Data Center Construction Cost Statistics 2026
By Axis Intelligence Research
Co-author: Sophie Winslow | Last updated: August 3, 2026 | License: CC BY 4.
Building a standard data center costs $11.3 million per megawatt globally in 2026, according to JLL’s forecast. That figure describes a shell. Fitting the same megawatt out for AI adds as much as $25 million more, which means the building most cost benchmarks measure is the smaller half of the bill.
Quick Answer
JLL forecasts the average global shell-and-core construction cost at $11.3 million per MW in 2026, up 6% from $10.7 million in 2025. Turner & Townsend identifies a 7–10% construction premium for AI-grade facilities over standard ones of equivalent IT capacity. Stacking JLL’s $25 million per MW AI tenant fit-out on top gives an Axis Intelligence Research figure of $37.3 million per operational AI megawatt — the ACCI™ (AI Construction Cost Premium Index) at 3.30x.
Key Findings
- According to JLL’s 2026 Global Data Center Outlook, average global shell-and-core construction cost is forecast at $11.3 million per MW in 2026, up from $7.7 million per MW in 2020.
- Axis Intelligence Research calculates the ACCI™ at 3.30x, meaning a fully equipped AI megawatt costs 3.30 times a standard shell-and-core megawatt — a dollar premium of $26.0 million per MW.
- Axis Intelligence Research finds that tenant fit-out represents 67.1% of the delivered cost of an AI megawatt, which is why shell-only benchmarks understate AI facility economics by roughly two-thirds.
- Axis Intelligence Research finds that JLL’s implied structure share of the data center supercycle (37.7%) and the structure share disclosed by Microsoft and Alphabet on their own earnings calls (36.7%) converge within one percentage point, despite being derived by entirely independent methods.
- According to Turner & Townsend’s Data Centre Cost Index 2025–2026, Tokyo remains the most expensive construction market globally at $15.2 per watt, ahead of Singapore at $14.5 and Zurich at $14.2.
How Much Does It Cost to Build a Data Center in 2026?
The question has no single answer, and the first duty of any source is to say why before quoting a number. “Cost” does five different jobs depending on who is answering.
Shell-and-core is the building: structure, envelope, HVAC rough-in, electrical backbone. Powered shell adds the utility interconnect. Mechanical and electrical fit-out brings cooling and power distribution. Tenant fit-out is what the occupier installs — liquid cooling manifolds, rack infrastructure, high-voltage distribution. Active IT, the GPUs and networking, sits outside almost every published construction benchmark, including every figure on this page.
Get the scope straight, then read the number.
The Global Shell-and-Core Benchmark
JLL’s 2026 Global Data Center Outlook puts the global average shell-and-core cost at $11.3 million per MW for 2026, a 6% increase on 2025.
| Year | Global average, shell-and-core | Change | Source |
|---|---|---|---|
| 2020 | $7.7M per MW | — | JLL |
| 2025 | $10.7M per MW | 7% CAGR from 2020 (JLL) | JLL |
| 2026 (forecast) | $11.3M per MW | +5.6% | JLL |
Two things about this benchmark that most citations of it omit, and that change how it should be used.
First, the scope. JLL’s figure describes a single-tenant, 50 MW, air-cooled facility, shell and core only. It is not a hyperscale AI campus, and it is not a colocation build. Quoting $11.3M/MW for a 300 MW liquid-cooled project is a category error.
Second, the arithmetic. JLL states a 7% compound annual growth rate for 2020–2025. Axis Intelligence Research computes 6.8% from the published endpoints of $7.7M and $10.7M. The difference is rounding in the source, not a dispute — we flag it because anyone recomputing our series should land on 6.8% and know why.
Sophie Winslow, Industrial Technology: A benchmark with an undisclosed baseline is a number you can be argued into agreeing with. Fifty megawatts, single tenant, air-cooled — that is a specific building, and it is not the building anyone is announcing in 2026. Every time this figure gets quoted against a gigawatt campus, someone is either being careless or is pricing a bid they intend to win and renegotiate. Ask what shape of facility the average describes before you let it into a pro forma.
What Is the ACCI™ (AI Construction Cost Premium Index)?
The gap between a standard megawatt and an AI-operational one is the single most consequential number in this market, and no published source states it end to end. JLL publishes the shell. Turner & Townsend publishes the AI construction premium. JLL publishes the fit-out. Nobody stacks them.
Axis Intelligence Research does, and calls the result the ACCI™.
ACCI™ v1.1 — formula and inputs
ACCI = (standard shell × (1 + AI construction premium) + AI tenant fit-out)
÷ standard shell
| Cost layer | Standard | AI-optimized | Source |
|---|---|---|---|
| Shell and core | $11.3M/MW | $12.26M/MW | JLL 2026 + Turner & Townsend premium |
| AI construction premium | — | +8.5% (midpoint of 7–10%) | Turner & Townsend 2025–2026 |
| Tenant technology fit-out | Not applicable | $25.0M/MW | JLL 2026 |
| All-in | $11.3M/MW | $37.26M/MW | Axis Intelligence Research |
| ACCI™ multiplier | 1.00x baseline | 3.30x | Axis Intelligence Research |
ACCI™ = 37.26 ÷ 11.3 = 3.30x. The dollar premium is $26.0 million per operational AI megawatt.
Sensitivity and the scope asymmetry we are not hiding
Turner & Townsend publishes a range, not a point. Running the ACCI™ at both ends: a 7% premium gives 3.28x, a 10% premium gives 3.31x. The index is insensitive to that input — it moves by 0.03x across the full published range, because the fit-out term dominates. That is worth stating plainly, because it tells a user that arguing about the construction premium is arguing about the wrong variable.
The asymmetry that does matter: the ACCI™ compares an AI megawatt including tenant fit-out against a standard megawatt excluding it. That is deliberate — the comparison answers “what does it cost to make a megawatt do AI work versus conventional work,” and a standard cloud megawatt does not carry a $25M fit-out. But it is not a like-for-like construction comparison, and presenting it as one would be misleading. For readers who want the like-for-like figure, it is the Turner & Townsend premium itself: 1.085x shell-to-shell. Both readings are published. We will not merge them into a single number that means neither thing.
Versioning note: ACCI™ v1.0 was published June 28, 2026 at the same 3.30x reading. v1.1 changes no inputs and no result. It adds the disclosed scope asymmetry, the shell-only companion reading, and the sensitivity range. Historical readings are not rewritten.
Where Does the Money Actually Go?
The building is the cheap part, and the ratio is the finding.
Axis Intelligence Research calculates that of the $37.26M delivered cost of an AI megawatt, $25M — or 67.1% — is tenant fit-out. Shell and core is 32.9%.
That single ratio explains a pattern that confuses people reading construction data: US data center construction put in place, which counts structures only, keeps growing more slowly than hyperscaler capital expenditure. It is not a contradiction. Two-thirds of an AI megawatt was never going to appear in a construction statistic.
The structure share convergence
Here is a cross-check that no publication has run, and it is the strongest evidence we have that both sides of this market are being measured correctly.
Route one, from JLL’s own forecasts. JLL projects roughly 100 GW of new global capacity between 2026 and 2030, and a total investment supercycle approaching $3 trillion. At $11.3M per MW, 100,000 MW of shell-and-core costs $1.13 trillion — 37.7% of the $3 trillion.
Route two, from company disclosures. On their second-quarter 2026 earnings calls, Microsoft said roughly two-thirds of capital spending goes to shorter-lived assets like CPUs and GPUs, and Alphabet said about 60% goes to servers. The two-firm mean implies a structure, networking and land share of 36.7%, a figure developed on our data center construction spending page.
Two entirely independent methods — a real estate forecaster’s capacity model and two companies’ capital allocation disclosures — land 1.0 percentage point apart.
| Method | Implied structure share | Basis |
|---|---|---|
| JLL capacity × cost benchmark | 37.7% | 100 GW × $11.3M/MW ÷ $3T |
| Microsoft + Alphabet disclosures | 36.7% | Two-firm unweighted mean |
| Divergence | 1.0 point | Axis Intelligence Research |
What this buys a reader: confidence that roughly 37 cents of every data center dollar is a building, and 63 cents is equipment. Neither source says that. Both imply it.
Sophie Winslow: Convergence like this is not proof, and I would not present it as proof. Two methods can share an upstream assumption and agree for bad reasons — JLL’s $3 trillion and the hyperscaler capex figures are not fully independent, since JLL is partly modelling the same companies’ announced spending. What the agreement does establish is that no one has made an order-of-magnitude error. In a market where the announced numbers change monthly, that is worth more than it sounds.
Data Center Construction Costs by Market
Turner & Townsend’s Data Centre Cost Index 2025–2026 is the only index focused exclusively on data center construction, now in its ninth year, covering 52 markets and drawing on a survey of 250 industry experts.
| Market | Cost to build | Region | Source |
|---|---|---|---|
| Tokyo | $15.2 per watt | APAC | Turner & Townsend |
| Singapore | $14.5 per watt | APAC | Turner & Townsend |
| Zurich | $14.2 per watt | Europe | Turner & Townsend |
| Silicon Valley | $13.3 per watt | North America | Turner & Townsend |
| London | $12.0 per watt | Europe | Turner & Townsend |
| Frankfurt | $11.6 per watt | Europe | Turner & Townsend |
Turner & Townsend attributes the top three positions to land availability, limited contractor pools, and workforce dynamics rather than to materials pricing.
[Older data] This index was published in November 2025 and covers the 2025 cost year. Turner & Townsend publishes annually in November, so the 2026–2027 edition does not yet exist. These are the most recent published per-watt benchmarks available; readers pricing 2026 work should treat them as a floor given the inflation expectations below.
Cost Inflation Is Decelerating, Not Reversing
| Metric | Value | Period | Source |
|---|---|---|---|
| Traditional data center cost inflation | 5.5% | 2025 | Turner & Townsend |
| Same metric, prior year | 9.0% | 2024 | Turner & Townsend |
| All-sector construction inflation | 4.2% | 2025 | Turner & Townsend |
| Respondents reporting 6–15% bid price rises | 47% | Past year | Turner & Townsend |
| Respondents reporting rises above 15% | 21% | Past year | Turner & Townsend |
| Respondents expecting 5–15% increases | 60% | 2026 | Turner & Townsend |
| Respondents expecting above 15% | 21% | 2026 | Turner & Townsend |
Traditional data center inflation running at 5.5% against 4.2% for construction generally means data centers are still a premium sector, but the gap has narrowed sharply from 2024. Turner & Townsend’s own framing is worth taking seriously: their survey found power and supply chain capacity, not cost, is what concerns the sector.
What Constrains Delivery: Power, Not Price
JLL reports that US average equipment lead times have stretched to 42 weeks, 83% above 2019 levels, and that 57% of projects in 2025 experienced delays of at least three months. JLL identifies speed to power as the primary criterion driving site selection, ahead of community support, latency and proximity to customers.
Grid connection timelines in constrained European markets have moved beyond anything a construction schedule can absorb: JLL Research puts the wait at roughly eight years in London and ten years in Amsterdam for a new 50 MW facility.
The U.S. Department of Energy’s Electricity Demand Growth Resource Hub exists because this constraint is not resolving through market forces. DOE maintains more than 30 programs supporting data center energy needs, and the Secretary of Energy Advisory Board’s working group on powering AI and data center infrastructure issued 16 recommendations on grid modernization. The federal apparatus is a reasonable proxy for the severity of the bottleneck; agencies do not build resource hubs for problems that are clearing on their own.
DOE also published the LBNL assessment of data center electricity demand growth, which remains the underlying federal reference for load-growth projections.
Sophie Winslow: Forty-two weeks of equipment lead time against an eighteen-month build is not a procurement problem, it is a sequencing problem, and the two get priced very differently. The schedule no longer starts at groundbreaking. It starts when the switchgear order is signed, which for a lot of 2026 projects was 2024. Anyone quoting cost per megawatt without stating the assumed energization date is quoting a number with a hole in it.
Methodology
Collection. Every figure was retrieved during the production session from the source named in the accompanying CSV, with URL and retrieval date logged. JLL’s 2026 Global Data Center Outlook and Turner & Townsend’s Data Centre Cost Index 2025–2026 pages were fetched and read directly. DOE resource pages were verified live.
ACCI™ v1.1. Formula, all three inputs, both readings, and the sensitivity range are stated in full above. A competent outsider can recompute 3.30x from the published inputs.
Structure share convergence. Route one: 100,000 MW × $11.3M ÷ $3,000,000M = 37.7%. Route two: unweighted mean of Microsoft’s ~66.7% and Alphabet’s 60.0% silicon share, inverted, = 36.7%. Both routes and the divergence are disclosed. We do not claim independence beyond what the sources support — JLL’s supercycle model partially reflects the same companies’ announced capital plans, so the two routes share some upstream information.
Currency and unit conversion. Turner & Townsend publishes in US dollars per watt; JLL publishes in millions of US dollars per megawatt. One dollar per watt equals one million dollars per megawatt. No exchange rate conversion was applied.
Limitations. Global averages conceal wide market variation; an $11.3M/MW average spans roughly $8M to $15M+ depending on market. All figures are shell-and-core unless explicitly labelled all-in. Active IT hardware is excluded throughout. The Turner & Townsend index reflects the 2025 cost year and is flagged as older data above. The ACCI™ compares scopes that are deliberately asymmetric, as disclosed.
What we did not compute. We do not publish a cost-per-square-foot series. Square-foot benchmarks depend on rack density assumptions that vary by a factor of ten between air-cooled and AI facilities, which makes a single figure actively misleading rather than merely imprecise.
About This Dataset
Contents: cost benchmarks by year and market, AI premium components, ACCI™ readings, cost inflation survey results, delivery constraint metrics, and the structure share convergence calculation.
Temporal coverage: 2020–2030 (forecasts labelled).
Spatial coverage: Global, 52 markets. License: CC BY 4.0.
Citation: Axis Intelligence Research, Data Center Construction Cost Statistics 2026, 2026.
Cite This Research
APA: Axis Intelligence Research & Winslow, S. (2026, August 3). Data center construction cost statistics 2026: $11.3M per MW and the fit-out that doubles it. Axis Intelligence. https://axis-intelligence.com/data-center-construction-cost-statistics/
MLA: Axis Intelligence Research, and Sophie Winslow. “Data Center Construction Cost Statistics 2026: $11.3M per MW and the Fit-Out That Doubles It.” Axis Intelligence, 3 Aug. 2026, axis-intelligence.com/data-center-construction-cost-statistics/.
Chicago: Axis Intelligence Research and Sophie Winslow. “Data Center Construction Cost Statistics 2026: $11.3M per MW and the Fit-Out That Doubles It.” Axis Intelligence, August 3, 2026.
Frequently Asked Questions
How much does it cost to build a data center in 2026?
JLL forecasts an average global shell-and-core cost of $11.3 million per MW in 2026. That benchmark describes a single-tenant, 50 MW, air-cooled facility. An AI-optimized megawatt including tenant fit-out reaches approximately $37.3 million per MW, per the Axis Intelligence Research ACCI™ calculation.
What is the ACCI™?
The AI Construction Cost Premium Index is an Axis Intelligence Research metric measuring the cost of a fully equipped AI megawatt against a standard shell-and-core megawatt. It reads 3.30x as of Q2 2026, a dollar premium of $26.0 million per MW.
Why are AI data centers more expensive to build?
Two reasons, and the second dominates. Construction itself carries a 7–10% premium for AI-grade facilities of equivalent IT capacity, per Turner & Townsend. But tenant technology fit-out adds up to $25 million per MW on top, per JLL — which Axis Intelligence Research calculates as 67.1% of the delivered cost of an AI megawatt.
How much of a data center’s cost is the building itself?
Axis Intelligence Research finds that roughly 37% of data center investment is structures, networking and land, with the remainder in equipment. JLL’s capacity and cost forecasts imply 37.7%; Microsoft’s and Alphabet’s own capital allocation disclosures imply 36.7%.
Which is the most expensive market to build a data center?
Tokyo, at $15.2 per watt, according to Turner & Townsend’s Data Centre Cost Index 2025–2026, followed by Singapore at $14.5 and Zurich at $14.2. Turner & Townsend attributes this to land availability, contractor pool size and workforce dynamics.
Are data center construction costs still rising?
Yes, but more slowly. Turner & Townsend measured 5.5% inflation for traditional data centers in 2025, down from 9.0% in 2024, against 4.2% across construction generally. Looking forward, 60% of their surveyed industry leaders expect increases of 5–15% in 2026 and 21% expect more than 15%.
What delays data center projects most in 2026?
Power, not cost. JLL reports US equipment lead times at 42 weeks, 83% above 2019 levels, with 57% of 2025 projects delayed at least three months. Turner & Townsend’s survey found power and supply chain capacity, rather than construction cost, is the sector’s leading concern.
How does construction cost relate to total US data center spending?
They measure different things. This page covers unit cost per megawatt. Aggregate construction dollars flowing into the US sector — Census put-in-place data, currently $50.7 billion annualized — are covered on our data center construction spending page.
Related Research
- Data Center Construction Spending — Census put-in-place data and the DCSR™ capex conversion ratio
- AI Data Center Cost per MW — unit economics and CEIR™
- Data Center Electricity Cost by State — DCECI™ operating cost rankings
- Liquid Cooling Data Center Statistics — cooling architecture and PUE
- AI Data Center Energy Consumption Statistics — TWh by operator and region
- AI Data Center Market Size — market sizing and capex
Correction, August 3, 2026: an earlier version of this article stated Alphabet’s 2026 capital expenditure guidance as $175–190 billion and a Big Four 2026 total of $680–725 billion. Alphabet raised its guidance to $195–205 billion on July 29, 2026. The hyperscaler capital expenditure table has been removed from this article, which covers unit construction cost; current capex figures are maintained on the AI Capex Tracker. The 2025 Big Four capex base previously stated as $378 billion is a cash property, plant and equipment figure excluding finance-lease right-of-use assets, and is now labelled as such.