AI Data Centers Hit a "Power Crunch": A Full Breakdown of the Electricity Gap and Which Semiconductor Stocks Face the Biggest Risk?
The pricing logic of the AI supply chain is undergoing a subtle but important shift. Over the past two years, the market has grown accustomed to inferring the pace of AI capital expenditure from GPU supply, CoWoS packaging capacity, and HBM yields. Now, what determines when a batch of computing power can actually be switched on is often not the chip delivery timeline, but when the substation is built and when the grid connection permit is approved.
According to a Benzinga report, Morgan Stanley estimated in an October 5 research note that by 2028, U.S. data centers will still face a net power shortfall of approximately 32GW after accounting for various mitigation measures, equivalent to a supply deficit of about 34%. More noteworthy is its tiered assessment: Nvidia and Broadcom are relatively protected, while second-tier suppliers in memory, optical modules, power management, and analog chips are more likely to passively bear order delays and inventory fluctuations due to data center project postponements.
This is a judgment that directly translates macro infrastructure constraints into differences in individual stock exposure, and it marks the first time in this round of AI trading that a clear "same demand, different pace" differentiation framework has emerged.
Key Takeaways
The power gap is a net figure, not a gross figure. Morgan Stanley's 32GW estimate has already deducted mitigation measures such as self-built generation, gas turbines, and nuclear power coordination, so it measures the portion that is genuinely difficult to fill.
Grid connection queues are the hardest constraint. According to Lawrence Berkeley National Laboratory's "Queued Up" 2026 edition, as of the end of 2025, more than 2,060GW of generation and storage capacity remained in queue awaiting grid connection in the U.S.
Leading chipmakers have scheduling power. When power is insufficient, major customers prioritize the deployment of core GPUs and custom ASICs, and can migrate computing power to regions with available electricity, which keeps the recognition pace for Nvidia and Broadcom relatively stable.
Second-tier components bear the timing risk. Memory, optical modules, power management, and analog devices typically ship according to complete-system production schedules. Once a rack cannot be powered on, these components are the most likely to be rescheduled.
Delay does not equal disappearing demand. The power bottleneck changes the temporal distribution of revenue recognition, not the long-term demand curve for AI computing power. But for investors who evaluate performance quarterly, timing itself is risk.
Power Is Replacing Computing Power as the New Bottleneck
How the Gap Is Calculated
The 32GW figure only makes sense when placed back in context. It is a net gap, meaning that even after Morgan Stanley assumes the industry will deploy self-built generation, gas peaking, fuel cells, and nuclear coordination, roughly one-third of power demand still has no confirmed source. By comparison, other analyses citing Morgan Stanley's earlier estimates suggest that new data center demand from 2026 to 2028 will be approximately 68GW, of which about 30GW has already been built or contracted, leaving a gap of approximately 38GW with no clear answer. The two figures use different methodologies, but point to the same conclusion: the pace of expansion on the power supply side can no longer keep up with the pace of orders on the computing side.
The U.S. Energy Information Administration's Short-Term Energy Outlook provides confirmation on the demand side. According to its electricity consumption forecast, total U.S. electricity consumption will reach 4,135 billion kWh in 2026 and 4,211 billion kWh in 2027, setting records for two consecutive years, with commercial sector electricity consumption projected to grow 3.3% and 2.7% respectively, accounting for approximately 63% and 56% of total electricity increment. Data centers are the primary source of this increment.
Grid Connection Queues and Local Approvals Become the Real Gatekeepers
Insufficient power is not just about inadequate generation capacity; it is more about the time lag in transmission, distribution, and approval processes. Grid connection applications typically take five to seven years to complete studies and construction, while hyperscale customers want capacity within two to three years. This temporal mismatch cannot be resolved through purchase orders.
Texas provides a real-world example. A Power Magazine report shows that Texas Governor Abbott ordered a pause on new data center connections to the ERCOT grid on August 3, 2026, to conduct case-by-case audits, affecting approximately 49.8GW of pending load, close to one-fifth of the roughly 253GW development pipeline nationwide; ERCOT's large load interconnection queue is approximately 474GW, about 90% of which comes from data centers. BloombergNEF estimates that if delays persist, the cumulative cost impact on related projects could reach $8 billion to $15 billion by the first quarter of 2027.
What characterizes such events is that they do not change any company's technology roadmap, but directly alter delivery timelines. For a supply chain that recognizes revenue quarterly, this is a substantive financial variable.
Why Nvidia and Broadcom Are Relatively Protected
Scarce Power Flows to Core Computing First
When a campus's available power falls below planned levels, the operator's prioritization is fairly clear: first power on the accelerators and networking equipment that directly generate token revenue, and defer the rest. This is precisely the core logic behind Morgan Stanley's view that the two leading companies face less impact. Its research notes that Nvidia and Broadcom have strong visibility into the final deployment locations of their chips, can geographically adjust shipment direction, and coordinate with data center and power infrastructure parties, giving them the ability to redirect scarce computing power to projects with available electricity. The firm also did not lower its 2027 forecasts for either company due to power constraints.
Financial data currently still supports this judgment. According to Nvidia's second-quarter earnings announcement filed with the U.S. Securities and Exchange Commission, the company reported revenue of $96.2 billion in the second quarter of fiscal year 2027, up 106% year-over-year, with data center revenue of $89 billion, up 117% year-over-year, and third-quarter guidance of $108 billion. Broadcom's third-quarter 8-K filing shows quarterly revenue of $29.6 billion, up 86% year-over-year, with AI semiconductor revenue of $16.7 billion, up 221% year-over-year, and guidance for fourth-quarter AI semiconductor revenue to accelerate to $21.7 billion.
Chipmakers Begin to Directly Intervene in the Power Segment
More telling is that leading companies are no longer just waiting for customers to solve power supply issues. According to Nvidia's announcement regarding the PORTS-Pike campus in Ohio, the company is providing credit guarantees for land, power, and plant construction for the campus developed by SB Energy, locking in an initial 4.25 IT-GW of capacity with an option for the remaining 3.75 IT-GW, for a total planned capacity of 8 IT-GW, while investing $1.5 billion in SB Energy. The project is expected to come online in phases starting in 2028, with supporting grid investment of no less than $4.2 billion. Jensen Huang said in the announcement that land, power, and plants have become critical in the AI era.
Directing credit resources toward power generation and distribution is essentially using the balance sheet to buy insurance for one's own shipment pace. This capability is only available to a handful of companies with extremely abundant cash flow, and thus constitutes a structural differentiation within the supply chain. Expansion on the demand side has not stopped either. According to reports, Anthropic is advancing sizable computing infrastructure deployment with multiple partners, and related background can be found in Anthropic's AI infrastructure investment plan.
Why Memory, Optical Modules, and Second-Tier Components Are More Easily Deferred
Memory Bears the Most Direct Scheduling Risk
Memory is the segment with the greatest elasticity in this cycle and the most dependent on complete-system production scheduling. According to Micron's fourth-quarter fiscal year 2026 earnings announcement, the company reported quarterly revenue of $54.23 billion, full-year fiscal 2026 revenue of $133.19 billion, with core data center business unit quarterly revenue of $18 billion, and provided first-quarter fiscal 2027 guidance of $61.5 billion plus or minus $1.5 billion. Management also noted that strategic customer agreements provide confidence in demand sustainability.
These figures explain the source of market tension. Extremely high growth base and extremely high gross margins mean that any scheduling change will be amplified on both the performance and valuation fronts. Memory makers typically ship according to customers' complete-system launch plans, and once racks cannot be powered on due to unavailable power supply, the pace of component receipt will be renegotiated. For more details on Micron's latest earnings, please refer to this earnings analysis.
Optical Modules and Power Management Are on the Same Production Line
Optical interconnect and power management devices face the same type of problem. The shipment pace of 800G and 1.6T optical modules is strongly correlated with switch and rack deployment. Revenue recognition for optical component suppliers like Coherent and Lumentum depends on whether downstream data centers can be commissioned on schedule. The same applies to power management and analog devices, which serve the complete-system electrical architecture rather than individual accelerators, and therefore are more tightly bound to "whether power is available."
Morgan Stanley's warning lands precisely at this layer: if deployment is delayed, these second-tier suppliers face not only revenue deferral but also potential inventory digestion pressure and order adjustments. For a segment that has just experienced rapid capacity expansion, this is a more realistic risk than slowing demand.
Server and Rack Manufacturers Are in the Middle
The situation for complete-system manufacturers is relatively complex. They both enjoy order explosions and bear delivery node risks. Dell Technologies' second-quarter earnings announcement shows the company reported revenue of $47 billion in the second quarter of fiscal year 2027, up 58% year-over-year, with AI server orders reaching a record $60.9 billion, AI backlog rising to $95 billion, and full-year AI-optimized server revenue guidance raised to $74 billion.
The larger the backlog, the more sensitive the delivery pace is to power availability. In other words, these companies' order books are real, but the timing of converting orders into revenue is not entirely up to them. The same logic applies to other AI infrastructure suppliers, and the recent background of HPE's stock performance can also be understood within this framework.
Supply-Side Bottlenecks Formed by Grids, Transformers, and Cooling
Transformer and Electrical Equipment Lead Times Have Become Industry Common Knowledge
The power bottleneck does not exist only on the generation side. According to a pv magazine USA report, U.S. power transformer delivery lead times have extended to approximately four years, with prices rising about 80% over the past five years; between 2019 and 2025, generator step-up transformer demand grew 274%, and substation transformer demand grew 116%. Supply constraints in grain-oriented electrical steel and copper are the main obstacles to capacity expansion. Hitachi Energy has invested approximately $1 billion in capacity expansion, with its South Boston plant planned to begin production in 2028, and Siemens has also invested $421 million to build a transformer plant in Charlotte, but analysts generally believe the supply-demand imbalance will persist for several years.
Electrical equipment manufacturers' order books confirm this. Eaton's second-quarter earnings announcement shows the company reported quarterly revenue of $8.5 billion, up 21% year-over-year, with Electrical Americas trailing twelve-month orders growing 41% organically and Electrical Global backlog growing 103% year-over-year.
Cooling and Data Center Thermal Management Upgrade in Tandem
Increasing rack power density has turned cooling from a supporting element into primary equipment. According to Vertiv's second-quarter earnings announcement, the company reported quarterly net sales of $3.274 billion, up 24% year-over-year, and raised its full-year 2026 revenue guidance to $13.8 billion to $14.2 billion. Management stated that AI and general computing demand continues to strengthen, and each technology iteration makes deployment more complex.
This means the benefit logic for cooling and electrical infrastructure suppliers differs from that of chip second-tier suppliers. The former's bottleneck lies in its own capacity expansion speed, while the latter's bottleneck lies in whether downstream can be powered on schedule.
Gas, Nuclear, and Renewables Together Fill the Gap
The supply-side response is forming along three paths. Gas turbines are the fastest. According to GE Vernova's second-quarter
