Goldman Sachs is directing attention to the physical infrastructure underpinning the artificial intelligence boom, arguing that access to high-voltage power has become a tighter constraint than the capital required to purchase chips. The bank said demand for AI infrastructure continues to outpace available capacity and that the supply-demand gap for compute is widening, based on recent discussions with technology company management teams.

In Goldman’s assessment, the bottleneck is shifting from chips toward power. Operators that have already secured grid capacity may be better positioned to expand, while power availability, rather than chip supply, could determine the pace of new data centre construction through 2027. The bank said leading hyperscalers — the largest operators of cloud and AI data centres — are already moving to secure power contracts for 2028 and beyond because grid capacity for 2026 and 2027 is essentially fully booked, a horizon that shows how far in advance power access must now be planned.

Goldman said a likely slowdown in model training is being offset by increased investment in AI safety and pacing monitoring. These activities require additional computing resources, keeping overall compute demand elevated despite a potential moderation in training activity.

The bank also said capital expenditure is rising because of component inflation and the need for more chips. At the same time, revenue generated per unit of power has increased by around two-thirds over the past year compared with earlier price points — a metric that ties operator revenue directly to the electricity it consumes. In Goldman’s view, having enough capital to buy GPUs is no longer sufficient by itself; access to high-voltage energy infrastructure increasingly determines which companies can expand.

The report described the situation in concise terms: money can buy the GPUs, but power decides who gets to switch them on.

Goldman also identified a structural shift in enterprise software spending. It said data and security platforms are among the clearest beneficiaries of AI investment because AI systems depend on clean, organised and secure data. Chief information officers are increasingly directing budgets toward foundational data layers and cybersecurity tools, with the aims of protecting data integrity and containing the risks associated with unsanctioned “shadow AI” use — AI tools adopted outside official IT approval — inside organisations.

By contrast, application software is increasingly being built for AI agents rather than human users. Goldman described this transition as application software going “headless.” The change allows vendors to add more powerful AI features to existing platforms and supports industry growth for now, although the bank cautioned that it could eventually alter how software companies price and monetise their products.

The shift in spending points to a widening gap between foundational platforms and application vendors. Goldman’s emphasis on core platform returns suggests that proven monetisation of compute may be viewed more favourably than speculative new revenue lines, although the bank did not frame this as a forecast of market performance.

Among the companies it highlighted, Goldman said Meta continues to generate two to three times return on investment per megawatt of power in its core business — a measure of returns relative to the electricity consumed — while exploring additional sources of incremental revenue. Google, meanwhile, is using what Goldman described as a full-stack advantage combining custom silicon, cloud, security and a vast distribution base to support core monetisation. Sales of Google’s TPU chips are one component of that broader portfolio strategy.

Goldman concluded that AI monetisation provides substantial optionality. Even so, it said core platform monetisation remains among the higher-confidence and higher-return uses of compute. In that context, established platforms that already generate strong returns on their power may remain the more dependable beneficiaries while capacity remains scarce. The practical markers to track from here, based on the bank’s observations, include data centre construction schedules through 2027, how far ahead operators secure power contracts for 2028 and beyond, and whether enterprise budgets continue shifting toward data and security foundations.