Top models are experiencing a "mutation in capability" with computing power requirements "systematically exceeding supply." According to Morgan Stanley, "the level of market optimism may not be sufficient."

When AI's explosive growth meets systemic supply bottlenecks, Morgan Stanley says that the current market's optimism about this AI revolution may still severely underestimate its true explosive power and depth. On April 11th, according to Wind Trading Station, Morgan Stanley's latest research report's core conclusion is: leading large language models (LLM) are experiencing "non-linear leaps in capabilities," and the demand for computing power is showing a systemic trend of surpassing supply. From early January to March 2026, the global weekly token usage surged from 6.4 trillion times to 22.7 trillion times, a 250% increase in just three months, and some LLM service providers have been forced to impose usage limits on users. Morgan Stanley predicts that the future growth rate of computing power demand will be about three times that of NVIDIA's predicted supply CAGR, and the shortage of computing power will persist and intensify in the long term. Energy is another "time bomb." Morgan Stanley's models predict that between 2025 and 2028, U.S. data centers will face an estimated 55 gigawatts of power shortage. Previously, $18 billion worth of data center projects have been directly canceled, and another $46 billion worth of projects have been delayed. Even with various "quick power supply" solutions such as natural gas turbines, fuel cells, bitcoin mining site conversions, etc., the net power shortage may still reach 18% to 30% of the total deployment scale of U.S. data centers during the same period. The impact of AI on the labor market is already evident. A survey by Morgan Stanley shows that in the five industries most affected by AI, 11% of positions have been eliminated in the past 12 months, while 12% of vacancies have not been filled, and only 18% of new positions have been recruited, resulting in a net reduction of about 4% in positions. The report estimates that 90% of occupations will be affected to some extent by AI automation or enhancement. The "leap in large model capabilities" is more extreme than expected Morgan Stanley's report lists the "non-linear leap in the capabilities of cutting-edge large models" as one of the most important themes for 2026, and cites a large amount of data to confirm the judgment that the situation is far more extreme than the market expected. The latest analysis by the third-party organization METR shows that the best large models can now independently complete complex tasks for more than 15 hourswhereas according to existing technology scaling laws, the current level should be around 8 hours. The actual capabilities have far exceeded the expected trajectory. Multiple independent data sources confirm this trend: - The continuous tracking index of Artificial Analysis shows that AI capabilities are still rapidly advancing; - OpenAI's CEO Sam Altman publicly warned at an AI summit in India: "The world is not ready yet, highly capable models are coming soon"; - Researchers have developed a cancer vaccine for their pet dog using DNA sequencing and DeepMind's AlphaFold tool; - A reader experiment conducted by New York Times technology editor Kevin Roose showed that 54% of readers prefer articles generated by AI rather than human-written ones; - Cutting-edge LLMs already have the ability to solve open challenges in the field of physics; - There have been reports that an unreleased model represents a "leap in capabilities" in software programming, academic reasoning, and network security. The report also references Leo Aschenbrenner's paper "Situation Awareness," predicting: "Surprising possibility of achieving AGI (Artificial General Intelligence) by 2027. From GPT-2 to GPT-4 in four years, we have leaped from preschool level to smart high school student level...Where would we be if we made the same intelligence leap again? It is very likely that we would have models that surpass doctors and top experts in all professional fields." The gap between supply and demand for computing power: 250% growth in Token usage behind a 3x demand gap If the leap in large model capabilities is the "demand side engine," the severe shortage of computing power supply is the "supply side ceiling." Morgan Stanley lists the "systemic surpassing of computing power demand over supply" as the most core market theme for 2026. The report states that the data is very straightforward: - According to actual tracking data from the OpenRouter platform, from early January to March 2026, global weekly token usage jumped from 6.4 trillion times to 22.7 trillion times, an increase of about 250% in three months; - The popularity of intelligent AI tools (represented by OpenClaw) has significantly accelerated the demand side explosion; - Several LLM service providers have already begun to set token usage limits for users to cope with uncontrollable demand growth; - Morgan Stanley predicts that the overall growth rate of computing power demand will be about three times that of NVIDIA's computing power supply CAGR prediction; - Three major driving forces behind the demand overlap: the continuous expansion of AI usage scenarios, the non-linear increase in AI task complexity, and the accelerated widening of AI adoption. At the application level, software programming is the single largest use case in token consumption among all LLM usage scenarios, and this field is dominated by proprietary (closed-source) models. Morgan Stanley's "Intelligence Factory" model reveals another key logic: as chip generations transition from Blackwell to Rubin GPUs, the average token price is expected to drop by over 70%the rapid decline in the cost of AI usage will further stimulate an explosion on the demand side, creating a positive feedback loop of self-reinforcing demand. As an example, a data center with a capacity of around 250 megawatts, using Blackwell GPUs, with an electricity cost of $100 per megawatt-hour, running GPT-4 queries, could generate a profit margin of about 60% for top large model developers. Morgan Stanley expects that the real computing power demand will be about three times the previously predicted model. Against this backdrop, any company that can break through the bottleneck of computing power expansion will see significant benefits. This includes not only the chip manufacturing supply chain but also core components of storage, optical network equipment, and data centers. Morgan Stanley highly regards these "merchants of compute" and believes that they will directly benefit from this systemic supply-demand imbalance. Energy is the lifeblood of AI: 55 gigawatt gap and the race for "off-grid" solutions Electricity has become the most critical physical constraint for the expansion of AI infrastructure. Based on its in-depth analysis model of "AI power supply," Morgan Stanley has reached alarming conclusions. Between 2025 and 2028, U.S. data center developers will face an estimated power supply gap of about 55 gigawatts. Meanwhile, $18 billion worth of data center projects have been directly canceled due to community opposition and concerns about rising electricity prices, and another $46 billion worth of projects have been delayed. Multiple obstacles constraining data center growth have been fermenting simultaneously: competition for grid access resources, shortage of power equipment, lack of labor, and local political resistance. In the face of this gap, Morgan Stanley has outlined four types of "Time to Power" solutions: - Natural gas turbines: could alleviate a 15-20 gigawatt gap with a 90% success rate; - Bloom Energy fuel cells: could alleviate a 5-8 gigawatt gap with a 90% success rate; - Deploying data centers at existing nuclear power stations: could alleviate a 5-15 gigawatt gap with a 75% success rate; - Repurposing bitcoin mining farms as data centers: could alleviate a 10-15 gigawatt gap with a 90% success rate. However, even when the probabilities of all these solutions are weighted and combined, Morgan Stanley's baseline estimate shows that by 2028, the net power shortage will still be equivalent to 18% to 30% of the total deployment scale of U.S. data centers during the same period. From a strategic perspective, Meta has already begun to take proactive actionproviding funding for a commercialization project with Terrapower's sodium-cooled fast reactor and directly investing in the power infrastructure in Louisiana. Morgan Stanley believes this may be a strategic signal that the tech giants are starting to systematically control energy infrastructure to ensure the lifeblood of computing power. The impact of employment is beginning to show, and the economic value of AI adoption exceeds 25% of S&P 500 pre-tax profit Morgan Stanley's survey data and model calculations reveal the early and profound impact of AI on the labor market. In the five industries most significantly affected by AI (consumer goods retail, real estate management and development, transportation, medical devices and services, automotive and parts), Morgan Stanley's on-site research shows: - In the past 12 months, AI has directly led to the elimination of 11% of positions; - Another 12% of positions that become vacant are not being refilled; - Only 18% of new positions have been recruited, resulting in a net reduction rate of about 4%; - It is worth noting that smaller companies have the weakest recruitment of new positionsthis may reflect the more flexible and rapid application of AI by small businesses. From a macro perspective, Morgan Stanley estimates that 90% of occupations will be affected to some extent by AI automation or enhancement, with the impact not typically being "wholesale job elimination" but rather "restructuring the task structure within positions." From a quantitative economic value perspective, the AI adoption TAM (target addressable market) calculated by Morgan Stanley is also astonishing: - The potential cost reduction brought about by "AI automation" corresponds to a value TAM exceeding 25% of the expected adjusted pre-tax profit of the S&P 500 index in 2026; - This "AI automation" cost reduction amounts to over 40% of total employee compensation expenditure; - The value contribution of intelligent AI (software level) and embodied AI (physical level, represented by Siasun Robot & Automation) is almost evenly distributed; - In terms of industry distribution, the AI adoption economic potential in consumer goods retail, real estate management, transportation, and medical devices fields is relatively higher compared to pre-tax profit. The "moat" disrupted by AI: What assets can truly retain value in the AI era? As AI capabilities accelerate, the core question becomes increasingly urgent: in a world where AI can replicate almost everything, what assets truly have defensive value? Morgan Stanley's report cites investor Michael Bloch's framework and proposes a key distinction: "AI compresses the time required to complete things, but it cannot compress the time required for things to happen naturally. This difference is the most important filtering criterion for current investments." Based on this, asset types that possess truly defensive moats include five categories: - Continuously accumulated proprietary datanot static datasets, but dynamic data generated through defensible business operations; - Network effectseach new user adds value to other users; having a network advantage with accumulated liquidity, as AI reduces barriers to creating competitors, will become more pronounced; - Regulatory licensesbanking licenses take years, FDA approvals take years, regulatory barriers widen with increasing AI capabilities, rather than narrowing; - Capability for large-scale capital deploymentwhen the bottleneck shifts from software to physical infrastructure, the ability to mobilize large amounts of capital becomes a core advantage of the times; - Physical infrastructurefactories, power plants, data centers...physical laws set an unbreakable limit on time, and the lead advantage of early adopters widens each passing month. The report further lists eight categories of assets that may appreciate in the era of "transformational AI," including: real estate with physical scarcity (land for AI infrastructure, industrial real estate), AI application adopters with pricing power, luxury goods and unique services, platforms with network effects, real and unique human experiences, regulatory franchises, proprietary data and brands, and critical semiconductor assets (advanced chip manufacturing plants, ASML's EUV lithography machines, rare earth processing capabilities). This article is reprinted from the "Wall Street See News" app, author: Dong Jing, GMTEight editor: Song Zhiying.