Insights Investing
The AI Capital Paradox: Why Controlling Distribution Beats Manufacturing Intelligence
Podcast · Listen (11 min)
AI Prices Fell 13x as CapEx Nears $800B: Who Owns the Moat? | Stock Market
In brief
DeepValues Research, “The AI Capital Paradox: Why Controlling Distribution Beats Manufacturing Intelligence” (Nov 16, 2026): The rapid acceleration of artificial intelligence infrastructure has created an unprecedented economic disconnect: the unit price of machine intelligence is collapsing while the capital expenditures required to manufacture it are surging.
Source: https://research.deepvalues.ai/insights/2026-11-16-notebook-76e8407d/ · Markdown
Contents
- 1. The Genesis Paradox: High Capex and Collapsing Unit Costs
- 2. Moat Durability: Why Creating Intelligence Yields No Pricing Power
- 3. Capital Allocation and Balance Sheets: Funding the Capex Gap
- 4. The Real Commanding Heights: Distribution, Tollbooths, and Financial Channels
- 5. Productivity Disconnect and Corporate Restructuring: Owner Earnings vs. Realized Profit
- Key numbers
- What would change our view
- Sources
The rapid acceleration of artificial intelligence infrastructure has created an unprecedented economic disconnect: the unit price of machine intelligence is collapsing while the capital expenditures required to manufacture it are surging. From a disciplined value-investing perspective, raw intelligence is rapidly commoditizing into a high-capex loss leader subject to severe price deflation and hyper-competitive capital overbuild. Durable economic moats and attractive returns on invested capital are not accruing to upstream creators of foundation models, but rather to owners of downstream distribution networks, structural tollbooths, and machine-native financial rails. For long-term capital allocators, navigating this landscape requires distinguishing temporary operational bottlenecks from permanent tollbooths, ensuring capital is deployed into businesses possessing genuine pricing power rather than capital-intensive manufacturing operations.
1. The Genesis Paradox: High Capex and Collapsing Unit Costs#
The upstream pipeline of model production—encompassing architectural design, pre-training, evaluation, and inference delivery, collectively termed "Genesis" by Will Manidis and Joe Roberts in The Commanding Heights—is undergoing severe microeconomic margin compression. Over the past twelve months, the unit price for equivalent machine intelligence has fallen by approximately 13-fold. Yet, over the exact same period, aggregate capital commitments by major cloud hyperscalers have escalated dramatically: combined capital expenditures reached over $400 billion last year, are approaching $800 billion this year, and are projected to cross $1 trillion next year.
This divergence highlights a structural supply-demand dynamic unique to synthetic intelligence. According to benchmark tracking data from Epoch AI, the minimum cost required to achieve equivalent model performance has dropped by approximately 47% per quarter since 2023. This rate of cost deflation represents a far steeper decline than historic technology rollouts, including fiber optic networks during the dot-com era, cellular infrastructure, the electric telegraph, or 19th-century transcontinental railroads.
While unit output costs are falling, the capital required to build frontier capability continues to compound. McKinsey & Company forecasts that total global investment in data center infrastructure will reach $7 trillion by 2030. Industry-wide capital expenditures are growing at an annual rate of 70%, drastically outpacing hyperscaler operating cash flow growth of just 23% per year. Furthermore, the capital required to train frontier models increases by approximately 3.5-fold annually, and compute demand expands by 5-fold per year, easily outstripping the 3.4-fold annual growth rate in available chip compute supply.
Viewed through the analytical lens of Benjamin Graham and Warren Buffett, individual cloud providers are engaging in rational microeconomic behavior: halting capital spending risks forfeiting future call options on existing, highly profitable search, advertising, and cloud profit pools to direct competitors. However, when aggregated across the industry, this rational individual behavior results in severe collective capital overbuild. Businesses that demand relentless capital expenditures while suffering from structural price deflation inherently destroy long-term shareholder value, regardless of how rapidly top-line aggregate compute volume expands.
2. Moat Durability: Why Creating Intelligence Yields No Pricing Power#
A foundational tenet of value investing is that structural competitive advantages—economic moats—must protect excess returns on invested capital over multi-year horizons. In model creation, technological leads are proving fleeting, preventing foundation model builders from establishing durable pricing power.
Economic history demonstrates that pioneering a breakthrough technology rarely guarantees long-term capital retention. Johannes Gutenberg invented movable type but ultimately defaulted into bankruptcy; United States transcontinental railroads fundamentally transformed American commerce, yet financial returns accrued to holders of land rights-of-way, financial syndicates, and organizational networks rather than locomotive manufacturers; Bell Labs invented the transistor, but the economic surplus settled downstream in semiconductor fabrication, system architecture, and consumer applications.
Today, the capability lag between open-source alternatives and state-of-the-art closed models has compressed to approximately four months. As technical capability diffuses, frontier performance quickly turns into a low-cost commodity. For example, Chinese model builder DeepSeek released its V4 model with output pricing roughly 97% lower than equivalent U.S. frontier models. Data from February 2025 to February 2026 shows Chinese open-weight models capturing a 41% share of global model downloads, surpassing the 36.5% share held by U.S. models. Furthermore, Stanford University’s 2026 Artificial Intelligence Index Report highlights that the performance gap between top U.S. and Chinese models has narrowed to approximately 2.7%.
Even as base intelligence commoditizes, commercial usage patterns remain highly skewed. Inference telemetry indicates that the top 10% of enterprise users consume 8.3 times more inference per active user than the median 10%, a dramatic widening from the 2.6-fold gap recorded in January. Because open-source alternatives rapidly replicate top-tier capabilities at a fraction of the cost, foundation models increasingly function as commoditized loss leaders. As Manidis and Roberts explicitly observe in The Commanding Heights, "In the coming world, you either own the channel or you pay the person who does." Under Graham and Buffett principles, a technological lead that erodes within months cannot be classified as an economic moat.
3. Capital Allocation and Balance Sheets: Funding the Capex Gap#
To bridge the growing deficit between a 70% annual capex growth rate and a 23% operating cash flow growth rate, hyperscalers and model labs are relying heavily on balance sheet expansion, debt issuances, private credit markets, and complex structured finance vehicles.
Concurrently, top-tier model developers are demonstrating top-line revenue growth, primarily driven by enterprise software integrations. OpenAI has projected an annualized revenue target of $70 billion by the end of 2026, while Anthropic disclosed in October that its single-quarter revenue rose 45%, approaching $17 billion.
At the macroeconomic level, American household wealth, pension fund solvency, and credit expansion have become deeply intertwined with AI equity valuations and ongoing infrastructure spending. Because retirement accounts and public equity markets rely on this infrastructure spending to sustain market multiples, central monetary authorities and fiscal policy makers face structural incentives to backstop financial markets, accepting persistent currency debasement and inflation as an implicit tax on holders of fiat currency.
This dynamics is further complicated by a geopolitical asymmetry in capital returns. U.S. model laboratories are bound by commercial market realities, requiring high revenue multiples to support public equity valuations and domestic retirement fund expectations. Conversely, Chinese entities such as DeepSeek and Alibaba release low-cost, open-weight models unburdened by the mandate of sustaining domestic equity market returns. By distributing near-frontier intelligence at near-zero marginal prices, Chinese open-source ecosystems directly undercut the U.S. commercial funding loop.
Simultaneously, model development is bifurcating into commercial market applications and state-backed "Sovereign AI." Sovereign AI focuses on national security, defense, and non-commercial strategic power. Operating within sensitive compartmented information facilities (SCIFs) and managed by dedicated laboratory subsidiaries, these sovereign systems are evaluated on geopolitical statecraft rather than cash-flow generation. Highlighting this transition, NVIDIA’s revenue from sovereign infrastructure sales to foreign governments and state entities surpassed $30 billion in the fiscal year ending January 2026, representing more than a three-fold increase year-over-year.
4. The Real Commanding Heights: Distribution, Tollbooths, and Financial Channels#
The term "Commanding Heights" originated in Vladimir Lenin’s 1922 address to the Fourth Congress of the Communist International, where he justified permitting private trade under the New Economic Policy (NEP) because the state retained control over key economic infrastructure nodes. Applied to modern technology markets by Manidis and Roberts, a true Commanding Height must possess three structural traits: absolute scarcity, operational control, and the continuous ability to extract economic rents as underlying transactional volume expands.
Temporary hardware bottlenecks or transient model leads do not constitute Commanding Heights. A true tollbooth retains rent-extraction authority long after initial market shortages clear. Warren Buffett demonstrated this understanding through his historical media investments. In 1973, Berkshire Hathaway purchased a $10.6 million stake in the Washington Post, which grew to $221 million by 1985 because the publication held an absolute distribution monopoly over regional advertising. However, after deploying $344 million into dozens of regional daily newspapers through 2012, Berkshire sold its entire newspaper portfolio in 2020 for less than half its purchase price. The core value-investing insight is clear: economic moats are not permanent properties of corporate brands, but functions of distribution dominance. When digital platforms shifted distribution control, historic media franchises suffered irreversible capital impairment.
In digital ecosystems, dominant tollbooths are exemplified by platforms like Google Ads, which taxes web-wide commercial customer acquisition. Recognizing this dynamic early, Google acquired advertising platform DoubleClick and purchased Android for $50 million in 2005, securing the primary distribution gateway for mobile internet traffic. Intelligent capital allocation requires identifying the modern equivalents of DoubleClick and Android, rather than funding capital-intensive foundation model manufacturers.
This structural control extends directly into machine-native financial channels. As autonomous software agents increasingly execute economic transactions—purchasing compute, acquiring data, and paying API fees—they require low-cost, programmable, cross-border settlement rails. Dollar-denominated stablecoins are rapidly becoming the native payment architecture for these agentic workflows.
In 2025, adjusted stablecoin transaction volume surpassed $11 trillion, rivaling total Visa network volume. By September 2026, stablecoin circulation reached over $300 billion. Furthermore, in controlled tests across 36 frontier AI models, stablecoins were selected as the preferred settlement mechanism in 53.2% of autonomous payment scenarios. By bypassing legacy banking clearinghouses and credit card processing fees, stablecoins establish a machine-native payment layer embedded directly into autonomous software workflows. Entities controlling stablecoin issuance, reserve management, and regulatory compliance gateways sit at a structural Commanding Height, collecting recurring tolls on machine transaction velocity.
5. Productivity Disconnect and Corporate Restructuring: Owner Earnings vs. Realized Profit#
A central challenge for fundamental analysts is the disconnect between micro-level employee productivity gains and macro-level enterprise profit expansion. According to McKinsey & Company’s Global AI Enterprise Adoption and EBIT Impact Survey (2026), 88% of surveyed organizations regularly utilize AI within at least one business function. However, 60% of enterprise respondents report zero net EBIT impact attributable to AI investments. Moreover, while 80% of individual employees report personal efficiency gains, only 37% state that these efficiency gains translate into tangible contributions to enterprise operating profit.
This productivity disconnect highlights the critical distinction between GAAP EBIT and Warren Buffett’s framework of Owner Earnings (defined as Net Income plus Depreciation and Amortization minus required Capital Expenditures). Enterprise AI adoption frequently fails to expand Owner Earnings because continuous software licensing costs, ongoing model maintenance capex, and integration overhead absorb individual efficiency gains, shifting the economic surplus from enterprise balance sheets to downstream infrastructure providers.
This structural dynamic aligns with Ronald Coase’s 1937 theory of the firm, which posits that business organizations exist to reduce internal transaction and coordination costs. As autonomous AI agents reduce internal coordination friction, enterprise boundaries shrink. Corporate entities are restructuring into lean executive hubs focused on capital allocation and strategic oversight, while routine operational workflows are delegated to machine execution.
This corporate restructuring is driving a K-shaped labor market split. For example, Indian IT services provider Tata Consultancy Services (TCS) reduced its workforce by 23,460 employees in FY2026 while generating a 7.5% increase in adjusted operating profit. Across the broader Indian technology services sector, which employs nearly 6 million professionals, only 16 net new positions were created per 100 IT graduates in FY2026.
This disruption aligns with Balaji Srinivasan’s concept of the "Paper Belt," which explains how developed nations construct credentialing barriers, occupational licenses, and immigration walls to insulate domestic white-collar labor, while uncredentialed and international workers remain exposed to direct algorithmic labor substitution. Human case studies illustrate this geographical friction:
- Ravi (Bangalore): A 14-year implementation manager at a multinational tech services firm who advanced to oversee complex client workflows. As his firm deployed automated agentic review pipelines, his role shifted to approving software outputs. His daily approvals provided the reinforcement training data used to refine the software system. Once the agent workflow reached full reliability, his position was eliminated. While his former employer's operating margin expanded and its parent stock appreciated in New York, his personal employment path was interrupted.
- Malena (Buenos Aires): A digital graphic designer who received overseas client payments in dollar-linked stablecoins via a mobile application to protect her savings against domestic currency depreciation. The design software tools that enhanced her individual productivity ingested her design portfolio, stylistic choices, and edit histories. The underlying models were subsequently retrained and offered directly to clients at lower price points. As clients migrated to direct model interaction, her freelance earnings dissolved, leaving downstream infrastructure operators and payment rails with the remaining economic surplus.
A Stanford University labor study reveals that by mid-2026, employment rates for young professionals aged 22 to 25 in high-AI-exposure roles dropped 19% below pre-adoption baseline projections. Standardized entry-level corporate positions are contracting rapidly as marginal processing costs approach electricity base rates.
For long-term value investors, the practical takeaway is decisive: avoid allocating capital to capital-intensive originators of commoditized intelligence whose relentless capex requirements impair return on invested capital. Capital should instead be directed toward businesses that command unassailable distribution channels, essential financial settlement rails, and structural economic tollbooths. We invite disciplined investors to explore our full library of open-source company research on the DeepValues Research site or evaluate any security's intrinsic value in the DeepValues app.
Key numbers#
- Unit Cost Deflation: ~13x annual decline in unit price for equivalent machine intelligence (Past 12 months).
- Model Performance Cost Decline: ~47% quarterly drop in minimum cost to reach equivalent model performance (Epoch AI benchmark, 2023–2026).
- Hyperscaler Capital Expenditures: >$400 billion (prior year), ~$800 billion (current year), projected >$1 trillion (following year).
- Data Center Investment & Cash Flow Growth: $7 trillion total projected data center spending by 2030; annual capex growth rate of 70% vs. operating cash flow growth of 23% (McKinsey).
- Model Training & Compute Costs: ~3.5x annual increase in frontier model training cost; 5x annual increase in compute demand vs. 3.4x annual increase in available chip compute supply.
- OpenAI Annualized Revenue Target: $70 billion projected by end of 2026.
- Anthropic Revenue Growth: 45% single-quarter revenue growth in October, reaching ~$17 billion.
- Enterprise Inference Consumption Gap: Top 10% enterprise users consuming 8.3x more inference than median 10% (up from 2.6x in January).
- Open-Source Capability & Download Share: ~4-month performance lag behind frontier closed models; DeepSeek V4 output pricing ~97% lower than US frontier equivalents; 41% global download share for Chinese models vs. 36.5% for US models (Feb 2025–Feb 2026).
- US-China Model Performance Gap: ~2.7% performance gap between top US and Chinese models (Stanford 2026 AI Index).
- NVIDIA Sovereign Infrastructure Revenue: >$30 billion (>3x YoY growth) in the fiscal year ending January 2026.
- Stablecoin Transaction Volume & Circulation: >$11 trillion adjusted transaction volume (2025); >$300 billion in circulation (September 2026); selected in 53.2% of payment test scenarios across 36 frontier models.
- IT Services Employment Impact: TCS reduced workforce by 23,460 employees while adjusted operating profit rose 7.5% (FY2026); 16 net new jobs per 100 IT graduates in Indian tech services sector out of ~6 million workforce (FY2026).
- Youth Employment Suppression: Employment rate for 22-to-25-year-olds in high AI exposure roles 19% below expected baseline by mid-2026 (Stanford study).
- Enterprise Financial Realization Gap: 88% of organizations using AI regularly, but 60% report no EBIT impact; 80% of individuals report efficiency gains, but only 37% report operating profit contribution (McKinsey 2026 Survey).
- Buffett Newspaper Investment: $10.6 million investment in Washington Post (1973) grew to $221 million (1985); total newspaper purchases of $344 million up to 2012 sold in 2020 for less than half cost.
- Google Android Acquisition: Purchased for $50 million in 2005.
What would change our view#
- Durable Technological Decoupling: Frontier closed-source model developers maintaining a multi-year performance, safety, and reliability lead over open-source alternatives, successfully restoring pricing power and halting unit-cost deflation.
- Sustained Enterprise Return on Investment: Broad enterprise surveys demonstrating direct, widespread transmission of worker productivity gains into net EBIT margin expansion across a clear majority of corporate adopters.
- Regulatory or Infrastructure Bottlenecks on Distribution: Regulatory interventions, antitrust mandates, or architectural shifts that prevent major distribution platforms, operating systems, and financial networks from charging rent on autonomous agent activity.
- Capital Expenditure Discipline and Cash-Flow Alignment: Hyperscalers voluntarily slowing infrastructure capital expenditure growth to align with underlying operating cash flow growth (~23% per year), restoring rational capital allocation to cloud data center construction.
Sources#
- The Commanding Heights, by Will Manidis and Joe Roberts (October 29)
- AI's Ultimate Winners May Not Be the Builders of Intelligence, video analysis hosted by Da Fei on the YouTube Channel Best Partners ("最佳拍档")
- 2026 Artificial Intelligence Index Report, by Stanford University
- Global AI Enterprise Adoption and EBIT Impact Survey (2026), by McKinsey & Company
- Benchmark Tracking Report on AI Compute and Model Costs, by Epoch AI
Watch
Source links
Watch & listen
- AI Prices Fell 13x as CapEx Nears $800B: Who Owns the Moat? | Stock Market Podcast · EN
- AI智慧單價一年暴跌13倍巴菲特指標,收過路費的竟是穩定幣 YouTube · 中文
- AI智慧單價一年暴跌13倍巴菲特指標,收過路費的竟是穩定幣|美股市場 Podcast · 中文