacific together accounted for 6.1 GW across 541 sites. The average US site is roughly four times larger than its European or APAC counterpart.
AI Data Center Cost Structure: What Eats the Budget per MW?
The single most important thing to understand about data center budgets is that the building — the concrete, steel, and walls — is a small fraction of what you spend. Electrical and power infrastructure absorbs 40–50% of a standard facility’s cost. For AI-optimized builds, mechanical, electrical, and plumbing systems together consume approximately 75% of the total Guaranteed Maximum Price, according to industry benchmarks cited by Archdesk (2026).
Here is the cost breakdown for an AI-optimized 100 MW facility, based on Axis Intelligence Research synthesis of Turner & Townsend, JLL, and Epoch AI data:
| Cost Component | Share of Construction Budget | Per-MW Cost Estimate |
|---|---|---|
| Electrical & power infrastructure | 40–48% | $4.8M–$7.2M |
| Mechanical / cooling systems (liquid-cooled) | 29–33% | $3.5M–$5.0M |
| Shell, structure & envelope | ~15% | $1.8M–$2.3M |
| Security, fire suppression, networking | ~7% | $0.8M–$1.1M |
| Utility interconnection (substation, fiber) | Variable | $5M–$25M total project |
Source: Axis Intelligence Research cross-analysis of Turner & Townsend Cost Index 2025–2026, JLL Global Data Center Outlook 2026, Archdesk (April 2026).
Cooling is the sharpest cost wedge in 2026. Standard air-cooled heat rejection costs roughly $1.8M per MW. Liquid-cooled AI halls — necessary for rack densities of 40–80 kW per rack — cost $4.5M–$5.2M per MW once you account for CDUs, secondary pumping loops, and dry coolers, according to Archdesk’s April 2026 analysis. That $2.7M–$3.7M delta per MW is the hidden tax of running dense GPU clusters.
Equipment lead times compound the problem. Medium-voltage switchgear is currently running a 22-month lead time. High-capacity transformers are 18 months. The global average for critical components sits at 33 months, according to industry benchmarks cited in construction guides. Owners who start procurement 18–22 months out avoid the worst of this. Those who don’t end up paying the true premium of AI data center construction: not the $/MW, but the months of delayed revenue.
A 60 MW facility delayed by one month loses approximately $14.2M in revenue opportunity, based on AI capacity leasing rates of $150,000–$250,000 per MW per month in major US markets.
Total Cost of Ownership per MW: CapEx vs. OpEx
Construction cost is a one-time number. Total cost of ownership is the number that determines whether a data center is actually a business.
Epoch AI’s May 2026 model of a 1 GW AI data center (GB200 NVL72 servers, US hyperscaler ownership and operation) yields the clearest public breakdown available:
- Upfront CapEx: $38B total ($38M per MW)
- Annual OpEx: $0.9B ($900K per MW per year)
- Annualized TCO (CapEx spread over asset lifespans): $8.5B per year ($8.5M per MW per year)
- Servers & GPUs: $5B per year (60% of annualized TCO)
- Energy: $0.6B per year (7% of annualized TCO)
The key insight from Epoch AI’s analysis is counterintuitive: energy is not the biggest cost driver of an AI data center — servers are, by a factor of roughly 8x. The narrative that electricity bills are strangling the economics of AI infrastructure is overstated. Electricity at $0.0834/kWh — the weighted US industrial rate Epoch AI uses — combined with a PUE of 1.14 and a 71% utilization rate, yields roughly $626K per MW per year. That is less than what you spend amortizing the facility itself.
This changes the calculus on power procurement. Paying a small premium for reliable, behind-the-meter power — from nuclear, gas, or renewables — is a reasonable trade against GPU downtime caused by grid instability. Microsoft’s Three Mile Island unit restart (837 MW, 20-year agreement with Constellation Energy), Google’s SMR agreement with Kairos Power (up to 500 MW), and Amazon’s 960 MW nuclear PPA with a Pennsylvania plant are not charity. They’re supply-chain insurance on the one input that actually matters: uptime for the GPU investment.
Hyperscaler Capital Expenditure on Data Centers in 2026
The four largest US hyperscalers — Amazon, Alphabet, Microsoft, and Meta — collectively committed to approximately $725B in capital expenditure for 2026, according to Axis Intelligence Research’s analysis of Q1 2026 earnings reports. That is a 77% increase over $410B in 2025, which was itself a record.
The numbers by company, as reported in earnings:
| Company | 2026 CapEx Guidance | 2025 CapEx (actual) | YoY Change |
|---|---|---|---|
| Amazon (AWS parent) | ~$200B | ~$125B | +60% |
| Alphabet (Google) | $175B–$185B | ~$91B | +92–103% |
| Microsoft | ~$190B (calendar year) | ~$90B | +111% |
| Meta | $125B–$145B | ~$72B | +74–101% |
| Combined | ~$725B | ~$410B | +77% |
Sources: Company Q1 2026 earnings calls; Financial Times compilation; Axis Intelligence Research synthesis, June 2026. Figures are full-year capex, majority directed at AI data centers, GPUs, and power. Not all CapEx is data center-specific; estimates vary by analyst.
Microsoft’s CFO Amy Hood separately disclosed that $25B of the company’s additional spend reflects rising memory chip and component costs — not additional MW. That is a distinction worth holding: the headline CapEx number includes hardware inflation, not just build expansion. The company remains capacity-constrained even at these spending levels, with an Azure backlog it cannot fulfill.
Alphabet’s Q1 2026 capex of $35.67B in a single quarter more than doubled year-over-year, alongside a Google Cloud backlog that jumped to over $460B. Amazon’s Q1 alone came in at $44.2B.
Goldman Sachs now estimates combined hyperscaler capex at $5.3 trillion across fiscal years 2025–2030 — from $4.5 trillion before Q1 earnings. The five-year baseline across compute, data centers, and power is $7.6 trillion. These are not projections that benefit from rounding.
AI Data Center Electricity Cost and Consumption Statistics
The electricity economics of AI data centers break into two problems: what a facility costs to power, and what that cost does to the grid around it.
For a 100 MW facility running at full capacity with a PUE of 1.14, annual electricity draw is approximately 996 GWh per year (100 MW × 8,760 hours × 1.14). At the US weighted-average industrial rate of $0.0834/kWh (from Epoch AI / EIA data), that is $83M in annual electricity cost. At premium coastal rates of $0.15/kWh, the same facility costs $149M per year in power alone. Regional rate differences are not incidental — they explain, more than anything else, why Northern Virginia, Texas, and parts of the Midwest continue to attract the largest builds.
Globally, the IEA’s Energy and AI Report found that data center electricity consumption reached 415 TWh in 2024, representing 1.5% of global electricity demand. The United States alone accounted for 45% of that — 180 TWh, or 4.4% of US total electricity consumption in 2023 per Lawrence Berkeley National Laboratory’s 2024 US Data Center Energy Usage Report (the most detailed US-specific dataset by scope, covering 2014–2028). LBNL projects US data center consumption will rise to 325–580 TWh by 2028, representing 6.7–12% of US electricity depending on adoption pace.
The IEA’s April 2026 update found that electricity demand from data centers surged 17% in 2025 — versus 3% growth in global electricity demand overall. AI-focused facilities grew even faster. The IEA’s base case projects global data center consumption to reach 945 TWh by 2030, roughly the current electricity consumption of Japan.
One concrete manifestation of that trajectory: PJM capacity market clearing prices for the 2026–2027 delivery year hit $329.17/MW — more than ten times the $28.92/MW cleared in 2024–2025, with rapid data center growth identified as a primary driver (from ArXiv analysis citing FERC data, 2026). The consequences flow directly to residential ratepayers near data center clusters. Bloomberg’s analysis found wholesale electricity prices were up 267% over five years in areas near large data center concentrations.
| Year | Global DC Electricity (TWh) | US DC Electricity (TWh) | Global DC as % of Global Demand |
|---|---|---|---|
| 2020 | ~270 | ~58 (est.) | ~1.1% |
| 2023 | ~360 (est.) | 176 | ~1.3% |
| 2024 | 415 | 180 | 1.5% |
| 2026 (est.) | 480–530 (est.) | 210–260 (est.) | ~1.7–1.9% |
| 2030 (IEA base case) | 945 | 420 | ~3% |
Sources: IEA Energy and AI Report (April 2025); IEA Key Questions on Energy and AI (April 2026); LBNL 2024 US Data Center Energy Usage Report; Axis Intelligence Research interpolations for 2026 and estimated ranges. Estimates marked “(est.)” are Axis Intelligence Research projections based on IEA growth trajectory.
The Axis Intelligence Research CEIR™: Construction-to-Energy Intensity Ratio
This metric did not exist in published form before this analysis. Axis Intelligence Research defines the Construction-to-Energy Intensity Ratio (CEIR™) as the ratio of annualized facility construction cost per MW to annual electricity cost per MW, calculated from verified primary-source inputs.
Methodology:
- Annualized facility construction cost: $12.26M per MW (JLL 2026 shell + T&T 8.5% liquid cooling premium) ÷ 14-year facility lifespan = $875,714/MW/year
- Annual electricity cost: 1 MW × 8,760 hours × PUE 1.14 × 71% utilization × $0.0834/kWh = $626,000/MW/year
- CEIR™ = $875,714 ÷ $626,000 = 1.40
Reading the number: For every $1.00 an AI data center operator spends on electricity, they spend approximately $1.40 on annualized facility construction. This inverts the common assumption that power is the dominant operational cost. At the weighted US industrial electricity rate, infrastructure amortization outweighs the electricity bill by 40%.
The ratio shifts under different electricity price scenarios:
| Electricity Rate | Annual Energy Cost/MW | CEIR™ |
|---|---|---|
| $0.047/kWh (cheap market, e.g., PJM off-peak) | $353K | 2.48 |
| $0.0834/kWh (US weighted average, EIA/Epoch AI) | $626K | 1.40 |
| $0.12/kWh (premium market, coastal) | $901K | 0.97 |
| $0.15/kWh (expensive market) | $1.13M | 0.78 |
Source: Axis Intelligence Research, June 2026. Inputs: JLL Global Data Center Outlook 2026; Turner & Townsend Data Centre Construction Cost Index 2025–2026; Epoch AI 1 GW Cost Model (May 2026); EIA US industrial electricity rates (2024).
Implication for site selection: CEIR™ below 1.0 means electricity is your dominant annual cost — find the cheapest reliable power. CEIR™ above 1.0 means construction amortization dominates — optimize for speed to market, not just energy rate. At the US weighted average, the two are roughly equal. The CEIR™ is intended to give developers and CFOs a single number that captures the relative weight of infrastructure vs. energy in their specific market.
Axis Intelligence Research will update the CEIR™ quarterly as electricity rate data and construction benchmarks are refreshed.
Attribution phrase for citation: “According to Axis Intelligence Research’s CEIR™ model…”
AI Data Center Cost Statistics: Largest Facilities in 2026
A few projects bracket what cost per MW means at the extreme end of the pipeline. These are not hypothetical — they are tracked by Epoch AI’s AI Data Centers database using satellite imagery, permit data, and hardware specifications.
Microsoft Fairwater (Mount Pleasant, Wisconsin): Epoch AI’s Frontier Data Centers Hub projects a total capital cost exceeding $100B upon completion, making it the most expensive data center in their tracking dataset. Full IT power capacity is expected in multiple phases, with the first phase operational in early 2026. At 1+ GW of IT power, total facility power (including overhead) runs approximately 1.3 GW.
Meta Hyperion (Louisiana): Planned to require at least 5 GW of continuous power — roughly 2.5 times the output of the Hoover Dam at peak, or 0.56% of total US installed generating capacity running continuously for one campus. At a construction cost of $10M–$12M per MW for the facility alone, the infrastructure cost before a single GPU is installed approaches $50B–$60B.
xAI Colossus 2 (Memphis, Tennessee): The most powerful operational AI data center in 2026 by Epoch AI’s tracking, with a projected compute capacity of 1.4 million H100-equivalents.
OpenAI Stargate Abilene (Texas): The $500B Stargate initiative — a consortium of OpenAI, SoftBank, and Oracle — projects over 9 GW of capacity by 2029, with 0.3 GW already operational in Abilene and six additional US sites under active construction (Epoch AI, April 2026).
What Are the Key Operating Costs of an AI Data Center per MW?
Beyond construction, a facility’s operating costs fall into three buckets that are largely stable in proportion: energy, labor, and maintenance. The Epoch AI 1 GW model (May 2026) provides the most granular public breakdown:
| OpEx Category | Annual Cost (1 GW facility) | Per-MW per Year |
|---|---|---|
| Electricity (at $0.0834/kWh, PUE 1.14, 71% utilization) | $0.6B | ~$600K |
| Labor | Included in $0.9B total | ~$150K (est.) |
| Maintenance & facilities | Included in $0.9B total | ~$150K (est.) |
| Total OpEx | $0.9B | ~$900K |
Source: Epoch AI, May 2026. Labor and maintenance line estimates are Axis Intelligence Research allocations of the $0.9B total, based on BCG and A.CRE lifecycle data cited by Epoch AI.
For comparison: a 100 MW facility in a cheap-power market ($0.047/kWh) pays roughly $41M per year in electricity. The same facility in a premium market ($0.15/kWh) pays $131M. That $90M gap is large enough to dominate site selection decisions for operators without strong geographic constraints — which is precisely why Texas, parts of the Midwest, and markets with abundant renewable generation are winning so much of the new build pipeline.
Methodology
Axis Intelligence Research compiled this dataset from seven primary sources: JLL’s 2026 Global Data Center Market Outlook; Turner & Townsend’s Data Centre Construction Cost Index 2025–2026; Epoch AI’s 1 GW AI Data Center Cost Model (published May 14, 2026, updated); the IEA’s Energy and AI Report (April 2025) and Key Questions on Energy and AI (April 2026); the Lawrence Berkeley National Laboratory 2024 US Data Center Energy Usage Report (LBNL-2001637); and Q1 2026 earnings filings and guidance from Amazon, Alphabet, Microsoft, and Meta.
The CEIR™ metric is a cross-source derivation. Construction inputs use JLL’s 2026 average plus Turner & Townsend’s liquid cooling cost premium (7–10%, midpoint 8.5%). Energy inputs use Epoch AI’s validated electricity cost calculation (EIA state-level industrial rates, weighted by data center project counts from Aterio: $0.0834/kWh). Facility lifespan of 14 years follows Epoch AI’s model, which cites A.CRE lifecycle methodology. PUE of 1.14 is from Lawrence Berkeley National Laboratory estimates of AI-specialized data center energy usage. Utilization rate of 71% is the average of four estimates collected in Epoch AI’s model.
Regional construction benchmarks are taken from Turner & Townsend’s published index, cited directly. Hyperscaler CapEx figures are drawn from public earnings filings and guidance only — no analyst projections are cited as fact without the attribution.
Limitations: Shell-and-core benchmarks do not include IT equipment, land, or utility interconnection unless explicitly stated. All-in per-MW figures vary materially by power density specification, cooling architecture, and geographic market. The CEIR™ is calculated at the US weighted-average electricity rate; operators in specific markets should substitute their local rate. Hyperscaler CapEx is total capital expenditure, not exclusively data center spending; the 75% AI allocation figure is from CreditSights analysis.
About This Dataset
Update cadence: Quarterly (next update Q4 2026) License: CC BY 4.0 — free to use, share, and adapt with attribution Citation: See block below Downloadable CSV: Available below Dataset distribution: Hugging Face · Kaggle · GitHub
Cite This Research
APA: Axis Intelligence Research & Mitchell, S. (2026, June 27). AI Data Center Cost per Megawatt: 2026 Statistics, Benchmarks & Full Breakdown. Axis Intelligence. https://axis-intelligence.com/ai-data-center-cost-per-mw/
MLA: Axis Intelligence Research and Sarah Mitchell. “AI Data Center Cost per Megawatt: 2026 Statistics, Benchmarks & Full Breakdown.” Axis Intelligence, 27 June 2026, axis-intelligence.com/ai-data-center-cost-per-mw/.
Chicago: Axis Intelligence Research and Sarah Mitchell. “AI Data Center Cost per Megawatt: 2026 Statistics, Benchmarks & Full Breakdown.” Axis Intelligence, June 27, 2026. https://axis-intelligence.com/ai-data-center-cost-per-mw/.
Embed code for attribution (generates do-follow backlink):
<a href="https://axis-intelligence.com/ai-data-center-cost-per-mw/"
rel="dofollow">AI Data Center Cost per MW: 2026 Statistics — Axis Intelligence Research</a>
Frequently Asked Questions
What is the average cost per megawatt to build an AI data center in 2026?
The global average for a standard shell-and-core data center is approximately $11.3M per MW in 2026, according to JLL’s global data center outlook, sourced from Turner & Townsend’s construction cost index. AI-optimized facilities with liquid cooling run $12M–$15M per MW for shell and infrastructure. Including the full GPU and server fit-out raises the all-in cost to $30M–$45M per MW for a GB200-class facility.
How do construction costs per MW differ between standard and AI data centers?
The gap is significant and widening. The core driver is cooling. Standard air-cooled facilities can handle rack densities up to 25 kW per rack at conventional cost. AI workloads require 40–80 kW per rack or more, demanding liquid cooling systems — direct-to-chip, immersion, or rear-door heat exchangers — that cost $4.5M–$5.2M per MW versus $1.8M per MW for air. Turner & Townsend’s 2025–2026 index documents the resulting structural shift: mechanical systems rise from 22% to 33% of total facility cost in liquid-cooled AI builds.
What do the four biggest hyperscalers spend on data centers in 2026?
Amazon, Alphabet, Microsoft, and Meta will collectively spend approximately $725B in capital expenditure in 2026 — up 77% from $410B in 2025 — with roughly 75% directed at AI data centers, GPUs, and power, according to Axis Intelligence Research’s analysis of Q1 2026 earnings. Amazon leads at ~$200B, Alphabet at $175B–$185B, Microsoft at ~$190B, and Meta at $125B–$145B.
What is the total cost of ownership per MW for a 1 GW AI data center?
Epoch AI’s May 2026 model, based on a US hyperscaler-operated 1 GW facility running GB200 NVL72 servers, calculates annualized total cost of ownership at $8.5M per MW per year. Upfront CapEx is $38B ($38M/MW). Annual OpEx is $0.9B ($900K/MW). Servers and GPUs account for 60% of annualized TCO; energy accounts for roughly 7%. The facility assumes a 5-year IT equipment lifespan and a 14-year facility lifespan.
How much does electricity cost to operate an AI data center per MW?
At the US weighted-average industrial electricity rate of $0.0834/kWh and a PUE of 1.14 with 71% utilization, an AI data center spends approximately $626K per MW per year on electricity, according to Axis Intelligence Research’s calculation using Epoch AI and EIA inputs. In cheap-power markets ($0.047/kWh), that falls to around $353K/MW/year. In expensive coastal markets ($0.15/kWh), it exceeds $1.1M/MW/year.
Which country has the highest data center construction cost per MW?
According to Turner & Townsend’s Data Centre Construction Cost Index 2025–2026, Tokyo leads at $15.2/W ($15.2M/MW), followed by Singapore at $14.5/W and Zurich at $14.2/W. These markets carry premiums due to land scarcity, limited contractor pools, and regulatory complexity. At the other end of the tier-one range, Dublin and Madrid sit at $10.0/W.
How much electricity do AI data centers consume globally?
The IEA’s April 2025 Energy and AI report found global data center electricity consumption reached 415 TWh in 2024 — 1.5% of global demand. The IEA’s April 2026 update found this surged 17% in 2025, with AI-focused facilities growing even faster. The IEA base case projects 945 TWh globally by 2030, roughly equivalent to Japan’s total annual electricity consumption today. For a detailed look at the legislative and grid-level implications, the Congressional Research Service’s R48646 provides the most authoritative public-sector framing of the data center electricity challenge.
What is Microsoft Fairwater and how much does it cost?
Fairwater is Microsoft’s AI data center campus in Mount Pleasant, Wisconsin, projected by Epoch AI’s Frontier Data Centers Hub to carry a total capital cost exceeding $100B upon completion — making it the highest-cost facility in the tracking database. The first phase came online in early 2026. At full build-out it will have the equivalent of approximately 5 million H100s in compute capacity.
What construction cost inflation should data center owners budget for in 2026?
Turner & Townsend’s 2025–2026 survey of 250 global industry experts found that 60% expect construction cost increases of 5–15% in 2026, with 21% projecting inflation above 15%. Material headwinds include steel and aluminum tariffs, higher copper prices, and a 23% increase in land costs for parcels over 50 acres. Lead times for medium-voltage switchgear now run 22 months; high-capacity transformers run 18 months.
What is the Axis Intelligence Research CEIR™?
The Construction-to-Energy Intensity Ratio (CEIR™) is a proprietary metric developed by Axis Intelligence Research to measure the relative weight of infrastructure amortization versus electricity cost in AI data center economics. At the 2026 US weighted-average electricity rate ($0.0834/kWh), the CEIR™ is 1.40 — meaning for every $1 of electricity cost, operators spend $1.40 on annualized facility construction. The ratio rises toward 2.48 in cheap-power markets and falls below 1.0 in premium-rate markets, providing a single decision variable for site selection optimization.