Goldman: Power, Not Chips, Now Sets AI Data Centre Pace

Goldman Sachs says the constraint on AI expansion has moved from chips to electricity. In a note published 12 October 2026, the bank said grid capacity for 2026 and 2027 is essentially booked out, pushing leading hyperscalers to lock in power contracts for 2028 and beyond. Revenue generated per unit of power has climbed by around two-thirds over the past year, and Meta is earning two to three times return on investment per megawatt in its core business. Compute supply continues to trail demand, the bank said.

Context — Why Does Power Now Matter More Than Chips for AI?

The report's own comparison is the one that matters: capital to buy GPUs was previously treated as the gating factor, and Goldman now says it is not sufficient on its own. Access to high-voltage energy infrastructure decides who can expand. That is a change in the identity of the bottleneck rather than in the size of the buildout.

The catalyst chain runs through demand first. Goldman said new infrastructure requirements and resilient enterprise spending are widening the gap between compute supply and demand, based on its discussions with technology company management teams.

Model training was expected to slow, which would ordinarily ease pressure on compute. Goldman said that relief is being offset by rising investment in AI safety and pacing monitoring, work that itself consumes computing power. The net effect keeps overall compute demand elevated even as one workload class decelerates.

Grid interconnection is the slowest link in that chain. Because capacity for the next two calendar years is spoken for, hyperscalers cannot simply buy their way into 2026 or 2027 supply. They are contracting further out instead, which converts a near-term procurement problem into a multi-year power strategy.

Capital expenditure is rising at the same time, on component inflation and on the need for more chips. Two cost lines are therefore moving together: the hardware bill and the energy bill. Goldman's framing puts energy in the decisive seat.

Data — What the Numbers Show

The report's headline figures are ratios and returns rather than dollar totals. Revenue per unit of power is up around two-thirds versus earlier price points over the past year. Meta generates two to three times return on investment per megawatt in its core business. Grid capacity for 2026 and 2027 is essentially fully booked, with contracts now extending to 2028 and beyond.

Metric	Report's figure
Revenue per unit of power	up around two-thirds over the past year
Meta return on investment	2x to 3x per megawatt, core business
Grid capacity, 2026-2027	essentially fully booked
Power contracting horizon	2028 and beyond

Before and after, the same operator tells the story. A hyperscaler holding secured grid capacity can convert capital into live compute. A hyperscaler without it holds the same capital and the same chip orders but cannot energise them on the same timetable.

Goldman did not disclose dollar capex totals, per-megawatt pricing, or the identity of the counterparties behind the 2028 power contracts. It also gave no megawatt figures for any named company. Those gaps matter because the ratio of revenue to power is the metric the report leans on most heavily.

On software, the report's comparison is structural rather than numerical. Data and security platforms sit on the beneficiary side, while application vendors face a change in how their products are consumed and priced.

Analysis — Which Sectors and Tickers Carry the Exposure

Goldman's own list of highlighted companies is short. Meta is credited with strong per-megawatt returns in its core business while it explores new sources of incremental revenue. Google is described as relying on a full-stack advantage combining custom silicon, cloud, security and a large distribution base, with TPU chip sales one part of a broader portfolio rather than the whole thesis.

The second-order read runs through utilities and grid operators, independent power producers, and the electrical equipment supply chain, all of which sit upstream of any energised data centre. The report does not name those companies, so the exposure is directional rather than a list.

In software, the mechanism Goldman describes is application software going headless — increasingly built for AI agents rather than human users. The bank said this lets vendors push more powerful AI features into existing platforms and supports industry growth for now. Its caution is that the shift could eventually change how software companies price and monetise products. Seat-based pricing is the model most directly exposed, because agents do not occupy seats.

The counter-argument is worth stating plainly. If training genuinely slows and safety and monitoring work proves less compute-hungry than Goldman assumes, the demand gap could narrow faster than the note implies, easing the power constraint and shifting advantage back toward whoever holds the cheapest capital. The report does not quantify that scenario.

Positioning follows the constraint. Capital is being directed toward operators with secured grid capacity and toward foundational data and security layers, and away from application vendors whose pricing models depend on human user counts. Goldman said chief information officers are steering budgets toward foundational data layers and cybersecurity tools to protect data integrity and contain the risks of unsanctioned "shadow AI" use inside organisations.

Outlook — What to Watch Next

The first thing to watch is the 2028 power contracting cycle. Goldman said leading hyperscalers are already moving aggressively there, and the pace of those announcements is the clearest available signal on who has secured the next tranche of capacity. The report gives no dates for that cycle.

The second is the training-versus-safety balance. Goldman expects model training to slow but sees safety and pacing monitoring investment rising. If safety spending fails to offset the training deceleration, the compute supply-demand gap narrows and the power premium compresses.

The third is software pricing disclosure. Application vendors going headless have not yet restated how they charge. Any change to per-seat models would confirm the mechanism Goldman flagged as a future risk rather than a current one.

Goldman concluded that AI monetisation offers substantial optionality, while core platform monetisation remains among the higher-confidence and higher-return uses of compute. The practical implication is that established platforms already earning strong returns on their power stay the more dependable beneficiaries while capacity remains scarce.

Frequently Asked Questions

What does Goldman Sachs mean by power being the AI bottleneck?

It means capital alone no longer secures compute. Goldman said grid capacity for 2026 and 2027 is essentially booked out, so a company can hold the funding and the chip orders but still wait for an energised connection. High-voltage energy infrastructure now decides who can expand, which pushes hyperscalers to sign power contracts for 2028 and beyond.

Why is application software going headless a risk for vendors?

Headless means software built for AI agents rather than human users. Goldman said this lets vendors embed stronger AI features into existing platforms and supports growth for now. The risk is pricing: models built around human user counts do not map cleanly onto agents, so monetisation may need to change. The bank did not say when or how.

What does the report say about Google's and Meta's AI returns?

Goldman said Meta generates two to three times return on investment per megawatt in its core business while pursuing incremental revenue. Google is described as using a full-stack advantage spanning custom silicon, cloud, security and a large distribution base, with TPU chip sales forming one part of a wider portfolio strategy rather than the central case.

Bottom Line

Goldman says secured grid capacity, not chip budgets, now determines which operators can energise AI compute through 2027.

Disclaimer: This article is for informational purposes only and does not constitute investment advice. CFD trading carries high risk of capital loss.