The Neocloud Bible: How to Build an AI Data Center

An AI data center is a facility engineered for GPU compute at rack densities of 40 kW and far beyond, compared with the industry mean of 7.6 kW per rack (Uptime Institute, 2025). It is defined by three departures from traditional design: liquid cooling as a baseline, power as the binding site-selection constraint, and east-west network fabric as the performance backbone. Build costs run $20M+ per MW for the facility alone, and the GPUs inside cost roughly twice that again.

This post covers what makes an AI data center different → the build sequence (power, cooling, network, structure) → cost and timeline benchmarks → EU compliance → why the build model itself is changing.

Why an AI data center is a different building

Most data centers were designed for a world that no longer exists. The Uptime Institute's 2025 global survey puts the mean rack density across the industry at 7.6 kW, with 5–9 kW the most common band. An NVIDIA DGX H100 node alone draws 10.2 kW. One server. More than the average rack.

And that was the gentle generation. A GB200 NVL72 rack runs at roughly 120–132 kW. The Rubin Ultra "Kyber" rack arriving in H2 2027 is specified around 600 kW. Per rack.

That's not an upgrade path. It's a different category of building — different power distribution, different cooling, different floor structure, different fire suppression assumptions. Operators who try to retrofit a 2015-era facility into hosting NVIDIA Blackwell-class hardware discover this in the engineering study, usually right before the project dies.

So the honest starting point for anyone asking "how do I build an AI data center" is this: stop thinking about it as a data center with more GPUs. Start thinking about it as an industrial power-and-cooling plant that happens to host compute.

Step 1: Secure power before anything else

Site selection used to start with fiber and land. Now it starts with megawatts.

In the US, projects that reached commercial operation in 2024 had spent an average of 55 months in the interconnection queue, according to Lawrence Berkeley National Laboratory's Queued Up report (2025). Four and a half years. In Europe it's worse: FLAP-D hub grid connections average 7–10 years (IEA, 2025), the Netherlands reports waits up to a decade, and Dublin has paused new data center connection agreements until 2028 entirely.

Which means your shortlist starts from substation capacity and queue position, not postcode. Increasingly it also includes generating your own power. Gas turbines, batteries, behind-the-meter renewables. The industry has shifted from "wait for the grid" to "bring your own electrons," a move we covered in detail in why data centers are done waiting for the grid.

One more AI-specific wrinkle most builders miss: step loads. Synchronized GPU training phases can swing facility draw 30–60% within milliseconds (McKinsey, 2025). Your UPS, harmonic filtering, and distribution have to be sized for the swing, not the average.

Step 2: Design the power architecture for the workload, not the habit

Traditional enterprise design defaults to 2N everywhere. AI changes the math.

GPU training clusters tolerate checkpoint-restart, so many operators now run N or N+1 on the compute blocks and reserve 2N for the network, storage, and control plane (Uptime Institute, 2025). Inference fleets serving production traffic push back toward full redundancy, because a dropped inference request is a dropped customer. The right answer depends on what the GPUs do, which is why redundancy topology should be a workload decision. We walk through the full N / N+1 / 2N logic in our modular data center power architecture guide.

For the gigawatt-scale Rubin era, NVIDIA and its OCP partners are pushing 800 VDC rack power distribution. If you're designing a facility today with a 10-year horizon, the DC-power question belongs in the base design review, not a future retrofit study.

Step 3: Cooling, where liquid is the baseline now

Air cooling has a physics ceiling around 40 kW per rack. Engineering analyses cite roughly 41 kW as the practical hard limit (Introl, 2025). H100 was the last generation you could broadly air-cool. Everything after it isn't a judgment call.

Direct-to-chip cold plates handle roughly 100–175 kW per rack and are now the default for Blackwell-class systems; the GB200 NVL72 puts about 115 kW of its load on liquid by design. That pulls a chain of requirements into the base building: coolant distribution units, facility water loops, leak detection, manifold-ready racks. Retrofitting any of these into a live air-cooled hall is expensive and slow. Designing them in from day one is neither. Our liquid cooling guide for data centers covers the direct-to-chip vs immersion decision in depth.

There's a regulatory bonus here too. Liquid cooling delivers higher-grade, reusable heat, which matters enormously once German heat-reuse mandates apply (more on that below) and makes waste heat recovery a revenue conversation instead of a disposal one.

Step 4: Build the network for east-west traffic

AI traffic is GPU-to-GPU. Collective operations, all-reduce, parameter synchronization: the bandwidth that matters flows sideways between racks, not north-south to users.

The fabric market has flipped accordingly. When Dell'Oro Group began tracking AI back-end networks in late 2023, InfiniBand held over 80%. By Q3 2025, Ethernet took more than two-thirds of AI back-end switch sales, with 800 Gbps ports already the majority and 1,600 Gbps expected by 2027 (Dell'Oro, 2025). For the building, this translates into dense fiber pathways, tight rack adjacency to preserve copper NVLink domains, and physical room for rail-optimized topologies.

Step 5: Check the floor

A GB200 NVL72 rack weighs roughly 1.4–1.5 metric tons on a footprint of about 0.8 m². That's a point load near 1,875 kg/m², when typical raised floors are rated around 1,000 kg/m². The fix is reinforced slab-on-grade, which is exactly how purpose-built AI facilities and factory-built modules are constructed. A legacy raised-floor hall is, structurally, the wrong building.

Water deserves a line here too. Evaporative-cooled campuses can draw millions of gallons a day, but closed-loop direct liquid cooling nearly eliminates ongoing consumption: a one-time fill, recirculated. In water-stressed permitting environments, that's the difference between approval and refusal.

What an AI data center costs

The numbers, with sources, because adjectives don't survive procurement:

Read that last row twice. The building is a minority of total capex, but it's the schedule gate for the majority. Every month of construction delay strands GPU capital worth roughly twice the building's cost. The cheapest thing you can buy in an AI data center project is time. Our modular data center cost breakdown shows where the facility money actually goes.

How long it takes, and why the answer is changing

A traditional hyperscale-class build runs 18–24 months of construction (IEEE Spectrum, 2025), but actual delivery now averages closer to four years once grid connection is included (IEA, 2025). Meanwhile the GPUs ship in weeks.

That mismatch is why the build model itself has flipped. McKinsey (2025) finds prefabricated and modular construction already makes up 40–60% of a typical data center project, and that standardized prefab builds cut delivery timelines by up to 50%. Factory-built modular data centers are a category-level speed advantage over traditional construction: the module is manufactured in parallel with site works and grid paperwork, then delivered with power, cooling, and fire suppression pre-integrated.

The hyperscale world has noticed. Microsoft's Fairwater facilities are standardized, replicated liquid-cooled designs stamped out as identical copies across sites. Crusoe, prime contractor on the 1.2 GW Stargate campus in Abilene, went further and opened a dedicated manufacturing plant in Colorado in March 2026 to mass-produce prefabricated modular AI factories. When the company building the world's largest AI campus decides the future is factory-built modules, the procurement signal is hard to miss.

When power is the bottleneck, the building must never be. That's the entire modular thesis, and the full modular data center guide covers how the delivery model works end to end.

The EU compliance layer

Building an AI data center in Europe means building into a reporting and efficiency regime that's already in force. Two instruments matter, and they are not the same thing.

The EU Energy Efficiency Directive (EED) is a transparency framework: under Delegated Regulation (EU) 2024/1364, any data center with ≥500 kW installed IT power must report energy use, PUE, WUE, heat-reuse factor, and renewables share annually to the European database. Reporting, not caps.

Germany's EnEfG is binding national law. Data centers commissioned on or after 1 July 2026 must achieve PUE ≤1.2 and reuse at least 10–15% of their waste heat, with the heat-reuse bar rising for later commissioning dates. Existing facilities face PUE ≤1.5 by July 2027 and ≤1.3 by July 2030.

A PUE of 1.2 is effectively unreachable with legacy air-cooled architecture in a German climate. Liquid-cooled, factory-engineered modules with heat-recovery interfaces arrive at that number by design rather than by retrofit. The full regulatory picture — country by country, deadline by deadline — is in our EU data center regulations guide.

Who's actually building all this: the neocloud context

The buyers driving AI data center demand increasingly aren't hyperscalers — they're neoclouds, GPU-first cloud providers built around AI workloads. CoreWeave guided to up to $35B of capex for 2026. Nebius is building greenfield capacity across Finland, the US, and Western Europe. Crusoe builds where the power is and manufactures its own modules.

The competitive logic across all of them, as one industry earnings roundup put it in 2026, has shifted "from GPU race to power wars." Advantage belongs to whoever has secured megawatts and can energize them fastest. How that business actually makes money (pricing, contracts, unit economics) is its own story, and we tell it in how the neocloud business works. And if you've already got the GPU allocation and need the path from loading dock to production, start with so you have GPUs, then what?

The build sequence, compressed

Power first: queue position, on-site generation, step-load tolerance. Then cooling: liquid as baseline, heat reuse as design input. Then network: Ethernet east-west fabric, fiber-dense pathways. Then structure: slab-on-grade, 1.5-ton racks. Then compliance: EED reporting from 500 kW, EnEfG thresholds if you're in Germany. And running through every step, one question: does this path get GPUs producing revenue this year or next year?

If the answer is next year, change the path. The GPUs depreciate either way.

FAQ

What is an AI data center?

An AI data center is a facility purpose-built for GPU compute at rack densities of 40 kW to 600+ kW, versus the 7.6 kW industry mean (Uptime Institute, 2025). It differs from traditional data centers in three fundamentals: liquid cooling as the design baseline, power availability as the primary site constraint, and east-west network fabric optimized for GPU-to-GPU traffic.

How much does it cost to build an AI data center?

AI-optimized facilities cost $20M+ per MW for the building, with GPU and IT fit-out adding up to ~$25M/MW on top (JLL, 2026). At a 1 GW campus scale, chips cost roughly twice the facility: about $20B versus $10B (Epoch AI, 2025). Shell-and-core averages $10.7M/MW globally as of 2025.

How long does it take to build an AI data center?

Traditional construction takes 18–24 months, but total delivery averages closer to four years once grid connection is included (IEA, 2025). US interconnection queues average 55 months (LBNL, 2025); EU hub markets run 7–10 years. Factory-built modular data centers compress the construction portion by 40–60%, with some AI modules delivered in 3–9 months.

What rack density does an AI data center need?

H100-class systems run 40+ kW per rack. GB200 NVL72 racks draw 120–132 kW. NVIDIA's Rubin Ultra systems arriving in H2 2027 are specified around 600 kW per rack. Any facility designed for AI compute should plan liquid cooling from roughly 40 kW per rack upward, because air cooling cannot go meaningfully beyond that.

Do AI data centers need liquid cooling?

Above roughly 40 kW per rack, yes. Direct-to-chip liquid cooling handles 100–175 kW per rack and is the default for Blackwell-generation hardware. The GB200 NVL72 carries about 115 kW of its thermal load on liquid by design. Liquid cooling also produces higher-grade reusable heat, which supports EU and German heat-reuse compliance.

What EU regulations apply to AI data centers?

The EU Energy Efficiency Directive requires annual reporting of PUE, WUE, energy use, heat reuse, and renewables share for any facility with ≥500 kW installed IT power (Delegated Regulation 2024/1364). Germany's EnEfG goes further: binding PUE ≤1.2 and ≥10–15% waste-heat reuse for data centers commissioned from 1 July 2026.

What is a neocloud?

A neocloud is a GPU-first cloud provider offering AI compute (GPU-as-a-Service) rather than general-purpose IT: companies like CoreWeave, Nebius, Lambda, and Crusoe. Neoclouds are among the fastest-growing buyers of AI data center capacity, and their main competitive constraint is secured power and speed to energization.

‍