"Tail risks" of components such as ASIC, storage, and optical modules: Power shortage in the US means racks are installed but can't be powered

The bottleneck of U.S. AI infrastructure is shifting from a “chip shortage” to a “power shortage,” and it is the tail end of the supply chain—such as ASIC, storage, optical modules, and power management—that is most threatened by insufficient electricity, rather than NVIDIA itself.

A recent report by Morgan Stanley states that U.S. data centers are expected to face an approximate 34% net power shortfall between 2026 and 2028, equivalent to 32GW. The report believes that the performance forecasts for NVIDIA and Broadcom for 2027 will not be materially affected for now, but power shortages may result in chips not being deployed as planned, in turn creating risks of order delays, cancellations, and inventory adjustments for downstream components.

The key distinction lies in the fact that NVIDIA GPUs have higher power output per unit, and leading chip companies have greater visibility into the final deployment location; in contrast, ASICs are more sensitive to power resources, while products like storage, optical, and power management are more susceptible to customer stocking and project delays.

According to an earlier article from Wallstreet Insights, Oracle’s 1.3GW “Project Lighthouse” in Wisconsin is a real-world reflection of this risk: After the resumption of transmission approval, full power supply for the project may be delayed until as late as October 2028, and in a pessimistic scenario, even spring 2029. Even if AI servers are already racked, without electricity, they cannot be converted into actual computing power and revenue.

NVIDIA and Broadcom: Higher Demand Visibility, Greater Pressure on ASIC

Morgan Stanley believes that NVIDIA and Broadcom’s management teams have already incorporated land, power, and site (LPS) constraints into their guidance, and both companies have high visibility into where chips will ultimately be deployed, with a global layout reducing reliance on the U.S. market alone.

Product efficiency is also crucial. NVIDIA GPUs can generate more tokens per gigawatt, making them advantageous for deployment when electricity is limited. By contrast, ASICs generate lower output per unit of power, requiring more LPS resources and thus are more easily affected by power bottlenecks.

NVIDIA CEO Jensen Huang and Broadcom CEO Hock Tan have both recently emphasized that power resources have become a more challenging constraint in AI infrastructure deployment than chips and storage.

Storage, Optical Modules, and Power Management: Order Risks Concentrated at the Tail End

The report argues that storage, optical components, power management, and analog chips are more susceptible to project delays. These categories previously benefited from the expansion of AI infrastructure and tight supply-demand, but if end projects are delayed, customers may first consume existing inventory, followed by reductions in subsequent orders.

ASICs also face deployment efficiency challenges. In times of power shortages, customers will pay more attention to the computing output per unit of power, so less efficient ASIC projects are more likely to be postponed.

Broadcom’s change in guidance also signals that market expectations are converging: In March, the company expected FY27 AI revenue to far exceed $100 billion, but the latest guidance in September is $115 billion. While the absolute level remains high, further upward adjustment has become limited.

Self-Supplied Power and Overseas Expansion: Still Difficult to Fill the Power Gap

The report projects that new capacity from behind-the-meter gas turbines and engines is around 19GW between 2026 and 2028 in a baseline scenario, with Bloom Energy fuel cells expected to contribute about 6GW; nuclear power and the transformation of crypto mining sites can also provide supplementary capacity, but overall, it remains insufficient to close the gap.

Overseas expansion can only partially alleviate the pressure. Morgan Stanley has already revised the projected U.S. share of global computing power from 60% down to 55%, but slow approvals in Europe and geopolitical risks in the Middle East are limiting the shift in demand.

Meanwhile, the U.S. is also facing a shortage of skilled workers and local opposition. CSIS estimates that by 2030, the U.S. will need to add over 140,000 skilled workers, but the current labor force can only support about 10 to 20GW of new natural gas capacity installations per year.

For the AI industry, the ability to manufacture chips is no longer the only bottleneck; securing sufficient electric power and fully deploying that capacity is now becoming the next major hurdle that determines the speed at which orders are fulfilled and the profitability of the supply chain.