Decoding the AI Stack: Analyzing Bottlenecks, Moats, and Value Accrual for Long-Term Investors

Value accrues across the artificial intelligence technological stack not merely through absolute physical scarcity, but where technological bottlenecks intersect with pricing power, recurring volume expansion, switching costs, and high incremental return on invested capital (ROIC). While Chamath Palihapitiya’s six-layer stack framework (Infrastructure, Chips, Data, Models, Execution, Application) accurately highlights the economic commoditization of foundation models and the emergence of critical hardware chokepoints, long-term investors must distinguish between absolute technical scarcity and sustainable equity value creation. As hyperscalers transform into capital-heavy industrial enterprises projected to spend over $700 billion in 2026 on compute and grid infrastructure, evaluating intrinsic value requires analyzing free cash flow durability, capital intensity risks, and proprietary workflow feedback loops rather than relying solely on physical bottlenecks.

1. Deconstructing the AI Stack: Bottlenecks Versus Economic Value Accrual#

Analyzing artificial intelligence through a layered architecture provides essential clarity for fundamental investors. Chamath Palihapitiya categorizes the ecosystem into six distinct strata: Infrastructure (power, land, cooling, minerals, fabs), Chips (GPUs/ASICs, High Bandwidth Memory [HBM], packaging, networking), Data (proprietary, sensory, robotics data), Models (foundation intelligence), Execution (agents, orchestration, APIs, actuators), and Application (end-user software and systems). The central analytical concept connecting these layers is the "fulcrum asset"—an irreplaceable component that every superior layer must pass through and that cannot be easily substituted.

However, a critical distinction exists between an absolute technological bottleneck and a high-yielding economic moat. An asset can sit directly in the technological critical path without capturing proportional financial returns for shareholders. For long-term investors evaluating economic moats through a Graham and Buffett framework, physical scarcity alone does not guarantee long-term compounding of owner earnings.

This dynamic is clearly illustrated by comparing ASML and NVIDIA. ASML possesses an absolute, irreplaceable technological bottleneck in extreme ultraviolet (EUV) lithography systems; advanced semiconductor fabs cannot substitute ASML’s machinery. Yet NVIDIA has generated far greater equity value appreciation. ASML sells finite, highly complex, capital-intensive machines to a concentrated group of buyers, limiting automated revenue scaling when global token consumption doubles. Conversely, NVIDIA pairs silicon bottlenecks with the CUDA software toolchain, system architectures, NVLink networking, and deep ecosystem integration. NVIDIA directly captures the expanding volume elasticity of AI inference and training.

To evaluate economic value accrual across the stack, investors should apply a comprehensive qualitative formula rather than measuring scarcity in isolation:

$$\text{Value Accrual} \approx \text{bottleneck} \times \text{pricing power} \times \text{volume growth} \times \text{recurring demand} \times \text{switching costs} \times \text{incremental ROIC}$$

Durable economic returns materialize only when an operational chokepoint is combined with volume elasticity, high switching costs, and strong incremental returns on invested capital.

2. Infrastructure and Capex Intensity: The Industrial Revolution Beneath AI#

The foundational layers of the AI stack are undergoing a structural transition from asset-light software models to capital-heavy industrial operations. Critical bottlenecks at the base of the stack reveal extreme geographic and corporate concentrations:

• Semiconductor Material Scarcity: Japanese materials suppliers control approximately 78% of global semiconductor photoresist sales by company nationality, with Shin-Etsu, JSR, Tokyo Ohka Kogyo, and Fujifilm dominating specific high-purity chemical markets.
• High-Purity Quartz (HPQ): Semiconductor-grade silicon ingot production relies on fused-quartz crucibles derived from exceptionally pure quartz mined in Spruce Pine, North Carolina. Sibelco (producing IOTA material across two pure ore bodies) and The Quartz Corp represent the only suppliers economically producing advanced, premium HPQ at scale.

Beyond physical raw materials, electrical utility infrastructure presents a severe operational constraint. The International Energy Agency (IEA) projects that global data-center electricity consumption will nearly double, expanding from approximately 485 TWh in 2025 to roughly 950 TWh by 2030. Hyperscaler capital expenditures are estimated to exceed $700 billion in 2026, primarily driven by data-center expansion.

Crucially, raw power generation is not the primary near-term limit. Research from Johns Hopkins University demonstrates that grid delivery equipment—specifically transformers, switchgear, uninterruptible power supply (UPS) systems, transmission interconnections, and specialized cooling infrastructure—will constrain data-center buildouts before generation capacity itself becomes the primary bottleneck. The immediate economic scarce asset is deliverable power delivered at the requisite voltage, location, and reliability.

From a value investing perspective, this transition severely impacts corporate balance sheets and return metrics. As hyperscalers deploy over $700 billion in annual capex, multi-billion-dollar physical assets enter balance sheets as property, plant, and equipment (PP&E). Once operational, straight-line depreciation schedules hit income statements regardless of immediate revenue realization. If monetizable token demand or enterprise software adoption lags behind physical buildouts, Return on Invested Capital (ROIC) and Free Cash Flow (FCF) conversion rates collapse under the weight of unamortized infrastructure. This dynamic directly mirrors historical infrastructure overbuilds, such as the telecom fiber crash involving WorldCom and Global Crossing, where massive capital expenditures created severe overcapacity and destroyed equity value despite the ultimate validity of long-term internet adoption.

3. Chips, Memory, and Model Economics: Commoditization vs. Moving Frontiers#

While theoretical models assume hardware bottlenecks naturally dissipate over time as manufacturing capacity normalizes, semiconductor supply constraints remain acute. High Bandwidth Memory (HBM) and advanced packaging represent multi-year structural bottlenecks rather than temporary friction. The IEA projects HBM supply shortages to persist through at least the end of 2027. Market share remains highly concentrated; Counterpoint data for Q2 2026 estimates HBM revenue share at:

• SK Hynix: 50%
• Samsung: 33%
• Micron: 18%

Modern hardware moats do not rely on standalone compute dies, but on full system-level integration: GPU/ASIC dies integrated with HBM via interposers, advanced packaging, high-speed networking (such as NVLink), power delivery, and specialized cooling. This system integration becomes even more vital as workloads shift from model training to inference. McKinsey projects that inference will represent approximately 60% of total AI demand by 2030. Consequently, the competitive battlefield is shifting directly toward memory bandwidth, capacity, latency, thermal efficiency, and power consumption per useful token output—solidifying the structural moats of integrated hardware vendors.

Concurrently, the economics of foundation intelligence models exhibit rapid cost degradation. Data from the Stanford AI Index indicates that the cost of running a model at GPT-3.5-level MMLU performance fell from approximately $20 per million tokens in November 2022 to roughly $0.07 in October 2024—a 280-fold cost reduction over 18 months. Furthermore, OECD studies show quality-adjusted language-model prices declined nearly 80% between January 2024 and April 2026.

Despite this price collapse, foundation models have not yet become pure commodities. Basic intelligence is commoditizing—with the Stanford 2026 AI Index showing top frontier models (Anthropic, xAI, Google, OpenAI) separated by only 25 Arena Elo points, and closed models leading open-weight alternatives by just 3.3% to 3.4% as of March 2026. However, enterprise customers continue to pay a premium for frontier capabilities. Menlo research shows frontier models' share of enterprise AI spend rose from 35% in Q3 2025 to 63% in Q3 2026.

The model layer operates as a moving economic frontier: base capabilities rapidly commoditize, while state-of-the-art reasoning maintains pricing power, forcing model developers to absorb ongoing, massive capital investments just to maintain frontier parity.

4. Execution Planes, Physical AI, and the Jevons Paradox#

As foundation models commoditize, economic value shifts toward the execution and control plane. This middleware layer encompasses dynamic model routing, durable state memory, identity management, permissioning, evaluation, tool integration, and orchestration. Framed through corporate balance sheets, enterprise software buyers do not purchase raw foundation tokens; they purchase completed business outcomes. The execution layer operates as institutional "middleware"—analogous to Kubernetes, identity access management, and observability in cloud computing—locking in enterprise switching costs and protecting profit margins even as underlying base models fluctuate. The 2026 LangChain production survey confirms that multi-model enterprise architectures are now standard, with primary deployment bottlenecks shifting from token unit costs to execution reliability, safety, and task completion.

Similarly, generic web training data is commoditizing relative to proprietary, outcome-based feedback loops. Sustainable economic moats are created when software applications embed directly into domain-specific operational workflows where actions generate verifiable real-world evaluations.

A prime illustration of this dynamic is Harvey in the legal sector. Harvey does not compete as a fragile foundation model wrapper selling generic token generation. Instead, it embeds directly into legal workflows, combining legal systems of record, distribution, regulatory compliance, client trust, and proprietary feedback loops (contract analysis $\rightarrow$ attorney revision $\rightarrow$ final court filing/resolution). Because Harvey owns the end-to-end workflow and the resulting outcome data, it builds high switching costs and defends its pricing power, rendering the underlying choice of foundation model secondary to the business outcome achieved. Similar outcome-based feedback loops are emerging in specialized domains such as coding (patch $\rightarrow$ test result $\rightarrow$ deployment outcome), medicine (diagnosis $\rightarrow$ therapy $\rightarrow$ patient health outcome), and finance (invoice $\rightarrow$ clearing $\rightarrow$ dispute resolution).

In physical AI systems, capital allocation dynamics differ fundamentally from pure software enterprise economics. Software scales at near-zero marginal cost, whereas physical robotics requires extensive bill-of-materials (BOM) capital. McKinsey estimates that precision actuators represent 40% to 60% of a humanoid robot's BOM, compared to 5% to 10% for battery modules. Furthermore, CRU estimates that each humanoid robot requires 2 to 3 kg of NdFeB (neodymium-iron-boron) rare-earth magnets. While physical AI components represent long-dated strategic optionality, data-center compute and memory assets generate immediate, verifiable cash flows today.

Understanding this structural capex expansion requires applying the Jevons Paradox. A 100-fold drop in unit token costs does not shrink total spending on compute; instead, lower unit prices trigger an explosive expansion in aggregate token consumption. As single query-response interactions are replaced by multi-agent reasoning loops executing continuous background simulations and automated tool calls, aggregate hardware compute demand accelerates despite collapsing unit prices.

For long-term fundamental investors evaluating capital allocation, sustainable value capture requires avoiding pure physical scarcity in favor of businesses exhibiting high incremental ROIC, pricing power, and embedded workflow feedback loops. Investors interested in deeper valuation models and fundamental owner-earnings calculations can explore our open-source company research on the DeepValues Research website or evaluate any stock's intrinsic value and margin of safety using the DeepValues app.

5. Key numbers#

• 78%: Japanese firms' approximate market share of semiconductor photoresist sales by company nationality.
• >$700 Billion: Estimated 2026 hyperscaler capital expenditures according to the IEA, predominantly for data-center expansion.
• 485 TWh to ~950 TWh: Projected rise in data-center electricity consumption from 2025 to 2030 (IEA).
• End of 2027: Expected persistence of High Bandwidth Memory (HBM) supply shortages (IEA).
• 50% / 33% / 18%: Estimated Q2 2026 HBM revenue share for SK Hynix, Samsung, and Micron, respectively (Counterpoint).
• $20 to $0.07 per Million Tokens: Cost decline for running GPT-3.5-level MMLU performance between November 2022 and October 2024, representing a 280x reduction (Stanford AI Index).
• ~80%: Decline in quality-adjusted language-model prices from January 2024 through April 2026 (OECD).
• 25 Arena Elo Points: Narrow performance spread separating Anthropic, xAI, Google, and OpenAI frontier models as of March 2026 (Stanford 2026 AI Index).
• 3.3%–3.4%: Performance gap between the best closed-source model and best open-weight model in March 2026 (Stanford AI Index).
• 35% to 63%: Increase in frontier models' share of enterprise AI spending between Q3 2025 and Q3 2026 (Menlo research).
• 40%–60% vs. 5%–10%: Estimated proportion of a humanoid robot's Bill of Materials (BOM) represented by actuators versus battery modules (McKinsey).
• 2–3 kg: Estimated NdFeB rare-earth magnet requirement per humanoid robot (CRU).

6. What would change our view#

• Rapid Obsolescence of CUDA and HBM Architectures: Acceleration of alternative interconnects or memory architectures that successfully unseat HBM and NVLink, undermining system-level networking moats before 2027.
• Hyperscaler Capex Retrenchment: A sharp downward revision in the >$700 billion annual hyperscaler capital expenditure trajectory due to disappointing return on invested capital (ROIC) or severe utility grid interconnection delays.
• Rapid Monetization and Stagnation of Open-Weight Models: The collapse of the 3.3–3.4% frontier performance gap, causing enterprise customers to abandon premium closed models in favor of localized, cheap open-weight models.
• Accelerated Physical AI Deployment: Mass commercialization and high-volume manufacturing of humanoid robotics reaching millions of units well before 2030, shifting immediate cash flow value capture to actuation and magnet supply chains.

7. Sources#

Decoding AI Value, Chamath Palihapitiya (2025 Annual Letter and Related Framework) Stanford AI Index (Stanford University, 2024–2026 Data) International Energy Agency (IEA) Electricity and Capex Reports Counterpoint Research (Q2 2026 Semiconductor & HBM Market Analysis) McKinsey & Company Robotics BOM Research CRU Group Materials & Rare Earths Research OECD Language Model Price Studies (2024–2026) Menlo Ventures Enterprise AI Spending Research (2025–2026) Johns Hopkins University Infrastructure & Data Center Power Research LangChain Production AI Survey (2026)