Morgan Stanley: AI Correction Technical, Not Fundamental

Woofun AI reports that Morgan Stanley released a comprehensive 120-page research document titled "Playing the AI Infrastructure Dip" on July 27, addressing the significant correction observed in the global AI sector since late June. The report’s central thesis asserts that the recent market volatility is driven by technical factors—specifically the digestion of overcrowded positions, deleveraging by collateral financing, and a reversal in momentum factors—rather than any breakdown in the fundamental logic of the industry. This distinction is critical for investors attempting to determine whether the current downturn signals a cyclical peak or merely a temporary technical adjustment. By isolating these technical drivers from the underlying business case, the report provides a framework for evaluating the long-term viability of AI infrastructure investments amidst short-term market noise.

Concerns regarding enterprise AI budget sustainability have emerged as companies impose limits on employee AI Token usage, raising questions about revenue growth for large model providers.

However, Morgan Stanley’s analysis of numerous enterprise-level AI use cases demonstrates that these fears are unfounded due to the extremely high return on investment (ROI). Research indicates that an average AI call saves approximately $55 in labor costs, while the cost of completing an enterprise task through Agent collaboration ranges from only $2 to $5. This dynamic results in an ROI exceeding 10 times the input cost. For tools delivering such substantial returns, adoption is not a discretionary budget decision but a matter of core competitiveness. Companies that fail to actively deploy AI capabilities will face increasingly significant competitive disadvantages, making the current spending base for Token usage by corporate employees highly efficient despite its low absolute value.

The evolution of GPU generations further supports the argument that profit margins in data centers will improve, allowing for significant reductions in Token prices without harming profitability. Morgan Stanley’s Intelligence Factory model estimates that net profit margins from Token sales in data centers utilizing Blackwell GPUs are currently around 58%. As newer generations are deployed, these margins are projected to rise substantially: Rubin GPUs are expected to achieve approximately 80% profit margins, while Feynman GPUs could reach 90%. This structural improvement in efficiency implies that Hyperscalers have ample room for cost reduction. Consequently, they can lower Token prices by about 75% while maintaining their current profit margins, thereby stimulating further demand through a virtuous cycle of lower costs and higher adoption rates.

The release of Kimi K3 has prompted capital markets to reevaluate the assumption that lower-cost training in China might reduce the annual AI Capex of U.S. Hyperscalers, which amounts to trillions of dollars. Morgan Stanley argues that this efficiency gain will not weaken demand but rather reinforce the structural judgment that demand far exceeds supply. Drawing on the historical precedent of William Stanley Jevons, who observed that James Watt’s improvements to the steam engine increased coal efficiency yet caused Britain’s total coal consumption to soar, the report applies the Jevons Paradox to the current AI revolution. As hash rate efficiency improves and the cost per Token decreases, more use cases, users, and frequent calls emerge, ultimately driving up overall computing consumption rather than reducing it.

To quantify the severity of this supply-demand imbalance, Morgan Stanley cites data indicating that Google executives recently stated the company may need to double its hash rate every 6 months, equating to a 1,000-fold increase within 5 years. From the supply side, NVIDIA’s compound annual growth rate (CAGR) for AI chip sales from 2025 to 2028 is projected to be around 140%. Even if this growth rate is extended over five years, the total amount of hash rate delivered would cover less than 10% of Google’s single-company demand forecast. This stark disparity illustrates that even the world’s largest hash rate providers, operating at their highest historical growth rates, could only meet a fraction of the demand from one major tech company, highlighting a profound structural deficit in available computing capacity.

Physical constraints are identified as the primary bottleneck preventing hash rate infrastructure from being delivered as needed, collectively referred to by Morgan Stanley as the 3P framework: People, Power, and Politics. Regarding People, there is a structural shortage of skilled workers required for data center construction, including electricians, welders, and plumbers. In terms of Power, the waiting time to get connected to the grid in some areas has stretched to 5–7 years, becoming the biggest single time bottleneck for data center operations. Politically, data center construction faces increasing resistance at multiple levels, with a structural reversal in trends. States that previously competed to offer generous incentives are now suspending, conditioning, or revoking tax incentives, with these issues becoming part of gubernatorial campaign platforms and expected to be major voter concerns in the November elections.

The regulatory landscape is further complicated by federal efforts to establish a national "data center tariff." The House of Representatives is reviewing the Ratepayer Protection Act, the first federal attempt to legislate the sharing of infrastructure construction costs. This act requires state utility companies to consider creating "heavy load standards" that force data centers to cover the costs of grid upgrades. In March, major tech firms including Amazon, Google, Meta Platforms, Microsoft, Oracle, and xAI signed the White House Ratepayer Protection Pledge, voluntarily committing to protect existing consumers from the impacts of data center construction costs. The proposed legislation would legalize these voluntary commitments, effectively creating a nationwide data center electricity surcharge system, thereby adding another layer of complexity to the political and regulatory environment surrounding AI infrastructure development.

Morgan Stanley acknowledges the validity of concerns regarding physical constraints but classifies them as 'speed bumps' rather than structural barriers, emphasizing the critical importance of "Time to Power." Quantitative analysis reveals that U.S. data centers will require approximately 68 GW of electricity between 2026 and 2028. After accounting for existing facilities (15 GW) and contracted grid capacity (15 GW), there is a potential gap of 38 GW. With grid connection waiting times reaching 5–7 years in some areas, the time value of power access represents the area with the greatest pricing distortion in the current market. Data center operations are constrained by power supply, and alternative solutions that can provide power within 1–3 years offer significant time arbitrage value compared to the lengthy grid connection process.

To address this bottleneck, Morgan Stanley identifies two main paths: repurposing Bitcoin mining farms and deploying rapid power generation solutions. Bitcoin mining farms already possess considerable grid connection capacity and physical land that can be directly converted into data center use, amounting to 10–19 GW. Rapid deployment power generation solutions, such as natural gas turbines (15–20 GW) and fuel cells (5–8 GW), can provide a 1–3 year advantage over grid connection. Even when including probability-weighted calculations for various "Time-to-Power" solutions, including direct nuclear power supply, there remains a net gap of about 1 GW under a moderate scenario, which expands to 11 GW in a pessimistic scenario. This analysis underscores that Hyperscalers lack not just Capex, but physical space with accessible power, known as a "Powered Shell."

The valuation of Powered Shell providers appears significantly undervalued relative to their strategic importance.

Woofun AI reports that the current EV/Watt value for these providers ranges only from $2 to $4. This includes companies such as TeraWulf, Cipher Mining, HUT 8, Riot Platforms, Applied Digital, and Galaxy Digital. In contrast, established data center operators like Equinix and Digital Realty trade at a benchmark range of 20–25x EV/Watt. Based on this comparison, Morgan Stanley assigns a discounted target valuation of 15x EV/Watt to these transitioning companies. This valuation gap suggests that the market has not fully priced the critical role of Powered Shell providers in overcoming physical constraints, presenting a potential opportunity for investors as the AI infrastructure build-out accelerates.