Morgan Stanley 120-Page Research Report Deep Dive: How to View the AI Market Correction?

Original Title: "How to View the AI Market's Major Pullback, Deep Dive into Morgan Stanley's 120-Page Research Report"

Since late June, the global AI sector has experienced a significant pullback.

Whether this is a temporary technical correction or a signal of a cyclical peak is a question worth pondering.

On July 27, Morgan Stanley released a 120-page in-depth report exploring this issue, titled "Playing the AI Infrastructure Dip."

The report provides a clear assessment: the main drivers of the pullback are technical factors, including position crowding digestion, margin financing deleveraging, and momentum factor reversal, rather than a disruption of fundamental logic.

I have carefully read this report, gained insights, and will now interpret the main content for you.

Capital Market Concern 1: Tokenmaxxing Narrative Reversal?

Many companies have begun imposing budget limits on employees' AI Token usage, leading to doubts about the sustainability of revenue growth for large-scale model companies.

This concern is unfounded.

First, the current base of employee Token spending is extremely low, yet the ROI is very high.

Morgan Stanley's research on a large number of enterprise-level AI use cases shows that the average cost savings per AI invocation is around $55 in labor costs, and the average cost of completing an enterprise task through Agent collaboration is only $2-5, with an ROI exceeding 10 times.

For tools with an ROI multiple exceeding 10x, their adoption is not a budget decision but a core competency issue. Enterprises that do not actively deploy AI capabilities will face increasingly significant competitive disadvantages.

Second, GPU generational evolution will boost data center profit margins, meaning that Tokens can significantly decrease in price in the future without impairing profitability, further unleashing demand.

The Morgan Stanley Intelligence Factory model estimates that Blackwell-based data center Token sales have a net profit margin of around 58%.

After the Rubin and Feynman intergenerational GPU deployment, profit margins increased to around 80% and 90%, respectively.

This means that the Hyperscalers can lower the Token price by approximately 75% while keeping the profit margin unchanged, leaving ample room for the cost curve to shift downwards, achieving both price reduction and profitability.

(2) Capital Market Concern 2: Has Kimi Moment Falsified Capital Expenditure Rationality?

The release of Kimi K3 has prompted the capital markets to reexamine a core assumption: If China can train frontier models at a lower cost with similar performance, will the annual trillion-dollar AI Capex of American Hyperscalers face a reduced return on investment?

Goldman Sachs believes that the extreme pursuit of efficiency by large AI companies in China and the U.S. will not weaken the demand for computing power. On the contrary, it strengthens the structural judgment of "demand far exceeding supply."

In the 19th century, economist William Stanley Jevons observed that the improvement of the Watt steam engine significantly increased coal combustion efficiency. However, Britain's total coal consumption skyrocketed because the steam engine, now economically viable, was deployed in far more factories and mines than before.

He thus proposed the famous Jevons Paradox: when the efficiency of using a resource increases, the total consumption of that resource does not decrease but instead increases.

The Jevons Paradox similarly applies to the current AI revolution: improved computing power efficiency reduces the unit Token cost, and lower costs mean more scenarios, more users, and higher-frequency calls, ultimately driving up total computational consumption.

Goldman Sachs cited a set of data to quantify the severity of this supply-demand imbalance:

Google executives recently stated that the company may need to double computing power every 6 months, achieving a 1000-fold increase within 5 years.

However, from the supply side, NVIDIA's AI chip sales CAGR from 2025 to 2028 is around 140%. Even if this rate is extrapolated over 5 years, the cumulative delivered computing power would be less than 10% of Google's single-company demand forecast.

In other words, even if the world's largest computing power provider is operating at its historical maximum growth rate, it can only cover a fraction of a single customer's needs.

(3) Capital Market Concern 3: Does Supply-Side Constraint Constitute a Hard Cap?

The third concern of the capital market is: even if demand-side certainty is sufficient, will real-world constraints prevent computing power infrastructure from being delivered on demand?

Morgan Stanley refers to the real-world constraints as the 3P: People, Power, Politics.

People: Skilled labor types required for data center construction (electricians, welders, pipefitters) are in a structurally short supply.

Power: Grid interconnection queue times have extended to 5-7 years in some areas, becoming the largest single time bottleneck for data center commissioning.

Politics: Data center construction is facing multi-level political resistance from local to federal levels, and the political winds are undergoing a structural reversal.

Over the past few years, states have competed to offer generous incentive policies to attract data center establishments, but the trend has now reversed, with states pausing, adding conditions, or outright revoking data center tax benefits.

Issues such as slowing down data center development, protecting residents from rising energy bills due to data center infrastructure costs, etc., are increasingly becoming part of governors' election platforms and are expected to be key voter issues in the November elections.

At the same time, the federal government is attempting to establish a nationwide "data center tariff."

The House is reviewing the Ratepayer Protection Act, which is the first federal attempt to legislate cost sharing for infrastructure development, requiring state utility companies to consider creating a "large load standard" that would have data centers pay for grid upgrades.

Previously (in March), Amazon, Google, Meta, Microsoft, Oracle, xAI, and others have signed the White House's Ratepayer Protection Pledge, voluntarily committing to protect existing consumers from the impact of data center infrastructure costs.

The bill will formalize this voluntary commitment, essentially creating a national data center electricity surcharge system.

Goldman Sachs acknowledges the validity of this concern but characterizes it as "speed bumps" rather than structural barriers.

With grid interconnection queue times exceeding 5 years in some areas and data centers facing increasing pressure for "self-generation," on-site self-generation is becoming a core solution.

（4）Time to Power: The Underestimated Electricity Time Arbitrage

Goldman Sachs conducted a quantitative assessment of the power gap in U.S. data centers.

The conclusion is that the U.S. data center power demand is approximately 68GW in 2026-2028, with a potential gap of 38GW after deducting facilities under construction (15GW) and contracted grid capacity (15GW), and grid interconnection queue times in some areas already reaching 5-7 years.

In this context, Goldman Sachs believes that the time value of power access (Time to Power) is the direction with the greatest current market pricing deviation.

The underlying logic is very clear: data center deployment is constrained by power supply → 5-7 year grid interconnection queue → alternative solutions capable of providing power within 1-3 years have significant time arbitrage value.

Goldman Sachs identified two core "unbottlenecking" paths:

· Bitcoin Mining Farms: These companies already have significant grid interconnection capacity and physical land that can be directly converted for data center use, totaling 10-19GW.

· Rapid Deployment Supportive Generation Solutions: Compared to grid interconnection, they can provide a 1-3 year time advantage, with natural gas turbines contributing 15-20GW and fuel cells contributing 5-8GW.

Even after probability-weighted calculations incorporating "Time-to-Power" solutions like natural gas turbines, fuel cells, direct supply from nuclear power plants, and converting Bitcoin mining farms, there is still a net gap of about 1GW in the base case scenario, expanding to 11GW in the pessimistic scenario.

In other words, the current most scarce resource for Hyperscalers is not Capex but physically available power, i.e., "Powered Shell."

A research by Goldman Sachs suggests that the current market has not fully priced in such Powered Shell Provider targets.

To quantify this undervaluation, Goldman Sachs compared it to a traditional renewable energy PPA, as detailed in the table below:

The EV/Watt for these Powered Shell Providers is currently in the range of $2-4, including companies such as TeraWulf, Cipher Mining, HUT 8, Riot Platforms, Applied Digital, Galaxy Digital, among others.

Based on a reference of mature data center operators (such as Equinix, Digital Realty) at 20-25x EV/Watt, Goldman Sachs has assigned a discounted target valuation of 15x EV/Watt to these transitioning companies.

Original Article Link