AI Infrastructure:
The $690 Billion Sprint

01 Scale of the Buildout

AI infrastructure spending has entered a phase with no historical parallel. The combined 2026 capex of Amazon, Microsoft, Google, Meta, and Oracle — approximately $690 billion — exceeds the GDP of Switzerland. It represents a 56% increase over 2025 and a 4x increase from 2022 levels.

This is not speculative R&D spending. It is industrial-scale deployment of compute, power, cooling, and networking infrastructure to serve inference workloads that are already generating revenue. The shift from model training (concentrated, episodic) to inference serving (distributed, continuous) has fundamentally changed the economics — and the capital intensity is permanent.

02 Who Is Spending What

The spending is concentrated but the strategies diverge. Each hyperscaler is making distinct bets on compute architecture, power sourcing, and market positioning.

A critical detail: Microsoft has disclosed an $80 billion Azure backlog that it cannot fulfill due to power constraints. GPUs are sitting in warehouses waiting for data centers to be energized. This is not a demand problem — it is a supply problem, and it defines the investment opportunity.

Financing the buildout

Hyperscalers raised $108 billion in debt during 2025 alone, with projections suggesting $1.5 trillion in cumulative debt issuance will be needed over coming years. Capital intensity has reached 45-57% of revenue — historically unprecedented for technology companies that traditionally ran at 10-15%. This creates a multi-decade infrastructure financing opportunity similar to telecom in the 1990s, but at 5-10x the scale.

03 The Power Bottleneck

Power has replaced GPUs as the binding constraint on AI infrastructure deployment. The grid cannot keep up with demand, and the gap is widening.

The numbers tell a stark story. AI racks require 30-60 kW each — 10-20x the power density of traditional server racks. A single large AI training cluster consumes 100+ MW, equivalent to powering a small city. The IEA projects data center electricity consumption will surge from 460 TWh in 2024 to over 1,000 TWh by 2026 and 1,300 TWh by 2035.

Grid connection crisis

• Virginia (Loudoun County): The world's largest data center market. Power connection waiting lists stretch to 2028-2029. New major connections are effectively closed.
• Texas: Some developers are required to build substations costing $100M+ just to connect to the grid.
• Ireland: EirGrid has imposed a moratorium on new data center grid connections in Dublin until 2028.
• Power transformer lead times: 128 weeks and growing. The bottleneck isn't just generation — it's the physical grid infrastructure to deliver power.

04 GPU Economics & Supply

NVIDIA dominates the AI accelerator market with approximately 80-90% market share in data center GPUs. The H100/H200 and B100/B200 series are the de facto standard for both training and inference workloads.

The inference economics shift

The market is transitioning from training-dominated (large, concentrated clusters, episodic usage) to inference-dominated (distributed, continuous, cost-sensitive). This changes the hardware calculus:

• Training: Performance is paramount. Customers pay premium for the fastest chips. NVIDIA's moat is widest here.
• Inference: Cost-per-token matters most. AMD, custom ASICs (Google TPU, AWS Trainium), and even older GPUs become competitive. This opens the market.

By 2027, inference is projected to represent 70-80% of AI compute spending, up from ~50% today. This structural shift creates opportunities for alternative hardware providers and for operators who can optimize inference workloads on mixed hardware fleets.

05 Data Center Market Structure

The global data center market has bifurcated into two distinct segments with different economics and competitive dynamics.

Hyperscale (self-built)

Amazon, Google, Meta, and Microsoft are increasingly building their own data centers rather than leasing from third parties. This vertical integration is driven by the need for custom power, cooling, and networking architectures optimized for AI workloads. Hyperscale self-build now represents the majority of new capacity additions.

Colocation / third-party

Companies like Equinix, Digital Realty, CyrusOne, and QTS serve enterprises that lack the scale or expertise to self-build. This segment is consolidating rapidly, with three major dynamics:

• Power scarcity premium: Colo operators with secured power capacity are seeing record-high lease rates — up 20-40% in power-constrained markets.
• AI-ready requirements: Customers increasingly demand liquid cooling, high-density racks (30+ kW), and 100G+ networking. Retrofitting existing facilities is expensive.
• Long-term contracts: Average lease terms have extended from 3-5 years to 7-10+ years as enterprises lock in capacity. This improves cash flow visibility but reduces flexibility.

06 The Nuclear Pivot

With grid connections constrained and renewable intermittency limiting AI workloads that require 24/7 baseload power, hyperscalers are making aggressive moves into nuclear energy — both existing plants and new technologies.

The logic is straightforward: a 1 GW nuclear plant provides 8,760 GWh/year of baseload power at a marginal cost of $20-30/MWh — enough to power 10-15 large AI data center campuses indefinitely. Compare this to grid power at $50-120/MWh with zero availability guarantees. Nuclear is the only technology that can provide gigawatt-scale, 24/7, carbon-free power for AI.

07 Where Value Accrues

Not all AI infrastructure spending creates equal value. The $690B is distributed across a supply chain with very different margin profiles and competitive dynamics.

The value chain shows a clear pattern: the higher the margin, the harder the moat. NVIDIA's 75% gross margin reflects a near-monopoly in AI training silicon. Power infrastructure margins are lower but access is increasingly scarce. Construction is commoditized but volume is enormous.

08 Revenue & Return Framework

The AI infrastructure buildout creates investable opportunities across multiple time horizons and risk profiles. Below is a framework for evaluating returns across the value chain.

09 Risk Analysis

10 Outlook & Positioning

The AI infrastructure buildout is a multi-decade trend, not a cycle. Even if AI model improvement slows, the existing trajectory of inference workload growth, enterprise migration, and applications deployment will sustain infrastructure demand well into the 2030s.

12-month view

The fundamental insight: $690 billion in annual spending creates a supply chain where the constraining resources — power, cooling, land with grid access — appreciate faster than the compute itself. The investment opportunity is not in buying GPUs; it is in owning the infrastructure that GPUs depend on to operate.