The building takes two years. The power takes five. And the four biggest buyers on earth will spend $725 billion this year on exactly this problem. That is data center construction in 2026 in three sentences, and it explains the strange market underneath the AI boom: shells without megawatts, megawatts without shells, GPUs waiting on both, and a grid queue where 77% of the projects that ever entered never came out. Building an AI data center is entirely doable, people do it every week, but the honest version of “how” is mostly about power, procurement, and patience, not concrete.
This guide walks the full pipeline the way we walk it at MillionMiner Colocation Hosting: seven stages from raw requirement to racks running, with the real 2026 numbers at each one, the cost per megawatt with nothing hidden, the timeline mismatch that kills projects, and the three alternatives that are usually the smarter answer. We run our own facilities, host other people's hardware, and source capacity, sites, and equipment for clients on both sides of the market, so this is the map we actually use, published.
This guide walks the full pipeline the way we walk it at MillionMiner Colocation Hosting: seven stages from raw requirement to racks running, with the real 2026 numbers at each one, the cost per megawatt with nothing hidden, the timeline mismatch that kills projects, and the three alternatives that are usually the smarter answer. We run our own facilities, host other people's hardware, and source capacity, sites, and equipment for clients on both sides of the market, so this is the map we actually use, published.
The short answer
- Cost: standard shell-and-core averages $11.3M per MW in 2026 (JLL); an AI-optimized facility runs $20-37M per MW all-in, before GPUs, servers, or networking. A 1 GW AI campus pencils near $38 billion.
- Time: 18-30 months to build, but roughly 5 years average in the US grid interconnection queue, 8-10 in parts of Europe. Power, not construction, is the critical path.
- The pipeline: land with real power → utility commitment → density design → procurement → fit-out → commissioning → operations. Most failed projects die at stages 1-2.
- The alternative most teams actually need: sourcing existing capacity, hosted racks, powered shells, or operating sites, in weeks to months instead of years.
The 7 stages of building an AI data center
Data center construction looks like a real-estate project and behaves like a supply-chain project wearing a hard hat. The seven stages below are the whole game, in the order they must be won:
Notice what the pipeline is not: it is not “pour foundation, raise walls, install servers.” The construction itself, stage 5, is the most predictable part of the entire venture. The stages that decide whether your megawatts ever exist are the two at the top, which is why the rest of this guide spends its time where the projects actually live and die.
Stage 1-2: finding land that actually has power
The scarce asset in 2026 is not land, and it is not even capital; it is powered land, sites where a utility will actually deliver the megawatts. A parcel without an interconnection path is a field; a powered shell with 20 MW energized is one of the most sought-after assets in industrial real estate. The numbers explain the obsession: the average US interconnection request now takes roughly five years from submission to commercial operation, double the wait of 15 years ago, and of every project that entered the US queue between 2000 and 2019, only 13% had reached operation by the end of 2024; 77% withdrew. In Europe it is harsher still: a new 50 MW connection quotes around 8 years in London and 10 in Amsterdam.
Site selection therefore starts at the substation, not the property line: existing utility capacity, queue position, transmission proximity, and, increasingly, the option to generate behind the meter while the queue grinds, the playbook our PPA guide covers and the one Bitcoin miners wrote first by siting at stranded generation. Water rights and cooling constraints belong in this stage too, since siting decides most of a facility's water profile before any engineer touches a drawing. And this stage is exactly why sourcing exists as a service: a database of sites, powered shells, and capacity that already cleared the queue is worth more than any amount of enthusiasm for greenfield.
How much does it cost to build a data center in 2026?
The benchmark number: $11.3 million per megawatt for standard shell-and-core construction, per JLL's 2026 global outlook, up from $7.7M in 2020, a 7% compound annual climb. AI density changes the equation entirely. The shell itself carries a 7-10% premium for structure and power routing, liquid cooling plant runs $4.5-5.2M per MW against $1.8M for air, tenant AI fit-out adds up to $25M per MW, and the all-in range for a fully built AI facility lands between $20 and $37 million per megawatt, roughly 3.3x the standard build. Line by line:
| Cost component | Standard facility | AI-optimized facility | Notes |
|---|---|---|---|
| Shell-and-core | ~$11.3M / MW | +7-10% premium | JLL 2026 global average |
| Cooling plant | ~$1.8M / MW (air) | $4.5-5.2M / MW (liquid) | DLC mandatory at AI density |
| Electrical systems | 40-45% of budget | 40-45% of budget | The largest single share |
| Mechanical share | ~22% of budget | ~33% of budget | Turner & Townsend index |
| Utility interconnection | $5-25M / project | $5-25M+ / project | Distance to capacity decides |
| Tenant fit-out | Modest | Up to $25M / MW | Manifolds, PDUs, rack infra |
| All-in build | ~$8-12M / MW | $20-37M / MW | ~3.3x standard |
| Active IT (GPUs, servers, network) | Excluded | Excluded, and often exceeds the building | No benchmark includes it |
Benchmarks per JLL 2026 Outlook and Turner & Townsend's 2025-26 Cost Index; regional variation runs up to 40%, and per-square-foot costs now average near $1,000 (Cushman & Wakefield). Reproduce the table with attribution and a link.
Scale that against the market and the capital numbers stop feeling abstract: the four largest hyperscalers are deploying roughly $725 billion in 2026, up 77% year over year, about three quarters of it aimed at AI facilities, chips, and power; US data center construction spending alone hit $85 billion, nearly doubling since 2024. The lesson buried in the boom is the one every first-time builder learns expensively: the building is not the big line item. The silicon inside it is, which is why the procurement stage deserves its own section.
How long does it take to build a data center?
Construction itself runs 18 to 30 months from concept to commissioning for a hyperscale build, up from about 12 months before switchgear lead times and permitting tightened. But the construction schedule is not the project schedule:
The three-year gap between an 18-month build and a five-year queue is where announced megawatts go to die: design fees spent, contracts signed, and the energization date drifting annually to the right. The famous counter-example proves the rule rather than breaking it: xAI's Colossus went from empty building to 100,000 GPUs hashing in 122 days, and it did so precisely by not building, taking over an existing Memphis factory shell and bringing generation to the site rather than waiting on a queue. Strip the branding and that is a sourcing story: the fastest AI data center ever stood up was the one where the land, the shell, and the power path already existed. That, in one anecdote, is the entire case for sourcing capacity before committing to greenfield.
Ready to Start Mining?
Free worldwide DDP shipping. Professional hosting from $0.055/kWh.
Stage 4: the equipment that actually sets the schedule
Procurement, not civil works, is the modern schedule-setter, and it splits into two very different shopping lists. The facility list is the long-lead heartbreaker: high-voltage transformers, switchgear, generators, chillers, and CDUs quote in quarters and sometimes years, which is why experienced developers order electrical equipment before drawings finish and why the construction squeeze is really an equipment squeeze. The IT list is the expensive one: the GPUs, AI servers, networking, and storage that no construction benchmark includes and that, for an AI hall, routinely exceed the cost of the building around them. An 8-GPU node draws about 10 kW and a rack of them 30-60, numbers that flow backward into every electrical drawing, so the density decision (which GPU generation at which rack class) has to be locked at stage 3, not discovered at delivery.
Two procurement rules save the most money in practice. Match the hardware to the workload before matching it to the building: the GPU tier that suits your models decides the cooling class, not the other way around. And sequence deliveries against energization honestly: silicon depreciating in a warehouse while the substation waits is the single most expensive form of optimism in this industry, and it is the procurement-coordination problem a single sourcing contact exists to solve.
Two procurement rules save the most money in practice. Match the hardware to the workload before matching it to the building: the GPU tier that suits your models decides the cooling class, not the other way around. And sequence deliveries against energization honestly: silicon depreciating in a warehouse while the substation waits is the single most expensive form of optimism in this industry, and it is the procurement-coordination problem a single sourcing contact exists to solve.
Build, colocate, rent, or source: the honest decision
Here is the section most construction guides omit, because most are written by people selling construction. Building is one path of four, and for the majority of teams asking the question, it is the wrong one:
Rent if your utilization is spiky or under roughly 40%: the cloud GPU math beats ownership there and no facility should exist for a workload that idles. Colocate if you have sustained workloads and want to own hardware without owning a building: GPU hosting delivers racked, powered, cooled capacity in weeks at a published rate. Source if you need megawatts at a scale colocation menus do not cover, or on a timeline the queue will not give you: existing capacity, powered shells, and operating sites change hands constantly, and a sourcing partner with a live database and its own operating experience compresses years into a procurement cycle. Build when you are deploying at a scale and duration where $20-37M per MW amortizes, you can carry five-year power risk, and control is worth the premium. The four are not rivals; mature AI operations usually run three of them at once.
The other side: how data centers make money
Every buyer's problem above is an owner's opportunity, and the economics are worth understanding from both chairs. Data centers earn as wholesale leases (whole halls or buildings to single tenants on long terms), retail colocation (per-rack or per-kW with services layered on), powered-shell leases (the building and the interconnection, tenant brings the fit-out), or managed hosting (capacity plus operations as one service, the model our hosting operations run for both miners and AI hardware). In a market where energized megawatts trade at a premium and 77% of queue entrants never arrive, holders of land with grid position, half-built shells, surplus capacity, or even idle transformers are sitting on exactly what the demand side cannot manufacture. That is the two-sided logic of MillionMiner AI Data Center Colocation: tell us what you need and we source it, or bring us your land, power, capacity, or equipment and we place it.
What this guide cannot decide for you
Three honest limits. Every figure here is a benchmark, not a quote: regional costs vary up to 40%, scope definitions (IT load vs facility power, shell vs fit-out) move numbers by multiples, and your utility's answer outranks any industry average, so treat the tables as screening tools for the conversations that actually price a project. Second, this guide compresses disciplines, grid engineering, construction law, tax structuring, that need their own professionals; it makes you a better client, not a developer. Third, our position: we operate facilities, sell AI hardware, and earn on sourcing mandates, so we benefit when you conclude that existing capacity beats greenfield; the queue statistics making that case are LBNL's, not ours, and they are checkable in Sources.
The bottom line
How do you build an AI data center in 2026? Secure powered land, win a utility commitment, design for your true rack density, procure the long-lead iron early, and budget $20-37M per megawatt before a single GPU, on a timeline where the grid, not the contractor, sets the date. Whether you should is the better question: under 40% utilization, rent; sustained workloads without facility ambitions, colocate; megawatt-scale needs on real deadlines, source existing capacity; hyperscale patience and capital, build. The market's strangest feature is also its biggest opportunity: the assets everyone needs already exist, scattered across owners who do not know what they are worth, and connecting the two sides is faster than concrete every single time. That is the business MillionMiner AI Data Center Colocation is in, and this guide is the map we use to do it.
Frequently asked questions
How much does it cost to build a data center in 2026?
Standard shell-and-core construction averages $11.3 million per megawatt globally (JLL 2026), with mainstream all-in builds at $8-12M per MW. AI-optimized facilities run $20-37M per MW once the density premium, liquid cooling ($4.5-5.2M/MW), and tenant fit-out (up to $25M/MW) are included, roughly 3.3x a standard build. None of those figures include the GPUs, servers, or networking, which for AI halls often cost more than the building.
How much does a 1 MW data center cost?
As a screening range: $8-12 million built to standard density, $20-37 million built AI-ready, plus $5-25 million of utility interconnection scope depending on distance to grid capacity, plus the IT hardware itself. Small facilities also carry worse economics per MW than hyperscale builds because fixed costs (design, permitting, interconnection studies, commissioning) do not shrink with the building, which is why sub-5 MW needs are usually served better by colocation than construction.
How long does it take to build a data center?
The building: 18-30 months from concept to commissioning for a hyperscale project, up from about 12 months before the current squeeze on switchgear and permits. The power: roughly 5 years average in the US interconnection queue, and 8-10 years quoted for new 50 MW connections in London and Amsterdam. Power is the critical path; construction rarely is.
Why are AI data centers so much more expensive than normal ones?
Density. AI racks draw 40-130+ kW versus 5-10 kW for standard IT, which forces liquid cooling ($4.5-5.2M per MW against $1.8M for air), heavier electrical distribution (40-45% of budget), stronger floors, and tenant fit-out that can reach $25M per MW for manifolds, PDUs, and rack infrastructure. Turner & Townsend's index shows mechanical systems climbing from 22% to 33% of total facility cost in liquid-cooled AI builds.
What is the grid interconnection queue and why does it matter?
It is the utility's waiting line for new large connections, and it is the single most project-killing fact in the industry: US requests now average about five years from submission to commercial operation, and of everything that entered the queue from 2000 to 2019, only 13% reached operation by end-2024 while 77% withdrew. Projects routinely spend design and procurement budgets on megawatts that never arrive, which is why powered land and existing capacity trade at premiums.
How did xAI build Colossus in 122 days?
By not building a data center. Colossus went into an existing Memphis factory shell that xAI acquired, with generation brought to the site rather than waiting on a full grid interconnection, and 100,000 GPUs came online in 122 days. It is the strongest recent proof that the fastest path to AI capacity is securing assets that already exist, a sourcing strategy, not a construction achievement, even though it is usually told as one.
Can I buy an existing data center instead of building?
Yes, and in 2026 it is often the rational path: operating facilities, powered shells, and partially built projects trade regularly, and acquiring energized capacity skips the five-year queue entirely. The trade-offs are inherited design (density ceilings, cooling class) and competitive pricing, since everyone else has read the same queue statistics. Sourcing services exist to surface these assets, match them to requirements, and manage the diligence that a construction-first team typically lacks.
What equipment does a data center need?
Two lists. Facility equipment: high-voltage transformers, switchgear, generators, UPS, chillers or CDUs and liquid cooling loops, PDUs, and fire suppression, where transformers and switchgear are the long-lead items that set schedules. IT equipment: the GPU servers, networking fabric, and storage, which no construction benchmark includes and which for AI facilities routinely exceed the building's cost. The density decision, which GPU class at which kW per rack, has to precede the electrical design, not follow it.
How do data centers make money?
Four main models: wholesale leases of whole halls to single tenants on long terms; retail colocation sold per rack or per kW with services; powered-shell leases where the tenant installs its own fit-out; and managed hosting, where capacity and operations are sold as one service. In the current market, owners of energized capacity, grid-positioned land, or surplus equipment hold assets the demand side cannot quickly manufacture, which is why supply-side listings are as valuable as demand-side mandates.
Do I need to build a data center, or should I rent capacity?
Run the utilization test first: below roughly 40% sustained use, hourly cloud rental wins and no facility should exist for the workload. Sustained workloads without facility ambitions belong in colocation or GPU hosting, racked in weeks at published rates. Megawatt-scale requirements on real deadlines are usually best served by sourcing existing capacity. Building wins only at hyperscale duration and control requirements, with the capital to carry $20-37M per MW and the patience to carry the queue.
What should I look for in a data center site?
Power first, always: existing substation capacity, interconnection queue position, transmission proximity, and realistic behind-the-meter options. Then the environmentals that follow siting: water availability and cooling constraints, climate, natural-hazard exposure. Then the classics: fiber routes, land expansion room, tax and permitting climate, and community posture. A mediocre parcel with a real power path outranks a perfect parcel without one in every case that matters.
Sources and notes
Construction benchmarks: JLL 2026 Global Data Center Outlook ($11.3M/MW shell-and-core global average, up from $7.7M in 2020 and $10.7M in 2025; tenant AI fit-out to $25M/MW), sourced via Turner & Townsend's Data Centre Construction Cost Index 2025-2026 (AI shell premium 7-10%; mechanical share 22%→33% in liquid-cooled builds); AI all-in range $20-37M/MW and the ~3.3x premium per 2026 industry benchmark compilations; liquid vs air cooling plant ($4.5-5.2M vs $1.8M per MW), electrical systems at 40-45% of budget, interconnection scope $5-25M+, regional variation to 40%, and $/sq ft near $1,000 (Cushman & Wakefield) per the same 2026 benchmark literature; 1 GW AI campus ~$38B per 2026 development analyses. Timelines: hyperscale construction 18-30 months (up from ~12) per JLL/T&T-derived reporting; US interconnection queue ~5 years average, 13% completion / 77% withdrawal for 2000-2019 entrants, per Lawrence Berkeley National Laboratory research; London 8 years and Amsterdam 10 years for new 50 MW connections per the same 2026 reporting. Market scale: big-four hyperscaler capex ~$725B in 2026 (+77% YoY, ~75% AI-directed) per Q1 2026 earnings analyses; US data center construction spending $85.3B per 2026 industry data. xAI Colossus: 122 days to first 100,000 GPUs in an acquired Memphis facility, per contemporaneous 2024-25 reporting. Rack density and cooling figures cross-reference our GPU hosting requirements research. All figures are planning benchmarks, not quotes; scope definitions (IT load vs facility power, shell vs fit-out vs active IT) materially change any per-MW number and should be fixed before comparing bids. We operate data center and hosting infrastructure, sell AI hardware, and provide compute sourcing services; that commercial position is disclosed in text. Hero: USDA, public domain; body: PtiBzh, CC0, via Wikimedia Commons. Diagrams and tables: original MillionMiner graphics, free to reuse with attribution and a link to this page. Informational content, not investment, legal, or engineering advice.
Sources and notes
Construction benchmarks: JLL 2026 Global Data Center Outlook ($11.3M/MW shell-and-core global average, up from $7.7M in 2020 and $10.7M in 2025; tenant AI fit-out to $25M/MW), sourced via Turner & Townsend's Data Centre Construction Cost Index 2025-2026 (AI shell premium 7-10%; mechanical share 22%→33% in liquid-cooled builds); AI all-in range $20-37M/MW and the ~3.3x premium per 2026 industry benchmark compilations; liquid vs air cooling plant ($4.5-5.2M vs $1.8M per MW), electrical systems at 40-45% of budget, interconnection scope $5-25M+, regional variation to 40%, and $/sq ft near $1,000 (Cushman & Wakefield) per the same 2026 benchmark literature; 1 GW AI campus ~$38B per 2026 development analyses. Timelines: hyperscale construction 18-30 months (up from ~12) per JLL/T&T-derived reporting; US interconnection queue ~5 years average, 13% completion / 77% withdrawal for 2000-2019 entrants, per Lawrence Berkeley National Laboratory research; London 8 years and Amsterdam 10 years for new 50 MW connections per the same 2026 reporting. Market scale: big-four hyperscaler capex ~$725B in 2026 (+77% YoY, ~75% AI-directed) per Q1 2026 earnings analyses; US data center construction spending $85.3B per 2026 industry data. xAI Colossus: 122 days to first 100,000 GPUs in an acquired Memphis facility, per contemporaneous 2024-25 reporting. Rack density and cooling figures cross-reference our GPU hosting requirements research. All figures are planning benchmarks, not quotes; scope definitions (IT load vs facility power, shell vs fit-out vs active IT) materially change any per-MW number and should be fixed before comparing bids. We operate data center and hosting infrastructure, sell AI hardware, and provide compute sourcing services; that commercial position is disclosed in text. Hero: USDA, public domain; body: PtiBzh, CC0, via Wikimedia Commons. Diagrams and tables: original MillionMiner graphics, free to reuse with attribution and a link to this page. Informational content, not investment, legal, or engineering advice.
