Powering It: Time-to-Power as the Binding Constraint

Chapter 62 ended at the rack. The next question is not how many accelerators have been ordered, but when a particular site can operate them. For some large AI campuses, the scarce quantity is not generation capacity but time-to-power: deliverable megawatts at the required location, reliability level, and date. That claim is not universal. The binding item can still be silicon, financing, land, permits, network equipment, or demand. It must therefore be stated for a site, product, and date.

Three phrases often compress this transition too far. A campus self-generates rather than waits for the grid only when an on-site project has passed its commercial, regulatory, and commissioning gates. A dense rack may make liquid cooling as the default a sensible product decision without making liquid cooling mandatory for every data center. And a large computational load is something a grid operator must actively manage only through a defined model, telemetry, protection, and operating agreement. This chapter turns each slogan into a quantity that can be checked.

Power is a dated delivery contract

A statement such as “this campus has one gigawatt” is incomplete. It might refer to a utility request, a generator nameplate, a power-purchase agreement, the point-of-interconnection limit, the facility design, or measured IT load. Those numbers are not interchangeable. The useful question is: how many net, firm megawatts can reach accepted IT equipment, under the stated contingency, by a stated date?

The schedule is also not a row of independent queues whose largest duration can simply be selected. Site control may precede an interconnection study; a substation design may wait on that study; cooling construction may proceed in parallel; commissioning waits for several branches. The completion time is a critical-path calculation:

$$T_{\mathrm{ready}} = \max_{p \in \mathcal{P}} \sum_{a \in p} D_a .$$

where:

• $T_{\mathrm{ready}}$ is elapsed time from the declared project start to acceptance of usable IT capacity;
• $\mathcal{P}$ is the set of dependency paths from that start to acceptance;
• $p$ is one dependency path;
• $a$ is one activity on that path; and
• $D_a$ is that activity's duration, including both waiting and execution.

The equation is deterministic. A real plan should also report scenario or simulation results such as P50 and P90 readiness dates. Percentiles cannot be added mechanically: the P90 of every activity does not generally produce the P90 of the project, especially when durations are correlated.

tasks = {
    "interconnection": {"months": 18, "after": []},
    "substation": {"months": 14, "after": ["interconnection"]},
    "onsite generation": {"months": 12, "after": []},
    "cooling plant": {"months": 10, "after": []},
    "commissioning": {
        "months": 3,
        "after": ["substation", "onsite generation", "cooling plant"],
    },
}

finish = {}
paths = {}
pending = set(tasks)
while pending:
    ready = sorted(
        name for name in pending
        if all(parent in finish for parent in tasks[name]["after"])
    )
    if not ready:
        raise ValueError("dependency graph contains a cycle")
    for name in ready:
        parents = tasks[name]["after"]
        parent = max(parents, key=lambda item: finish[item]) if parents else None
        start = finish[parent] if parent else 0
        finish[name] = start + tasks[name]["months"]
        paths[name] = (paths[parent] if parent else []) + [name]
        pending.remove(name)

critical = paths["commissioning"]
print("Critical path: " + " -> ".join(critical))
print(f"Ready after: {finish['commissioning']} months")

The numbers above are illustrative, not a forecast. Change a duration or a dependency and the result changes visibly. That is more useful than repeating that “megawatts take years.”

From a source to usable IT power

Calendar readiness is only half the problem. A facility also needs a common power boundary. One compact upper bound is

$$P_{\mathrm{IT,max}}(t) = \min\!\left( \frac{P_{\mathrm{source}}(t)}{\operatorname{PUE}(t)}, \frac{P_{\mathrm{sub}}(t)}{\operatorname{PUE}(t)}, \frac{P_{\mathrm{dist}}(t)}{\operatorname{PUE}(t)}, Q_{\mathrm{cool}}(t) \right), \qquad N_{\mathrm{rack}} \le \left\lfloor \frac{P_{\mathrm{IT,max}}(t)}{P_{\mathrm{rack}}} \right\rfloor .$$

where:

• $P_{\mathrm{IT,max}}(t)$ is the maximum supported IT real power at time $t$;
• $P_{\mathrm{source}}(t)$ is net firm real power available from the contractually usable grid and on-site sources, without double counting;
• $P_{\mathrm{sub}}(t)$ and $P_{\mathrm{dist}}(t)$ are the accepted substation and downstream distribution limits;
• $\operatorname{PUE}(t)$ is power usage effectiveness (PUE), the ratio of total facility power to IT equipment power, measured over the same boundary and interval;
• $Q_{\mathrm{cool}}(t)$ is cooling capacity expressed as the IT-power load it can support under the stated supply temperatures and failure case;
• $N_{\mathrm{rack}}$ is the supported whole-rack count; and
• $P_{\mathrm{rack}}$ is the chosen design power per rack.

The source, substation, and distribution terms must use net firm real power, the same contingency, and the same averaging interval. Cooling capacity needs the same ambient and redundancy assumptions. The rack quotient is a planning bound, not a prediction of consumption: rack nameplate, measured coincident draw, diversity allowance, and design redundancy must be reported separately.

PUE measures facility-support overhead; it does not measure useful compute. Uptime Institute's 2025 survey reported a respondent-weighted average annual PUE of 1.54 and little movement for six years, but the sample contains many facility types and climates (Donnellan et al. 2025). That result does not prove that efficiency gains are exhausted, nor does it make a fleet-wide company average directly comparable with an individual site.

Interconnection queues are not delivery forecasts

The word queue hides two different populations. Lawrence Berkeley National Laboratory's Queued Up dataset follows proposed generation and storage in a generator interconnection queue; it is not a queue of data centers seeking load service (Rand et al. 2025). It can explain why new supply may arrive slowly. It cannot, by itself, establish the wait for a particular large load.

Large-load requests have their own evidence problem. A developer may request service at several sites before choosing one. A request may lack site control, equipment orders, financing, or a completed study. FERC has explicitly warned that speculative and duplicate requests can distort forecasts and has asked grid operators to use observable readiness milestones (Rosner 2025). A queue total should therefore be accompanied by counts at each maturity gate; on its own, it is not a delivery forecast.

A useful status ladder is therefore: large-load request; validated site and commercial readiness; study; signed agreement; required upgrades under construction; approval to energize; observed energized load. Record both MW and date at every rung. The same discipline applies to equipment. DOE reports that large power transformers are customized, hard to transport, and recently quoted at lead times up to or beyond 36 months, with maxima as high as 60 months (U.S. Department of Energy 2024). That supports a risk range for the specified transformer; it does not justify assigning one universal lead time to every voltage class and design.

Behind the meter changes the boundary

On-site generation can shorten one branch of the critical path. It cannot erase fuel delivery, emissions permits, water, noise, equipment lead time, protection studies, or obligations for backup service. Nor does “behind the meter” answer who supplies the load when an on-site unit trips. The electrical one-line and tariff define the boundary; the slogan does not.

Report every project using the same evidence ladder: announcement, contract, permit or license, financing, construction, commissioning, and operation. A signed power-purchase agreement is not operating capacity. A closed loan is not regulatory approval. A reactor restart is not a routine maintenance date.

The Christopher M. Crane Clean Energy Center shows why this distinction is necessary. DOE closed a $1 billion loan in November 2025 to help finance an 835 MW restart, subject to NRC licensing approvals (U.S. Department of Energy 2025). As of August 7, 2026, the NRC still listed the plant's license status as SAFSTOR and described the restart as a licensing, restoration, inspection, and oversight process (U.S. Nuclear Regulatory Commission 2026). The project is real and advanced; it is not yet an operating source. Apply the same test to small modular reactor (SMR) agreements and on-site gas proposals: record the achieved gate, not the largest announced GW number.

This is also why “self-generates” should not be read as “independent of the power system.” A campus may rely on the grid for start-up, maintenance, contingency reserve, export, or emergency operation. Its firm capacity must survive the declared outage case, and its cost must include the service and network capacity it retains.

Cooling follows a product and facility contract

NVIDIA specifies that a full GB300 NVL72 rack requires up to 142 kW of input power and uses liquid cooling (NVIDIA 2026). This is a product-specific maximum design requirement. It is not measured average draw, annual energy, or total facility load.

At this density, the vendor's architecture makes direct liquid cooling part of the rack design. The broader engineering choice still depends on component heat flux, rack density, coolant and air supply temperatures, the fraction of heat captured by liquid, local climate, serviceability, and the existing building. ASHRAE's retrofit guidance recommends hybrid cooling for dense AI systems: liquid removes the concentrated component heat while air removes residual heat (ASHRAE 2026). Cold plates, rear-door heat exchangers, and immersion are alternatives or complements, not a universal maturity ladder.

The acceptance plan must extend beyond “liquid-ready.” It should specify the coolant distribution unit, supply and return temperatures, flow and pressure, water chemistry, material compatibility, filtration, leak detection, residual air load, controls, and behavior after a pump or power failure. A cooling design passes only when it removes the measured heat while maintaining the equipment limits and redundancy target.

The grid needs modeled load behavior

Large computational loads are unusual not merely because they are large, but because many power-electronic devices and synchronized workloads can respond together. On May 4, 2026, NERC issued a Level 3 Essential Action Alert after observing customer-initiated large load reductions and significant oscillations occurring in seconds, leaving little or no time for real-time operator response (North American Electric Reliability Corporation 2026). NERC's statement covers computational loads that include AI, cryptocurrency, and other applications. It does not establish that every event was an AI training job, or that every AI campus behaves identically.

Integration therefore needs a testable operating contract. The grid model should cover steady state, transient response, harmonics, protection, fault ride-through, maximum pickup and shed, ramp rate, and rebound after curtailment. Commissioning should validate those models. Operations need time-synchronized telemetry, event records, communication paths, and a clear operating envelope. Firmware, workload controls, UPS behavior, or protection changes must trigger review when they change the modeled response.

A data center can sometimes offer flexibility by shifting work, reducing power, or using storage and on-site supply. That service is valuable only when the curtailment quantity, notice, duration, recovery, and availability are contracted and tested. Workload deadlines, checkpoint cost, safe ramp limits, and the rebound peak all reduce the theoretical flexibility. An LBNL workshop found genuine opportunities alongside high costs, regulatory fragmentation, and a need for standard controls and incentives (Kirchstetter et al. 2025).

The rules are still moving. FERC's December 2025 PJM order proposed Firm Contract Demand and Non-Firm Contract Demand services for co-located loads (Federal Energy Regulatory Commission 2025). On June 18, 2026, FERC also directed all six regional grid operators under its jurisdiction to justify or reform large-load interconnection tariffs (Federal Energy Regulatory Commission 2026). These were orders and ongoing compliance proceedings as of August 7, 2026, not proof that one national interconnection design had already taken effect.

Energy per result needs a declared boundary

Aggregate energy estimates establish scale, not site readiness. The IEA estimates that data centers used about 415 TWh in 2024, around 1.5 percent of global electricity consumption; its base case reaches about 945 TWh in 2030 (International Energy Agency 2025). LBNL estimates 176 TWh for US data centers in 2023 and models 325--580 TWh in 2028 (Shehabi et al. 2024). The latter is a conditional scenario envelope, not a confidence interval. TWh per year, average GW, connected MW, and generator nameplate answer different questions.

For one workload and reporting window, define

$$e_{\mathrm{result}} = \frac{\displaystyle \int_{t_0}^{t_1} P_{\mathrm{fac}}(t)\,dt} {N_{\mathrm{accept}}} .$$

where:

• $e_{\mathrm{result}}$ is facility energy per accepted result;
• $t_0$ and $t_1$ delimit the declared measurement window;
• $P_{\mathrm{fac}}(t)$ is measured facility power attributable under the published allocation rule; and
• $N_{\mathrm{accept}}$ is the number of outputs that meet the declared quality threshold and service-level objective in that window.

“Per query” is not a sufficient method. A reproducible report identifies the model and runtime, hardware, request mix, input and output sizes, cache state, batching, retries, throughput, latency distribution, and quality threshold. It states whether the meter covers accelerator, server, cluster IT, or the whole facility; how idle capacity and shared services are allocated; and whether training amortization, embodied energy, external networks, and user devices are excluded. It reports total energy and the distribution, not only a median.

Google's production study, for example, estimated 0.24 Wh for the median Gemini Apps text prompt in May 2025 under its allocation method. Its boundary included active accelerators, host CPU and memory, allocated idle capacity, and facility overhead, while excluding training, external networking, user devices, and data storage (Elsworth et al. 2025). This is useful evidence for that product, fleet, workload distribution, and month. It is not a constant for “AI,” and its median does not determine fleet total energy.

How to verify time-to-power

A defensible review can fit on one evidence register. Freeze the electrical one-line and bill of materials revision. State MW at the point of interconnection, facility MW, and IT MW; separate firm and conditional capacity. For every dependency, record the planned and achieved date, current status, evidence owner, source document, acceptance criterion, and P50 and P90 forecast. Link permits, studies, agreements, purchase orders, factory tests, site tests, and operating telemetry to the item they support.

Then test the failure boundary. Model the required contingency, loss of grid or on-site generation, transformer or coolant-loop failure, load step, curtailment, and rebound. Give every failed milestone a recovery scenario, decision date, and accountable owner. This turns “one gigawatt campus” from a press-release number into an auditable delivery claim.

Constraint arrow

Power does not replace the constraints above it. It joins them. A rack can be delivered before its substation, a generator can be contracted before its permit, and a campus can be energized before its load behavior is accepted. The usable cluster is bounded by the intersection of those schedules and capacities. That intersection determines how much of the serving stack can run and at what reliability and cost (cross-ref Chapter 76).

The durable skill is to turn a power claim into a dated evidence chain: source to substation, substation to rack, rack to cooling, model to operating record. The next chapter, Chapter 69, starts after those gates pass. It asks what reliability means when the energized machine contains enough components that something is almost always failing.

Further reading