Will AI CapEx Pay for Itself?

And yet, even in this best case — the one where AI becomes the most transformative technology since the internet, where enterprise adoption follows the fastest S-curve in history, where revenue compounds above 60% annually through 2027 — the infrastructure capex creates a financial hole so deep that it takes a decade to climb out. The outcome depends not just on revenue growth but on a second, independent variable: capital intensity — how many infrastructure dollars are needed per dollar of AI revenue. Under my central case (moderate revenue growth, base efficiency), the gross infrastructure return (GP / cumulative capex — see Section 7.2 for methodology) reaches only 8.0% by 2035, corresponding to an operating-level return of roughly 4-5% after estimated opex. Depreciation breakeven occurs around 2034. Cumulative gross profit minus cumulative capex remains negative (−

The infrastructure cost is borne primarily by hyperscalers, but increasingly by the foundation model companies themselves. Anthropic and OpenAI are no longer pure compute renters — Anthropic has announced plans to spend ~$50B on infrastructure over five years (Fluidstack partnership, 2025; Reuters) while maintaining cloud arrangements with AWS, Google Cloud, and Azure; OpenAI launched the $500B Stargate project (SoftBank/Oracle/MGX) alongside continued Azure usage. Both run hybrid models: owned infrastructure for training, cloud for burst capacity and inference. The application layer — Cursor, Harvey, Glean — remains capital-light, renting API access and earning software margins. The value created by AI flows upward through the stack, but the capital required to create it is concentrated at the infrastructure and model layers.

For the hyperscalers to earn an adequate return on this investment, it is not enough for their AI cloud businesses to grow. They need the AI wave to lift their entire existing businesses — advertising, enterprise software, e-commerce — in ways that go beyond what is currently measurable. So far, the clearest beneficiary has been the advertising segment, through two distinct channels. First, a supply-side effect: AI demonstrably improves ad targeting, ranking, and conversion. Meta’s end-to-end automated ad solutions reached $60 billion in annualized throughput by late 2025 (Q3 2025 prepared remarks), with its GEM ranking model driving a 3.5% lift in ad clicks on Facebook and >1% conversion gain on Instagram (Q4 2025 earnings); advertisers enabling Advantage+ creative features reported a 22% improvement in return on ad spend (Meta business materials). Google’s Smart Bidding Exploration delivered a 19% average increase in conversions for participating campaigns (Google Ads materials). Second, a demand-side effect: the AI investment boom has created a wave of VC-funded startups with massive customer acquisition budgets, and a significant share of that marketing spend flows to Google and Meta’s ad platforms — AI companies buying ads, not AI making ads better. Both effects are real, though the supply-side channel appears larger and more durable. The AI infrastructure rental business itself is the other clear beneficiary. Whether AI drives transformative growth across the full breadth of hyperscaler operations remains the open question.

The most important finding: capital intensity matters more than revenue for determining whether AI investment pays off. Low intensity with strong revenue achieves capex payback by ~2031. High intensity never breaks even regardless of revenue. But in a bust, the choice of intensity barely matters — contractual commitments (datacenter leases, GPU purchase agreements, power contracts) create a spending floor that forces hyperscalers to keep deploying capital regardless of demand. Even Low intensity doesn’t pay back until ~2033 in Scenario C, because the transition floor locks in years of spending that far exceeds what weak revenue would justify. NVIDIA and the hyperscalers sit on opposite sides of this dimension — NVIDIA benefits from high intensity (more GPU purchases), while hyperscalers benefit from low intensity (same revenue from less spending). Every dollar of efficiency that improves cloud AI profitability is a dollar NVIDIA doesn’t collect.

Crucially, an adequate infrastructure ROIC does not guarantee adequate shareholder returns. Front-loaded losses and back-loaded gains destroy net present value: at a 10% cost of capital, every cell in the 3×3 matrix has negative NPV through 2035. Even after accounting for terminal value, only A×Low — the single most optimistic cell — produces a positive NPV (+$324B). The combined Big 4 hyperscaler market caps embed an estimated ~$3 trillion AI premium; the model’s best-case NPV covers only ~11% of that.

The market prices for AI-related equities reflect a level of optimism that goes beyond even the most bullish combination. NVIDIA at $4.45 trillion implies a probability-weighted average annual cash yield of 1.2% over the next decade — below the risk-free rate on Treasury bonds in every cell of the 3×3 matrix. But a market capitalization is not a consensus valuation. It is a marginal price set by the last buyer, multiplied by total shares outstanding. The bulk of NVIDIA’s shareholders built their positions years ago at far lower prices and are sitting on enormous unrealized gains. A $4.45 trillion market cap does not mean that every market participant believes NVIDIA is worth $4.45 trillion. It means the marginal buyer — who may represent a small fraction of daily volume — is willing to pay that price, and the existing holders have no urgency to sell. This distinction matters for interpreting what the market is “saying” about AI valuations.

What follows is the data, the model, and the projections. I lay out every assumption explicitly so the reader can adjust any input and see how the conclusions change.

Part I: The Known Data

1.1 AI Company Revenue — What Can Be Verified

The foundation model companies have disclosed enough data to construct a reliable revenue trajectory. All figures below are annualized revenue run rates unless noted — not classic SaaS ARR (recurring contracts), since much of this revenue is usage-based API spend, which may be more volatile. Sources are industry reporting (Bloomberg, Reuters, tech press) and company communications; these are not audited GAAP disclosures.

Anthropic

Additional details (sourced from conference remarks and industry reporting — not audited disclosures):

• $6B added in February 2026 alone (Dario Amodei, Morgan Stanley TMT Conference)
• Claude Code: reported

Steady-state benchmarks:

• High (0.5x): AI remains 25% more capital-intensive than cloud. Training of frontier models continues to require massive concentrated compute; new architectures prevent full efficiency gains.
• Base (0.4x): Matches AWS steady-state. Normal technology maturation path.
• Low (0.25x): Aggressive efficiency. Per-token inference costs decline 5-10x/year (see Section 5.3 decomposition), custom silicon dramatically reduces per-unit costs, and this efficiency translates to actual capex reduction rather than being reinvested in more capacity.

2024-2027 capex is committed and fixed across all combinations:

Scope: Hyperscaler-only (Google, Microsoft, Meta, Amazon). Excludes neocloud capex (CoreWeave, Oracle AI, Lambda, etc.) — see Section 6.3 for neocloud comparison.

From 2028 onward: AI CapEx = max(AI Revenue × Intensity, 70% of prior year’s AI CapEx)

The 30% maximum YoY decline cap reflects the physical reality of infrastructure commitments. Datacenter construction takes 18-24 months, GPU purchase agreements are signed quarters in advance, and power contracts span multiple years. Hyperscalers cannot cut spending by 50-70% in a single year, even if they wanted to. The intensity formula phases in over 2028-2030 as committed contracts roll off — it becomes the binding constraint once spending has declined enough that contractual momentum no longer matters.

This transition rule has minimal impact on high-revenue or high-intensity combinations (A×High capex is unaffected) but significantly raises capex for low-revenue and low-intensity combinations. In the extreme case (C×Low), the floor binds for six consecutive years (2028-2033), adding $732B of capex above what the pure formula would produce. This captures the capex trap: in a bust, hyperscalers are locked into spending that far exceeds demand, precisely because the commitments were made 2-3 years earlier when the outlook was brighter

Part IV: The Revenue Model (Bottom-Up)

4.1 ARR vs Actual Revenue — A Critical Distinction

ARR (Annualized Run Rate) is the exit-month revenue × 12. Actual revenue is the sum of all months’ revenue during the year. For fast-growing companies, these diverge significantly because revenue ramps through the year.

Example: OpenAI started 2025 at ~$500M/month ($6B ARR) and ended at ~

4.3 Revenue Double-Counting Note

Revenue flows through multiple layers and gets counted at each level:

End user pays Cursor: 

The capex intensity numbers provide an additional macro check: at B×Base (

Scope note: These figures are hyperscaler-only (Google, Microsoft, Meta, Amazon). Total industry AI capex including neoclouds is ~8-10% higher — see Section 6.3.

5.2 Depreciation Schedules

AI infrastructure comprises multiple asset classes with different useful lives, as disclosed in hyperscaler 10-K filings:

Disclosed Useful Lives by Hyperscaler

Asset Composition Estimates (derived from 10-K property disclosures and industry analysis)

Sources: Meta reports 41.5% servers/network and 28.6% buildings in PP&E. Microsoft reports 44% compute and 44% buildings. Amazon reports 29% servers and 31% land/buildings. Google reports 56% technical infrastructure. Industry estimates (McKinsey, Goldman Sachs) suggest ~60% IT equipment for AI-specific datacenter builds. The 50/10/35/5 split used here is a blended estimate across hyperscalers, weighted toward AI-specific composition (higher server share than corporate-average PP&E).

Under blended depreciation, each capex vintage depreciates at 14.0% per year (years 1-5) vs 20.0% under 5-year straight-line — a 30% reduction in early-year depreciation. However, building depreciation (2.3%/year) persists for 15 years, creating a long tail. This report uses 5-year straight-line as the conservative base case; blended depreciation is shown for comparison throughout.

B×Base — Central Case

† Transition floor binding: 2028 capex = max(

Peak 5yr SL depreciation: $385B/year in 2029. Peak blended:

Three structural forces drive intensity decline from the committed-period peak:

1. Inference cost per token at fixed capability is declining 5-10x/year. But this headline number requires decomposition. a16z coined “LLMflation” by comparing the cheapest model at a given MMLU score over time ($60/Mtok for GPT-3 in Nov 2021 → $0.06/Mtok for Llama 3.2 3B today — 1,000x in 3 years). Epoch AI finds a wider range: 9x to 900x/year depending on capability level, with a median of ~50x/year. An academic decomposition (arxiv 2511.23455) isolates the drivers:

   • Hardware price-performance: ~1.3x/year (30% improvement per Moore’s Law trajectory)
   • Algorithmic efficiency: ~3x/year (quantization, distillation, speculative decoding, architecture improvements)
   • Economic competition + model shrinkage: the remainder — smaller open-source models competing with proprietary ones, and API pricing wars compressing margins

   Critical caveat (Jevons paradox): The same paper finds that absolute benchmarking costs stayed flat or increased despite per-token price collapse, because achieving higher performance requires larger models with longer reasoning traces. Cost per token falls, but tokens per task rises. This means per-token efficiency gains do NOT translate 1:1 into capex intensity reduction — users consume more compute as it gets cheaper, partially offsetting the efficiency gain at the infrastructure level.

2. Capital stock maturity: Early years require greenfield data center construction (land, power, cooling, networking) — a one-time cost. Later years are incremental GPU refreshes within existing facilities. The marginal cost of adding compute declines sharply.

3. Workload shift from training to inference: Training requires massive, concentrated compute bursts (thousands of GPUs for months). Inference is continuous but more capital-efficient per dollar of revenue generated. As deployed models stabilize and inference dominates, capital intensity falls.

The NVIDIA-Hyperscaler paradox: Falling intensity is good for hyperscalers (lower capex per dollar of revenue → better ROIC) but bad for NVIDIA (less GPU purchasing). In the 3×3 matrix, NVIDIA’s best outcome (A×High: 1.3%/yr) is the hyperscalers’ worst within Scenario A (ROIC 7.5% at 2035). They sit on opposite sides of the intensity dimension — every dollar of efficiency that makes cloud AI more profitable is a dollar NVIDIA doesn’t collect.

Who benefits from falling intensity — and who doesn’t

The three forces driving intensity decline are not a single “learning curve” available to all participants equally. They differ in who captures the benefit:

The third force is the critical differentiator. Greenfield datacenter construction costs $8-12B per GW-scale facility (land acquisition, power grid connection, cooling systems, networking backbone, physical security). Once built, a GPU refresh cycle within an existing shell costs a fraction of that. An incumbent hyperscaler adding compute in 2031 pays only for new GPUs and installation; a new entrant pays for the entire facility plus the GPUs. This is a structural cost asymmetry that does not diminish with time — if anything, it widens as the best power sites and fiber corridors are locked up.

The implications for each layer of the value chain:

Incumbent hyperscalers (Google, Microsoft, Amazon, Meta): The

Foundation model companies (Anthropic, OpenAI): Hybrid position — building owned infrastructure (Anthropic’s planned $50B Fluidstack partnership, OpenAI’s $500B Stargate) while continuing to rent from hyperscalers. Their owned datacenters follow the same maturity curve but lagged by 2-3 years. The strategic logic: capture the capital stock maturity benefit themselves rather than paying hyperscaler margins on compute indefinitely. The risk: they are taking on capex risk that was previously externalized, and their balance sheets are thinner than the hyperscalers’.

Application layer (Cursor, Harvey, Glean): Completely shielded from infrastructure economics. They rent API access and benefit from everyone else’s efficiency gains — as incumbent infrastructure matures, API prices fall, and application margins improve. Best risk-adjusted position in the value chain: upside from AI adoption without infrastructure risk.

Historical precedent: Every major infrastructure overbuild follows this pattern. In the fiber optic bust (1997-2001),

6.2 Who Bears the CapEx?

The capex burden is shifting. Hyperscalers still bear the majority, but foundation model companies are increasingly building their own infrastructure — a strategic shift from the pure-rental model of 2023-2024.

Anthropic: planned ~$50B infrastructure partnership (Fluidstack, 2025; Reuters) + AWS Rainier (500K-1M Trainium2 chips) + Google Cloud TPUs + $30B Azure deal. OpenAI: $500B Stargate project (SoftBank/Oracle/MGX,