The three levels of transformation in AI storytelling

Since 2026, the global AI narrative has undergone a significant marginal shift, with at least three levels of narrative change.

First Level: Divergence in Scaling Laws.

In recent years, the core driver of AI investment has been the empirical rule of Scaling Laws: larger models, more data, stronger compute, and better performance. However, cracks are emerging in this pattern:

First, physical constraints such as power supply and transformer components.

Second, data bottlenecks, as publicly available high-quality text data for pretraining are depleting.

Third, diminishing marginal returns on investment. While the trend of Scaling Laws continues, and increasing investment still makes sense, the marginal gains in model capability per unit of input may be decreasing.

Therefore, beyond building compute infrastructure, expanding algorithms is becoming another major focus, such as shifting toward reasoning expansion (Test-Time Compute) (e.g., Chain-of-Thought reasoning, inference-time scaling), post-training, architectural efficiencies (like Linear Attention, State Space Models), and edge intelligence (such as SLMs).

Second Level: From “Reward” CAPEX to Return on Investment Anxiety.

According to recent guidelines, major US tech companies have announced that their AI-related capital expenditures will exceed $700 billion by 2026. However, the market has shifted from rewarding “capital expenditure” to worrying about “slow monetization.” Interpreting this scale of investment through two reference frames:

(1) Historical reference: In 2025, US tech firms’ capital expenditure as a percentage of GDP reached about 1.9%, and in 2026 it is expected to rise above 2%, nearly equivalent to the total of major infrastructure projects in the 20th century: early 2000s nationwide broadband expansion (~1.2% of GDP), 1949’s power grid expansion, Apollo moon landing, and 1960s interstate highway system (~0.6% each). Currently, AI infrastructure investment intensity in the US is at an extremely high level in economic history.

(2) Corporate cash flow perspective, which has recently become a primary concern: According to estimates, the five largest US cloud providers will allocate about 90% of their operating cash flow to capital expenditures in 2026 (up from 65% in 2025). Some companies even expect capex to exceed operating cash flow, potentially turning free cash flow negative in 2026, making earnings warnings more realistic. Meanwhile, debt issuance is rapidly increasing, with market expectations that US tech giants may issue up to $400 billion in bonds in 2026, including 100-year bonds, which also draws market attention.

Third Level: Deeper Concerns About AI Disruption, Impacting Multiple Industries Recently.

This evolution of the narrative follows a clear progression: from transforming search and information access, to revolutionizing software applications and business processes, and finally to macro paradigm shifts closely tied to AI development stages.

First is the Chat era, transforming search and information retrieval. From the advent of ChatGPT to early 2025, AI mainly existed as conversational assistants—answering questions, generating text, aiding search. The impact was relatively mild, not directly replacing specific business software or jobs. Market narratives focused on “who can train the best models” and “who provides underlying compute.”

Second is the Agent era, transforming software applications and business workflows. In early February, Anthropic launched ClaudeCowork, marking AI’s shift from “generative responses” to “autonomous execution of cross-functional workflows,” triggering a sharp sell-off in software stocks (“SaaSpocalypse,” fears of SaaS demise), and spreading to finance, alternative asset management, legal services, commercial real estate, and transportation.

Third is the era of comprehensive AI, a forward-looking extrapolation. Substack article “THE 2028 GLOBAL INTELLIGENCE CRISIS” offers little novelty but is well-written and sharp, discussing “Ghost GDP” and white-collar replacement, sparking debates on macro paradigm shifts. When AI replaces not just specific industries or assists labor but directly replaces the labor factor itself, traditional macroeconomic models may face disruptive challenges.

The conventional paradigm is a closed loop of “production → distribution → consumption → reproduction,” where “people” are both producers and consumers—sources of supply-side factors and demand. This creates a five-sector economic cycle. But if AI directly replaces labor in the comprehensive AI era, several outcomes may occur:

① From the factor perspective, the importance of labor diminishes further, while models, data, and compute (essentially capital factors) become more critical.
② From the supply side, AI fundamentally alters the supply curve, significantly lowering marginal costs, rapidly increasing supply elasticity, and pushing economies of scale to the limit.
③ From the demand side, alienation of labor income may impact income and demand structures, potentially distorting traditional supply-demand and investment-savings relationships. Economic cycles, distribution mechanisms, and macroeconomic paradigms could face upheaval, leading to further restructuring of financial systems and social contracts.

These marginal shifts in AI narratives mean markets are no longer just “buying stories.” On one hand, there’s concern that AI won’t deliver (slow monetization); on the other, fear that AI will be disruptive (transformative). How to understand this seemingly contradictory mindset?

Logically, the issues pointed to by these three levels of narrative are real and can be logically extrapolated, raising serious questions. But the more critical issue is the timing and ultimate boundaries of these changes, which are very difficult to predict in advance. Currently, markets are linearly extrapolating based on panic, pricing in the worst-case scenarios.

One key reason may also be overestimation of valuations and the fragility of trading structures, amplifying panic. Before this correction, AI-related sectors were at historically high valuations, and even the software sector was not cheap. Under the trigger of narratives, valuations released collectively.

Conversely, many affected companies still show resilience in fundamentals. Leading software firms’ latest earnings reports demonstrate steady revenue growth and improved margins. Some have deep customer integrations, high switching costs, and data and compliance barriers. If AI can be internalized as an added value, they might even benefit.

Regarding macro paradigm shifts, there are counterarguments. First, the “Jevons Paradox” suggests efficiency gains often lead to demand explosions rather than mere substitution; even with large productivity improvements, the resulting “deflationary dividends” (lower prices) could stimulate new demand and industries. Second, AI may create new jobs currently unimaginable, and society’s adaptability often exceeds model predictions. Third, in tasks involving regulation, physical-world interaction, complex human relationships, and high non-standardization, AI’s substitution costs are higher than market panic assumes, and institutional, legal, and social inertia naturally slow down progress.

Therefore, AI-driven change warrants serious attention, but the process may not be abrupt. The timing, boundaries, and uncertainties of transformation actually present differentiated and structural opportunities. From a dynamic and structural perspective, investors should shift from “buying a basket of AI” to “more refined selection of targets.” Especially after the panic subsides and valuations normalize, the key questions are: which changes are likely to happen, which won’t, which will come first, which later, what are substitutes, what are complements? Differentiation will likely widen further.

We recommend focusing on several screening perspectives:

(1) Hardware layer: identify “strong constraints.” With capital expenditure expectations already aggressive, the marginal gains from hardware are diminishing. The market no longer rewards capex. Focus on supply constraints with the strongest bargaining power, especially segments with slow past capacity expansion, longer future cycles, and few substitutes—such as storage, power grid components, transformers, advanced packaging, fiber optics. Supply bottlenecks imply stronger bargaining power.

(2) Model layer: competition is fiercer. Besides model weights, screening should prioritize:

• Do they have exclusive private data for differentiated training?
• Do they possess low-cost inference infrastructure?
• Do they have the engineering capability to quickly turn model capabilities into closed-loop solutions and applications?
  Look for “model capability + data flywheel + business barriers.” Since mid-2025, the correlation of US big tech stocks has dropped from about 0.8 to 0.2. Recent market focus on Anthropic and ByteDance indicates ongoing differentiation at the model level.

(3) Application layer:
Prioritize targets that can be quickly implemented and have proven AI value conversion—applications that directly quantify cost reduction and efficiency gains (ROI), those that can rapidly integrate into core workflows, and vertical AI-native applications.

For recently volatile SaaS sectors, the market may gradually distinguish between “shallow SaaS vulnerable to AI replacement” and “fundamental data and execution platforms indispensable in the AI era.” Applications that occupy key data nodes or execution links (e.g., security, compliance, data pipelines, transaction settlement) or that can be AI-internalized and empowered may be “mispriced” opportunities.

(4) The differences in AI development routes between China and the US are another important perspective. The macro implications differ:

• China emphasizes “compute efficiency first,” relying more on algorithm optimization, open-source ecosystems, and engineering to improve efficiency. Despite current compute shortages, domestic alternatives may be more advantageous from an investment perspective, with large models also a key focus.
• The US, with a higher service sector share and expensive white-collar labor, faces stronger short-term substitution and deflationary pressures. Long-term, whether AI can revive US manufacturing remains uncertain.
• China, with a large manufacturing base and unique advantages like electricity, views AI mainly as a tool to improve total factor productivity rather than just labor substitution. Structural opportunities lie in scenario richness and transforming the producer services sector. This means Chinese AI investment logic will focus more on “industrial empowerment” and “hardware-software integration.”

Over the past few years, tracking and understanding the AI chain has provided investors with clear “cognitive alpha.” The AI revolution remains undoubtedly the most important theme of the era. But as valuations of related stocks soar and new AI leaders prepare for IPOs, narrative shifts are likely to accelerate, posing greater challenges for investors.

This article is from Huatai Securities.

Risk Warning and Disclaimer

The market carries risks; investments should be cautious. This does not constitute personal investment advice and does not consider individual user objectives, financial situations, or needs. Users should consider whether the opinions, views, or conclusions herein are suitable for their specific circumstances. Investment is at your own risk.