The market has finally realized that AI disruption is no longer a distant threat. On February 14th, according to information from trading desks, Morgan Stanley stated in its latest research report that as AI models advance non-linearly and with acceleration, the market's pricing of disruption risks is beginning to exhibit a domino effect.

Just one month ago, the market believed approximately 4% of the MSCI Europe Index weighting faced AI disruption risk; one week ago, this proportion rose to 7%; and by February 13th, this figure had jumped to 24% (including the banking sector).

The report pointed out that Morgan Stanley believes as cutting-edge AI model capabilities breach a critical threshold—with GPT-5.2 reaching or surpassing human expert levels on 71% of professional tasks—investors must re-examine their asset allocation logic. Morgan Stanley has shifted its stance from neutral to a cautious view on cyclical stocks relative to defensive stocks, and noted that European credit markets offer inexpensive downside hedging opportunities. It highlighted utilities, semiconductors, defense, and tobacco as among the most resilient safe havens.

The bank emphasized the need to reassess which assets cannot be "replicated" by AI—these will become the value anchors in the new era. In an age where intelligence and labor can be infinitely replicated, true value will revert to things that cannot be copied—physical assets, regulatory barriers, network effects, human experiences, and proprietary data.

The astonishing leap in AI capability: 71% of professional tasks have been conquered. Humans are not adept at understanding non-linear change, and the progress of AI models is a classic example of non-linear acceleration. Morgan Stanley stated that data shows a remarkable pace of advancement: Grok 4, launched in July 2025, scored 24% on the GDPVal test, meaning the model achieved human expert level on 24% of real-world professional tasks; just five months later, GPT-5.2, released on December 12, 2025, saw its score soar to 71%.

What is GDPVal? It is a metric that measures AI model performance on real-world knowledge work, encompassing actual tasks performed by experienced professionals across various industries. OpenAI's research found that cutting-edge models complete these tasks approximately 100 times faster and at about 1/100th the cost of industry experts. The report stressed that even more震撼ing breakthroughs are imminent. If the scaling laws for Large Language Model (LLM) training continue to hold in 2026—which Morgan Stanley believes is likely—multiple US frontier LLMs are expected to launch in the first half of 2026, with capabilities far exceeding current models. The reason is simple: the five major US LLM developers are currently training their next-generation models using roughly 10 times the compute power of current models.

The domino effect of disruption risk: from software to banking. The speed of change in market perception is equally astounding. Morgan Stanley's tracking shows that the market initially began questioning whether revenue growth in the software industry could slow sharply in the coming years, but soon this concern spread like dominoes to broader economic disruption risks—changes in competitive landscapes, employment impacts, deflationary pressures, etc. This recalls the market psychology evolution during the early COVID-19 pandemic in 2020: in January, it was just demand and supply chain risks; by February, it expanded to travel & leisure, industrials, banking, and other sectors; by March, it evolved into a full-scale market sell-off, ultimately triggering major policy actions. Currently, Morgan Stanley estimates that approximately 10% of the MSCI Europe Index weighting (excluding banks) is perceived by the market as facing substantive AI disruption concerns; including banks, this reaches 24%. Concerns regarding the banking sector are relatively new, focusing mainly on broader economic deflation and employment issues, and (to a lesser extent) AI-related competition for deposits.

Notably, these "disruption stocks under market debate" have seen their peak price-to-earnings ratio fall from 24 times at the start of 2025 to 16.4 times today. But Morgan Stanley warns, referencing the valuation trajectory of "undisputed disruption stocks" (falling from 24.7x to 11.1x), that there may be further downside for valuations.

Who can survive the AI era? Facing this disruption storm, Morgan Stanley provides an evaluation framework, combining five dimensions to judge the resilience of sectors and individual stocks:

1. AI Exposure: Is the entity being disrupted, a "debated disruption target," an enabler, or protected? 2. Business Nature: Does it provide services, physical assets, commodities, or compute power? 3. Cyclicality: Is it a cyclical stock, defensive stock, or other? 4. Investor Positioning: Current positioning levels. 5. Individual Stock Momentum: Overlaying fundamental factors.

Based on this framework, Morgan Stanley considers the most resilient sectors to be, in order: utilities, semiconductors, defense, tobacco, and personal & household goods. Morgan Stanley stated that European utility companies almost entirely occupy the top 20 spots on the most disruption-resistant list. Common characteristics of these companies include: providing physical infrastructure that AI cannot replicate, belonging to defensive industries, and being relatively underweight in the current environment. Conversely, service-intensive sectors like software, commercial services, media & entertainment, travel & leisure, as well as sectors like transportation, diversified financials, and banking, are considered to face the highest pressure from the diffusion of disruption risk.

Eight asset classes that cannot be replicated by AI. Simultaneously, Morgan Stanley emphasized that once AI reaches transformative levels, the value of asset classes that cannot be "replicated" by AI will rise. This is a key framework for understanding future asset allocation:

A. Physical Scarcity: Real estate, energy and power assets, transportation infrastructure, data centers, minerals & metals, water resources, casino licenses in limited jurisdictions, theme park land, cruise port and terminal rights, spectrum licenses, fiber optic cable networks, etc. B. AI Adopters with Pricing Power: The bar for demonstrating pricing power is rising. C. Unique Luxury Goods, Properties, and Services. D. Network Effects: Large tech platforms, online marketplaces, healthcare businesses with patient relationships. E. Genuinely Unique Human Experiences: Media businesses with strong brands, sports assets/teams, music and other performances that value the human element. F. Regulatory Scarcity: Businesses with various licenses, approvals, and protected franchises. G. Proprietary Data and Brands: AI adopters with proprietary datasets and IP libraries. H. A Range of Semiconductor Assets: Such as leading-edge process nodes, ASML's EUV lithography, TSMC's manufacturing expertise, rare earth processing for chips.

Credit markets: inexpensive downside protection. Although AI disruption concerns have begun affecting parts of the credit market, particularly the leveraged loan space, European investment-grade spreads remain near post-global financial crisis lows. Even as equity implied volatility has been rising, credit volatility remains unusually subdued. However, if AI disruption concerns spread to more sectors (coupled with anticipated acceleration in issuance), it could begin to challenge the resilience of credit markets. Morgan Stanley believes the credit options market offers investors a good entry point to position for spread widening. Given Europe's relatively low tech exposure, overall yields still at high levels, policy support, and economic growth resilience, these hedging tools offer particularly attractive value.

Compute power demand gap: an invisible supply crisis. On the flip side of AI disruption is the insatiable demand for compute infrastructure. Multiple data points indicate that the growth rate of compute demand far exceeds current supply forecasts:

A Google executive recently stated the company may need to double its compute power every six months, "reaching 1000x in 4-5 years." For comparison, Morgan Stanley forecasts a compound annual growth rate of about 210% for Nvidia's compute sales from 2025-2028; extrapolating over 5 years, cumulative compute growth would be approximately 300x—significantly below Google's projected need of over 1000x.

OpenRouter data shows that from late November 2024 to late November 2025, average weekly token demand grew over 2200%. Token usage is a direct proxy for compute demand.

More critically, the computational intensity per individual LLM query is rising rapidly. Research firm METR points out that the average "work" duration performed by AI per customer query is doubling approximately every 7 months.

According to the report, even with a constant number of clients, this growth implies that compute demand will increase at a rate significantly higher than Nvidia's projected ~120% CAGR. Morgan Stanley stated this supply-demand imbalance is already visible in the market:

CoreWeave was able to renew leases for older-generation Nvidia GPUs (Hopper) at around 95% of their original price, far above the price implied by economic depreciation of chips over time; The "powered shell" leasing deal guaranteed by Google for Anthropic and FluidStack delivered an unlevered capital return of approximately 18.5% for bitcoin miner Hut 8, equivalent to a power access premium of about 300%.