Methodology and assumptions
How to read this model.
This atlas represents a simplified but physically coherent data-center architecture. It is an educational model with representative configurations; it is not a blueprint for any specific facility and does not describe any operator's actual design.
Who this is for
Curious non-specialists first: investors, policymakers, journalists, students and people working in tech who want to understand what an AI data center physically is. Engineers (electrical, mechanical, civil, network) are a second audience: they should find the system logic, vocabulary and orders of magnitude right, and can open Pro mode for full specs. It is not a design, sizing or permitting tool.
How accurate are the numbers?
| Kind of number | Example | Expect it to be |
|---|---|---|
| Vendor spec | Blackwell transistors, NVL72 GPU count | Exact as published |
| Reported | Queue size, deal sizes, capex ranges | Right within ≈10–30%; dated, will change |
| Typical | Pipe sizes, voltages, rack kW, floor loads | A real range used in practice; any given site may sit outside it |
| Estimate | Live electrons, water and signal rates, calculator | Physics-based, order-of-magnitude right (≈±25–50%) |
| Illustrative | 3D layout, campus scale, equipment counts | Logically correct, not to scale, not a real site |
Rule of thumb: trust the direction and the order of magnitude everywhere, trust exact figures only where they are tagged as a vendor spec, and check dated items (deals, queues, prices) against the source before quoting them.
What is real and what is illustrative
Every one of the 91 components carries specifications labeled representative (a reasonable value for the class), illustrative (a teaching value), typical (common practice) or varies. Ranges are preferred to single numbers. Vendor figures for the Blackwell generation are vendor-published or widely reported. Supplier lists are categories, not endorsements, and supply-chain regions are representative geographies, not supplier claims.
The 3D campus is not to scale. Equipment is enlarged so it reads at campus distance, and roughly one world unit stands for two meters. Placement follows real logic: power enters from the transmission side and steps down toward the halls; fiber enters from the opposite side through diverse routes; heat rejection sits downwind of the halls; water treatment sits beside the cooling plant. Flows are drawn as Manhattan routes between the components they connect, so moving equipment in the data moves its flows.
Fact-check status
Every headline figure carries a status: a vendor specification, a value reported by a named source, an estimate made by this model, or a historical record.
| Claim | Value | Status | Source |
|---|---|---|---|
| B200 transistors / process | 208B, TSMC 4NP, two dies | Vendor spec | NVIDIA Blackwell architecture materials |
| NVLink 5 bandwidth | 1.8 TB/s per GPU; 72-GPU domain in NVL72 | Vendor spec | NVIDIA |
| B200 / GB200 HBM | Up to 192 GB HBM3e, ≈8 TB/s | Vendor spec | NVIDIA |
| GB200 NVL72 rack power | ≈120–132 kW | Reported | Potter (IFP, 2024); industry reporting |
| Rack kW by GPU generation | Hopper ≈40–60 kW; Blackwell ≈130–250 kW | Reported | Crucible Capital Part 2 (2025) |
| Air-cooling practical limit | ≈20–30 kW/rack (≈50 kW with RDHx) | Reported | Potter (2024); Crucible Part 2 |
| Average PUE trend | ≈2.5 (2007) to ≈1.5; hyperscale ≈1.1 | Reported | Uptime Institute via Potter (2024) |
| Global data-center electricity | ≈1–1.3% of world electricity (2022) | Reported | IEA via Potter (2024) |
| Large transformer lead time | ≈50–100+ weeks | Reported | Crucible Capital Part I (2025) |
| Grid interconnect, >75 MW | ≈2–5 years | Reported | Crucible Capital Part I (2025) |
| Uptime tier availability | 99.671% to 99.995% | Reported | Uptime Institute (public tier descriptions) |
| Gas plant capex | ≈$720–1,680/kW | Reported | EIA via Crucible Part I |
| SMR capex (early units) | ≈$4,000–6,000/kW, estimated | Reported | Crucible Part I; industry estimates |
| Pod thermal behavior | Throttle ≈87 °C, shutdown ≈100 °C | Model estimate | This atlas (teaching parameters) |
| WUE by cooling method | 0.15–1.8 L/kWh | Model estimate | This atlas (illustrative central values) |
| All-in kW per GPU | ≈1.4–2.0 kW | Model estimate | Derived from reported rack power ÷ GPUs |
| Facility / IT capex per MW | ≈$12M / ≈$40M | Model estimate | Illustrative; IT figure as reported by Crucible Part 2 |
| 1961 microwave tower sabotage, Utah | Three relay towers dynamited | Historical | Hu, A Prehistory of the Cloud (2015) |
The scenario model
- Facility power = IT power × PUE. Annual energy uses an average IT draw of (0.3 + 0.7 × utilization) × capacity, because idle accelerators still draw power.
- Cooling overhead is (facility − IT) with a cooling share that depends on the cooling method; UPS losses ≈3.5% of IT; transformer losses ≈1% of facility.
- Site water uses an illustrative WUE per cooling method: Air 0.3 L/kWh, Chilled water 1.8 L/kWh, Direct-to-chip 1 L/kWh, Rear door 1.3 L/kWh, Dry / economizer 0.15 L/kWh, Immersion 0.15 L/kWh. Real values vary strongly with climate and season, and upstream water used to generate electricity is excluded.
- Operational carbon = facility MWh × grid intensity (location-based, scope 2 framing). Embodied carbon uses illustrative placeholders per MW; it is not an audited inventory.
- Capex uses ≈$12M per MW for an AI-ready facility and ≈$40M per MW for current GPU racks (as reported). Cost per GPU-hour annualizes facility capex over 20 years and IT over 5 years.
The live pod model
The operations view runs a small energy balance for one rack-scale system and multiplies by 16 racks. Heat into the liquid loop is 90% of rack power; coolant leaves at supply + Q ÷ (ṁ · cp), with 100 L/min at 30 °C supply (45 °C in warm-water mode). GPUs sit a thermal resistance above the coolant, throttle near 87 °C and shut down near 100 °C. Training is modeled as synchronous compute-then-communicate swings; inference as bursty demand. These are teaching parameters, not vendor data.
Redundancy tiers
| Tier I | Tier II | Tier III | Tier IV | |
|---|---|---|---|---|
| Availability target | 99.671% | 99.741% | 99.982% | 99.995% |
| Power paths | 1 | 1 | 1 active + 1 alternate | 2 active |
| Component redundancy | N | N+1 | N+1 | 2N (fault tolerant) |
| Concurrently maintainable | No | No | Yes | Yes |
| Generator fuel on site (common practice) | ≈8 h | ≈24 h | ≈72 h | ≈96 h |
Summarized from public descriptions of the Uptime Institute tiers and common design guides. Certification requires Uptime's own review.
Language rules
The atlas avoids claiming that every AI cluster uses one network technology, that all hyperscalers or neoclouds share a cooling topology, or that illustrative megawatt, density, water or cost figures are audited data. Cloud, hyperscaler and neocloud are useful categories, not fixed or mutually exclusive ones.
Accessibility and performance
Everything in the 3D view is also available as semantic HTML: the inspector, the text map, comparison tables and glossary. Flows pair every color with a distinct particle shape and line pattern. Reduced-motion and high-contrast modes are in the display settings. Rendering uses instancing for racks, tiles and particles, three quality tiers, adaptive resolution, and a text fallback when WebGL is unavailable.
Sources
Public vendor documentation for accelerator specifications; Uptime Institute and IEA public material for tiers and global energy context; and the reading list on the Field notes page (Crucible Capital's Datacenter Series, Brian Potter's essay for the Institute for Progress, and Tung-Hui Hu's A Prehistory of the Cloud). Source placeholders on each component are structured for future citations.