Capacity & Density Envelope Study — Reference Colocation Campus, hall DH-02

1. The answer

The binding constraint as found is rack feed / tap-off rating (electrical). platform needs 205 [174–236] A (E0) per rack; installed tap-offs rated 63 A (E5).

Selection objective: Maximum deployable compute. On an AI site the revenue attached to a rack dominates the capital cost of enabling it, so capacity is ranked first and capex is the tie-break. Section 9 shows every architecture against all three objectives so the ranking can be overruled with the evidence in view.

2. Headroom ladder

Each row is the constraint that binds at that point, what it costs to move it, and how long that takes. Reliefs are assumed to run in parallel, so the schedule is governed by the longest lead time rather than the sum.

Headroom ladder for the recommended architecture. A step's price follows what it adds (units, MVA, kW, racks) where the library is priced per unit; a fixed lump is marked as not scaled. A component with no price is named, never counted as zero.

Steps marked NOT TAKEN are shown because they are the next thing that binds, not because they are recommended. Their cost is excluded from the totals below.

Rungs 1, 2, 4, 18 are charged but add no racks on their own. Each relieves one of several constraints that limit the hall to the same figure at that point; without it a later rung cannot deliver its gain. They are kept, and costed, because dropping any one of them lowers the final answer — that is checked, not assumed. The EUR-per-kW column is left blank for them for the same reason: no racks are unlocked at that step.

3. Time to power

Capacity without a date is not a decision. Each relief on the ladder carries a lead time; reliefs are assumed to run in parallel, so the racks unlocked by the first N steps become available at the longest lead time among those N. One item sets the date, and it is almost never the construction work.

Energisation curve for the recommended architecture.

Time to power across the architectures compared.

4. Site and baseline

Baseline as supplied. E5 means supplied by the client — a declared rating, contract term or design value as given to us, not measured or surveyed by us. E3 marks a figure that is estimate-grade by nature; E0 marks a library default.

5. Electrical capacity analysis

A gate either passes (it does not limit the hall) or fails (it permits no racks until relieved). Rack counts carry the weakest evidence class among the constraints evaluated.

• No new or increased connection capacity is available from the DSO in this metro before the 2030s. The existing firm capacity is therefore the asset, and the whole question is how much compute can be produced from it.
• GPU racks present step loads; confirm the UPS block accepts the step, not just the steady load.

6. Thermal and cooling analysis

A gate either passes (it does not limit the hall) or fails (it permits no racks until relieved). Rack counts carry the weakest evidence class among the constraints evaluated.

• No capex relieves this constraint. If it binds, the architecture is wrong for the platform and a different scenario must be selected.
• TCS water quality per OCP: conductivity <1500 uS/cm, pH 8.0-10.5, 50 um inline filtration, sidestream <5 um at 10% of flow.
• Pressure drop budget and pipe DN are not modelled at screening level; a hydraulic model is required before design.
• A rack whose residual air load cannot be removed cannot be placed at all, however much liquid cooling is installed.
• The ASHRAE W45 rating on the platform is an equipment survival rating, not an operating point; operators specify 27-30 degC for 500-700 W parts.

7. Physical constraints

A gate either passes (it does not limit the hall) or fails (it permits no racks until relieved). Rack counts carry the weakest evidence class among the constraints evaluated.

• Raised floor: 120 kW+ racks normally require load-spreading plates or removal of the raised floor and deployment on slab.
• TIA-942 Rated-3 design value is 12 kPa (~1223 kg/m2).

8. Power and thermal interaction

Power and thermal are not independent budgets. Every kilowatt the cooling architecture consumes is a kilowatt of grid capacity that cannot be sold as compute, so the choice of cooling architecture directly changes the electrical envelope. The table shows the same hall under each architecture.

A better cooling architecture buys compute out of the same grid connection.

9. Economics

Cost basis: Class 5 (order of magnitude): -50% / +100%. Excludes IT hardware, migration and lost tenancy revenue. A total marked with ≥ is a lower bound: a cost it needs has no price. A dash means no rack can be deployed in that architecture and no relief exists, so the figure is undefined rather than zero.

The same scenarios ranked under each objective. The recommendation above uses the stated objective; where another objective selects a different architecture, that is a judgement for the client, not for the model.

The ladder's cumulative figure covers the relief works only; the headline adds the per-rack conversion inside the hall and the CDUs the model assumed so the liquid scenarios could be solved (the hall has no CDU schedule on file; those units must be bought).

Operating cost is stated separately from capital. It covers energy and, where on-site generation is installed, that generation's declared operating cost; it excludes the non-energy operating cost of the cooling plant.

What each price in this study rests on. The accuracy class above follows the weakest line, not the average.

10. Sensitivity

One-at-a-time sensitivity on the recommended architecture, evaluated against the relieved case rather than the hall as found. Inputs are ranked by how far each moves the answer; the ones at the top are the ones worth paying to measure.

11. Risk register

12. Recommendation and decision gates

Proceed with Full DLC plus behind-the-meter supply. The same hall with on-site firm supply where the grid connection binds. The only scenario whose capacity is set by what can be built rather than by the queue. On the modelled inputs this unlocks 62 racks of NVIDIA GB300 NVL72 against 0 as found.

• Gate 1 — obtain the DSO position in writing on firm capacity and any flexibility product before committing capital. In most target metros this is the shortest path to a definitive answer and it costs nothing but time.
• Gate 2 — commission a structural assessment of the hall floor and a hydraulic model of the secondary loop. Both are cheap relative to the decision and both can invalidate it.
• Gate 3 — place long-lead orders (transformer, busway, CDU) only after Gates 1 and 2 clear; lead time, not construction, governs the programme.
• Gate 4 — instrument the hall and reconcile this model against twelve months of measured data before the second phase. That reconciliation is what moves these numbers from estimated to field-validated.

13. Scope, basis and limitations

This study is an engineering opinion supported by a documented model. It is not a design package and confers no design liability. The following are outside scope:

• Stamped design drawings and any deliverable requiring a professional engineering signature.
• CFD of the hall (available as a priced add-on or a later phase).
• Equipment selection, bill of materials, pricing or procurement.
• Structural sign-off — the floor-loading check here is a screening calculation only.
• Electrical protection and arc-flash studies.
• Commissioning, migration planning and tenancy/lease strategy.

Every quantity in this report carries an evidence class; Appendix A lists the assumptions and sources relied upon and the provenance digest of each headline quantity. No number here is a measurement of a physical asset unless it is labelled E5 or above — and E5 means supplied by the client (a declared rating, contract term or design value, or the client's own data as given), not measured or surveyed by us.

14. Model accuracy against instrumented sites

Two different things get called accuracy. The ranges shown throughout this report are input uncertainty propagated through the model — they say how much the answer moves when an input is uncertain. This section is the other one: how far this engine's outputs have been reconciled against data from instrumented sites, which is the only thing that can tell you whether the model itself is right.

Bias is the median of observed / predicted, so a positive figure means the model has run conservative. An output with no observations is reported as uncalibrated rather than omitted.

Reconciliation is how a figure in this report moves from estimated to field-validated. It requires site data we do not have and cannot generate: interval metering at the relevant point, a BMS trend, or a commissioning record. Where this study recommends instrumentation, that is what it is for — the client gets a better answer next time, and the model gets an observation it can be held to.

Appendix A — Assumptions, provenance and validation status

Every number in this report carries an evidence class. The class of a computed number is never higher than the weakest of its inputs, so a model applied to measured data yields a modelled number and never a measurement.

E5 means supplied by the client — a declared rating, contract term or design value, or the client's own data as given. It is not a measurement we made, and it has not been verified by a site survey.

A.1 Assumptions relied upon

• Excludes IT hardware, migration and lost tenancy revenue. Public retrofit benchmarks span 2-12 MEUR per MW; this default is neither a quotation nor a benchmark.
• Library default, not a price. Replace with a quotation before this figure supports a capital decision.
• Class 5 (order of magnitude): -50% / +100%
• Placeholder re-expressed per MVA from the earlier 1.2 MEUR lump at the reference hall's 9 MVA added (3 x 5 MVA, +60%). Not a market price. Excludes civils and outage management.
• Placeholder per nameplate kW of UPS added.
• Placeholder re-expressed per run from the earlier 450 kEUR lump for the reference hall's 8 runs. Not a market price.
• Placeholder covering the unit supplied and mechanically installed only. Hydraulic integration, controls and commissioning are separate components.
• Tap-off units are a reported shortage item
• Frankfurt: no new large connections before 2030; Amsterdam queue ~10 years
• Design value from the building record, not a survey
• 12 months of half-hourly metered data
• Legacy 400 A busway; modern AI halls use 800-1000 A
• Vendor claims of 1.05-1.15 are unproven; 1.15-1.25 used for design per docs/02 §7
• Rating quoted at 5 K approach; site offers 24 K. Linear derate is a screening approximation - confirm against the manufacturer's performance curve.
• Cold plates on CPU/GPU/HBM/NVSwitch; residual air load must still be removed
• Assumes positions can be released from existing tenancy
• Legacy low-temperature plant retained for residual air load, so free-cooling benefit is largely forgone
• Placeholder re-expressed per kW of cooling added from the earlier 900 kEUR lump at the reference hall's 3,400 kW (+50% of 6,800 kW). Not a market price.
• Placeholder. Replace with the site's actual contracted tariff.

A.2 Sources

• Lenovo Press LP2357, GB300 NVL72, 2026 [V]
• NVIDIA NVL72 AI Factory Reference Architecture, 2026 [V]
• Vertiv, quantifying PUE when introducing liquid cooling (modelled), 2023 [V]
• Uptime Institute Global Survey, 2025

A.3 Provenance of headline quantities