The race for leadership in artificial intelligence may ultimately turn on whether a company can secure adequate energy infrastructure, financial resources, and the physical data-centre capacity to deploy cutting-edge hardware at scale — rather than on chip access alone, according to a research note from Morgan Stanley.
A 32-gigawatt gap between electricity demand and available supply at US data centres through 2028 — even once additional generation options are factored in — forms the central finding of a Morgan Stanley equity research note examining power constraints facing the semiconductor-infrastructure buildout. That figure accounts for roughly 34 per cent of the power demand forecast for the period.
The document noted that Nvidia's own findings pointed to a mismatch between chip output and the infrastructure available to house it, with power supply falling short of what demand requires — a gap that analysts say chip manufacturing growth alone cannot close without parallel advances in electricity access, suitable land, and ready-to-operate data centre capacity.
Questions about whether AI infrastructure spending is outrunning the power grid's capacity have drawn attention to a broader access gap, with smaller technology firms potentially facing steeper barriers to entry than their larger, better-capitalised counterparts.
A Morgan Stanley report outlined how grid connections for data centres are becoming concentrated among a narrow group of operators that carry strong credit ratings and offer utilities predictable revenue streams, with smaller players finding it harder to secure the same access. Firms backed by stronger credit profiles and higher electricity demands were reported to be gaining grid access ahead of others, a pattern that analysts said could place operators with fewer resources at a structural disadvantage in the queue.
According to projections in the research, hyperscalers and specialist cloud operators are on course to represent around 94 per cent of new data-centre capacity coming online worldwide in 2026, with that share tapering to roughly 64 per cent in 2027 and 57 per cent in 2028. Morgan Stanley noted in its statement that such a concentration of build-out activity by those players has the potential to place smaller operators under growing competitive strain over time.
Processor manufacturers face a related exposure: when installation backlogs at data centres leave newly built chip capacity without a home, buyers may choose to push back delivery schedules or walk away from purchase agreements altogether, according to the analysis.
Projections for Nvidia and Broadcom hitting their 2027 targets look achievable, according to analysts at Morgan Stanley, who point to those companies' clear line of sight into what customers plan to order and what infrastructure will be in place to support it. Manufacturers of memory chips, optical components, and other parts further down the supply chain face heightened exposure if the pace of deployment eases, the analysts added.
Beyond 2028, the gap between available electricity and data-centre demand across the United States is forecast to keep widening, with projections placing the cumulative shortfall at around 68 GW by 2029.
Supply constraints of this kind have, in the past, pushed technology firms to broaden their data and computing networks beyond US borders, with Asian markets — India among them — drawing particular attention as candidate locations. Analysts at Morgan Stanley's Asia data-centre division put potential demand migration from US hyperscalers toward Asia at around 14 GW by 2030, though the breakdown by individual country — including India's portion — was left unaddressed in the findings, which also noted that overseas growth by hyperscalers is not viewed as a remedy for domestic power constraints in the United States.
The Morgan Stanley research identified a structural constraint facing the AI sector: investment in chips and advances in hardware capability are unlikely to expand usable computing power unless electrical grid capacity and broader physical infrastructure develop in parallel, the note concluded.
