Projecting Compute Spending

Quick Assessment

Overview

The AI infrastructure buildout from 2024-2026 represents a capital investment cycle that exceeds, in nominal dollar terms, the annual spending rates of previous major infrastructure projects. In Q4 2025 and Q1 2026 earnings calls, the five largest hyperscalers — Amazon, AlphabetOrganizationGoogle DeepMindComprehensive overview of DeepMind's history, achievements (AlphaGo, AlphaFold with 200M+ protein structures), and 2023 merger with Google Brain. Documents racing dynamics with OpenAI and new Front...Quality: 37/100, MicrosoftOrganizationMicrosoft AIMicrosoft invested $80B+ in AI infrastructure (FY2025) with a restructured $135B stake (27%) in OpenAI, generating $13B AI revenue run rate (175% YoY growth) and 16 percentage points of Azure's 39%...Quality: 44/100, MetaOrganizationMeta AI (FAIR)Comprehensive organizational profile of Meta AI covering $66-72B infrastructure investment (2025), LLaMA model family (1B+ downloads), and transition from FAIR research lab to product-focused GenAI...Quality: 51/100, and Oracle — collectively announced approximately $700 billion in planned capital expenditures for 2026.¹²³⁹¹⁰ This represents a 58% increase over the $443 billion spent in 2025, which was a 73% increase over 2024.

For two consecutive years, Wall Street consensus estimates for AI capex were substantially lower than actual spending: analysts projected approximately 20% annual growth for both 2025 and 2026, while actual spending exceeded 50% growth in each period.¹¹

To contextualize the $700 billion figure: it represents approximately 2.1% of projected 2026 US GDP flowing from five companies into infrastructure buildout in a single year — more than four times the combined annual capital expenditure of the entire publicly traded US energy sector.¹²

This spending trajectory influences AI governanceParameterAI GovernanceThis page contains only component imports with no actual content - it displays dynamically loaded data from an external source that cannot be evaluated. dynamics: the physical infrastructure being deployed determines the compute available for training frontier models, the geographic distribution of AI capabilities, and the economic incentives shaping competitive dynamics between AI developers.

Historical Context

Comparison to Previous Infrastructure Cycles

Major historical infrastructure investments provide context for the current AI buildout, though direct comparisons require accounting for inflation, project duration, and economic scale differences:

The telecom buildout of 1996-2001 offers the most relevant comparison. During that period, telecommunications companies invested approximately $650 billion (nominal) in fiber optic networks and wireless infrastructure, driven by projected internet growth.¹⁵ The buildout resulted in significant overcapacity, with dark fiber capacity utilization below 3% by 2002.¹⁷ However, the infrastructure eventually supported subsequent internet growth, suggesting that timing and demand forecasting accuracy matter more than whether infrastructure is ultimately utilized.

Key differences from the telecom cycle:

• Concentration: Current AI spending is concentrated among 5-6 companies versus dozens of telecom carriers
• Revenue visibility: Telecom had established usage-based revenue models; AI monetization pathways remain less certain
• Technology obsolescence: AI accelerators face 3-5 year replacement cycles versus 15-20 year telecom infrastructure lifespans¹⁸
• Capital efficiency: Hyperscalers maintain strong balance sheets and operating cash flow; many telecom companies were highly leveraged startups

Quantitative Analysis

Company-by-Company Capex Breakdown

Master Investment Table

Amazon — $200B in 2026

Amazon CEO Andy Jassy announced $200 billion in capital expenditures for 2026 during the company's Q4 2025 earnings call, primarily allocated to AWS infrastructure.¹ This guidance was $50 billion above the Wall Street consensus estimate of $150 billion. AWS posted $35.6 billion in Q4 2025 revenue, representing 24% year-over-year growth (the fastest growth rate in 13 quarters).¹ The company added approximately 4 gigawatts of computing capacity in 2025 and projected doubling that capacity addition by the end of 2027.¹

Alphabet — $185B in 2026

Alphabet provided capex guidance of $175-185 billion for 2026 during its Q4 2025 earnings call, nearly doubling the $91.4 billion spent in 2025 and far exceeding the $52.5 billion spent in 2024.² The midpoint guidance of $180 billion was 55% above the analyst consensus estimate of approximately $119.5 billion.¹¹ CFO Anat Ashkenazi stated that approximately 60% of spending goes to servers and technical infrastructure, with 40% allocated to data centers and networking equipment.² GoogleOrganizationGoogle DeepMindComprehensive overview of DeepMind's history, achievements (AlphaGo, AlphaFold with 200M+ protein structures), and 2023 merger with Google Brain. Documents racing dynamics with OpenAI and new Front...Quality: 37/100 Cloud revenue reached $17.7 billion in Q4 2025, and Alphabet's annual revenues exceeded $400 billion for the first time.²

Microsoft — $145–150B in 2026 (Estimated)

Microsoft has not issued formal full-year guidance for calendar 2026 capex. The company spent $49 billion on capital expenditures in the first half of fiscal year 2026 (July-December 2025), with Q4 calendar 2025 capex at $37.5 billion, representing 65% year-over-year growth.³ Wall Street analysts project full calendar year 2026 estimates ranging from $145 to $165 billion based on current trajectory.¹¹ Microsoft continues to invest alongside OpenAIOrganizationOpenAIComprehensive organizational profile of OpenAI documenting evolution from 2015 non-profit to commercial AGI developer, with detailed analysis of governance crisis, safety researcher exodus (75% of ..., with plans to acquire approximately $135 billion in equity stake in OpenAI as part of a corporate restructuring, while OpenAI has committed to purchasing $250 billion in computing resources from Microsoft through 2030.²⁰

Meta — $115–135B in 2026

Meta announced 2026 capex guidance of $115 to $135 billion during its Q4 2025 earnings call, compared to $72.22 billion in 2025.⁹ CEO Mark Zuckerberg stated the company is "sprinting toward personal superintelligence" and building multiple gigawatt-scale data centers across the United States.⁹ Meta has announced partnerships with Vistra, Oklo, and TerraPower for nuclear energy supply, positioning the company as one of the largest corporate purchasers of nuclear energy globally.²¹ The company's Reality Labs division, which includes AI research, reported $5.8 billion in operating losses in Q4 2025.⁹

Oracle — $50B in FY2026

Oracle revised its fiscal year 2026 capital expenditure guidance upward to $50 billion during its Q3 FY2025 earnings call, representing a dramatic acceleration for a company historically focused on software.¹⁰ To fund this expansion, Oracle announced plans to raise $45–50 billion through combined debt and equity offerings in 2026.²² Oracle is a key partner in the Stargate project alongside OpenAI and SoftBank, with total remaining performance obligations reaching a record $523 billion.¹⁰ However, the company's total debt has increased to approximately $175 billion, and free cash flow turned negative at -$13.1 billion in the most recent quarter.¹⁰

xAI — $30B+ and Accelerating

xAI closed a $20 billion Series E funding round in January 2026, bringing total funding to approximately $30 billion.¹⁹ The company's Colossus facility in Memphis, Tennessee, has expanded to 2 gigawatts of total capacity housing 555,000 Nvidia H200 GPUs, purchased for approximately $18 billion at an estimated $32,000 per unit.²³ This represents the largest single AI training cluster globally as of Q1 2026. The company constructed the initial facility infrastructure in 19 days by building on-site natural gas power generation rather than waiting for utility grid interconnection, which typically requires 2-4 years.²³

OpenAI / Stargate — $500B Commitment by 2029

The Stargate project, a joint venture between OpenAI, SoftBank, and Oracle, announced plans to invest up to $500 billion in AI data center infrastructure in the United States by 2029.²⁴ As of March 2026, the project had secured nearly 7 gigawatts of planned capacity and over $400 billion in committed investment according to project statements.²⁴ The first Stargate data center in Abilene, Texas, became operational in Q4 2025, with five additional facilities under construction as of Q1 2026.²⁴ However, the timeline for full deployment and the binding nature of the $500 billion commitment remain uncertain, with some analysts expressing skepticism about the feasibility of the stated timeline.²⁵

Infrastructure Allocation

CreditSights estimates that approximately 75% of hyperscaler capex — roughly $525 billion in 2026 — is allocated directly to AI infrastructure: GPUs, servers, networking equipment, and data centers specifically designed for AI workloads.²⁶ The remaining 25% covers traditional cloud computing infrastructure, real estate, corporate facilities, and other capital needs. This allocation varies by company, with newer entrants like xAI approaching 95% AI-focused spending while more diversified players like Amazon may allocate 60-70% to AI.²⁶

Supply Chain Impacts

Semiconductor Market Trajectory

Global semiconductor sales reached $791.7 billion in 2025, representing 25.6% year-over-year growth according to the Semiconductor Industry Association (SIA).⁵ The SIA projects sales will exceed $1 trillion in 2026, a milestone arriving approximately four years ahead of earlier industry projections.⁵ McKinsey projects $1.6 trillion in annual semiconductor sales by 2030 in its base case scenario.³¹

Nvidia maintains approximately 90% market share in AI accelerators used for training large language models.³² The company's fiscal year 2025 revenue (ended January 2025) reached $130.5 billion, representing 114% year-over-year growth, with data center revenue comprising $116.2 billion of the total.³³ Wall Street analysts estimate Nvidia's calendar year 2026 revenue at approximately $280-324 billion, maintaining gross margins near 70%.³²³³

Capital Efficiency Metrics

The capital intensity of AI infrastructure varies significantly across companies:

These ratios indicate that all five major hyperscalers are spending more on capex than they generate in operating cash flow, necessitating either debt issuance, asset sales, or reduced shareholder returns. Oracle's 250% capex/OCF ratio is particularly notable, requiring substantial external financing.¹⁰

For context, the capital-intensive energy sector typically maintains capex/revenue ratios of 15-25%, while the telecom sector during its 1996-2001 buildout peaked at approximately 35% capex/revenue.¹⁵³⁴

Depreciation and Replacement Cycles

Hardware depreciation significantly impacts the actual economics of AI infrastructure investments. Most hyperscalers use 3-5 year depreciation schedules for server equipment:³⁵

If the $700 billion in 2026 capex is depreciated over 4 years on average, these companies face $175 billion in annual depreciation expenses starting in 2027, which must be covered by revenue from AI products and services to maintain profitability. Additionally, technological advancement in AI accelerators creates functional obsolescence risk: H100 GPUs from 2023 are already being replaced by H200 and GB200 systems with 2-3x performance improvements, potentially rendering older hardware economically obsolete before the end of its depreciation schedule.³⁶

Funding Mechanisms

The $700 billion in 2026 capex is funded through multiple sources:

The shift from share buybacks to capex represents a significant reallocation of capital. In 2021, these five companies spent $149 billion buying back their own stock; in Q4 2025, combined buybacks fell to $12.6 billion, the lowest level since Q1 2018.⁸ This reallocation implies management teams believe AI infrastructure investments will generate higher returns than returning capital to shareholders, though it also reduces financial flexibility.

Debt issuance is accelerating: combined debt offerings from these companies reached $108 billion in 2025, with projections suggesting $150-200 billion in 2026.²² Oracle's planned $50 billion raise is particularly significant, increasing the company's total debt from $96 billion to approximately $175 billion.¹⁰²² Interest coverage ratios remain healthy for most companies given low debt levels relative to cash flow, but Oracle's negative free cash flow raises questions about debt serviceability if AI revenue growth disappoints.¹⁰

Economic Analysis

Return on Investment Requirements

The central analytical question for this investment cycle: what level of AI revenue is required to justify $700 billion in annual capex, and is that revenue achievable?

ROI Framework

For cumulative $5 trillion in AI infrastructure capex through 2030 to generate a 10% annual return on invested capital (ROIC), the AI industry must produce $1 trillion in incremental annual revenue by 2030, assuming:

• 40% gross margins (below current software SaaS margins of 70-80% but above hardware margins of 20-30%)
• 20% operating margins after R&D and sales costs
• Resulting in $200 billion in annual operating income
• Which, over 5 years, equals approximately $1 trillion in cumulative operating income

JPMorgan's analysis suggests the technology industry must collect an incremental $650 billion in annual revenue — approximately three times Nvidia's entire 2025 revenue — to earn a reasonable investment return on the cumulative capex.⁶ This calculation assumes:

• Some existing revenue is already captured by these companies
• Additional revenue must come from new AI products and services
• Revenue must be incremental (not cannibalized from existing businesses)

Alternative ROI Scenarios

Revenue Pathway Analysis

For context on the $1 trillion annual revenue target, current AI revenue data points include:

Current 2025 AI revenue of approximately $90 billion would need to grow to $1 trillion by 2030, implying a compound annual growth rate (CAGR) of approximately 61% — faster than cloud computing's 40% CAGR during 2010-2020 but slower than mobile app revenue's 85% CAGR during 2008-2015.⁴⁰

Potential Revenue Sources

Multiple revenue pathways could contribute to the $1 trillion target:

Enterprise productivity enhancement: If AI tools improve knowledge worker productivity by 20% and are priced to capture 30% of that value, the addressable market in the US alone (100M knowledge workers × $80K average compensation × 20% improvement × 30% capture rate) equals approximately $480 billion annually.⁴¹

Consumer AI services: ChatGPT reached 200 million weekly active users in late 2025.⁴² If consumer AI reaches 1 billion paying users globally at $20/month average, that represents $240 billion in annual revenue.

Autonomous systems: Self-driving technology, robotics, and automation represent large potential markets, though deployment timelines remain uncertain. McKinsey estimates the autonomous vehicle market could reach $300-400 billion by 2035.⁴³

Drug discovery and materials science: AI-accelerated research in pharmaceuticals and materials could generate substantial value, though revenue recognition may be delayed by regulatory approval timelines.

Code generation: GitHub Copilot and similar tools are used by 1.8 million developers as of Q4 2025.³⁹ If 30 million of the world's ~50 million developers adopt AI coding tools at $30/month, that represents $11 billion annually.

However, several factors complicate revenue projections:

1. Cannibalization: Some AI revenue may cannibalize existing software and services revenue (e.g., AI-powered customer service reducing human support costs but also reducing revenue for customer service software vendors)

2. Commoditization pressure: Open-source models and competition may drive per-query pricing down faster than volume increases

3. Adoption friction: Enterprise adoption faces organizational change management barriers beyond technological readiness

4. Regulatory constraints: Data privacy regulations, copyright issues, and potential AI-specific regulations may limit deployment

5. Market concentration: If a small number of applications (e.g., ChatGPT, coding assistants) generate most revenue, the broad infrastructure buildout may exceed actual demand

Bull and Bear Cases

Optimistic Scenario

Arguments supporting revenue achievement:

• Horizontal technology: AI applies across industries. If AI adds 2% to revenue for the Fortune 500 (combined revenue ≈$15 trillion), that generates $300 billion in economic value, a portion of which flows to AI providers.⁴⁴
• Workforce cost displacement: If AI reduces US workforce costs by 3% across sectors where it applies (estimated ≈$5 trillion in applicable labor costs), that represents $150 billion in savings, creating willingness-to-pay for AI services.
• Recurring revenue: Subscription models and API usage create recurring revenue streams versus one-time software licenses.
• Platform effects: Companies that establish AI platforms benefit from network effects and ecosystem lock-in.
• Replacement cycles: Hardware obsolescence creates recurring capex demand, benefiting semiconductor manufacturers and extending the investment cycle.
• International expansion: Current revenue projections focus primarily on US/Europe; expansion to developing markets could add substantial volume.

Pessimistic Scenario

Arguments suggesting revenue targets may not materialize:

• Current losses: OpenAIOrganizationOpenAIComprehensive organizational profile of OpenAI documenting evolution from 2015 non-profit to commercial AGI developer, with detailed analysis of governance crisis, safety researcher exodus (75% of ... projects losses exceeding $14 billion in 2026 despite $30 billion in projected revenue, and cumulative losses potentially exceeding $100 billion through 2030.³⁷ This suggests unit economics may not support profitable scale.
• Rapid obsolescence: 3-5 year hardware replacement cycles require rapid payoff; if revenue growth doesn't materialize before next-generation hardware arrives, stranded asset risk increases.
• Concentration risk: If a small number of applications dominate AI revenue, infrastructure may be overbuilt relative to actual demand concentration.
• Margin compression: Intense competition and open-source alternatives may drive AI service pricing down faster than volume grows.
• Demand uncertainty: Consumer willingness-to-pay for AI services beyond novelty phase remains unclear; enterprise adoption may plateau at lower levels than projected.
• Regulatory intervention: Antitrust concerns about hyperscaler dominance, data privacy regulations, or copyright disputes could limit monetization pathways.
• Technology plateau: If capability improvements slow, the value proposition may not support premium pricing.

Historical Comparison: Telecom Bubble Outcomes

The 1996-2001 telecom buildout provides instructive parallels and contrasts:

Similarities:

• Massive infrastructure investment driven by projected demand growth
• Wall Street analyst enthusiasm and upward revision of forecasts
• Technology providers (Cisco, Lucent, Nortel) experiencing explosive revenue growth
• Debt-fueled expansion by infrastructure providers
• Competition driving aggressive buildout to secure market position

Differences:

• Telecom featured many competitors; AI infrastructure is concentrated among 5-6 players with stronger balance sheets
• Telecom had clear usage-based revenue model; AI monetization pathways less established
• Telecom equipment had 15-20 year lifespans; AI hardware faces 3-5 year obsolescence
• Regulatory environment was deregulatory (Telecommunications Act of 1996); AI faces potential increased regulation

Outcomes:

• Telecom carriers invested $650 billion (nominal) during 1996-2001¹⁵
• Resulted in $2 trillion in destroyed market value during 2000-2002 crash¹⁵
• Dark fiber utilization fell below 3% by 2002; massive overcapacity¹⁷
• However, infrastructure eventually supported subsequent internet growth (cloud computing, streaming, mobile)
• Suggests timing and demand forecasting accuracy matter more than whether infrastructure is ultimately useful

Energy and Power Constraints

AI infrastructure at the projected scale requires substantial electrical power generation and distribution capacity. This constraint may limit the pace of actual deployment regardless of capital availability.

Power Demand Projections

For context, 1 gigawatt serves approximately 700,000 US homes.⁴⁷ The projected 15-20 GW of incremental 2026 demand equals the residential power consumption of 10-14 million homes, or roughly the entire state of Pennsylvania.

Grid Interconnection Queue

The US power grid interconnection queue — the backlog of projects waiting for permission to connect to the grid — reached 2,600 gigawatts as of Q4 2025, with average wait times of 4-5 years.²⁹ Data center projects comprise approximately 15% of the queue, or ~390 GW.²⁹ This queue length substantially exceeds the physical construction timelines for data centers themselves (18-24 months), making power connection the critical path for many projects.

Key constraints:

• Transformer shortages: Lead times for large power transformers extend to 2-3 years⁴⁸
• Transmission capacity: Many regions lack transmission capacity to deliver power to proposed data center locations
• Permitting delays: Environmental reviews and local opposition extend project timelines
• Utility planning cycles: Utilities plan major projects 5-10 years ahead, creating mismatch with rapid AI infrastructure timeline

Power Supply Strategies

Hyperscalers are pursuing multiple strategies to address power constraints:

Nuclear Energy Partnerships

Nuclear partnerships face substantial uncertainty:

• Regulatory approval timelines are unpredictable (Three Mile Island restart requires NRC approval)
• Small modular reactor (SMR) technology is unproven at commercial scale
• Capital costs are high ($6-10 billion per 1000 MW for new nuclear)⁵²
• Public opposition and safety concerns may delay or block projects

On-Site Gas Generation

xAI's approach — building on-site natural gas power generation — bypasses grid interconnection delays but faces other constraints:

• Requires natural gas pipeline access
• Subject to air quality permits (can take 12-18 months)⁵³
• Environmental groups oppose new fossil fuel infrastructure
• Economics depend on natural gas pricing (currently $3-4/MMBtu but historically volatile)⁵⁴

As of Q1 2026, xAI's 2 GW Colossus facility operates on on-site gas generation, while the company plans to transition partially to renewable energy for future expansion.²³

Renewable Energy Contracts

Hyperscalers have collectively contracted for approximately 50 GW of renewable energy capacity (wind and solar) through power purchase agreements.⁵⁵ However, renewable energy faces intermittency challenges: AI training workloads require continuous power, while solar and wind generation varies. Battery storage technology is improving but adding 4-12 hours of storage at multi-gigawatt scale remains expensive ($300-500/kWh for utility-scale batteries).⁵⁶

Water Consumption

Data center cooling consumes substantial water, particularly for liquid cooling systems increasingly required for AI accelerator density. Estimated water consumption:

• Air cooling: ~1-2 liters per kWh
• Evaporative cooling: ~4-6 liters per kWh
• Direct liquid cooling: ~0.5-1 liters per kWh

A 1 GW data center operating at 50% capacity factor using evaporative cooling consumes approximately 17-26 million liters per day, or 6-10 billion liters annually — equivalent to the daily water consumption of 140,000-200,000 people.⁵⁷

Water constraints are particularly acute in:

• Arizona: Data center construction faces increasing water permit scrutiny⁵⁸
• Northern Virginia: Loudoun County (data center capital) experiencing water stress⁵⁹
• Texas: Drought conditions limiting new data center projects in some regions⁶⁰

Companies are investing in water recycling and air cooling alternatives, but these increase capital and operating costs.

Strategic Implications for AI Governance

The magnitude and pace of AI infrastructure investment creates several dynamics relevant to AI governanceParameterAI GovernanceThis page contains only component imports with no actual content - it displays dynamically loaded data from an external source that cannot be evaluated. and AI safety:

Capability Acceleration and Governance Windows

The infrastructure being deployed in 2026 will determine the compute available for frontier model training through 2028-2030. If hardware depreciation and replacement cycles are 3-5 years, the $700 billion in 2026 capex establishes the training capacity ceiling until 2029-2031.

This timeline matters for governance: if AGI or highly capable AI systems emerge within this infrastructure generation, the governance window for implementing safety measures is constrained to the period before these systems are trained. Physical infrastructure creates a degree of predictability: unlike software which can scale instantly, GPU clusters and data centers require 18-36 months from planning to operation.

Racing Dynamics and Safety Incentives

The scale of capital deployed creates pressure to generate returns rapidly, which may incentivize:

• Accelerated deployment of AI systems to begin revenue generation
• Reduced investment in safety testing and alignment research relative to capability development
• Competitive pressure where no individual company wants to slow down unilaterally

However, countervailing factors exist:

• Companies face reputation and regulatory risk from safety failures
• Safety incidents could trigger restrictive regulation, harming long-term business interests
• Major AI developers have made public commitments to safety research

The net effect remains uncertain and likely varies by company and competitive context.

Geographic Concentration and Geopolitical Dynamics

Approximately 60-70% of the $700 billion in 2026 hyperscaler capex is being invested in United States-based infrastructure, according to data center industry reports.⁶¹ This concentration has several implications:

• US dominance: Reinforces US position as primary location for frontier AI development
• Allied access: Close US allies (UK, Canada, EU) can access US-based AI systems but may face latency or data sovereignty constraints
• China considerations: China's AI infrastructure investment is substantial but not fully transparent; estimates suggest $150-200 billion in annual AI capex across Alibaba, Tencent, Baidu, and ByteDance⁶²
• Export controls: US government export controls on advanced semiconductors (export controlsPolicyUS AI Chip Export ControlsComprehensive empirical analysis finds US chip export controls provide 1-3 year delays on Chinese AI development but face severe enforcement gaps (140,000 GPUs smuggled in 2024, only 1 BIS officer ...Quality: 73/100) limit China's access to frontier AI hardware, though effectiveness remains debated⁶³

Compute Governance Leverage Points

Physical infrastructure creates potential governance intervention points:

1. Semiconductor manufacturing: TSMC produces >90% of advanced logic chips (3nm and below); single chokepoint for supply²⁸
2. HBM memory: Three suppliers (SK Hynix, Samsung, Micron) control production
3. Data centers: Physical facilities are regulatable at local, state, and national levels
4. Power supply: Energy regulators can influence or restrict data center power access
5. Chip design tools: EDA software (Cadence, Synopsys, Siemens) required for chip design

These chokepoints enable potential governance interventions such as:

• Compute allocation requirements (e.g., mandating safety testing access)
• Export controls and international agreements
• Energy policy tied to AI safety standards
• Licensing requirements for large-scale training runs

However, chokepoint governance faces challenges:

• International coordination difficulty
• Enforcement complexity
• Risk of stifling beneficial AI development
• Potential for domestic political opposition from industry

Governance Recommendations

Based on the spending trajectory and its implications, several governance approaches merit consideration:

1. Compute governancePolicyCompute GovernanceThis is a comprehensive overview of U.S. AI chip export controls policy, documenting the evolution from blanket restrictions to case-by-case licensing while highlighting significant enforcement cha...Quality: 58/100 frameworks: Establish international norms for responsible compute allocation and monitoring of large training runs

2. Safety-conditional deployment standards: Create industry standards (potentially with regulatory backstop) requiring safety testing proportional to system capabilities

3. Transparency requirements: Mandate disclosure of training compute used, model capabilities, and safety testing results for frontier systems

4. International coordination: Develop agreements among major AI nations to prevent race-to-the-bottom dynamics on safety standards

5. Power supply policy: Link data center power allocation to AI safety commitments and monitoring compliance

6. Antitrust analysis: Evaluate whether compute concentration among 5-6 companies creates governance leverage or reduces competition-driven safety corners-cutting

These recommendations are not universally endorsed and face substantial implementation challenges.

Structural Power and Adversarial Risks

For comprehensive analysis, see Compute ConcentrationRiskCompute ConcentrationAll six major AI infrastructure spenders (Amazon, Alphabet, Microsoft, Meta, Oracle, xAI) are US companies subject to CLOUD Act and FISA 702, giving the US government effective legal access to the ...Quality: 70/100, which covers:

• US jurisdictional monopoly (CLOUD Act, FISA 702) over global AI compute
• Cybersecurity and systemic vulnerability of concentrated infrastructure
• Military-commercial dual use and the trust problem it creates
• Collusion dynamics and historical precedents (DRAM, LCD, no-poach)
• Extreme concentration scenarios and democratic accountability gaps

The spending data documented above reveals a structural fact that extends beyond economics: all six major AI infrastructure spenders are US-headquartered companies subject to US federal jurisdiction. The CLOUD Act and FISA Section 702 give the US government effective legal authority to compel access to data — including model weights, training data, and inference queries — processed through any of these companies' infrastructure, anywhere in the world. No contractual data residency provision overrides these statutes.

This jurisdictional monopoly, combined with $700 billion in annual capex concentrated in 5-6 entities, creates a set of interlocking risks:

These dynamics are analyzed in depth on the dedicated Compute ConcentrationRiskCompute ConcentrationAll six major AI infrastructure spenders (Amazon, Alphabet, Microsoft, Meta, Oracle, xAI) are US companies subject to CLOUD Act and FISA 702, giving the US government effective legal access to the ...Quality: 70/100 risk page, which examines how the physical, legal, and economic characteristics of compute infrastructure create risks qualitatively different from those posed by software or data concentration.

International Context

China AI Infrastructure Investment

Chinese technology companies are investing heavily in AI infrastructure, though comprehensive data is less transparent than US public company disclosures:

China's total estimated AI infrastructure spending of $145-180 billion is lower than the US total but still substantial. Key differences:

• Domestic chip production: US export controls limit access to Nvidia H100/H200; Chinese companies rely on Huawei Ascend chips, SMIC-produced chips, and smuggled/stockpiled Nvidia GPUs⁶³
• Performance gap: Domestic Chinese chips lag Nvidia by approximately 1-2 generations in performance/watt⁶⁴
• Government coordination: Chinese government provides targeted subsidies and policy support for AI infrastructure⁶⁵
• Data advantages: Access to large domestic user base for training data

European Union Investment

EU AI infrastructure investment lags both US and China:

• Estimated $30-40 billion in combined public and private AI infrastructure investment in 2026⁶⁶
• Focus on sovereign AI capabilities and regulatory frameworks (AI Act)⁶⁷
• Smaller hyperscaler presence; reliance on US cloud providers
• Energy constraints particularly acute given natural gas dependency

The EU's primary contribution to AI governance is regulatory rather than infrastructural, with the AI Act establishing risk-based regulation frameworks.⁶⁷

Limitations and Uncertainties

This analysis faces several methodological limitations:

Data Quality and Availability

• Guidance vs. execution: Capex guidance from earnings calls represents plans, not binding commitments; actual spending may deviate
• Allocation uncertainty: The 75% AI infrastructure allocation is an estimate; actual allocation varies by company and quarter
• Private companies: xAI, OpenAI, and other private companies do not disclose comprehensive financial data
• International data: China capex estimates are particularly uncertain due to limited transparency

Revenue Projection Uncertainty

• Adoption rates: Enterprise and consumer AI adoption trajectories are highly uncertain
• Pricing dynamics: Competitive pressure and commoditization may impact achievable prices
• Cannibalization: Difficult to estimate how much AI revenue replaces existing revenue
• Application emergence: New AI applications may emerge that are not currently predictable

Technological Uncertainty

• Efficiency improvements: Algorithmic improvements may reduce compute requirements faster than anticipated
• Alternative architectures: New chip architectures or training methods may change economics
• Capability plateau: AI capabilities may plateau before reaching projected levels
• Breakthrough scenarios: Alternatively, capability breakthroughs may accelerate timelines

Power and Energy Assumptions

• Grid interconnection: Actual timelines may differ from projections
• Nuclear deployment: SMR technology and regulatory approval remain uncertain
• Renewable integration: Battery storage and grid flexibility may improve faster or slower than assumed
• Water constraints: May limit deployment in unexpected regions

Analytical Limitations

• This analysis focuses on economic and infrastructure dimensions; does not fully assess technical AI safety considerations
• Historical comparisons (telecom bubble, etc.) are imperfect analogies; AI may follow different dynamics
• Scenario analysis is not exhaustive; other outcomes are possible
• Governance recommendations are preliminary and require deeper policy analysis

References