The bottleneck holding back American AI infrastructure is shifting from a shortage of chips to a shortage of electricity, and the power deficit may threaten not Nvidia first, but the tail end of the supply chain: ASICs, memory, optical modules and power management.

A recent Morgan Stanley report estimates that US data centers will face a net electricity shortfall of roughly 34% between 2026 and 2028, equivalent to about 32GW. The report argues that earnings forecasts for Nvidia and Broadcom in 2027 remain largely unaffected, but power shortages could prevent chips from being deployed as planned, exposing downstream components to order delays, cancellations and inventory adjustments.

The key difference lies in output per unit of power. Nvidia's GPUs generate more tokens per gigawatt, and leading chipmakers have greater visibility into where their products will ultimately be deployed. ASICs, by contrast, are more sensitive to power constraints, while memory, optical and power management products are more vulnerable to customer inventory building and project delays.

Oracle's 1.3GW "Project Lighthouse" in Wisconsin is a real-world illustration of this risk: after transmission approvals were restarted, full-power supply to the project may be delayed until October 2028 at the earliest, and under a pessimistic scenario as late as spring 2029. Even if AI servers are already mounted in racks, without power they cannot be converted into actual compute capacity and revenue.

Nvidia and Broadcom: Higher Demand Visibility, Greater Pressure on ASICs

Morgan Stanley believes that management at both Nvidia and Broadcom has already incorporated land, power and shell (LPS) constraints into their guidance, while both companies have strong visibility into final deployment locations and a global footprint that reduces reliance on the US market alone.

Product efficiency matters just as much. Nvidia's GPUs can generate more tokens per gigawatt, giving them an advantage when power is limited. ASICs, with lower output per unit of power, must compete for more LPS resources and are therefore more exposed to power bottlenecks.

Nvidia CEO Jensen Huang and Broadcom CEO Hock Tan have both recently emphasized that power resources have become a harder constraint to coordinate in AI infrastructure deployment than chips or memory.

Memory, Optical Modules and Power Management: Order Risk Concentrates in the Tail

The report argues that memory, optical components, power management and analog chips are more vulnerable to project delays. These categories previously benefited from AI infrastructure expansion and tight supply-demand conditions; once end projects are postponed, customers may first draw down inventory and then cut subsequent orders.

ASICs also face deployment efficiency issues. When power is tight, customers place greater emphasis on compute output per unit of electricity, making less efficient ASIC projects more likely to be delayed.

Changes in Broadcom's guidance also show market expectations converging: in March the company expected FY27 AI revenue to far exceed US$100 billion, while its latest September guidance stands at US$115 billion. The absolute scale remains high, but room for further upward revisions has become limited.

Self-Generation and Overseas Expansion: Still Not Enough to Close the Power Gap

The report estimates that behind-the-meter gas turbines and engines will add about 19GW of capacity between 2026 and 2028 under a base case, with Bloom Energy fuel cells contributing roughly 6GW, while nuclear power and converted crypto mining sites can provide additional supply. Even so, this is still not enough to absorb the shortfall.

Overseas expansion can only ease part of the pressure. Morgan Stanley has cut its forecast for the US share of global compute from 60% to 55%, but slow European approvals and geopolitical risks in the Middle East limit how much demand can be shifted elsewhere.

At the same time, the US also faces a shortage of skilled labor and local resistance. CSIS estimates that by 2030 the US will need more than 140,000 additional skilled workers, while the existing workforce can support only about 10 to 20GW of new gas capacity per year.

For the AI industry, whether chips can be manufactured is no longer the only constraint. Whether enough power can be secured and actually put into operation is becoming the next threshold determining how quickly orders are fulfilled and how profitable the supply chain will be.