AI Stock Pullback No Cause for Alarm? Morgan Stanley's 120-Page Report Breaks Down 3 Major Concerns: The Real Undervalued Opportunity Is 'This'!

Since the end of June, global artificial intelligence (AI) stocks have undergone a significant pullback, stirring renewed debate over whether the AI investment boom has peaked. In response, Morgan Stanley (MS-US) released a comprehensive 120-page report on July 27 titled 'Playing the AI Infrastructure Dip,' aiming to clarify the situation for investors.

The report opens with a clear conclusion: the root of this pullback lies in technical factors—overcrowded positions unwinding, margin de-leveraging, and a reversal in momentum factors—not in any deterioration of AI fundamentals.

In other words, this is more of a 'reshuffling' on the capital side rather than a refutation of the industry's core logic.

Morgan Stanley then systematically dismantles three major bearish narratives circulating in the market and introduces a previously under-discussed but potentially the most mispriced investment theme: 'Time to Power.'

Concern 1: Are companies capping AI usage budgets, threatening cloud giants' revenue?

Recently, some companies have started imposing budget caps on employee AI token usage, raising concerns that this could slow revenue growth for large language model providers.

Morgan Stanley disagrees. The report notes that current enterprise spending on AI tokens is extremely low, yet the return on investment (ROI) is remarkably high.

Based on extensive research into enterprise AI use cases, the firm estimates that each AI call saves companies about $55 in labor costs, while completing a business task via AI agents costs only $2–$5 on average—yielding an ROI exceeding tenfold.

The firm argues that a tool with such high ROI is no longer a budgetary question but a matter of corporate survival and competitiveness. Companies that delay AI adoption will inevitably fall behind.

Moreover, as GPU generations advance, data center profitability will continue to rise. This means cloud providers can maintain or even improve margins even if token prices drop significantly, thereby unlocking further demand.

Morgan Stanley's Intelligence Factory model estimates that data centers using the current Blackwell architecture achieve a net margin of about 58% on token sales. Upgrading to next-generation Rubin and Feynman chips could push margins to approximately 80% and 90%, respectively.

This implies cloud providers could theoretically cut token prices by around 75% while maintaining the same profit levels. Price cuts and profitability are not mutually exclusive.

Concern 2: Does China's low-cost 'Kimi moment' undermine U.S. giants' capex narrative?

After the launch of Kimi K3, markets began questioning whether U.S. cloud giants' trillion-dollar AI capital expenditures face downward pressure on returns if Chinese teams can train similarly powerful models at a fraction of the cost.

Morgan Stanley answers no. The report argues that the efficiency race between U.S. and Chinese model training will not reduce demand for computing power; instead, it will reinforce the structural imbalance where supply lags far behind demand.

The report cites 19th-century economist William Stanley Jevons, who observed that after James Watt improved the steam engine, coal efficiency rose—but total coal consumption in Britain increased. Why? Efficiency lowered the barrier to adoption, leading to widespread deployment across more factories and mines.

This phenomenon, known as the 'Jevons Paradox,' holds that when the efficiency of using a resource improves, total consumption often rises rather than falls.

Morgan Stanley applies this logic to AI: higher compute efficiency lowers per-token costs, which in turn drives more use cases, more users, and higher usage frequency—ultimately increasing total compute demand.

The report highlights the staggering scale of this supply-demand gap. Google executives recently revealed the company may need to double its compute capacity every six months—equivalent to a 1,000x increase over five years.

On the supply side, NVIDIA (NVDA-US) is projected to grow AI chip sales at a CAGR of about 140% from 2025 to 2028. Even extrapolating this growth over five years, the cumulative compute delivered would still fall short of just one-tenth of Google's estimated demand.

In short, even the world's largest chip supplier operating at record speed can only meet a fraction of a single major customer's needs.

Concern 3: Will power, labor, and political resistance become hard limits on AI expansion?

Even if the first two concerns are unfounded, a third worry remains: even with strong demand, could real-world constraints slow data center construction?

Morgan Stanley categorizes these constraints as the '3Ps': People (labor), Power (electricity), and Politics.

On labor, data center construction relies heavily on skilled trades like electricians, welders, and pipefitters—all currently facing structural shortages.

On power, grid interconnection queues in some regions have stretched to five to seven years, becoming the biggest bottleneck to timely data center deployment.

On politics, data center expansion faces rising resistance at both state and federal levels. The policy wind is shifting.

In recent years, states competed to attract data centers with generous incentives. Now, the trend is reversing: more states are pausing, adding conditions, or even revoking tax breaks.

Issues like 'slowing data center growth to prevent rising electricity bills for residents' are becoming key campaign platforms in gubernatorial races, expected to draw voter attention in the November elections.

Meanwhile, at the federal level, the U.S. is considering a nationwide 'data center tariff' mechanism.

The U.S. House is currently reviewing the 'Ratepayer Protection Act,' which would require utility companies to establish a 'large power user standard,' making data centers bear the cost of grid upgrades.

In fact, as early as March, tech giants including Amazon (AMZN-US), Google (GOOGL-US), Meta (META-US), Microsoft (MSFT-US), Oracle (ORCL-US), and xAI signed the White House-initiated 'Ratepayer Protection Pledge,' voluntarily committing not to pass infrastructure costs onto consumers. The bill aims to formalize this pledge into law, effectively creating a national system of power surcharges for data centers.

Nonetheless, Morgan Stanley acknowledges these concerns are valid but classifies them as 'speed bumps,' not structural barriers.

The report notes that with grid interconnection queues exceeding five years in some areas and growing pressure on data centers to secure their own power, on-site power generation is becoming a core solution.

The truly undervalued opportunity: Time arbitrage in 'Time to Power'

Morgan Stanley quantifies the U.S. data center power gap.

Results show that between 2026 and 2028, U.S. data centers will need about 68 gigawatts (GW) of power. After subtracting 15 GW under construction and 15 GW already contracted for grid connection, a potential shortfall of 38 GW remains—while grid queues in some regions already stretch five to seven years.

Under these conditions, Morgan Stanley identifies 'Time to Power' as the most mispriced investment theme in the market. The logic is clear: data center deployment is constrained by power availability → grid interconnection takes 5–7 years → any alternative power solution deployable within 1–3 years offers significant time arbitrage value.

The report highlights two paths to accelerate bottleneck relief:

First, Bitcoin mining facilities, which already possess substantial grid interconnection capacity and land, can be repurposed for data centers, potentially contributing 10–19 GW.

Second, fast-deployable power solutions—such as natural gas turbines, which offer a 1–3 year time advantage over traditional grid connections—could contribute around 15–20 GW.