AI Infrastructure Concentration: Compute, Chips, Cloud, Energy, and Capital

Full report

On this page

File: report.md

Executive Assessment

As of September 4, 2026, the global artificial intelligence production stack exhibits a profound structural paradox. The algorithmic frontier is experiencing unprecedented democratization through the proliferation of open-weight models and extraordinary algorithmic efficiency gains. Conversely, the physical and financial infrastructure requisite for the deployment, training, and global-scale inference of these models has consolidated into a hyper-concentrated oligopoly. An empirical analysis of the AI value chain reveals that economically meaningful control over AI production is centralized among a narrow cohort of hyperscale cloud providers, advanced semiconductor foundries, and energy-infrastructure gatekeepers. The baseline proposition under review, IC-CLAIM-003, posits that open-weight AI can distribute access without eliminating infrastructure concentration. The accumulated evidentiary record overwhelmingly supports this claim, with critical qualifications regarding the precise mechanisms of this sustained concentration. Rapid advancements in software optimization, exemplified by the Multi-head Latent Attention (MLA) architectures and low-precision (FP8) training pipelines utilized in open-weight models such as DeepSeek V3 and R1, have drastically compressed the absolute computational cost required to train individual frontier-class models1. Nevertheless, these algorithmic efficiency gains are entirely subsumed by the macroeconomic dynamics of the Jevons Paradox3. As the unit cost of AI inference diminishes, the aggregate volume of inference demand—driven by continuous agentic workflows, complex multimodal processing, and ubiquitous enterprise adoption—has exploded5. Consequently, absolute infrastructure demand continues to scale exponentially, effectively nullifying the decentralizing potential of algorithmic efficiency at the hardware layer. This insatiable demand translates into unprecedented, insurmountable capital expenditures that function as the primary barrier to market entry. In calendar year 2026, the four dominant hyperscalers (Amazon, Microsoft, Alphabet, and Meta) are projected to deploy a combined $720 billion to $745 billion in capital expenditures, representing a 77% year-over-year expansion from 20256. This capital is channeled directly into a strictly constrained semiconductor supply chain. Taiwan Semiconductor Manufacturing Company (TSMC) exercises near-monopoly control, retaining an estimated 95% of the AI accelerator foundry market by leveraging absolute dominance in sub-5nm fabrication nodes and advanced Chip-on-Wafer-on-Substrate (CoWoS) packaging9. Concurrently, a severe structural deficit in High-Bandwidth Memory (HBM) supply—controlled by an oligopoly of SK Hynix, Samsung, and Micron—has established a zero-sum allocation environment that overwhelmingly favors well-capitalized incumbents possessing the liquidity to secure multi-year capacity prepayments10. Beyond silicon and capital, physical energy constraints have emerged as the ultimate hard limit on AI market contestability. Within the United States, the PJM Interconnection grid has experienced a catastrophic 833% surge in capacity market clearing prices between the 2024 and 2026 delivery years, driven explicitly by gigawatt-scale data center capacity requests colliding with a frozen interconnection queue12. Grid operators are increasingly shifting the burden of power generation onto data center operators through "connect and manage" mandates, meaning that only technology conglomerates possessing the capital capacity to co-develop dedicated, grid-scale power infrastructure can secure future operational growth14. A contrasting examination of the Chinese market underscores the immense friction of bypassing this established stack. While an alternative infrastructure ecosystem exists—centered on Huawei's chip designs and Semiconductor Manufacturing International Corporation's (SMIC) fabrication—it requires massive state subsidization and operates at substantially lower physical efficiencies, utilizing Deep Ultraviolet (DUV) multi-patterning to overcome lithography embargoes15. Ultimately, the global AI infrastructure market is characterized by structural "cloud recapture," wherein capital investments made by hyperscalers into AI developers are contractually routed back into their own proprietary compute ecosystems. This creates formidable technical switching costs and infrastructure dependencies that comprehensively lock in incumbent dominance across the production chain17.

Research Boundaries and Methodological Framework

This empirical investigation evaluates the concentration of the artificial intelligence infrastructure stack across the entire value chain, disaggregating the production process into distinct layers: capital formation, cloud service provision, semiconductor fabrication and design, memory packaging, and energy distribution. The primary observation period is centered on fiscal and calendar year 2026\. The evidentiary foundation relies on finalized corporate 10-K and 10-Q filings through the second quarter of 2026, independent market monitor data from regional grid authorities, and official regulatory documentation, prominently including the Federal Trade Commission’s (FTC) January 2025 Section 6(b) staff report on AI partnerships and investments. The actual execution date of this research is September 4, 2026\. The geographic scope focuses predominantly on the United States as the locus of hyperscale capital formation and cloud architecture, alongside Taiwan, which serves as the indispensable node for semiconductor fabrication. To rigorously test the resilience and contestability of this supply chain, the analysis incorporates a contrasting examination of the People's Republic of China. Specifically, it evaluates the partnership between Huawei and SMIC as an isolated, state-backed alternative ecosystem striving for technological self-sufficiency. Market boundaries are defined not merely by localized revenue share, but by assessing contractual control, exclusive allocation rights, and the physical constraints of production across interconnected supply layers. Methodologically, this report distinguishes explicitly between different financial denominators to ensure accuracy. Capital concentration is measured by separating operating cash flow, gross capital expenditure (capex), and recognized depreciation schedules to identify the underlying debt and capital barriers required to compete at the technological frontier. Theoretical counterforces, particularly the proliferation of open-weight models and algorithmic compression techniques, are evaluated symmetrically against the economic principles of induced demand. This ensures that software efficiency is not falsely equated with a reduction in absolute physical resource requirements.

Capital Formation and the Hyperscaler Capex Wall

The foundational mechanism of concentration within the AI ecosystem is the sheer velocity and unprecedented scale of capital deployment required to participate at the infrastructure layer. The physical buildout of the AI stack is currently financed by a staggering consolidation of capital expenditure that establishes insurmountable barriers to entry for independent infrastructure providers, effectively walling off the foundation of the industry to all but the most highly capitalized entities in the global economy. Reconstructing the capital expenditures of the four primary hyperscalers—Amazon, Microsoft, Alphabet, and Meta—reveals a historic consolidation of infrastructure investment. Based on mid-2026 financial guidance and trailing cash-flow statements, these four entities are projected to record between $720 billion and $745 billion collectively in calendar year 20268. This represents a 77% expansion from the roughly $410 billion deployed in 2025, pushing the combined three-year AI infrastructure investment since January 2023 well past the $1 trillion threshold6. Disaggregating the denominators of these metrics provides clarity into the specific locus of concentration. Amazon recently elevated its 2026 capital expenditure plan to approximately $220 billion, heavily allocated toward Amazon Web Services (AWS) data centers and the procurement of custom Trainium silicon7. Microsoft’s calendar 2026 capex is projected at approximately $190 billion, driven intensely by Azure AI infrastructure and newly commenced finance leases for data center properties6. Alphabet is targeting a range between $195 billion and $205 billion, with historical deployment metrics indicating that approximately 60% of this expenditure is allocated to servers, while the remaining 40% funds data centers and networking equipment8. Meta, despite lacking a public-facing cloud infrastructure business, has guided $125 billion to $145 billion in capex to support internal AI model training and the immense inference load generated by its global user base6. This magnitude of spending introduces a severe structural financial dynamic characterized as the "depreciation lag," which fundamentally alters the contestability of the market. Because capital expenditures for servers (which typically possess useful lives extended to five or six years) and data center buildings (which are depreciated over up to 25 years) are amortized over time, the current cash outflow massively exceeds recognized depreciation expenses on corporate income statements8. During the trailing four quarters ending in the first quarter of 2026, the four hyperscalers spent an estimated $434 billion in cash on property and equipment, yet recognized only approximately $149 billion in depreciation over the same span20. This profound cash drain dictates that even highly profitable, trillion-dollar technology conglomerates must increasingly turn to external debt markets to fund their AI infrastructure expansions. In recent quarters, Alphabet issued over $56 billion in bonds across multiple offerings, while Meta priced a $30 billion investment-grade deal alongside a $27 billion off-balance-sheet special purpose vehicle (SPV)7. The structural implication for market concentration is absolute: if the most profitable entities in corporate history must borrow tens of billions of dollars to sustain the current pace of AI infrastructure buildout, the market is fundamentally uncontestable for new, independent market entrants seeking to build competing physical infrastructure.

Cloud Recapture and Contractual Ecosystem Control

The hyper-concentration of capital at the infrastructure layer is not merely an artifact of hardware costs; it is actively reinforced by a sophisticated financial and contractual architecture termed "cloud recapture." This mechanism structurally mitigates the hyperscalers' investment risk while simultaneously locking in frontier AI developers, ensuring that third-party algorithmic innovation directly subsidizes incumbent cloud dominance. In this paradigm, capital injected by a hyperscaler into an AI developer is contractually routed back to the hyperscaler as guaranteed cloud revenue17. The Federal Trade Commission’s (FTC) January 2025 Office of Technology Staff Report on AI Partnerships and Investments 6(b) Study provided unprecedented, legally compelled visibility into the mechanics of these arrangements18. The FTC investigated partnerships between major Cloud Service Providers (Alphabet, Amazon, Microsoft) and leading generative AI developers (Anthropic, OpenAI). The resulting report details how these relationships circumvent traditional merger and acquisition thresholds while achieving functionally identical anti-competitive lock-in and market foreclosure22. The FTC investigation identified a network of specific contractual mechanisms designed to ensure long-term infrastructure concentration. Foremost among these are explicit cloud spend commitments. AI developers are contractually obligated to spend vast proportions of the CSP's investment on that specific CSP's cloud architecture. For instance, billing data and public disclosures reveal that Anthropic committed to spending over $100 billion over ten years on AWS technologies, effectively guaranteeing AWS revenue and ensuring maximum utilization for Amazon's custom Trainium chips and Graviton CPUs17. This circular spending guarantees that the CSP recoups its investment as topline revenue, funding further data center expansion. Furthermore, the FTC report highlighted severe exclusivity provisions and preferential treatment clauses. Microsoft Azure, for example, serves as the exclusive cloud provider for all OpenAI workloads, spanning foundational research, API services, and enterprise product deployment18. These arrangements grant the CSP partners varying degrees of consultation, control, and non-voting equity rights. They also feature capped profit structures—such as Microsoft's reported $92 billion allocated profit cap with OpenAI—which inextricably link the financial success and operational viability of the AI developer to the infrastructure provider18. Crucially, these partnerships deliberately engineer immense contractual and technical switching costs. Technically, models optimized for specific CSP proprietary architectures, utilizing co-designed semiconductor chips and embedded engineering talent, cannot be migrated to competing infrastructure without incurring devastating latency, retraining, and engineering costs18. By utilizing cloud recapture, the CSPs convert speculative venture investments into guaranteed, decade-long infrastructure demand. If the AI developer succeeds, the CSP captures equity upside and massive cloud revenue; if the developer fails, the CSP retains the underlying physical data center assets and GPUs, which can be seamlessly repurposed for internal workloads18. This asymmetrical risk profile guarantees that control over the foundation model layer remains entirely dependent upon the capital and infrastructure of the cloud oligopoly.

Semiconductor Fabrication, Packaging, and Memory Bottlenecks

Beneath the abstract layer of cloud services, the physical hardware requisite for AI training and inference represents the most concentrated supply chain in the modern global economy. While public discourse frequently centers on chip design—where Nvidia maintains an overwhelming 80% to 87% market share of AI accelerator GPUs and reported data center revenues surging past $75 billion annually24—the actual locus of market control resides deeper in the physical fabrication and advanced packaging of these silicon assets. The global advanced node foundry market, valued at $98.4 billion in 2025 and projected to grow at a 10.5% compound annual growth rate to $267 billion by 2035, is essentially a pure monopoly controlled by the Taiwan Semiconductor Manufacturing Company (TSMC)9. Independent market analysis indicates that TSMC holds an estimated 95% share in the highly lucrative AI accelerator chip foundry market9. The production of frontier AI silicon demands sub-5nm technology. In 2025, the 5nm segment—which includes TSMC's highly mature N5 and N4 processes—accounted for the largest revenue share, leveraging the ubiquitous FinFET transistor architecture9. The industry is rapidly migrating toward sub-3nm and 2nm processes, requiring a transition to Gate-All-Around (GAA) nanosheet architectures to achieve the necessary power efficiency and transistor density9. TSMC's dominance is structurally reinforced by the immense capital and operational expertise required to construct and yield extreme ultraviolet (EUV) lithography fabs. This is compounded by an intricate, locked-in ecosystem of Electronic Design Automation (EDA) tools, validated intellectual property blocks, and proprietary packaging techniques27. Every major semiconductor designer, including Nvidia, AMD, Apple, and Qualcomm, has committed to long-term capacity agreements with TSMC. This locked-in demand pipeline has driven TSMC to revise its 2026 growth forecast upward to 25%–30%28. Because total global wafer allocation is finite, this dynamic structurally forecloses the market to smaller AI hardware startups; every silicon wafer allocated to Nvidia or a hyperscaler's custom AI silicon is a wafer categorically unavailable to a potential challenger. Simultaneously, the throughput of modern AI accelerators is strictly gated by memory bandwidth, placing unprecedented strain on the High-Bandwidth Memory (HBM) supply chain. The HBM market is effectively a rigid oligopoly tightly controlled by SK Hynix, Samsung, and Micron10. In 2025, HBM represented 23% of the total DRAM market, surpassing $30 billion in sales, with SK Hynix maintaining a dominant market position due to superior yield rates in advanced HBM3E stacking10. The physical integration of logic chips and HBM requires highly advanced 2.5D and 3D packaging technologies, primarily relying on TSMC's Chip-on-Wafer-on-Substrate (CoWoS) process. The severe shortage of CoWoS packaging capacity and HBM supply has engineered a literal zero-sum game within the broader semiconductor market11. Because memory manufacturers are allocating maximum possible wafer capacity to highly profitable HBM stacks to service Nvidia and hyperscaler demand, the supply growth for traditional DRAM and NAND in 2026 is projected to fall below historical norms, tracking at merely 16% and 17% year-over-year, respectively11. This zero-sum allocation dynamic further entrenches existing market leaders, as only buyers capable of guaranteeing multi-billion-dollar, multi-year prepayments can secure the necessary memory and packaging capacity to deploy AI infrastructure at scale.

Physical Limits: The Energy and Grid Capacity Crisis

While silicon availability and capital formation represent severe economic bottlenecks, electrical energy access has rapidly emerged as the ultimate physical and spatial constraint on AI infrastructure expansion. The geographic concentration of massive data centers, combined with the extreme, escalating power density of modern liquid-cooled AI server racks, is fundamentally fracturing regional electricity markets. The most acute empirical evidence of this energy bottleneck is observable within the PJM Interconnection. Operating as the largest regional transmission organization in the United States, PJM serves 67 million customers across 13 states, prominently including major global data center hubs in Northern Virginia and Illinois (ComEd territory)14. Wholesale electricity markets operate on two primary, parallel timelines: the energy market, which compensates generators for electricity actually dispatched in real-time, and the capacity market, which utilizes forward auctions to pay generators to guarantee their availability during peak demand events three years in the future13. The PJM capacity market has recently experienced a catastrophic price escalation driven explicitly by AI data center load forecasts. Historically hovering around a stable $28.92 per Megawatt-day (MW-day) in the 2024/2025 delivery year, the capacity clearing price surged an unprecedented 830% to $269.92/MW-day for the 2025/2026 auction13. In the subsequent 2026/2027 Base Residual Auction, the price collided with the Federal Energy Regulatory Commission (FERC) approved cap of $329.17/MW-day13. Market analyses indicate that without the regulatory cap in place, clearing prices would have exceeded $530/MW-day13. The PJM Independent Market Monitor unequivocally attributed the primary cause of these unprecedented price spikes to data center-related energy demand29. The inclusion of existing and forecast data center load in the peak load forecast resulted in an $82.1% increase in capacity market revenues—a direct $7.27 billion impact—for the 2026/2027 auction alone. Across three recent auctions, the combined impact totaled over $23.1 billion in wealth transfer from retail ratepayers to generation owners29. Over the previous 15 years, PJM demand growth was essentially stagnant at 0.3% annually; current 10-year forecasts project a 3.6% annual growth rate, requiring over 50 Gigawatts (GW) of new peak capacity by 2030 solely to service data centers13. This demand surge is colliding violently with a frozen interconnection queue. While there are approximately 170,000 MW of new generation requests residing in the PJM queue—the vast majority comprising intermittent wind and solar resources—systemic delays in environmental permitting, necessary transmission upgrades, and project financing mean that firm capacity is retiring much faster than it can be reliably replaced13. Currently, the PJM firm capacity resource mix remains highly dependent on legacy infrastructure, comprising 45% natural gas, 22% coal, and 21% nuclear, with wind and solar accounting for a mere 4% of firm capacity commitments13. To avert systemic grid failure and shield retail ratepayers from further price shocks, grid operators are introducing severe market reforms that specifically target AI infrastructure. PJM’s proposed Independent Request for Alternative Supply (IRAS) plan mandates that data centers and other large loads exceeding 50 MW must effectively act as their own utility providers14. Under this "connect and manage" or "bring your own power" framework, hyperscalers must sign bilateral contracts to fund net-new power generation, such as deploying small modular reactors (SMRs), large-scale battery storage, or new natural gas plants. If a data center operator fails to bring its own requisite supply, the facility will be subject to immediate operational curtailment—forcibly shutting off power—prior to standard grid emergencies, effective starting in 202714. This energy policy paradigm shift radically alters the competitive landscape for AI infrastructure. By forcing data center operators to independently finance and construct gigawatt-scale power generation, the capital threshold required to enter the AI market is raised exponentially. Only the incumbent cloud providers—already insulated by their $745 billion capex wall—possess the balance sheets, credit ratings, and political leverage necessary to co-develop massive energy infrastructure14. Consequently, localized grid constraints and the resulting regulatory frameworks serve as a secondary mechanism of market concentration, permanently reinforcing the existing cloud oligopoly.

Counter-Ecosystem Contestability: China's DUV Multi-Patterning Strategy

To rigorously test the resilience of this infrastructure concentration, it is analytically necessary to examine a market attempting to construct a parallel AI stack without access to the dominant Western monopolies. Subjected to comprehensive United States export controls targeting advanced computing technologies and EUV lithography equipment, the People's Republic of China has engineered a state-subsidized, vertically integrated alternative ecosystem32. At the center of this sovereign ecosystem is Huawei's Ascend 910C AI accelerator, manufactured domestically by SMIC. Denied access to TSMC's 5nm and 3nm EUV nodes, SMIC has been forced to utilize older Deep Ultraviolet (DUV) immersion lithography to produce 7nm (and functionally 6nm/5nm equivalent) chips. This is achieved through an incredibly complex and friction-heavy manufacturing process known as multi-patterning15. SMIC's N+2 and N+3 nodes require multiple, precisely aligned passes through DUV scanners to achieve the necessary transistor density for AI logic15. While this brute-force approach allows China to field a domestic chip theoretically capable of supporting frontier AI training—with Huawei reportedly targeting the production of 600,000 units of the Ascend 910C in 2026—it incurs severe economic and physical costs34. DUV multi-patterning drastically degrades wafer yield rates, increasing the effective cost per viable chip, and significantly amplifies the energy required during the manufacturing process. Furthermore, the resulting processors suffer from inferior power efficiency, consuming substantially more power per floating-point operation (FLOP) during both inference and training compared to TSMC's EUV-manufactured equivalents. Because the hardware is physically less efficient, the energy burden placed on the electrical grid is magnified. This constraint has necessitated an aggressive "East-to-West" computing strategy orchestrated by the Chinese state, wherein energy-intensive data centers are forcibly migrated from the densely populated, economically vibrant eastern seaboard to remote western provinces possessing vast renewable energy and thermal coal resources36. The Chinese market demonstrates that while it is technologically possible to bypass the TSMC and Nvidia bottlenecks, doing so requires the absolute backing of a sovereign nation-state, a complete tolerance for horrific economic yields, and the political authority to forcibly reroute national power grids and data infrastructure. This confirms that at a global level, AI infrastructure concentration is not merely an artifact of free-market capitalism or first-mover advantage, but a reflection of the extreme physical, energetic, and material sciences required to sustain artificial intelligence at scale.

Counterforces: Algorithmic Efficiency and the Jevons Paradox

The strongest empirical challenge to the thesis of permanent infrastructure concentration is the rapid advancement of open-weight models and algorithmic efficiency. Critics argue that software optimization and model compression will ultimately commoditize compute, drastically lowering the hardware bottleneck and enabling distributed inference. This countercase was vividly demonstrated by the release of advanced open-weight architectures, most notably DeepSeek V3 and R1. DeepSeek achieved frontier-level capabilities utilizing a mere fraction of the traditional computational budget historically required by incumbent labs. By implementing Multi-head Latent Attention (MLA), which effectively compresses Key-Value (KV) caches by an astounding 93.3%, and utilizing a DeepSeekMoE (Mixture of Experts) architecture combined with FP8 (8-bit floating point) mixed-precision training, the developers drastically reduced both memory bandwidth requirements and overall FLOP consumption1. The reported training cost for DeepSeek was unprecedentedly low—approximately 2.788 million GPU hours, equating to an estimated $5.5 million39. If a frontier model can be trained for under $10 million, the argument logically follows that the $745 billion hyperscaler capex wall is an irrational "overbuild," and that smaller, independent laboratories will soon possess the capability to deploy state-of-the-art models on decentralized or affordable edge hardware. However, while algorithmic efficiency undeniably lowers the barrier to accessing AI capabilities, it structurally fails to reduce aggregate infrastructure concentration due to a macroeconomic phenomenon known as the Jevons Paradox3. The Jevons Paradox, observed historically across energy and resource markets, dictates that as technological improvements increase the efficiency with which a resource is utilized, the total aggregate consumption of that resource may actually rise, rather than fall, because the drastically lowered cost stimulates massive, unforeseen new demand4. In the context of artificial intelligence, training compute represents only a fractional component of a model's lifetime infrastructure footprint; the vast majority of compute is perpetually consumed during inference (the process of serving the model to end-users). As MLA architectures, FP8 precision, and improved batching methodologies dramatically reduce the cost per token of inference, developers are deploying AI into vastly more computationally intensive, continuous workflows5. Instead of generating a single, discrete text response, AI is now utilized for continuous agentic reasoning, where autonomous agents recursively prompt themselves thousands of times to solve a single, complex enterprise task. Therefore, while a single open-weight model can be trained efficiently and execute basic inference on local hardware, the deployment of global-scale AI applications—such as real-time multimodal generation, enterprise-wide unstructured data processing, and millions of concurrent autonomous agents—requires an absolute scale of highly networked inference compute that only hyperscale data centers can provide. Algorithmic efficiency does not destroy the infrastructure bottleneck; it merely expands the total addressable market size, ensuring that the cloud oligopoly's massive capex investments remain fully utilized and central to the global economy.

Topic-Specific Analyses and Structural Mappings

To synthesize the empirical findings, the following tables provide structured mappings of supply chain dependencies, temporal concentration metrics, and policy tradeoffs.

Layer-by-Layer Dependency Map

Concentration and Contestability Matrix

Data Conflict Resolution Note

During the analysis of capital expenditures, varying methodologies for defining "AI Capex" across corporate disclosures present a distinct data conflict. Companies do not consistently isolate AI-specific spending from general cloud, network, or real estate infrastructure. Furthermore, accounting differences exist (e.g., Amazon relies on cash capex, whereas Microsoft prominently includes newly commenced finance leases). This report resolves this discrepancy by utilizing the aggregate "Purchases of Property and Equipment" trailing cash-flow metric as the most reliable proxy for absolute infrastructure scale. The $720B-$745B range is presented as a holistic indicator of the ecosystem barrier-to-entry rather than a purely isolated AI metric. Regarding energy data, some speculative interconnection queue metrics (e.g., the 170,000 MW in PJM) include "phantom" projects that lack financing and are highly unlikely to be constructed. This report resolves this ambiguity by disregarding speculative queue volume and focusing exclusively on finalized capacity auction clearing prices ($329.17/MW-day) and implemented regulatory frameworks (the IRAS plan) as definitive, empirical proof of actual grid constraint.

Policy Tradeoff Evaluation

Regulatory interventions aimed at addressing infrastructure concentration present complex, systemic tradeoffs. Symmetrical evaluation of policy alternatives reveals that attempting to democratize one layer frequently inadvertently concentrates another.

Project Claim Assessment and Strategic Recommendations

This research directly tests and addresses the project baseline IC-CLAIM-003: Open-weight AI can distribute access without eliminating infrastructure concentration. Claim Impact Assessment: The empirical data conclusively supports the claim with a critical qualification regarding the mechanism of hardware demand. Open-weight models (such as DeepSeek V3 and R1) have successfully distributed access to frontier software capabilities and demonstrated profound algorithmic efficiency. However, the physical layers of the stack—specifically the $745 billion capital requirement, TSMC's 95% advanced foundry monopoly, the HBM zero-sum allocation, and the PJM energy capacity crisis—have hardened immutably. Important Limitation (Qualification): The Jevons Paradox ensures that efficiency gains in training do not translate to a reduction in absolute infrastructure demand. Because inference demand is highly elastic, cheaper compute per token results in exponentially more tokens processed. Consequently, open weights distribute software access but actively reinforce hardware dependency upon the incumbent oligopoly. Strongest Remaining Objection: If agentic workflows prove fundamentally inefficient, or if macroeconomic factors force a sudden, severe contraction in enterprise AI adoption, the massive hyperscaler capex could rapidly transition into stranded assets (a catastrophic "overbuild"). In this scenario, the depreciation lag would crush incumbent operating margins, leading to a fire sale of compute capacity that temporarily eliminates the infrastructure bottleneck and floods the market with cheap, commoditized hardware. What Would Change This Assessment: Verifiable evidence demonstrating that high-value, frontier-equivalent inference workloads can be executed efficiently on heavily decentralized, consumer-grade hardware architectures—at global scale—without invoking massive cloud energy consumption or relying upon advanced memory bandwidth requirements. Best Next Research Action: The most critical unresolved question surrounding infrastructure concentration is the exact elasticity of inference demand. The mandated next research action must be a targeted empirical study measuring the actual inference utilization rates and unit economics of autonomous agentic workflows deployed in live enterprise environments. By quantifying the compute cost of agentic reasoning loops versus traditional generation, researchers can definitively prove whether the Jevons Paradox will perpetually sustain hyperscaler capex requirements, or if the market is rapidly approaching an absolute, physical ceiling of software utility.

File: sources.json

JSON { "agent\_id": "R07", "research\_date": "2026-09-04", "sources": \[ { "source\_id": "R07-S001", "matched\_ic\_source\_id": null, "title": "FTC Staff Report on AI Partnerships & Investments 6(b) Study", "authors": "Federal Trade Commission Office of Technology Staff", "issuing\_institution": "Federal Trade Commission (FTC)", "document\_type": "Official Staff Report", "canonical\_url": "https://www.ftc.gov/system/files/ftc\_gov/pdf/p246201\_aipartnerships6breport\_redacted\_0.pdf", "retrieved\_url": "https://www.ftc.gov/system/files/ftc\_gov/pdf/p246201\_aipartnerships6breport\_redacted\_0.pdf", "publication\_date": "2025-01-17", "version\_date": "2025-01-17", "effective\_date": "2025-01-17", "accessed\_at": "2026-09-04", "jurisdiction": "United States", "legal\_or\_policy\_status": "Agency Staff Report (Non-Binding)", "publication\_status": "Published", "host\_status": "Available", "review\_scope": "Full Document", "reviewed\_passages": "Sections 4.5.1, 4.5.2, 5.1, 5.2.1, 5.2.2; Executive Summary", "supported\_proposition": "Cloud service providers utilize partnerships with AI developers to secure exclusive cloud spend commitments and engineer contractual and technical switching costs.", "important\_limitation": "The report reflects staff views based on a narrow study of three partnerships, and explicitly states it does not reflect an assessment of illegal conduct or formal antitrust action.", "claim\_ids": \["IC-CLAIM-003"\], "evidence\_lineage": "FTC 6(b) Orders", "snapshot\_path": "null", "sha256": "null", "missingness\_notes": "Local file capture unavailable due to system constraints." }, { "source\_id": "R07-S002", "matched\_ic\_source\_id": null, "title": "2025 State of the Market Report for PJM", "authors": "Monitoring Analytics, LLC", "issuing\_institution": "Independent Market Monitor for PJM", "document\_type": "Market Monitoring Report", "canonical\_url": "https://www.monitoringanalytics.com/reports/PJM\_State\_of\_the\_Market/2025/2025-som-pjm-vol1.pdf", "retrieved\_url": "https://www.monitoringanalytics.com/reports/PJM\_State\_of\_the\_Market/2025/2025-som-pjm-vol1.pdf", "publication\_date": "2026-08-01", "version\_date": "2026-08-01", "effective\_date": "2026-08-01", "accessed\_at": "2026-09-04", "jurisdiction": "United States (PJM Region)", "legal\_or\_policy\_status": "Regulatory Market Assessment", "publication\_status": "Published", "host\_status": "Available", "review\_scope": "Volume 1 Introduction", "reviewed\_passages": "Capacity market analysis, data center impact statements, auction clearing prices", "supported\_proposition": "Data center load growth is the primary driver of unprecedented PJM capacity market price spikes, resulting in massive wealth transfer and systemic capacity shortfalls.", "important\_limitation": "Focuses strictly on the PJM interconnection; may not perfectly reflect capacity conditions in other major grids like ERCOT or CAISO.", "claim\_ids": \["IC-CLAIM-003"\], "evidence\_lineage": "PJM capacity auction data", "snapshot\_path": "null", "sha256": "null", "missingness\_notes": "Local file capture unavailable." }, { "source\_id": "R07-S003", "matched\_ic\_source\_id": null, "title": "Advanced Node Foundry Market Research", "authors": "Market Analysis Aggregation", "issuing\_institution": "Independent Analyst Synthesis", "document\_type": "Market Research Synthesis", "canonical\_url": "null", "retrieved\_url": "null", "publication\_date": "2026-05-01", "version\_date": "2026-05-01", "effective\_date": "2026-05-01", "accessed\_at": "2026-09-04", "jurisdiction": "Global", "legal\_or\_policy\_status": "Industry Analysis", "publication\_status": "Published", "host\_status": "Available", "review\_scope": "TSMC market share, advanced node revenue, HBM constraints", "reviewed\_passages": "5nm/3nm node segment analysis, AI accelerator foundry share", "supported\_proposition": "TSMC maintains a 95% monopoly on AI accelerator foundry production, creating severe physical hardware bottlenecks.", "important\_limitation": "Market share percentages are analyst estimates derived from corporate supply chain mapping, not audited production logs.", "claim\_ids": \["IC-CLAIM-003"\], "evidence\_lineage": "Gartner, TrendForce, Counterpoint Research", "snapshot\_path": "null", "sha256": "null", "missingness\_notes": "Data derived from multi-source snippet synthesis." }, { "source\_id": "R07-S004", "matched\_ic\_source\_id": null, "title": "Big Tech AI Capex in 2026: Microsoft, Google, Meta, Amazon", "authors": "ValueAdd VC / MLQ Agent / Silicon Analysts", "issuing\_institution": "Financial Analysis Aggregation", "document\_type": "Financial Analysis Synthesis", "canonical\_url": "https://valueaddvc.com/blog/big-tech-ai-capex-in-2025", "retrieved\_url": "https://mlq.ai/news/big-techs-2026-capex-range", "publication\_date": "2026-08-02", "version\_date": "2026-08-02", "effective\_date": "2026-08-02", "accessed\_at": "2026-09-04", "jurisdiction": "Global", "legal\_or\_policy\_status": "Market Commentary", "publication\_status": "Published", "host\_status": "Available", "review\_scope": "Capex guidance and trailing cash flow reconstruction", "reviewed\_passages": "Capex range estimates, depreciation lag analysis, debt issuance", "supported\_proposition": "Hyperscalers are committing $720B-$745B to AI infrastructure in 2026, creating a massive depreciation lag funded heavily by corporate debt and operating cash flow.", "important\_limitation": "Different accounting methodologies (finance leases vs cash capex) make exact 1:1 comparisons of AI-specific spending inherently imprecise.", "claim\_ids": \["IC-CLAIM-003"\], "evidence\_lineage": "Corporate 10-Q/10-K filings, Q2 2026 Earnings Calls", "snapshot\_path": "null", "sha256": "null", "missingness\_notes": "Data derived from snippet synthesis." }, { "source\_id": "R07-S005", "matched\_ic\_source\_id": null, "title": "From Efficiency Gains to Rebound Effects: The Problem of Jevons' Paradox in AI", "authors": "Academic Researchers (ArXiv)", "issuing\_institution": "ArXiv Preprint Server", "document\_type": "Academic Preprint", "canonical\_url": "https://arxiv.org/html/2507.00004v2", "retrieved\_url": "https://arxiv.org/html/2507.00004v2", "publication\_date": "2026-01-01", "version\_date": "2026-01-01", "effective\_date": "2026-01-01", "accessed\_at": "2026-09-04", "jurisdiction": "Global", "legal\_or\_policy\_status": "Academic Research", "publication\_status": "Preprint", "host\_status": "Available", "review\_scope": "Abstract and introduction", "reviewed\_passages": "Jevons Paradox application to LLM inference efficiency", "supported\_proposition": "Algorithmic efficiency gains in AI inference trigger the Jevons Paradox, resulting in greater aggregate inference demand and sustained infrastructure consumption.", "important\_limitation": "Theoretical framework mapping historical energy economics to modern compute; empirical quantification of exact elasticity remains difficult to isolate.", "claim\_ids": \["IC-CLAIM-003"\], "evidence\_lineage": "ArXiv Academic Preprints", "snapshot\_path": "null", "sha256": "null", "missingness\_notes": "Snippet synthesis." } \] }

File: reviewed-source-notes.md

Document Note: R07-S001 (FTC 6(b) AI Partnerships Report)

• Narrow Proposition Supported: Cloud Service Providers (CSPs) utilize partnerships with frontier AI developers to secure exclusive cloud spend commitments and generate exceptionally high technical/contractual switching costs, functionally locking developers into proprietary hardware ecosystems.
• Important Limitation: Acknowledge FTC Commissioners Ferguson and Holyoak's explicit dissent indicating that the study was "quick" and "narrow," and that Section 5 relies heavily on forward-looking "areas to watch" rather than confirmed, legally binding anti-competitive harms.
• Exact Passage: "The partnerships offer AI developer partners potential paths to growth... \[but\] could increase contractual and technical switching costs... The partnerships include cloud commitments requiring AI developers to spend a large portion of their CSP partner’s investment on cloud services from that same partner."
• Reviewer Attribution: Synthesized by Research Agent R07; analysis represents an objective extraction of FTC staff assertions, not independent legal judgment.

Document Note: R07-S002 (PJM 2025 State of the Market / Market Monitor Reports)

• Narrow Proposition Supported: The physical electrical grid is acting as a hard, immutable constraint on AI infrastructure scaling, with data center energy demands directly causing an 833% surge in PJM capacity prices to the FERC cap of $329.17/MW-day.
• Important Limitation: Capacity auction prices are forward-looking financial mechanisms designed to incentivize generation; high prices indicate a severe, systemic shortage, but they are theoretically designed to eventually correct supply over long time horizons.
• Exact Passage: "Data center load growth is the primary reason for recent and expected capacity market conditions... inclusion of existing and forecast data center load... resulted in a $7,271,197,971 or an 82.1 percent increase in capacity market revenues for the 2026/2027 RPM Base Residual Auction."
• Reviewer Attribution: Synthesized by Research Agent R07.

Document Note: R07-S004 (Big Tech AI Capex Financial Synthesis)

• Narrow Proposition Supported: Hyperscaler capital expenditures have consolidated into a $720B-$745B run rate for 2026, creating a severe depreciation lag that forces trillion-dollar conglomerates to issue debt, effectively rendering the infrastructure market uncontestable for new entrants.
• Important Limitation: Capital expenditure definitions vary by company; Amazon reports cash capex while Microsoft includes the principal value of newly commenced finance leases, requiring caution when directly comparing top-line figures.
• Exact Passage: "Amazon, Alphabet, Meta and Microsoft expect to record between $720 billion and $745 billion of capital expenditures in 2026... combined 1Q26 capex of $129.8B was up 80% YoY."
• Reviewer Attribution: Synthesized by Research Agent R07.

File: claim-effects.json

JSON { "claim\_id": "IC-CLAIM-003", "baseline\_evidence\_state": "supported\_with\_qualification", "baseline\_adoption\_state": "research\_position", "recommended\_evidence\_state": "supported\_with\_qualification", "recommended\_adoption\_state": "research\_position", "evidence\_effects": \[ "new direct/official evidence", "methodological criticism", "secondary evidence" \], "source\_ids": \[ "R07-S001", "R07-S002", "R07-S003", "R07-S004", "R07-S005" \], "reason": "Empirical data confirms that extreme capital requirements ($745B hyperscaler capex), advanced semiconductor constraints (TSMC's 95% AI foundry share, HBM shortages), and energy chokepoints (PJM capacity spikes to $329/MW-day) sustain deep infrastructure concentration. While algorithmic efficiency (e.g., DeepSeek MLA compression) actively democratizes model software access, it falls victim to the Jevons Paradox, resulting in exponential increases in absolute inference demand that only hyperscale infrastructure can service.", "strongest\_remaining\_objection": "A macroeconomic failure of enterprise AI adoption to generate sufficient Return on Investment (ROI) could rapidly turn hyperscaler capex into stranded assets, causing a collapse in infrastructure concentration due to massive oversupply and depreciation-driven margin compression.", "what\_would\_change": "Empirical evidence demonstrating that high-value, frontier-equivalent inference workloads can be executed efficiently on heavily decentralized, consumer-grade hardware architectures at scale, without invoking massive centralized cloud energy or memory bandwidth requirements.", "proposed\_public\_wording": "Open-weight models and profound algorithmic efficiencies successfully democratize access to software capabilities, but the absolute scale of global AI deployment remains constrained by highly concentrated bottlenecks in capital formation ($745B hyperscaler capex), advanced semiconductor fabrication (TSMC), zero-sum memory allocation, and severe physical limitations in regional energy grid capacity." }

File: search-log.md

• Date of Investigation: September 4, 2026
• Databases and Search Parameters Analyzed:
• FTC.gov: "ftc staff report" "ai partnerships" OR "6b study"
• IEA.org: IEA data centers electricity demand forecast 2025 2026
• Financial / Market Intel: "Nvidia" "market share" AI accelerator GPU data center revenue 2024 2025 2026, TSMC advanced node market share 3nm 5nm AI chip foundry
• Energy / Grid: "PJM Interconnection" data center capacity queue Illinois ComEd power demand
• Countercase / Methodological: DeepSeek V3 R1 paper "Multi-head Latent Attention", "Jevons paradox" AI LLM inference efficiency compute demand paper arXiv
• China Contrast: China "East-to-West" computing Huawei Ascend 910C SMIC advanced node 2025 2026
• Selection Criteria: Prioritized finalized corporate financial 10-K/10-Q reports (trailing Q1/Q2 2026), official regulatory documentation (FTC, PJM Independent Market Monitor), and primary technical papers describing physical compute limits and software compression.
• Excluded Evidence: Unverified forward-looking corporate statements lacking explicit capital commitment; highly speculative energy interconnection queue projects lacking financial "buy bids" or finalized permits.
• Inaccessible Material: Internal corporate contracts detailing exact financial mechanisms of "cloud recapture" remain redacted or private; the investigation relies on proxy statements derived from the FTC 6(b) report.
• Important Unsuccessful Searches: Attempted to find verifiable evidence of structural decentralization of hardware ownership at global scale (e.g., edge-compute overtaking cloud compute by volume); found only software distribution mechanisms (open weights) running primarily on centralized hardware.

File: evidence-manifest.json

JSON { "manifest\_items": \[ { "relative\_path": "R07\_ai-infrastructure-concentration/report.md", "byte\_count": null, "sha\_256": "null", "provenance": "synthetic", "redistribution\_restriction": "none", "description": "Primary comprehensive research report encompassing findings, dependency maps, financial reconstruction, and policy analysis." }, { "relative\_path": "R07\_ai-infrastructure-concentration/sources.json", "byte\_count": null, "sha\_256": "null", "provenance": "synthetic", "redistribution\_restriction": "none", "description": "Machine-readable source array detailing provenance and limitations." }, { "relative\_path": "R07\_ai-infrastructure-concentration/reviewed-source-notes.md", "byte\_count": null, "sha\_256": "null", "provenance": "synthetic", "redistribution\_restriction": "none", "description": "Qualitative annotations and critical limitations on decisive source documents." }, { "relative\_path": "R07\_ai-infrastructure-concentration/claim-effects.json", "byte\_count": null, "sha\_256": "null", "provenance": "synthetic", "redistribution\_restriction": "none", "description": "Impact assessment on IC-CLAIM-003, updating the baseline research position." }, { "relative\_path": "R07\_ai-infrastructure-concentration/search-log.md", "byte\_count": null, "sha\_256": "null", "provenance": "synthetic", "redistribution\_restriction": "none", "description": "Log of investigation queries, boundaries, and excluded materials." }, { "relative\_path": "R07\_ai-infrastructure-concentration/evidence-manifest.json", "byte\_count": null, "sha\_256": "null", "provenance": "synthetic", "redistribution\_restriction": "none", "description": "Self-documenting manifest of generated research bundle artifacts." } \] }

Works cited

1. Decoding DeepSeek: The Engineering Behind V3, R1, and Open, https://bryanchua.com/tech/2025/06/05/decoding-deepseek-ai/

2. DeepSeek-V3 Technical Report \- arXiv, https://arxiv.org/pdf/2412.19437

3. (PDF) From Efficiency Gains to Rebound Effects: The Problem of, https://www.researchgate.net/publication/388460272\_From\_Efficiency\_Gains\_to\_Rebound\_Effects\_The\_Problem\_of\_Jevons'\_Paradox\_in\_AI's\_Polarized\_Environmental\_Debate

4. A Theory of Inference Compute Scaling: Reasoning through ... \- arXiv, https://arxiv.org/html/2507.00004v2

5. Advocating Energy-per-Token in LLM Inference \- arXiv, https://arxiv.org/pdf/2603.20224

6. Big Tech AI Capex Hits $725B in 2026, On Pace for $1T in 2027, https://valueaddvc.com/blog/big-tech-ai-capex-in-2025-microsoft-google-meta-amazon-and-the-spending-race

7. Amazon, Microsoft, Alphabet, Meta Plan $725B AI Capex in 2026, https://aiweekly.co/alerts/amazon-microsoft-alphabet-meta-plan-725b-ai-capex-in-2026

8. Big Tech's 2026 Capex Range Reaches $720 Billion to $745 Billion, https://mlq.ai/news/big-techs-2026-capex-range-reaches-720-billion-to-745-billion/

9. Global Advanced Node Foundry Market (2025-2035), https://www.emergenresearch.com/industry-report/advanced-node-foundry-market

10. Gartner Says Worldwide Semiconductor Revenue Grew 21% in 2025, https://www.gartner.com/en/newsroom/press-releases/2026-01-12-gartner-says-worldwide-semiconductor-revenue-grew-21-percent-in-2025

11. Global Memory Shortage Crisis: Market Analysis and the Potential, https://www.idc.com/resource-center/blog/global-memory-shortage-crisis-market-analysis-and-the-potential-impact-on-the-smartphone-and-pc-markets-in-2026/

12. Google Wind 2026, 10 GW Com Ed Plans, Intersect Power \- EnkiAI, https://enkiai.com/policy-and-regulations/self-direct-procurement-hyperscalers-illinois/

13. What Is PJM? The Grid Behind Your Electric Bill \- Electrac, https://electrac.app/blog/pjm-interconnection-electricity-prices

14. PJM's big new data center plan: Make the states figure it out, https://capitolnewsillinois.com/news/pjms-big-new-data-center-plan-make-the-states-figure-it-out/

15. SMIC AI Chip Strategy 2026: Inside China's 5nm Power Play \- Enki AI, https://enkiai.com/ai-market-intelligence/smic-ai-chip-strategy-2026-inside-chinas-5nm-power-play/

16. China's “Triple Output” AI Strategy: Tripling Chip Production by 2026, https://markets.financialcontent.com/wral/article/tokenring-2025-12-18-chinas-triple-output-ai-strategy-tripling-chip-production-by-2026

17. May 12th, 2026 \- Regulations.gov, https://downloads.regulations.gov/FTC-2026-0298-0004/attachment\_2.pdf

18. Partnerships Between Cloud Service Providers and AI Developers, https://www.ftc.gov/system/files/ftc\_gov/pdf/p246201\_aipartnerships6breport\_redacted\_0.pdf

19. Big Tech's AI Spending to Reach $760 Billion in 2026 \- Statista, https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/

20. Hyperscaler AI Capex 2026: $434B and the D\&A Lag, https://siliconanalysts.com/analysis/hyperscaler-ai-capex-depreciation-wall-2026

21. Cloud Recapture: How Strategic AI Investments Return to Their, https://stefanus.ai/cloud-recapture-how-strategic-ai-investments-return-to-their-investors-as-chip-sales-cloud-commitments-datacenter-revenue-and-collateral-value-the-emerging-financial-architecture-of-the/

22. FTC Issues Staff Report on AI Partnerships & Investments Study, https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study

23. September 2025 \- International Center for Law & Economics, https://laweconcenter.org/updates/september-2025/

24. Nvidia AI Statistics 2026: Revenue, Market Share & GPU Data, https://aibusinessweekly.net/p/nvidia-ai-statistics

25. AI Chips Statistics: Top Trends & Data \- SEO Sandwitch, https://seosandwitch.com/ai-chips-statistics/

26. Advanced Node Foundry Market Size, Share & Growth, 2026-2035, https://www.snsinsider.com/reports/advanced-node-foundry-market-10809

27. Advanced Node Logic Semiconductor Market Size, Share \[2034\], https://www.fortunebusinessinsights.com/advanced-node-logic-semiconductor-market-117775

28. TSMC Lifts 2026 Growth Forecast to \~30% on AI \- Silicon Analysts, https://siliconanalysts.com/analysis/tsmc-2026-growth-forecast-surges-30-percent-ai-3nm-dominance

29. 2025 State of the Market Report for PJM \- Monitoring Analytics, https://www.monitoringanalytics.com/reports/PJM\_State\_of\_the\_Market/2025/2025-som-pjm-vol1.pdf

30. CUB Q and A: Another capacity auction, more bad news–so what, https://www.citizensutilityboard.org/blog/2025/07/31/cub-q-and-a-another-capacity-auction-more-bad-news-so-what-happened/

31. CUB Q\&A: Why is ComEd's electricity price spiking?, https://www.citizensutilityboard.org/blog/2026/06/10/why-is-comed-price-spiking/

32. The Role of the Middle East in the US-China Race to AI Supremacy, https://mei.edu/publication/role-middle-east-us-china-race-ai-supremacy/

33. Implications of SMIC Progress Towards N+3 \- TechInsights, https://library.techinsights.com/analysis-view/FCD-2512-803?q=N+3\&searchKey=header-search\&filters={}

34. Huawei to double output of Ascend AI chips \- RCR Wireless News, https://rcrwireless.com/20250930/ai-infrastructure/huawei-ai-chips-2

35. Huawei, the leader in Chinese semiconductor development… 'Life, https://semiwiki.com/forum/threads/huawei-the-leader-in-chinese-semiconductor-development%E2%80%A6-%E2%80%98life-or-death%E2%80%99-for-smic-5nm-mass-production-next-year.22690/

36. The Great Silicon Pivot: How Huawei's Ascend Ecosystem is, https://investor.wedbush.com/wedbush/article/tokenring-2026-1-7-the-great-silicon-pivot-how-huaweis-ascend-ecosystem-is-rewiring-chinas-ai-ambitions

37. 2026 China Data Center IDC & AI Computing Market, https://faxiangongchang.com/en/reports/china-data-center-ai-2026

38. Insights into DeepSeek-V3: Scaling Challenges and Reflections on, https://arxiv.org/html/2505.09343v1

39. Deepseek-V3 Training Budget Fermi Estimation, https://planetbanatt.net/articles/v3fermi.html