US Data Center Power Shortfall to Reach 38GW: Why Morgan Stanley Favors Nvidia and Broadcom

American AI data centers are critical infrastructure for training large language models, powering inference services, and supporting cloud computing businesses. Building them requires huge volumes of GPUs, ASIC chips, memory, networking equipment, and reliable power supply. As global tech companies continue to increase AI capital expenditures, data center construction is expanding rapidly. However, grid capacity, transmission infrastructure, and the speed of power access may not grow at the same pace. Power availability is increasingly becoming a key factor that determines whether AI computing power can be deployed in practice.

In its latest research note, Morgan Stanley raised its forecast for the U.S. data center net electricity shortfall from 33GW to 38GW for 2026–2028, reflecting growing market concern about the power supply capability for AI infrastructure. At the same time, the firm continues to highlight Nvidia (NVDA) and Broadcom (AVGO) as important positive picks in semiconductors. It believes factors such as chip shipment visibility, overseas market expansion, and unit-power computing efficiency may help some leading companies maintain competitiveness in environments where electricity is limited.

This shift suggests that competition in the AI chip industry is moving beyond simply chasing higher computational performance. It is now also increasingly focused on computing efficiency, power availability, and capital expenditure returns. For investors, a 38GW electricity gap is not only about data center construction timelines. It could also change the competitive relationship between GPUs and ASICs and influence future demand across parts of the semiconductor supply chain, including memory, optical communications, and power management.

What Does the 38GW Power Shortfall Mean for U.S. Data Centers?

Morgan Stanley raised its forecast for the U.S. data center net electricity shortfall from 33GW to 38GW. That means the firm expects a bigger mismatch between electricity demand from AI data centers and the available power supply over the coming years. Here, 38GW refers to a forecast gap in power capacity, not the size of actual outages that have already occurred. It also should not be interpreted as the U.S. grid overall lacking 38GW of generation capacity.

GW (gigawatts) is a unit of power. 1GW equals 1,000MW. Large AI data centers typically require tens to hundreds of megawatts of power capacity, while hyperscale campuses may plan for even higher electricity demand. As a result, a potential gap of tens of gigawatts implies that many data centers that are already planned may need to wait for grid upgrades, adjust their construction locations, or adopt alternative energy solutions.

It’s important to note that data center power supply is influenced by multiple factors, including local grid interconnection capacity, substation construction, transmission lines, generation resources, and related approvals. Even if a region’s overall generation is sufficient, that doesn’t guarantee a specific data center can secure enough power within its planned timeframe. Therefore, the electricity shortfall mainly reflects a mismatch between supply and demand across both time and geography.

For the AI industry, power constraints may delay some computing capacity buildouts. That could affect GPU procurement, server deliveries, and launch timing for cloud services. Investors need to distinguish between data center planned capacity, actual approved power capacity, and computing capacity that is already in operation. These metrics are not identical.

Why Could a Power Shortage Change the GPU vs. ASIC Competitive Landscape?

GPUs and ASICs are two important chip categories in AI computing. GPUs have strong general-purpose parallel computing capabilities and can support model training, inference, and a wide range of computing tasks. ASICs are application-specific integrated circuits designed for particular use cases. They typically improve performance and efficiency by optimizing chip architecture and computation workflows.

When power supply is relatively sufficient, data center operators can weigh chip purchase prices, compute performance, software compatibility, and long-term operating costs. But when power capacity becomes a hard constraint, companies need to focus more on how many computing tasks a given unit of power can complete. For AI inference businesses, this can be expressed as the number of Tokens processed per watt of electricity.

Morgan Stanley believes Nvidia has a competitive advantage in token output efficiency per watt. In power-constrained environments, some ASIC products may face pressure on market share. The core of this view is that data center customers may prioritize computing platforms that produce higher real output within a limited power budget, rather than focusing only on chip procurement costs.

However, it’s not accurate to conclude that all ASICs are less efficient than GPUs. Google TPU, Amazon Trainium, and other custom AI accelerators may also offer strong efficiency advantages under certain models and workloads. The actual performance of different chips depends on model architecture, inference accuracy, batch sizes, memory configuration, software optimization, and system utilization.

So there is no fixed conclusion that applies to all scenarios in the GPU vs. ASIC competition. Morgan Stanley’s view is best understood as follows: under current product and deployment conditions, system-level efficiency for some Nvidia computing platforms may receive more attention. ASIC competitiveness still needs to be evaluated based on specific workloads.

Why Does Morgan Stanley Keep Favoring Nvidia (NVDA)?

Nvidia (NVIDIA, NVDA) is a major supplier in the global AI acceleration computing market. Its competitive advantage comes not only from the GPU chips themselves, but also from the CUDA software ecosystem, networking interconnects, server systems, and related development tools. As demand for AI model training and inference grows, Nvidia keeps rolling out new computing platforms to improve compute performance, memory bandwidth, and overall system efficiency.

Morgan Stanley continues to favor Nvidia primarily based on factors such as conservative earnings guidance, improved chip shipment visibility, overseas market expansion, and unit-power computing efficiency. For investors, shipment visibility means product deliveries and customer demand over the next period may be clearer, which can reduce some uncertainty in revenue forecasts. Still, orders and delivery expectations may change due to customers adjusting capital expenditures or shifts in the supply chain.

In a power-constrained backdrop, Nvidia’s system-level efficiency could become an important competitive edge. Data center operators need to compare not only the performance of a single GPU, but also the efficiency of the servers, networking, cooling systems, and software runtime. If a computing platform can complete more inference tasks with the same power capacity, it may increase a data center’s revenue potential.

At the same time, Nvidia faces long-term competitive pressure. Large cloud computing companies are continually developing custom AI chips to reduce computing costs and improve supply chain autonomy. Even if Nvidia’s revenue keeps growing, its market share, gross margins, and valuation level may not rise in sync. Therefore, when analyzing NVDA stock, investors still need to focus on data center revenue, product gross margins, customer capital expenditures, and competition from custom chips.

Why Does Broadcom (AVGO) Also Get Morgan Stanley’s Attention?

Broadcom (Broadcom, AVGO) is a major semiconductor and infrastructure software company with business exposure in areas such as custom AI accelerators, network switching chips, and high-speed interconnects. As large cloud computing enterprises expand AI data center construction, demand for custom chips and network infrastructure becomes an important growth opportunity for Broadcom.

Unlike Nvidia, whose key business largely centers on providing general-purpose GPU platforms, one of Broadcom’s important directions is helping large technology customers develop custom chips tailored to specific workloads. For cloud providers with lots of repetitive computing tasks, custom ASICs can reduce some computing costs through targeted optimization. In addition, large AI clusters require high-speed switching chips and network equipment to connect compute nodes, which also creates market opportunities for Broadcom.

Morgan Stanley’s simultaneous positive stance on Nvidia and Broadcom indicates it does not view GPUs and ASICs as completely opposing investment directions. Even if some ASIC products face share pressure in power-constrained environments, custom chips could still keep growing in specific customers and computing scenarios. Broadcom is also not solely dependent on an ASIC business. Its network semiconductor business can participate in the expansion of AI cluster scale.

That said, Broadcom’s custom chip business may be affected by the project timelines of large customers. If customers change their chip R&D plans, data center construction schedules, or capital expenditures, related revenue growth could fluctuate. As a result, the AVGO investment thesis needs to consider custom AI chip demand, network business growth, and the risk of customer concentration at the same time.

Why Are 2027–2028 Data Center Capex Expectations Being Questioned?

Morgan Stanley notes that the market is trying to align the 2027 forecast of about $1.5 trillion and the 2028 forecast of about $1.6 trillion in data center capital expenditures with actual power supply capability. These figures are market forecasts of future investment scale, not capital expenditures that have already been approved, signed, or actually occurred. The exact statistical scope also needs to be confirmed with the underlying research note.

AI data center capital expenditures typically include projects such as compute chips, servers, networking equipment, facility construction, and power and cooling infrastructure. Even if tech companies have strong investment intent, some hardware may not be deployed on schedule if the grid cannot provide enough capacity within the planned timeframe.

This issue can affect companies’ capital allocation decisions. For example, if a data center plans to begin operations in 2027 but power delivery is delayed until 2029, an operator might choose to postpone GPU and server purchases to avoid long idle time after equipment arrives early. For AI chips that update quickly, early procurement could also increase the risk of equipment depreciation.

Therefore, the market needs to distinguish among three different stages: companies announce capex plans, they actually purchase and build the infrastructure, and then computing equipment is officially put into operation and starts generating revenue. Only the final stage truly converts into ongoing operating cash flows from those compute assets.

Why Could Slower Data Center Construction Hit Storage Chips and Optical Modules?

The AI data center industry chain has long procurement and production cycles. Chip manufacturers, memory suppliers, and networking equipment companies usually plan production based on customers’ anticipated needs over the next several quarters. When the market expects rapid growth in AI infrastructure investment, customers may buy key components in advance to ensure future equipment availability.

But if data center construction is delayed due to power shortages, some planned components may not be deployed on time. As a result, companies may delay purchases, adjust inventories, or reschedule deliveries. Such changes could suddenly expose previously tight-supply products to slower demand pacing.

Storage chips are one area to watch closely. HBM and server DRAM demand are closely tied to the scale and deployment pace of AI accelerators. If the speed at which new GPU clusters go live slows down, procurement timing for related memory could also be affected. At the same time, the storage chip industry is cyclical. Company expansions and customers stocking up in advance can further amplify shifts in supply and demand.

Optical modules and high-speed interconnect components also depend on building large compute clusters. If data center deployments are delayed, orders for related networking equipment may be pushed back. For power management and analog chip companies, the actual construction progress of servers and power supply systems will also influence product demand.

However, delays in construction do not necessarily mean final demand disappears. Some orders may only be delayed deliveries rather than permanently canceled. The real impact still depends on supplier inventories, customer purchase agreements, and growth in demand in other regions. Therefore, investors need to distinguish between short-term order fluctuations and changes in the long-term market size.

Which Semiconductor Stocks Could Be Affected by AI Data Center Power Bottlenecks?

Different semiconductor companies occupy different positions in the AI supply chain, so the impact of power supply constraints varies across their businesses.

For NVDA and AVGO, power bottlenecks may create both opportunities and challenges. On one hand, customers may place more emphasis on energy efficiency, potentially increasing demand for high-performance computing platforms and optimized chips. On the other hand, if data center construction slows, actual chip deliveries may also be affected.

For MU and other memory suppliers, the key question is whether AI server purchases can keep growing, and whether early stocking up leads to inventory accumulation. For power infrastructure companies like VRT and ETN, growth in data center electricity demand could create a new equipment market, but relevant projects are still constrained by construction schedules, approvals, and investment budgets.

Therefore, it’s not accurate to assume that power shortages will automatically benefit power equipment stocks or automatically hurt all semiconductor companies. Each company’s order mix, technical strengths, and valuation levels still determine how much it will actually benefit.

How to Use Gate Stock to Track the AI Chip and Data Center Power Supply Chain?

For investors who want to track the AI infrastructure investment cycle, they can use Gate Stock to follow semiconductor companies such as Nvidia (NVDA), Broadcom (AVGO), and Micron Technology (MU). They can also pair this with data center power management, cooling equipment, and related tech sector ETFs to observe how different segments of the supply chain perform in the market.

Gate Stock provides a way to use USDT to trade U.S. stocks and ETFs. The specific tradable tickers, trading fees, and product rules depend on the platform’s actual pages.

When analyzing related stocks, investors should look at three dimensions at the same time: first, whether AI compute demand keeps growing; second, whether data centers can secure enough power and come online on schedule; and last, whether companies can convert orders and technical advantages into profits and cash flow. Relying only on data center capex forecasts is not enough to judge a stock’s long-term value.

Summary

Morgan Stanley raised its forecast for the U.S. data center net electricity shortfall in 2026–2028 from 33GW to 38GW. This reflects increasingly clear energy constraints on the expansion of AI infrastructure. Even if large tech companies keep raising capital expenditures, whether data centers can secure power as planned will still affect the real deployment scale of GPUs, ASICs, memory chips, and networking equipment.

Against this backdrop, chip competition may place even greater emphasis on compute efficiency per unit of power. Morgan Stanley continues to favor Nvidia and Broadcom, reflecting long-term confidence in high-performance computing platforms, custom chips, and high-speed networking demand. Still, the GPU vs. ASIC competition depends on the specific workload. You can’t judge future market share based on a single efficiency metric alone.

Meanwhile, if data center construction slows due to insufficient power supply, storage, optical module, and power management components that were previously pulled forward due to supply tightness may face inventory adjustments and order volatility. For investors, the focus going forward should be on actual deployed capacity, companies’ capex realization rates, chip energy efficiency, and supply chain inventory—not just the total amount in data center construction plans.

In the next stage of AI infrastructure investment, the key is no longer only how many chips can be purchased. It’s how much effective compute can be supported by limited power resources—and how much commercial revenue those compute capabilities ultimately generate.

FAQ

What does the 38GW electricity shortfall for U.S. data centers mean?

This is Morgan Stanley’s forecast of the net electricity supply-demand gap for U.S. data centers from 2026 to 2028. It reflects the potential for a sizable difference between planned electricity demand and available power capacity. It does not mean the U.S. has already experienced 38GW of actual outages.

Why could a power shortage benefit Nvidia GPUs?

When available electricity for data centers is constrained, operators may place more emphasis on how many AI computing tasks can be completed per watt. If Nvidia’s computing platforms deliver higher system-level energy efficiency under specific workloads, Nvidia could gain a competitive advantage. Still, the real outcome depends on chip configurations, software optimizations, and application scenarios.

Will ASIC chips lose market share due to a power shortage?

Some ASIC products may face competitive pressure, but you can’t generalize. Custom ASICs can also achieve high energy efficiency for specific AI tasks. Future market share depends on chip performance, cost, software ecosystem, and customer deployment strategies.

Why does Morgan Stanley favor both NVDA and AVGO?

Nvidia has advantages in GPUs, accelerated computing systems, and software ecosystems. Broadcom participates in the custom AI chip and high-speed networking market. The two companies benefit from AI infrastructure investment in different ways, and it doesn’t mean that only GPUs or only ASICs can grow.

Why could slower data center construction affect storage chips?

If data centers delay server deployment due to power constraints, customers may correspondingly postpone purchasing HBM, DRAM, and other components. If there was early stocking up beforehand, short-term inventory adjustments could further amplify order and price volatility.

Will a 38GW electricity shortfall cause AI investment to stop across the board?

Not necessarily. Companies can mitigate some constraints by adjusting construction locations, deploying in phases, improving compute energy efficiency, or adding energy infrastructure. However, these measures usually take time and money, so they can’t guarantee that all planned projects will be completed on schedule.