Capital's AI Frontier: The Fundamental Shift in Investment Focus from Human Capital to Leverage

When discussing AI's impact on investment, it's crucial to distinguish between 'paradigm deepening' and 'paradigm revolution.' The former involves optimizing parameters within existing investment frameworks, such as using AI to improve the efficiency of portfolio companies or assigning higher valuation multiples to AI firms. The latter, however, entails a complete rewriting of the standards for 'what constitutes investment and what defines a good return.' This research report uses the following criteria: 'Have the four core links of the investment chain—the investment target's essence (what to invest in), value anchor (what makes it valuable), sources of risk (what to fear), and return structure (how to profit)—undergone a fundamental shift? And have the 'rights acquired by investors' (ownership vs. access rights) changed?

Based on these criteria, we conclude: the watershed moment is not merely the 'old wine in new bottles' notion that 'AI companies deserve higher multiples.' The true revolution lies in the fact that when execution is no longer scarce, revenue becomes completely decoupled from headcount, and the most valuable digital products can be replicated at near-zero marginal cost, the essence of investment shifts from 'a collection of people' to 'leverage and boundaries.' The value anchor moves from traditional discounted future cash flows to 'verifiability, reusability, and convexity.' Simultaneously, the primary sources of risk are no longer execution and management but rather capability depreciation, model drift, and computational power concentration. Investors' pursuit also changes from sharing cash flows to seeking convex returns, and the rights ultimately acquired shift from sole ownership to access, orchestration, and settlement rights over 'boundary layers.

Before the AI wave swept in, the traditional investment paradigm was built upon six 'self-evident' assumptions. Firstly, **the investment target is a 'collection of people,' and human capital is the core and moat of the production function**. Enterprises were viewed as a combination of people, equipment, and organization, with revenue growth and valuation models revolving around 'hiring—training—pricing—replication.' Human capital was both a cost and a core asset that could depreciate, be leveraged, and command a premium, especially in labor-intensive sectors like service outsourcing and IT services, where this assumption formed the cornerstone of M&A and integration strategies. Secondly, **value was believed to be intrinsic to the target, quantifiable through Discounted Cash Flow (DCF) or comparable multiples**, which is a widely accepted valuation framework in finance, with multiples often seen as a shorthand for DCF.

Thirdly, **growth was inseparable from capital, and capital was primarily used to acquire human resources and support operations**, with economies of scale achieved through decreasing marginal costs. Software, with its near-zero replication costs and high gross margins, was considered the ideal investment endgame. Fourthly, **risk was viewed as classifiable and compensable by a risk premium** (as articulated by the Capital Asset Pricing Model, CAPM); diversification could reduce unsystematic risk, and higher risk corresponded to higher expected returns, forming the common framework of portfolio theory.

Fifthly, **returns followed a power-law distribution, with winners capturing value long-term through 'asset ownership,'** and primary exit paths being IPOs or M&A. The VC business model was built on the logic that 'a few big winners cover the majority of failures,' with investors sharing in the cash flows and valuations of successful ventures through equity. Finally, **information asymmetry was primarily addressed through due diligence**, which focused on the team, market, and financials, with 'team background + market size + unit economics' being the industry standard.

These six assumptions collectively wove an investment worldview centered on an 'ontology of human resources': first, there are employable, organized people; then, these people generate predictable cash flows; and finally, these cash flows are discounted by risk. The essence of investment was people, and the assets driven by them.

The rise of AI is shaking the aforementioned traditional assumptions through seven unprecedented mechanisms. Firstly, **execution costs are collapsing rapidly, and humans are being demoted from 'irreplaceable execution units' to 'capacity replaceable by computational power.'** The marginal cost of tasks such as content generation, code writing, customer service responses, draft writing, and data organization has plummeted, breaking the old paradigm's equation of 'more people = more capacity = more cash flow.' Secondly, **revenue is decoupling from headcount, giving rise to investable samples of 'minimal units, extremely high output.'** According to Epoch AI's calculations, as relayed by KuCoin/AIMPACT, Anthropic's annual revenue per employee reached $9 million, and OpenAI's was approximately $5.6 million [Verified · Title/second-hand recount confirmed · awaiting first-hand verification]. Meanwhile, Midjourney, with zero external funding and around 40 employees, is estimated to have achieved an annualized revenue of $500 million [Second-hand estimate · awaiting verification]. These cases collectively demonstrate that 'per capita productivity' has risen to a level where traditional 'per-head valuation' methods are rendered ineffective, requiring new anchors for asset value.

Thirdly, **the value anchor is migrating from 'cash flow/multiples' to 'capability × reusability × verifiability × boundaries.'** As generation and replication costs approach zero, scarcity shifts to verification, attribution, distribution, and trust. The core of investment due diligence is no longer 'how many people, how much revenue,' but rather 'is the unit economics verifiable, how high is the reusability, and are the boundaries enforceable?' For instance, by the end of January 2025, the median forward ARR multiple for AI public software companies was twice that of traditional software companies (11.2x vs. 5.5x), and AI companies accounted for approximately 70% of B2B Series A funding, with valuations averaging 40% higher [Verified · First-hand blog (author is a VC, data self-reported)]. This indicates that the market is paying a premium for AI's 'convexity and boundaries,' rather than merely for human substitution.

Fourthly, **the sources of risk are shifting from 'execution/human labor' to 'capability depreciation/model drift/computational power concentration/validation tax/liability.'** An AI-enhanced asset requires continuous monitoring and re-validation to maintain its value. Model upgrades, data distribution shifts, and API changes can all lead to rapid asset obsolescence. Therefore, investment risk is rewritten as an assessment of capability depreciation rates, drift rates, concentration on single model/compute providers, validation tax, and exposure to liability in case of errors. In the old paradigm, 'people' were relatively stable assets; in the new paradigm, 'models' are rapidly depreciating, replicable assets that may face regulatory scrutiny.

Fifthly, **the return structure is shifting from 'sharing cash flows' to 'sharing convexity.'** The characteristic of non-rivalrous digital products is that successful upside is almost infinite, while the marginal cost of failure is limited. When AI enables 'successful products' to be replicated at near-zero cost, the return function resembles an option rather than a bond, explaining why capital is willing to pay high multiples for high-uncertainty AI targets—it's buying convexity. Market data also shows that capital is betting on AI earlier and more concentratedly, exhibiting a trend of 'fewer bets, more capital' [Verified · Title/second-hand recount confirmed]. Sixthly, **capital intensity is making a comeback, shifting from 'asset-light software' to 'power + chips + data centers.'** Contrary to the intuition that 'AI makes everything light,' AI infrastructure is capital-heavy. TrendForce predicts that by 2026, the capital expenditure of the world's top nine cloud vendors will reach approximately $830 billion, a 79% year-on-year increase [Verified · Text/second-hand recount confirmed]. This implies that the application layer becomes extremely light, while the infrastructure layer becomes extremely heavy, polarizing capital allocation logic.

Finally, **the unit of pricing is migrating from traditional seats and hours to outcomes and credits.** The SaaS model, which charges per seat, is being eroded by intelligent agents; the more capable the agent, the fewer seats a customer needs. Bain's analysis shows that 35% of SaaS vendors bundle AI into higher seat prices, and 65% adopt a 'seat + usage' hybrid model, though none have fully transitioned to outcome-based pricing [Verified · Text/second-hand recount confirmed], despite monday.com having introduced a 'seats + credits' model. This means the basis for revenue measurement is being renegotiated, and investment valuations must ultimately follow these shifts in pricing units. These core mechanisms collectively constitute a revolutionary shift in the investment paradigm, far beyond optimization and deepening under old definitions.

Facing the impact of AI, we must redefine the fundamental elements of investment. Firstly, **the essence of the investment target is redefined as an 'AI-enhanced asset'**: a production unit capable of transforming 'rentable AI capabilities' into 'verifiable outcomes.' It comprises four elements: 'humans (defining goals/verifying results/assuming responsibility), AI capabilities (rentable, replaceable), proprietary flows (data/processes/memory/distribution), and boundaries (exclusivity, verifiability, attributability).' Investment is no longer directed at existing human capital or equipment, but at the leverage effect and boundary strength of this unit. The old definition asked 'how many people, assets, and cash flows does a company have,' while the new definition focuses on 'how many times this unit can amplify AI capabilities, how strong its exclusivity and verification over results can be established, and who bears responsibility.' This asset view directly inherits FR-13's insight: models and computational power are rented, while proprietary flows and trusted boundaries are owned; the core of investment lies in the latter.

Secondly, **the valuation benchmark is rewritten as 'Valuation ≈ [Verifiable Unit Economics × Reusability] + Convexity (Option Value) × Boundary Enforceability × Distribution/Trust Thickness.'** The traditional DCF model degenerates into a 'lower bound check,' no longer the primary anchor; multiples are no longer a valuation method in themselves but rather a market signal for pricing convexity and scarce boundaries. Simultaneously, **the definition of risk also shifts from past execution and market risks to 'capability depreciation rate + drift rate + concentration (model/compute/customer) + validation tax + liability exposure.'** Risk is negatively correlated with the 'control layer thickness' (verification, traceability, compliance, responsibility assumption), rather than positively correlated with asset size.

Furthermore, **the return structure transforms into 'Cash Flow (Lower Bound) + Convexity (Upper Bound) - Verification and Replication Costs.'** This return distribution becomes more power-law-like, potentially realized earlier, and highly dependent on the quality of a few 'boundary nodes.' Consequently, the benefits of portfolio-level diversification decrease, while the importance of single-point control quality rises. Finally, **the rights acquired by investors also shift.** When core capabilities (models, computational power, cloud) are rented, the informational value of 'ownership' depreciates. What investors truly acquire are access rights, orchestration rights, verification rights, and settlement rights over 'boundary layers.' Legal equity remains the vehicle, but it no longer represents the entirety of value. In this new investment paradigm, **the role of capital shifts from 'purchasing human labor and scale' to 'purchasing time and certainty'**: trading compute for iteration speed, data for accuracy, evaluation and compliance for trust, and distribution for compounding returns. The optimal use of investment is to transform 'uncertainty' into 'verifiability.

This revolution in investment paradigms is not without foundation; its evidence is already manifest in capital markets. From the perspective of structural capital reallocation, AI companies are absorbing the vast majority of global VC investment. OECD data shows that in 2025, AI enterprises attracted approximately 61% of global VC investment, totaling $258.7 billion [Verified · Title/second-hand recount confirmed · awaiting first-hand verification]. Furthermore, according to Reuters/PitchBook, AI's share of global VC soared from 28% to 57.9% within a year [Verified · Title/second-hand recount confirmed]. Data from TechCrunch and Carta also indicates that AI accounted for 41% of $128 billion in funding in 2025, showing a concentrated trend of 'fewer bets, more capital' [Verified · Title/second-hand recount confirmed]. This demonstrates that capital is flowing into the AI sector at an unprecedented pace and scale, reflecting a strong pursuit of new value anchors.

However, we must also confront the warnings against pure technological optimism. Firstly, 'old assets' have not disappeared; instead, they demonstrate resilience in new forms. For example, Nvidia became the first company to surpass $5 trillion in market capitalization in October 2025 [Verified · Multi-source cross-reference], indicating that competitive physical capital (chips, manufacturing, energy) remains the largest source of value, refuting the notion that 'AI renders physical assets irrelevant.' Secondly, **AI companies do not automatically possess high gross margins.** Bessemer's 'State of AI 2025' shows that AI companies' gross margins are around 60%, with some even as low as 25%, notably lower than traditional SaaS at 70-85% [Verified · Text/second-hand recount confirmed]. Inference and human review costs constitute a structural drag, directly refuting the popular narrative that 'AI-enhanced assets = higher gross margins.

Furthermore, **most enterprise AI investments have not yielded returns.** The MIT NANDA report indicates that approximately 95% of generative AI enterprise pilots failed to produce measurable P&L impacts [Verified · Text confirmed], and while the methodology of this study has subsequently faced debate, it undoubtedly serves as a directional warning. Concurrently, **valuations are decoupling from revenue, and bubble warnings are emanating from mainstream institutions.** The IMF and the Bank of England publicly warned in October 2025 that AI assets might be overvalued [Verified · Title confirmed]. A senior PE executive bluntly stated that AI startups 'are valued at extremely high multiples relative to their modest revenues.' Additionally, reports of massive deals between Nvidia and OpenAI shrinking expose the circular financing risk where 'suppliers invest in customers, and customers purchase supplier products' [Verified · Title confirmed · full text to be read]. If a portion of AI revenue is generated by AI investments themselves, returns might be systematically overestimated. These counter-examples collectively paint a more complex picture, indicating that the transformation of the investment paradigm is accompanied by new risks and uncertainties, requiring investors to maintain a sober perspective.

The fundamental shift in underlying investment logic is having profound implications for business models, organizational structures, product design, and even the entire capital market. At the **business level**, the logic of valuation and fundraising is transitioning from 'buying growth, buying headcount' to 'buying boundaries, buying convexity.' Key unit economics metrics are no longer confined to ARR or gross margin but expand to include inference costs, validation tax, reusability, cost-per-task, and convexity. This explains why companies with similar revenue scales can achieve vastly different valuation multiples in the AI era: the market is paying for 'verifiable non-replicability' rather than revenue itself. Simultaneously, the unit of pricing is migrating from traditional seats to outcomes or credits, because when intelligent agents can directly deliver results, input-based billing models are structurally eroded by their own success.

At the **organizational level**, AI is accelerating the 'dumbbell' trend in organizational structures. On one hand, the power of super-platforms (possessing computational power, models, and distribution channels) is becoming increasingly concentrated; on the other hand, highly miniaturized, trusted production units (with the ability to define problems, wield tools, verify results, and assume responsibility) are becoming extremely valuable. The intermediate execution layer, particularly human-intensive functions such as BPO, customer service, content generation, and junior analysis, faces repricing and even disintermediation. Concurrently, the 'novice apprenticeship path' is being hollowed out by AI, further reducing the investability of traditional human capital. For Private Equity (PE), the value creation playbook needs to be completely rewritten: the core of AI transforming the service industry is no longer simply 'replacing with a cheaper executor' but rather rebuilding verification and liability structures.

At the **product level**, product design must embed 'verifiability,' which includes elements such as evaluations (evals), transparency of sources and assumptions, observability of costs, reversibility of failures, and attributability of responsibility. These not only constitute the 'boundaries' of AI-enhanced assets but also become AI-specific standard items for investment due diligence, covering data sources and authorization, model governance, compute cost structures, and regulatory exposure. Products lacking an embedded chain of evidence will face systemic discounts in fundraising and pricing. For the entire **capital market**, asset classes are being reconstructed: financing models for infrastructure like compute, power, and data centers are expanding from equity to private credit; public market concentration continues to rise, leading index investors to passively bear concentrated AI risk. On a deeper level, when AI capabilities are rented, revenue is metered, and delivery is outcome-based, the word 'ownership' itself depreciates in value within the investment context; the true meaning of an asset lies in its 'boundaries.