The bottleneck after the chip

Equipment, power and usable service are separate milestones. To assess a delivery promise, identify the handoff controlling the schedule and the party bearing the cost if it slips.

The question behind the spending

A capital commitment is easy to place on a timeline. The date at which it becomes dependable computing capacity is harder to establish. A server can be ordered before a site is energized; electricity can be available before the equipment is commissioned. Our judgment is that these transitions deserve as much attention as the size of the spending announcement.

The IEA’s 2026 outlook places global data-centre electricity consumption at 485 TWh in 2025 and projects roughly 950 TWh in 2030. These are an estimate and a forecast, respectively. The scale is material; the figures do not establish when a particular customer will be able to use a particular building. International Energy Agency ↗

A named project: follow the Crane restart through three distinct records →

Follow the handoffs

Berkeley Lab’s work on large-load connections separates interventions across forecasting, interconnection, planning and procurement, operations, and cost allocation. That makes the delivery question more precise: which handoff is controlling the schedule, and who holds the evidence that it has cleared? Lawrence Berkeley National Laboratory ↗

An investor presentation, a utility agreement and a commissioning certificate answer different questions. Treating any one of them as a proxy for all the others creates a false sense of completion. An evidence record should name the stage, the accountable party, the date, the source and the remaining condition. When those records disagree, the disagreement is itself worth reporting.

Revenue is evidence of a different event

NVIDIA’s latest quarterly release belongs in the commercial record. Its revenue is evidence about the supplier’s business; it is not a measure of the utilization of every customer installation. Reading the supplier result alongside power-system evidence gives a more useful picture than treating either as the whole story. NVIDIA ↗

This distinction also guards against an easy overcorrection. A connection problem at one site does not establish that global AI demand is weak. Neither strong supplier sales nor a delayed project should be allowed to decide a question at a scale it cannot observe.

What would change the judgment

A series of documented, on-schedule energizations followed by usable capacity would weaken the concern for the projects observed. Repeated delays after equipment delivery would strengthen it. Both tests require project-level evidence, a stated comparison period and a denominator that includes successes as well as failures.

The available evidence does not provide a representative project panel or a historical success rate. A completion claim needs commitments matched to dated milestones and operating capacity; a large pipeline alone cannot establish it.

Three clocks, one operating date

A compute project has at least three clocks. Procurement measures when equipment can be bought and delivered. Infrastructure measures when the site can receive dependable power and remove heat. Commercial use measures when a customer has a workload ready to run. A project can be ahead of schedule on one clock and late on the one that determines its first useful output.

Consider an illustrative project with servers on site but an unresolved connection date. Buying more servers does not clear the connection. It may increase the amount of capital waiting for it. Now reverse the situation: an energized building without a deployable system leaves the physical asset available while the commercial service remains unavailable. These are different operating problems, even if both appear under the same broad label of AI investment.

This distinction changes where to look for negotiating power. A supplier whose equipment can be substituted may face a different commercial position from the party controlling a necessary, time-sensitive handoff. That is a conditional mechanism, not evidence that every grid asset or equipment supplier earns a premium. Spare capacity, alternative sites, contract terms and the cost of waiting can reverse the result.

Who pays while the clocks disagree?

The delay does not disappear because it is absent from an equipment supplier’s revenue. It may reach the site developer through carrying costs, the customer through an unavailable service, or a counterparty through contractual remedies. The allocation depends on agreements that aggregate spending figures do not disclose.

A useful comparison therefore starts with two projects that promise a similar service, then asks what happens if the controlling milestone slips. Which costs continue? Which payments depend on completion? Can equipment move elsewhere? Does an alternative connection change the economics? Without those answers, a ranking by announced capital expenditure can confuse scale with readiness.

What would weaken the bottleneck argument?

If operating records showed that new capacity consistently entered service on its expected timetable, and that delivery dependencies were routinely resolved before equipment arrived, this argument would become less useful for that population. Lower demand could also leave more capacity idle for commercial reasons rather than physical ones. A delivery explanation should not absorb every possible disappointment.

For any disclosed project, the useful test is whether commitment dates, connection milestones and operating capacity can be matched. Until that chain is visible, the aggregate growth outlook establishes the size of the question. It does not settle who will deliver, who will wait or who will bear the cost.