Morgan Stanley analyst highlights AI adoption challenges amid computing bottlenecks

Morgan Stanley analyst Stephen Byrd has laid out a sobering reality check for the AI boom: demand for computing power is growing so fast that the physical infrastructure underpinning it simply cannot keep pace. The firm’s research points to a US data center power demand of up to 74 gigawatts by 2028, against a supply landscape that would leave a shortfall of roughly 49 gigawatts.

The numbers behind the bottleneck

Global investment in AI-related infrastructure, including data center expansion, is estimated to approach $3 trillion through 2028.

Morgan Stanley revised its data center power demand forecast upward in late 2025, flagging a widening gap in compute capacity driven by what Byrd describes as non-linear AI improvements.

From early January 2026, weekly token consumption surged roughly 250%, escalating from 6.4 trillion to 22.7 trillion tokens per week.

Byrd notes that while AI tools do yield genuine productivity gains, foundational constraints like energy supply are forecasted to remain a critical bottleneck until at least 2027 to 2028.

Beyond electricity: the full stack of constraints

Morgan Stanley’s research identifies what it calls “intelligence bottlenecks,” a collection of interrelated constraints that together define how quickly AI can actually scale. Labor shortages rank prominently on that list. Political approvals for new power generation and transmission infrastructure add another layer of friction. Ordering a large power transformer today can mean waiting two to three years for delivery.

Where the money is flowing

The projected 49 GW shortfall has implications for grid operators and utilities, which will need to accommodate massive new loads without destabilizing service for existing customers. Energy companies with generation capacity that can be brought online relatively quickly stand to benefit, with natural gas, nuclear, and geothermal assets all part of the conversation.