"Tail risk" for components such as ASICs, storage, and optical modules: the US lacks powerracks are installed but can't get electricity.

Title context: "Tail risk" for components such as ASICs, storage, and optical modules: the US lacks powerracks are installed but can't get electricity. Text: The bottleneck in US AI infrastructure is shifting from "lack of chips" to "lack of power," and the power shortage may first threaten not Nvidia, but the tail of the supply chain including ASICs, storage, optical modules, and power management. Morgan Stanley's latest report says US data centers are expected to face a net power gap of about 34% from 2026 to 2028, equivalent to 32GW. The report believes that Nvidia and Broadcom's 2027 earnings forecasts are not materially affected for now, but the power shortage could prevent chips from being deployed as planned, bringing risks of order delays, cancellations, and inventory adjustments to downstream components. The core difference is that Nvidia GPUs produce more output per unit of power, and leading chipmakers have high visibility into final deployment locations; by contrast, ASICs are more sensitive to power resources, while products such as storage, optics, and power management are more vulnerable to customer stockpiling and project delays. According to a previous Wallstreetcn article, Oracle's 1.3GW "Project Lighthouse" in Wisconsin is a real-world reflection of this risk: after transmission approval restarted, full-power supply for the project may be delayed until as late as October 2028, with a pessimistic scenario even extending to spring 2029. Even if AI servers are already racked, without power they cannot be converted into actual compute power and revenue. Nvidia and Broadcom: Higher demand visibility, more pressure on ASICs Morgan Stanley believes that the management of Nvidia and Broadcom have already incorporated land, power, and data center space (LPS) constraints into guidance, while both companies have high visibility into the final deployment locations of chips, and their global footprint also reduces reliance on the single US market. Product efficiency also matters. Nvidia GPUs can generate more tokens per gigawatt, giving them a deployment advantage when power is limited. By contrast, ASICs produce less output per unit of power and need to compete for more LPS resources, making them more susceptible to power bottlenecks. Nvidia CEO Jensen Huang and Broadcom CEO Hock Tan both recently emphasized that power resources have become a harder constraint to coordinate than chips and storage in AI infrastructure deployment. Storage, optical modules, and power management: Order risk concentrates in the tail The report believes storage, 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 delayed, customers may first digest inventory and then cut subsequent orders. ASICs also face deployment efficiency issues. When power is tight, customers will place more emphasis on compute output per unit of power, making low-efficiency ASIC projects more likely to be postponed. Changes in Broadcom's guidance also show that market expectations are converging: in March the company expected FY27 AI revenue to far exceed $100 billion, and its latest September guidance was $115 billion. The absolute scale remains high, but room for further upward revisions has been limited. Self-generated power and overseas expansion: Still difficult to close the power gap The report estimates that in the base case, new behind-the-meter gas turbine and engine capacity from 2026 to 2028 will be about 19GW, Bloom Energy fuel cells are expected to contribute about 6GW, and nuclear power and crypto mining site conversions can also provide supplements, but overall this is still not enough to absorb the gap. Overseas expansion can also only ease part of the pressure. Morgan Stanley has cut its forecast for the US share of global compute power from 60% to 55%, but factors such as slow approvals in Europe and geopolitical risks in the Middle East limit the shift in demand. At the same time, the US also faces a shortage of skilled workers 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 natural gas installations per year. For the AI industry, whether chips can be produced is no longer the only constraint. Whether enough power can be obtained and actually put into operation is becoming the next threshold determining the pace of order fulfillment and supply chain profitability. This article is reprinted from "Wallstreetcn", edited by GMTEight: Feng Qiuyi.