Original author: Fei Binjie
Source: Alpha Engineer
Since late June, the global AI sector has experienced a significant correction.
Is this just a temporary technical pullback, or a signal that the cycle has peaked? This is a question worth a fortune.
Morgan Stanley released a 120-page in-depth report on July 27 exploring this issue, titled Playing the AI Infrastructure Dip.
The report provides a clear judgment: the primary drivers of the pullback are technical factors, including the unwinding of crowded positions, deleveraging of margin financing, and momentum factor reversal, rather than a breakdown in fundamental logic.
I carefully read this report and gained valuable insights—let me break down the key points for you.
(1) Capital market concern one: Reversal of the Tokenmaxxing narrative?
The first negative signal recently attracting market attention is the reversal of the "Tokenmaxxing" narrative.
Many companies have begun imposing budget limits on employees' AI token usage, raising questions about the sustainability of revenue growth for large model companies.
This concern is unfounded.
First, the current baseline of token expenditures by corporate employees is extremely low, and the ROI is extremely high.
Morgan Stanley's research on a wide range of enterprise AI applications shows that an average AI invocation can save approximately $55 in labor costs, while the average cost of completing an enterprise task through agent collaboration is only $2–$5, yielding an ROI of over 10x.
Tools with an ROI multiple of 10x or higher are not a matter of budget allocation—they are a matter of core competitiveness. Companies that do not actively deploy AI capabilities will face increasingly significant competitive disadvantages.
Second, the generational advancement of GPUs will increase data center profit margins, meaning that future tokens can be significantly reduced in price without compromising profitability, further unlocking demand.
According to the Morgan Stanley Intelligence Factory model, the net profit margin for data center tokens based on Blackwell is approximately 58%.
After the Rubin and Feynman generation GPUs were deployed, profit margins rose to approximately 80% and 90%, respectively.
This means that a Hyperscaler can reduce Token pricing by approximately 75% without affecting profit margins, as the downward shift in the cost curve is substantial enough to achieve both lower prices and sustained profitability.
(2) Capital market concern two: Does the Kimi moment disprove the rationale for capital expenditures?
The release of Kimi K3 has prompted capital markets to reevaluate a core assumption: if China can train frontier models with comparable performance at a lower cost, could the annual AI Capex of over $1 trillion by U.S. hyperscalers face downward pressure on returns?
Morgan Stanley believes that the extreme pursuit of efficiency by Chinese and U.S. large model companies will not weaken demand for computing power; on the contrary, it reinforces the structural view that demand far exceeds supply.
In the 19th century, economist William Stanley Jevons observed that while Watt’s improvements to the steam engine greatly increased coal combustion efficiency, total coal consumption in Britain surged instead, because the steam engine became economically viable and was deployed in far more factories and mines than before.
Thus, he proposed the famous Jevons Paradox: when the efficiency of using a resource improves, the total consumption of that resource increases rather than decreases.
Jevons Paradox also applies to the current AI revolution: increased computational efficiency reduces the cost per token, and lower costs lead to more use cases, more users, and more frequent requests, ultimately increasing total computational consumption.
Morgan Stanley 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 its computing power every six months, achieving a 1,000-fold increase over five years.
However, from the supply side, NVIDIA's AI chip sales CAGR from 2025 to 2028 is approximately 140%; even projecting this rate forward for five years, the cumulative computing power delivered would still be less than 10% of Google's single-company demand forecast.
In other words, even if the world's largest mining power supplier were operating at its highest historical growth rate, it would still only cover a tiny fraction of a single client's demand.
(3) Capital market concern three: Do supply-side constraints constitute a hard ceiling?
The third concern in the capital market is that even with sufficient demand-side certainty, could physical-world constraints prevent compute infrastructure from being delivered as needed?
Morgan Stanley refers to the constraints of the physical world collectively as the 3P: People, Power, Politics.
People: Skilled trades required for data center construction (electricians, welders, plumbers) are experiencing structural shortages.
Power: The grid connection waiting period has been extended to 5-7 years in some regions, becoming the single largest time bottleneck for data center commissioning.
Politics: The construction of data centers is facing multi-level political resistance from local to federal levels, and the tide is undergoing a structural reversal.
Over the past few years, states competed to offer generous incentives to attract data centers, but now the trend has reversed, with states beginning to pause, impose conditions on, or directly revoke tax incentives for data centers.
Issues such as slowing the growth of data centers and protecting residents' electricity bills from being affected by data center infrastructure costs are increasingly becoming part of gubernatorial campaign platforms and are expected to be key voter issues in the November election.
Meanwhile, at the federal level, efforts are underway to establish a national "data center tariff."
The House is considering the Ratepayer Protection Act, the first federal attempt to legislate cost-sharing for infrastructure development, requiring state utilities to consider establishing "large load standards" that would require data centers to pay for grid upgrades.
Previously (in March), Amazon, Google, Meta, Microsoft, Oracle, xAI, and others signed the White House's Ratepayer Protection Pledge, voluntarily committing to protect existing consumers from the costs of data center infrastructure.
The bill will legalize this voluntary commitment, effectively establishing a nationwide data center electricity surcharge system.
Morgan Stanley acknowledges the validity of this concern but characterizes it as "speed bumps" rather than structural barriers.
As grid interconnection wait times exceed five years in some regions and data centers face increasing pressure to "self-power," on-site power generation is becoming a core solution.
(4) Time to Power: The Underestimated Arbitrage of Electricity Timing
Morgan Stanley quantitatively assessed the power shortfall for data centers in the United States.
The conclusion is that U.S. data center electricity demand between 2026 and 2028 is approximately 68 GW; after deducting facilities under construction (15 GW) and grid capacity already contracted (15 GW), a potential shortfall of 38 GW emerges, with grid connection wait times reaching 5 to 7 years in some regions.
Under this context, Morgan Stanley believes that the time value of power access is the area with the largest current market pricing discrepancy.
The underlying logic is very clear: data center deployment is constrained by power supply → grid connection waiting time of 5–7 years → alternative solutions capable of providing power within 1–3 years offer significant time arbitrage value.
Morgan Stanley identifies two core "de-bottlenecking" pathways:
- Bitcoin mining farms: These companies already possess significant grid connection capacity and physical land that can be directly converted for data center use, totaling 10-19 GW.
- Power generation solutions with rapid deployment: Can provide a 1-3 year time advantage over grid connection, with gas turbines contributing 15-20 GW and fuel cells contributing 5-8 GW.
Even when factoring in "Time-to-Power" solutions such as natural gas turbines, fuel cells, direct nuclear plant supply, and Bitcoin mine conversions into the probability-weighted calculation, there remains a net gap of approximately 1 GW under the base case, expanding to 11 GW under the bear case.
In other words, what hyperscalers currently lack most is not capital expenditure, but physical space with available power—the “Powered Shell.”
Morgan Stanley research believes the current market has not fully priced these Powered Shell Provider assets.
To quantify this undervaluation, Morgan Stanley compared it against traditional renewable energy PPAs, as shown in the table below:
Currently, the EV/Watt for these Powered Shell Providers ranges only between $2 and $4, including companies such as TeraWulf, Cipher Mining, HUT 8, Riot Platforms, Applied Digital, and Galaxy Digital.
Based on a reference of 20-25x EV/Watt for established data center operators such as Equinix and Digital Realty, Morgan Stanley assigns a discounted target valuation of 15x EV/Watt to these transitioning companies.
