February of this year marked a critical turning point. The leap in AI coding capabilities has propelled the global volume of functional code into a phase of exponential expansion. However, under current physical AI technology constraints, the growth rate of total global physical production value and aggregate income lags far behind the expansion rate of AI-generated code. The world is likely to first experience a period characterized by code inflation, execution overcapacity, intensifying competition, and diminished returns on capital investment. By analyzing industries based on two dimensions—physical dependency and regulatory/emotional barriers—we can categorize them into four groups: vulnerable (low physical dependency, low regulatory/emotional barriers), reshaped (low dependency, high barriers), fortress (high dependency, high barriers), and beneficiaries (high physical dependency, low regulatory/emotional barriers). In the coming period, the performance gap between beneficiaries of physical scarcity and those vulnerable under code inflation is expected to widen, a divergence trend likely to persist. This represents a new factor that must be considered when evaluating market conditions and sector allocation. From a short-term market perspective, the A-share market structure, dominated by manufacturing and finance, is relatively less impacted by this wave of AI disruption compared to U.S. and Hong Kong stocks. The trends of capital inflows and positive market sentiment remain unchanged, suggesting the post-holiday spring rally may continue. Pricing power remains one of the core allocation themes for the first quarter.
The global volume of functional code has formally entered a phase of exponential expansion. If last year's Lunar New Year period, marked by the popularity of DeepSeek, fueled market imagination about AI applications, this year's explosion of Coding Agents has generated significant anxiety regarding the expansion of global code volume and the potential disruption of traditional software applications. On February 5, 2026, OpenAI and Anthropic released new models on the same day. The experience delivered by GPT-5.3 Codex and Claude Opus 4.6 was disruptive, indicating that AI's role, at least in coding, has evolved from an "assistive tool" to an independent executor. Information from OpenAI and Claude's official websites indicates that both GPT‑5.3 Codex and Claude Opus 4.6 are positioned as agentic models capable of planning, debugging, and making multi-step modifications within large codebases, moving beyond the passive assistance of "prompting and code completion." This implies that any workflow describable in language and expressible in code will be replaced by AI at an extremely rapid pace. Concurrently, it signifies that the global volume of effective code is entering a stage of exponential growth.
Under current technological conditions, the expansion of total social physical production value and aggregate income is significantly slower than the rate of code volume expansion. According to data from the IEA and Ember Energy, global electricity generation is projected to increase from approximately 30,000 TWh in 2024 to about 32,000 TWh in 2026, representing a compound annual growth rate of merely 3.3%. In contrast, data center energy consumption is expected to surge from around 600 TWh in 2024 to approximately 1,050 TWh in 2026 (under an optimistic scenario), with a compound annual growth rate as high as 32.3%. In the short term, energy supply growth clearly cannot keep pace with the expansion of code and token consumption; solving energy consumption and latency challenges through engineering becomes more critical than simply stacking computing power. The ratio of the total number of GitHub repositories (in millions) to global GDP (in trillions of U.S. dollars) was only 3.93 in 2023 but reached 5.38 by 2025. Based on further extrapolation of the code repository growth rate disclosed in GitHub's 2025 Octoverse report and IMF GDP forecasts, we estimate this ratio could further increase to 6.29 by 2026. Meanwhile, competition among global large language models has reached a fever pitch. The capabilities of leading-edge models have not diverged significantly despite increased computational investment; instead, the gap between them is narrowing. Competition on the revenue side is expected to intensify further, while marginal costs continue to rise amid widespread hardware price increases. Perhaps at some future point, the maturation of embodied intelligence will lead to rapid advances in resource acquisition and physical production capabilities for society, potentially establishing an effective distribution mechanism to curb the expansion of wealth inequality. However, these developments are difficult to achieve in the short term. Overall purchasing power in the short run evidently cannot keep up with the pace of product expansion based on effective code and the rate of computational cost consumption. We are likely to undergo a societal phase of code inflation, execution overcapacity, intensified competition, and impaired capital returns.
Which industries are vulnerable to code inflation? Which are beneficiaries of physical scarcity? By analyzing industries based on their physical dependency and regulatory/emotional barriers, we can segment them into four quadrants: 1) Vulnerable Zone (Low physical dependency, low regulatory/emotional barriers): AI-generated code and content have high applicability here; business models are relatively easily substitutable and lack exclusive data or know-how as a moat. Typical industries include basic code outsourcing, general SaaS, marketing, and public relations. 2) Reshaped Zone (Low physical dependency, high regulatory/emotional barriers): Business models are primarily digital but involve regulatory requirements like legal liability and financial risk control, or rely on human emotional connections such as psychological counseling and client trust, creating moats difficult to replace with pure code. AI acts as a "super lever" in these industries, with core value in aiding decision-making, improving human efficiency, and driving "workforce reduction and efficiency gains," rather than disruptive replacement. Typical industries include litigation, high-end strategic consulting, and asset management. 3) Fortress Zone (High physical dependency, high regulatory/emotional barriers): Involves monopolistic assets or scarce resources. Typical industries include core mining, military manufacturing, and transportation infrastructure, as well as premium baijiu, top-tier luxury goods, and trendy toys—sectors that must be coupled with physical goods and embody brand or emotional value. 4) Beneficiary Zone (High physical dependency, low regulatory/emotional barriers): Involves physical carriers related to electrification and computing centers, such as copper, aluminum, and energy metals; AI hardware like semiconductor manufacturing, PCBs, optical modules, and servers; and energy infrastructure including power equipment and transformers.
The performance gap between beneficiaries of physical scarcity and those vulnerable under code inflation is widening, and this divergence trend is expected to continue. Based on the aforementioned logic of "physical scarcity" beneficiaries and "code inflation" vulnerability, we have identified corresponding beneficiary and vulnerable portfolios in both Chinese and U.S. stock markets. In the U.S. market since the start of 2026, the cumulative performance gap between these two portfolios has widened by 64 percentage points. In the A-share market, benefiting from ample liquidity and capital inflows, the divergence is not yet pronounced. Compared to the end of 2025, the excess return of the "physical scarcity" beneficiary portfolio over the "code inflation" vulnerable portfolio in the A-share market has increased by only 3 percentage points. This is partly because some software and media stocks in A-shares and Hong Kong saw significant gains in January driven by hype around AI applications, contrasting sharply with the U.S. software and services sector's decline of approximately 20% year-to-date. However, as global market linkages strengthen and liquidity premiums diminish, we believe Chinese assets will ultimately reflect the divergence trend between "physical scarcity" and "code inflation."
The A-share market structure, dominated by manufacturing and finance, is relatively less affected by this round of AI impact compared to U.S. and Hong Kong stocks. Given that China's enterprise-facing software and service market is not particularly large to begin with, the real economy is likely to be much less affected in the early stages of AI-generated code inflation compared to North America. For example, as of February 13, 2026, software and services companies accounted for 22.8% of the U.S. stock market's capitalization. This proportion was 31.5% in the Hong Kong market and only 5.6% in the A-share market. Software and enterprise services represented the "third pillar" of the North American tech sector—besides mobile internet and biopharma—prior to the breakthroughs in generative AI, characterized by stable, mature business models, high barriers, high capital returns, and steady cash flows. However, they face the most significant impact from this wave of AI's disruptive innovation. Hong Kong's internet giants, focused primarily on consumer-facing businesses, are also attempting to restructure workflows with AI but are inevitably affected by global market linkages. In contrast to U.S. and Hong Kong markets, the A-share market is still predominantly composed of monopolistic finance, manufacturing, and energy sectors. In the process of code inflation, traditional resource and manufacturing sectors involved in building AI-era infrastructure may反而受益. These represent truly scarce physical assets and are likely to be important safe havens for global capital in the coming years.
The trends of capital inflows and positive market sentiment remain unchanged; the post-holiday spring rally is expected to continue. The inflow trend of allocation-oriented capital has not shifted. According to People's Bank of China data, household deposits in January increased by 339 billion yuan less year-on-year, while deposits of non-bank financial institutions saw a year-on-year increase of 256 billion yuan for the month. As high-interest deposits mature集中, the decline in January deposit data suggests the ongoing trend of deposit transformation (i.e., deposits shifting into investment products) is continuing. Ultimately, products like wealth management plans and savings-type insurance policies will channel a portion of funds into the equity market. Overall, the landscape of capital inflows and bullish sentiment remains intact. The minor market adjustment before this year's Spring Festival was likely related to substantial gains in January, significant volatility in overseas markets starting in February, and the extended holiday period this year, which heightened the risk-off demand for some capital ahead of the break. As of February 13, the readings for our constructed A-share investor sentiment index for the single day, 5-day moving average (MA5), and 10-day moving average (MA10) were 43.6, 55.7, and 65.5, respectively, showing a noticeable decline compared to January, with the single-day indicator hitting a low since 2025. As of the end of January, the latest average positioning level for actively managed private fund products sampled via CITIC Securities channels was 79.3%, a significant reduction of 5 percentage points from the 84.3% recorded in the week of January 23. These signs likely indicate that risk-off driven capital had already completed its position reduction before the Spring Festival holiday, thereby creating conditions for potential post-holiday capital replenishment.
Pricing power remains one of the core allocation clues for the first quarter. 1) The fundamental framework since our annual strategy has been based on the "re-rating of China's resource and traditional manufacturing pricing power" as the foundation. The core allocation logic focuses on sectors where China holds a clear share advantage, where overseas capacity reset costs are high or difficult, and where supply elasticity is somewhat influenced by domestic policies. Based on this logic, the recommended portfolio includes chemicals, non-ferrous metals, power equipment, and new energy as a base. This is supplemented by increased allocation to undervalued insurance and securities firms (providing some exposure to the low valuation factor) and an increased allocation exposure to the consumption chain (duty-free, aviation, hotels, scenic spots, freshly made tea beverages, etc.) and the property chain (high-quality developers, building materials, REITs, etc.). This is primarily based on the expectation of a market shift from last year's extreme differentiation to a moderate broadening this year, coupled with a mild recovery in domestic demand and prices. Even considering the sharp volatility in precious metals and commodities in early February, the nomination of Kevin Warsh, the rebound in the U.S. dollar index, and significant adjustments in cryptocurrencies and overseas small-cap tech stocks, the logic of this basic allocation framework remains unaffected. Pricing power is the most directly trackable catalyst and trading clue within this framework. 2) Simultaneously, the new allocation framework of "code inflation" and "physical scarcity" must be considered. The rapid global expansion of code volume driven by AI could significantly impact various businesses with low physical dependency and high market competition. Conversely, businesses with high physical dependency and high regulatory/emotional barriers may become scarcer. Under this framework, market investors might actively seek exposure to fortress-type assets currently insulated from AI disruption, while temporarily avoiding businesses highly susceptible to being overturned by AI's disruptive innovation—regardless of whether AI Agents can immediately replace them in the short term. In our current allocation framework, sectors related to resources, traditional manufacturing, energy, service consumption, and the property chain恰恰 fall into the category of fortress-type assets that are temporarily shielded from the impact of generative AI.
Comments