The explosive growth of artificial intelligence is colliding with systemic supply bottlenecks. Morgan Stanley suggests that the current market's optimism regarding this AI revolution may still be severely underestimating its true explosive potential and depth. According to a recent research report from Morgan Stanley, a core judgment is that leading large language models (LLMs) are undergoing a "non-linear capability leap," while compute demand has already shown a trend of systemically surpassing supply. From early January to March 2026, global weekly token usage surged from 6.4 trillion to 22.7 trillion, an increase of approximately 250% in just three months. Some LLM service providers have been forced to impose usage caps on users. Morgan Stanley predicts that the future growth rate of compute demand will be about three times the projected CAGR of NVIDIA's compute supply, indicating that compute shortages will persist long-term and intensify. Energy presents another critical challenge. Morgan Stanley's models forecast that US data centers will face an electricity supply gap of approximately 55 gigawatts between 2025 and 2028. Previously, data center projects worth $18 billion were cancelled outright, with another $46 billion in projects delayed. Even when considering various "rapid power" solutions—such as natural gas turbines, fuel cells, and conversions of Bitcoin mining sites—the net electricity shortfall could still amount to 18% to 30% of the total US data center deployment capacity during that period. The impact of AI on the labor market is already becoming apparent. A Morgan Stanley survey indicates that within the five industries most affected by AI, 11% of positions were eliminated due to AI over the past 12 months, with an additional 12% of vacancies left unfilled after becoming available. New hiring accounted for only 18%, resulting in a net job reduction rate of approximately 4%. The report estimates that 90% of occupations will be affected to some degree by AI automation or augmentation.
**Model Capability "Step-Change": The Situation is More Extreme Than Market Expectations** The Morgan Stanley report identifies the "non-linear leap in frontier model capabilities" as one of the most important thematic drivers for 2026, citing substantial data to support its view that the "situation is far more extreme than market expectations." Recent analysis from third-party agency METR shows that the most advanced large models can now independently complete continuous complex tasks lasting over 15 hours—whereas extrapolation based on existing scaling laws suggested the current capability level should be around 8 hours. Actual capabilities have significantly surpassed the trajectory predicted by theory. Multiple independent data points corroborate this trend: - Ongoing tracking metrics from Artificial Analysis show that AI capabilities are still advancing rapidly. - OpenAI CEO Sam Altman publicly warned at an AI summit in India: "The world is not ready; highly capable models are coming." - Researchers have utilized DNA sequencing and DeepMind's AlphaFold tool to develop a cancer vaccine for a pet dog. - An experiment by New York Times technology readers showed that 54% of readers preferred AI-generated articles over those written by humans. - Frontier LLMs have demonstrated the ability to solve open problems in the field of physics. - Reports suggest that an unreleased model represents a "step-change in capability" in software programming, academic reasoning, and cybersecurity. The report also references a prediction from Leo Aschenbrenner's paper "Situational Awareness": "There is a surprising possibility of achieving AGI (Artificial General Intelligence) by 2027. In the four years from GPT-2 to GPT-4, we leaped from a preschool level to a smart high school student level... If we traverse a similar intelligence span again, where will it take us? Likely to models capable of surpassing PhDs and top experts across all professional domains."
**The Great Compute Supply-Demand Gap: A 250% Token Growth Rate Masks a 3x Demand Differential** If the leap in model capability is the "engine on the demand side," then the severe shortage of compute supply is the "ceiling on the supply side." Morgan Stanley identifies "compute demand systemically outstripping supply" as the core market theme for 2026. The report states the data is highly illustrative: - According to actual tracking data from the OpenRouter platform, global weekly token usage increased from 6.4 trillion in early January 2026 to 22.7 trillion in March, a roughly 250% increase in three months. - The rapid adoption of agentic AI tools has significantly accelerated the explosion in demand. - Several LLM service providers have begun imposing token usage caps on users to manage uncontrollable demand growth. - Morgan Stanley predicts the overall compute demand growth rate will be about three times the projected CAGR of NVIDIA's compute supply. - Three parallel forces are compounding to drive demand: the continuous expansion of AI use cases, the non-linear increase in AI task complexity, and the accelerating breadth of AI adoption. In terms of specific applications, software programming is currently the single largest use case for token consumption among all LLM applications, and this domain is dominated by proprietary (closed-source) models. Morgan Stanley's "Intelligence Factory" model reveals another key dynamic: as chip generations transition from Blackwell to Rubin GPUs, the average token price is expected to fall by over 70%. This rapid decline in AI usage costs will further stimulate explosive growth on the demand side, creating a self-reinforcing positive feedback loop. As a concrete example: a data center of approximately 250 megawatts, utilizing Blackwell GPUs, with a power cost of $100 per MWh, running GPT-4o queries, can yield top-tier large model developers a profit margin of around 60%. Morgan Stanley expects actual compute demand to reach roughly three times previous model forecasts. In this context, any company capable of breaking through compute scaling bottlenecks stands to benefit significantly. This includes not only the chip manufacturing supply chain but also memory, optical networking equipment, and core data center components. Morgan Stanley is highly optimistic about these "Merchants of Compute," believing they will directly benefit from this systemic supply-demand imbalance.
**Energy is AI's Lifeline: The 55-Gigawatt Gap and the Race for "Off-Grid" Solutions** Electricity has become the most critical physical constraint on AI infrastructure expansion. Based on its in-depth "AI Power" analysis model, Morgan Stanley has reached a sobering conclusion. Between 2025 and 2028, US data center developers will face an electricity supply gap of approximately 55 gigawatts. Concurrently, $18 billion worth of data center projects have been cancelled directly due to community opposition and concerns over rising power prices, with another $46 billion in projects delayed. Multiple constraints on data center growth are simultaneously intensifying: competition for grid interconnection resources, shortages of electrical equipment, labor shortages, and local political resistance. To address this gap, Morgan Stanley outlines four categories of "Time to Power" solutions: - Natural Gas Turbines: Could alleviate 15–20 GW of the gap, with a 90% probability of success. - Bloom Energy Fuel Cells: Could alleviate 5–8 GW of the gap, with a 90% probability of success. - Deploying Data Centers at Existing Nuclear Plants: Could alleviate 5–15 GW of the gap, with a 75% probability of success. - Converting Bitcoin Mining Sites to Data Centers: Could alleviate 10–15 GW of the gap, with a 90% probability of success. However, even when combining the probability-weighted contributions of all these solutions, Morgan Stanley's baseline calculation indicates that the net electricity shortfall before 2028 will still equate to 18% to 30% of the total US data center deployment capacity during that period. From a strategic perspective, Meta has begun taking proactive steps—providing funding for Terrapower's commercialization project for sodium-cooled fast reactors and directly investing in Louisiana's power infrastructure. Morgan Stanley views this as a potential strategic signal that AI giants are beginning to systematically secure control over energy infrastructure to ensure the lifeline of compute power.
**Early Employment Impact: AI Adoption Economic Value Exceeds 25% of S&P 500 Pre-Tax Profit** Morgan Stanley's survey data and model calculations reveal the early yet profound impact of AI on the labor market. Within the five industries most significantly impacted by AI (Consumer Retailing, Real Estate Management & Development, Transportation, Medical Equipment & Services, Automobiles & Components), Morgan Stanley's field survey shows: - Over the past 12 months, AI has directly led to the elimination of 11% of positions. - An additional 12% of vacancies were left unfilled after becoming available. - New hiring accounted for 18%, resulting in a net reduction rate of approximately 4%. - Notably, smaller firms showed the weakest new hiring figures, possibly reflecting their more agile and rapid adoption of AI applications. From a macro perspective, Morgan Stanley estimates that 90% of occupations will be affected to some degree by AI automation or augmentation, typically not through the "complete elimination of jobs" but through the "reconfiguration of tasks within roles." Quantifying the economic value, Morgan Stanley's calculated TAM (Total Addressable Market) for AI adoption is also staggering: - The value TAM corresponding to the cost-reduction potential from "AI automation" exceeds 25% of the S&P 500 index's expected adjusted pre-tax profit for 2026. - This "AI automation" cost reduction is equivalent to over 40% of total employee compensation expenses. - The value contribution is nearly evenly split between agentic AI (software layer) and embodied AI (physical layer, represented by robotics). - In terms of industry distribution, the economic potential of AI adoption relative to pre-tax profit is highest in Consumer Retailing, Real Estate Management, Transportation, and Medical Equipment & Services.
**The "Moat" in an AI-Disrupted World: Which Assets Truly Retain Value?** As AI capabilities accelerate, a core question becomes increasingly urgent: in a world where AI can replicate almost everything, what kind of assets possess genuine defensive qualities? The Morgan Stanley report, referencing investor Michael Bloch's framework, proposes a crucial distinction: "AI compresses the time required to *do* things, but it cannot compress the time required for things to *happen* naturally. This distinction is the most important screening criterion in investing today." Based on this, the types of assets possessing a truly defensive moat include five categories: 1. **Proprietary Data that Accumulates Continuously** – Not static datasets, but dynamic data continuously generated through defensible business operations. 2. **Network Effects** – Where each new user makes the product more valuable for other users; the advantage of an already-liquid network becomes more pronounced as AI lowers the barrier to creating competitors. 3. **Regulatory Licenses** – Bank charters take years, FDA approvals take years; regulatory barriers widen, rather than narrow, as AI capabilities increase. 4. **Large-Scale Capital Deployment Capability** – When bottlenecks shift from software to physical infrastructure, the ability to mobilize large-scale capital itself becomes a core advantage of the era. 5. **Physical Infrastructure** – Factories, power plants, data centers... physical laws set an unbreakable lower limit on time, and the lead of first-movers expands with each passing month. The report further lists eight major asset classes likely to appreciate in the era of "transformative AI," encompassing: real estate with physical scarcity (AI infrastructure land, industrial properties), AI application adopters with pricing power, luxury goods and unique services, platforms with network effects, authentic and unique human experiences, regulatory franchises, proprietary data and brands, and key semiconductor assets (advanced node chip fabs, ASML's EUV lithography machines, rare earth processing capacity).
Comments