Will AI CapEx Pay for Itself?

A Quantitative Framework for Evaluating AI Infrastructure Returns

March 2026

---

Every major technology cycle follows the same pattern: a breakthrough creates genuine demand, capital floods in to build infrastructure, and the market debates whether the investment will pay off. This has played out with railroads in the 1840s, fiber optics in the late 1990s, and cloud computing in the 2010s. Each time, the technology eventually transformed the economy. But the investors who funded the buildout did not always get their money back.

The AI infrastructure buildout is the largest capital deployment in technology history. Hyperscalers will spend an estimated $2.6 to $5.6 trillion on AI-specific infrastructure between 2024 and 2035, depending on demand trajectory and capital intensity. NVIDIA, the dominant supplier of the GPUs powering this buildout, has been valued at $4.45 trillion — more than the GDP of most nations. The bull and bear cases have been argued extensively in both the financial press and on Wall Street. Bulls point to explosive revenue growth: Anthropic reportedly reaching $19 billion in annualized revenue run rate in just over a year (Bloomberg), OpenAI crossing $20 billion (Reuters), dozens of AI startups surpassing $100 million. Bears invoke the fiber optic bust, where $2 trillion in infrastructure was built for demand that took a decade to materialize.

I believe this debate cannot be settled by narratives about technological progress or predictions about artificial general intelligence. It should be settled by hard revenue dollars — what has actually been collected, what the current trajectory implies, and whether the math works.

This report takes a different approach. Instead of arguing for or against the AI thesis, I construct a quantitative framework with two independent dimensions — three revenue growth scenarios (demand) and three capital intensity paths (supply-side efficiency) — creating a 3×3 matrix of nine outcome combinations. Each is anchored to verifiable revenue data, actual corporate capex guidance, and historical precedents from previous infrastructure cycles. I model the full value chain — from NVIDIA’s GPU sales through hyperscaler cloud margins to the end-user application layer — and calculate the return on invested capital under each combination.

The findings are uncomfortable regardless of which scenario you believe.

Current trajectory is tracking the most optimistic scenario. AI revenue is growing at rates that match or exceed the cloud computing trajectory at a comparable stage. Anthropic and OpenAI are each adding over $1 billion per month in new revenue. Enterprise adoption is accelerating. The technology is unambiguously real and commercially viable.

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 (−$2.0 trillion) through the end of the projection window. Even the most optimistic combination — strong revenue growth with aggressive efficiency gains — shows a cumulative deficit exceeding $1.1 trillion through 2035. A transition mechanism (max 30% YoY capex decline, reflecting contractual momentum from datacenter and GPU commitments) makes these numbers worse than a pure formula model would suggest, particularly for bust scenarios where hyperscalers are trapped into spending that far exceeds demand.

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 $2.5B+ ARR, doubled since January 2026
• Revenue mix: ~70-75% API, ~10-15% consumer, ~13% Claude Code (reported)
• 500+ customers spending >$1M/year (reported)
• Gross margin: ~40-50% (reported); target 77% by 2028
• Targets cash-flow positive by 2028 (delayed from original 2027 target due to rising inference costs)
• Valuation: $380B (Series G, Feb 2026). Total raised: ~$64B.

OpenAI

Additional details (OpenAI communications and Reuters reporting — not audited):

• Actual 2025 revenue: ~$13B (Reuters) (full-year sum; lower than $20B ARR because revenue ramped from ~$500M/mo in Jan to ~$1.67B/mo in Dec — ARR is the exit rate, not the annual total)
• Revenue mix: ~75% consumer (ChatGPT), ~25% enterprise/API
• Gross margin: 33% (constrained by inference costs)
• 2026 projected loss: $14B
• Cumulative losses projected through 2029: $115B
• Company’s own targets: $100B revenue by 2028, $280B by 2030
• Valuation: $840B (Feb 2026). Investors: Amazon $50B, NVIDIA $30B, SoftBank $30B.

Cursor (Anysphere)

Additional: 60% enterprise, 40% individual. Valuation: $29.3B.

Other Notable AI Revenue (2025-2026)

50+ AI products above $100M ARR. 10+ above $1B ARR (Menlo Ventures, 2026).

Hyperscaler AI Revenue

1.2 Hyperscaler CapEx — Confirmed and Guided

Note: 2025 actuals exceeded early guidance for Google (+22%), Meta (+14%), and Amazon (+32%), suggesting upward revision pressure on 2026 as well. These are TOTAL capex figures (property/equipment additions). At ~75% AI share (industry consensus), 2025 AI capex was ~$308B and 2026 AI capex guidance implies ~$500-535B. The model uses $243B for 2025 AI capex and $488B for 2026 (Section 3.3) — these are hyperscaler-only figures using a conservative 64% AI share for 2025 (reflecting that a significant portion of Amazon’s $132B capex supports non-AI logistics and fulfillment infrastructure), deliberately below total industry actuals.

1.3 GPU Depreciation Policies — Actual Company Filings

The “value cascade” model supports longer life: frontier training (Year 1-2) → inference (Year 3-4) → batch/edge (Year 5+). Secondary market evidence: A100s (2020 vintage) still rent at $1.29-4.10/GPU/hr in 2025. T4s (2018) at $0.15-0.95/hr.

Bear case for shorter life: NVIDIA ships new architecture every 2 years (Hopper → Blackwell → Rubin). One analyst estimates ~$176B of understated depreciation across 2026-2028 industry-wide.

1.4 NVIDIA Financials

• Market cap: $4.45T (March 2026)
• Data center revenue: ~$170B run rate (2025)
• Gross margin: ~71% GAAP (FY2026); Q4 FY2026 ~75%. Model uses 63% as conservative assumption (see Section 3.1)
• Own capex: ~3% of revenue
• Share of hyperscaler capex: ~25-30%
• Revenue is a derivative of hyperscaler capex spend, not AI end-user revenue

---

Part II: The Growth Rate Model

2.1 Base Effect Correction

Comparing percentage growth rates across companies at different scales is misleading. The relevant metric is absolute dollars added.

Absolute Revenue Addition Rate:

Anthropic and OpenAI are adding nearly identical absolute dollars. The “1,148% vs 47%” growth rate comparison is an artifact of different starting bases.

2.2 Growth Rate at Each Scale Stage

OpenAI was faster to $1B and $2B (ChatGPT consumer virality). Above $2B, Anthropic is faster at every stage, and the advantage widens.

Structural explanation: OpenAI is ~75% consumer revenue (saturates faster). Anthropic is ~70-75% API/enterprise (78% of companies still in pilot phase — larger addressable runway).

2.3 Historical Growth Rates at $20B+ Scale

Average at $20B+ scale: 25-43% (non-monopoly). No company has sustained >50% growth past $20B except Meta.

---

Part III: The Assumptions

3.1 Assumptions Common to All Scenarios

3.2 Revenue Scenarios (Demand Side)

Revenue scenarios determine how fast AI revenue grows. These are independent of infrastructure efficiency.

Scenario A: “Cloud Replay” (40-45% probability)

Scenario B: “Slow Grind” (35-40% probability)

Scenario C: “Fiber Optic Bust” (15-20% probability)

3.3 CapEx Intensity Paths (Supply-Side Efficiency)

CapEx intensity = AI CapEx / AI Revenue. This measures how many infrastructure dollars are needed per dollar of AI revenue — the inverse of asset turnover. It declines over time as inference costs fall, capital stock matures, and workloads shift from training to inference.

These paths are treated as independent of the revenue scenario in the model — you can have great revenue growth with high or low intensity. This is a simplifying assumption. In practice, the dimensions interact in two ways:

1. Mechanical coupling during transition (2028-2030+): The transition floor (70% of prior year capex) creates a fixed spending minimum ($382B in 2028) regardless of revenue. Dividing the same floor by lower revenue produces higher realized intensity in bust scenarios — not because efficiency is worse, but because the denominator is smaller. Scenario C shows 2.7x realized intensity in 2028 while Scenario A shows 1.1-1.5x, even though the target intensity path is identical. The “intensity choice” only matters once revenue grows large enough that Revenue × Intensity exceeds the declining floor.

2. Possible steady-state correlation: One could argue that weak demand (Scenario C) should correlate with low intensity (because hyperscalers cut spending aggressively) or with high intensity (because they keep building despite weak returns, chasing AI leadership). The model treats them as independent to show the full range of outcomes — the probability weights in Part VII could be adjusted to reflect expected correlation if the analyst has a view.

Current observed intensity (committed period): 3.9x in 2025 ($243B capex / $63B revenue), declining to 3.7x in 2026 under Scenario A. These are far above steady-state because the committed capex front-loads infrastructure investment ahead of revenue. The paths below begin in 2028 when spending becomes intensity-driven — but actual capex cannot decline faster than 30% YoY (transition floor), so the intensity formula phases in gradually as committed contracts roll off.

Target Intensity Paths (formula input)

These are the formula inputs used to compute Revenue × Intensity. The transition floor (max 30% YoY decline, see below) means realized intensity exceeds these targets during the transition period. The target becomes the binding constraint only after the floor stops binding — immediately for A×High, but not until 2033 for C×Low.

Realized Intensity (actual capex / revenue, after transition floor)

During the committed period (2024-2027), intensity is observed and varies by revenue scenario because capex is fixed:

During the transition period (2028-2030), realized intensity depends on both revenue and the transition floor:

In Scenario A, the transition is smooth — intensity drops from 2.3x (2027) to 1.5x (2028) to 0.9x (2030). In Scenario C, the floor dominates: intensity barely drops from 4.7x (2027) to 2.7x (2028) and is still at 1.1x in 2030 — far above any target path. The intensity path choice only matters once revenue is large enough that Revenue × Intensity exceeds the declining floor.

Scenario C’s realized intensity stays at 2.7x through 2028 — the “capex trap.” The committed $1,414B is sunk regardless of demand, and the transition floor locks in continued high spending. Weak revenue keeps the ratio elevated. This is how infrastructure overbuild works: the denominator (revenue) fails to grow into the numerator (spending).

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 ~$1.67B/month ($20B ARR). The average monthly revenue was ~$1.1B, so actual 2025 revenue = ~$13B (consistent with Reuters reporting) — not $20B.

For a company growing at rate G per year, the approximation is: Actual Revenue ≈ (Start ARR + End ARR) / 2

Since capex figures are actual cash spent during the year, all ROI calculations in this analysis use actual revenue (not ARR) for consistency. ARR figures are shown for reference.

4.2 Company-Level Projections

Continuing: 2029 ARR $555B / Actual $488B. 2030 ARR $690B / Actual $622B.

The ARR-to-actual gap is ~25-30% in 2025-2026 under Scenario A (the fastest growth path) and narrows to ~10% by 2030 as growth decelerates. Under Scenarios B and C, the gap is smaller (7-15% in 2026) because slower growth means ARR approximates actual revenue more closely. Cumulative 2025-2030 actual revenue: ~$1,898B vs $2,223B ARR — a 15% overstatement if ARR is used incorrectly.

Basis for each projection:

• OpenAI: Back-solved from own $100B/2028 target. Assumes consumer growth slows, enterprise/API accelerates.
• Anthropic: Based on $1.3B/month current add rate, decelerating as base grows. Consistent with company’s $70B/2028 target.
• Google Cloud AI: AI-specific portion of $70B+ cloud run rate. Conservative — not all cloud revenue is AI.
• Microsoft AI: Growing from $13B+ disclosed AI run rate (Jan 2025); estimated ~$20B+ by early 2026 based on Azure AI growth trajectory. Copilot penetration still low (3.3% of 450M installed base).
• Dev Tools: Cursor at $2B and accelerating. Plus GitHub Copilot, Claude Code standalone, Windsurf, etc.
• Rest of Market: Long tail of 50+ companies above $100M ARR. Includes Scale AI, vertical AI (Harvey, etc.), AI infrastructure.

4.3 Revenue Double-Counting Note

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

End user pays Cursor: $20/mo
  → Cursor pays Anthropic API: ~$8/mo
    → Anthropic pays Google Cloud: ~$3/mo
      → Google pays NVIDIA: ~$1.50/mo

Total "reported ARR" across all layers: $32.50
Actual end-user spending: $20.00
Inflation factor: ~1.6x

The $690B total for 2030 may represent ~$430B in actual unique end-user spending. This matters for TAM analysis but not for company-level revenue projections (each company genuinely receives their reported revenue).

4.4 Three Scenario Revenue Paths

ARR (Annualized Run Rate)

Actual Revenue (used for all ROI calculations — see Section 4.1)

Current trajectory (March 2026) is tracking Scenario A.

4.5 Macro Fact-Check: AI Revenue as Share of GDP and IT Spending

A sanity check on the revenue scenarios — how large is AI relative to the broader economy?

AI Total Revenue as % of US GDP (~$30T in 2025, ~$35T in 2030, ~$40T in 2035, at ~3% nominal growth)

These figures include double-counting (~1.6x). Unique end-user spending is ~60% of these numbers (i.e., Scenario A 2030 ≈ 1.1% of US GDP in unique spending).

AI Revenue as % of Global Enterprise IT Spending (~$5T in 2025, ~$6.5T in 2030, Gartner estimates)

Comparison to historical technology adoption at similar scale:

Assessment: Scenario A’s $622B by 2030 (1.8% of GDP) would make AI roughly the size of today’s cloud computing market — ambitious but not physically implausible given current growth trajectory. Scenario B ($440B, 1.3%) is closer to the size of internet advertising — a more moderate but still large outcome. Scenario C ($178B, 0.5%) implies AI fails to become a major sector and remains niche. By 2035, Scenario A at ~2.6% of GDP would make AI one of the largest technology categories in the economy — comparable to all enterprise software today. This is the most aggressive assumption in the model.

The capex intensity numbers provide an additional macro check: at B×Base ($220B AI capex in 2030), hyperscaler AI infrastructure investment alone would represent ~0.6% of US GDP — comparable to the entire US semiconductor industry’s revenue today.

---

Part V: The CapEx Model

5.1 The 3×3 CapEx Framework

AI infrastructure spending is determined by two independent dimensions: revenue growth (demand) and capital intensity (efficiency). This creates a 3×3 matrix of nine outcome combinations (see Sections 3.2 and 3.3 for definitions).

2024-2027: Committed — Fixed Across All Combinations

2028-2035: AI CapEx = max(AI Revenue × Intensity, 70% of prior year’s AI CapEx)

From 2028 onward, annual AI capex is the greater of (a) that year’s AI revenue × intensity ratio, or (b) 70% of the prior year’s capex. The 30% max YoY decline cap reflects contractual momentum — datacenter builds and GPU orders committed in 2026-2027 cannot be cancelled overnight (see Section 3.3). Revenue determines how much infrastructure demand exists; intensity determines how efficiently that demand translates to spending; the transition floor determines how fast spending can actually decline.

Cumulative AI CapEx Through 2030 ($B)

Cumulative AI CapEx Through 2035 ($B)

The range spans from $2,644B (C×Low) to $5,569B (A×High). The transition floor compresses the lower bound significantly: without it, C×Low would be ~$1,900B, but contractual momentum prevents spending from falling that fast. The committed 2024-2027 spending ($1,414B) accounts for 25-53% of total — and the transition floor guarantees substantial 2028-2030 spending even in bust scenarios. Note: all Scenario C combinations have identical capex through 2030 ($2,250B) because the transition floor binds for all three intensity paths when revenue is low enough.

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($235B formula, $382B floor); 2029 = max($231B formula, $267B floor). The floor adds $183B of capex above the pure intensity formula, reflecting contractual momentum from 2026-2027 commitments.

Peak 5yr SL depreciation: $385B/year in 2029. Peak blended: $279B/year — 27% lower. Depreciation breakeven: ~2034 (5yr SL) or ~2033 (blended). The transition floor extends the overhang by one year vs a model without it, because the higher 2028-2029 capex feeds depreciation through 2033.

A×High — Worst Case

Peak depreciation: ~$543B/year in 2030 (5yr SL). Depreciation breakeven: never within projection window. Even Scenario A’s strong revenue cannot overcome the depreciation from sustained high capex.

The full 3×3 depreciation breakeven and capex payback matrices are in Section 7.3. Under blended depreciation, breakeven advances by ~1 year for Base and Low intensity paths; High intensity never breaks even regardless of method. One exception: for C×Low, blended depreciation is actually worse — the building depreciation tail (2.3%/year for 15 years) outlasts the 5-year schedule, so residual depreciation from 2026-2027 vintages continues through the mid-2030s even after servers have fully depreciated.

5.3 CapEx Intensity — The Efficiency Curve

CapEx intensity (AI CapEx / AI Revenue) is the key efficiency metric — the inverse of asset turnover. From 2028+, it follows the defined paths in Section 3.3. During the committed period (2024-2027), intensity is observed and varies by revenue scenario because capex is fixed while revenue differs:

Observed Intensity During Committed Period

Scenario C’s observed intensity stays at 4.7x through 2027 — the “capex trap.” The committed $1,414B is sunk regardless of demand, so weak revenue keeps the ratio elevated. This is how infrastructure overbuild works: the denominator (revenue) fails to grow into the numerator (spending).

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 $1,414B committed through 2027 is painful going in — but once sunk, it becomes a moat. By 2030+, these companies operate depreciated infrastructure at low marginal cost, and their effective intensity on incremental investment is at the low end of the curve. At 0.4x intensity and 62% hyperscaler gross margin (Section 3.1), each new dollar of capex supports $2.50 of revenue and generates ~$0.85 of hyperscaler GP — an 85% marginal ROIC. The cumulative deficit (−$2.0T in B×Base) reflects sunk costs that are irrecoverable; what matters for forward equity value is whether marginal returns are adequate, and by 2031+ they are. This is the cloud computing playbook: AWS had terrible cumulative ROI through 2015 but dominant marginal returns thereafter, because the infrastructure was built and competitors couldn’t replicate a decade of investment quickly.

New infrastructure entrants: Face greenfield intensity (~3-4x) while incumbents operate at ~0.5x. This structural disadvantage explains why neoclouds (CoreWeave, Lambda, Nebius) remain niche — partnering with hyperscalers or serving specialized workloads, not competing at scale. The capex trap is simultaneously a barrier to entry: the same $1,414B that depresses incumbent ROIC also prevents challengers from replicating the infrastructure at any reasonable cost of capital.

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), $2 trillion was invested, most builders went bankrupt, and the surviving companies or acquirers operated profitably on infrastructure purchased at pennies on the dollar. In cloud computing (2006-2015), AWS invested heavily for nearly a decade before reaching attractive profitability, but by 2020 earned 33% operating margins — extraordinary marginal returns on a cost base that competitors could not replicate in less time. In railroads (1840s-1870s), repeated bankruptcies destroyed equity holders but left the physical infrastructure intact, and survivors operated as natural monopolies for a century. In each case: (1) the builders often lost money, (2) the infrastructure survived and eventually produced strong marginal returns, (3) those returns accrued to incumbents and acquirers rather than new entrants, (4) the overbuild itself became the barrier to entry.

The open question for AI infrastructure: does the technology change fast enough that the physical infrastructure depreciates before generating adequate marginal returns? The “value cascade” model (Section 1.3) suggests no — frontier GPUs cascade to inference to batch/edge, retaining utility for 5+ years. But NVIDIA’s 2-year architecture cycle (Hopper → Blackwell → Rubin) creates pressure in the opposite direction. Amazon shortening GPU useful life from 6 to 5 years is the market’s current answer: somewhere in between.

5.4 Sensitivity: Depreciation by Useful Life Assumption

Annual depreciation in 2028 by GPU useful life (B×Base capex with transition floor, with Scenario B GP of $67B):

The qualitative conclusion holds across all nine combinations: regardless of depreciation assumption, intensity path, or revenue scenario, AI gross profit does not cover depreciation until the early 2030s at best (Low intensity) or never (High intensity).

---

Part VI: Value Chain Economics

6.1 Revenue and Margin by Layer

For every $1 of end-user AI spending:

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, $100B deployed immediately) + continued Azure. Both run hybrid models — owned infrastructure for frontier training, cloud for inference and burst capacity.

This shift has two implications: (1) foundation model companies are taking on capex risk that was previously externalized to hyperscalers, and (2) their gross margins will be pressured by depreciation as they own more of the stack. The application layer remains fully capital-light — the best risk-adjusted position in the value chain.

6.3 Neocloud CapEx — The Missing Layer

The ROIC analysis in Part VII uses hyperscaler-only capex (Google, Microsoft, Meta, Amazon — companies with audited 10-K filings). Total industry AI capex includes an additional ~8-10% from neoclouds (CoreWeave, Oracle AI, Lambda, Nebius, Crusoe, and others) that is excluded from the hyperscaler ROIC denominator because neocloud financials are mostly private and unauditable.

Estimated Neocloud AI CapEx

This is ~15-20% of cumulative hyperscaler AI capex ($3,568B in B×Base) — a meaningful, untracked layer.

Why This Matters for NVIDIA

Neoclouds are disproportionately NVIDIA-dependent. Hyperscalers are developing custom silicon alternatives (Google TPU, Amazon Trainium, Microsoft Maia, Meta MTIA) that displace NVIDIA GPUs for internal workloads. Neoclouds have no such alternative — their entire business model is renting NVIDIA GPUs.

¹ NVIDIA revenue as a fraction of total AI capex — which includes data center construction, power infrastructure, networking, memory, storage, CPUs, and custom silicon, not just GPU/accelerator purchases. NVIDIA’s share of the GPU/accelerator-only spend is significantly higher (~50-70% for hyperscalers, ~90%+ for neoclouds), but the total-capex denominator is the relevant metric for ROIC analysis since all capex must be recovered.

Neoclouds provide ~15% of total capex but ~26-29% of NVIDIA’s cumulative revenue. This disproportionality grows over time as hyperscaler NVIDIA share declines (custom silicon adoption) while neocloud NVIDIA share stays high.

Implication: NVIDIA’s share erosion model (Section 7.4) is driven primarily by hyperscaler custom silicon. The neocloud segment acts as a floor — a customer base that cannot easily switch away from NVIDIA. If neoclouds grow faster than hyperscalers (which is the current trajectory: CoreWeave growing >300%/yr), NVIDIA’s blended share erosion could be slower than modeled.

Scope note: The NVIDIA FCF analysis in Section 7.4 uses total industry AI capex (including neoclouds) as NVIDIA sells to all infrastructure buyers. Only the hyperscaler ROIC analysis (Sections 5.1, 7.1-7.3) uses the hyperscaler-only capex denominator.

---

Part VII: ROI Projections

7.1 Hyperscaler AI ROI — The 3×3 Matrix

Assumptions: 55% of AI revenue captured by hyperscalers. Gross margin: 40% (2025) → 62% (2030) → 73% (2035), fixed across all combinations (current competitive landscape maintained). All revenue figures are actual (not ARR). CapEx from the 3×3 framework (committed through 2027, then Revenue × Intensity from 2028+).

Since gross margin is held constant, hyperscaler gross profit depends only on revenue:

Annual Hyperscaler GP by Revenue Scenario ($B)

B×Base — Central Case (Detailed)

† Transition floor binding (see Section 3.3). Adds $183B capex above pure intensity formula.

• Depreciation breakeven: ~2034
• Cumulative GP − CapEx through 2035: −$1,978B

A×Low — Best Case for Hyperscalers (Detailed)

† Transition floor binding. Adds $178B capex above pure intensity formula.

• Depreciation breakeven: ~2032
• Annual GP exceeds annual capex from 2031
• Cumulative GP − CapEx through 2035: −$1,172B

3×3 Summary: Cumulative GP − CapEx Through 2035 ($B)

Every cell is negative. Even the best combination (A×Low) shows a cumulative deficit exceeding $1.1 trillion through 2035. The worst (A×High) shows −$3.3 trillion. The transition floor makes Scenario C particularly painful: all three intensity paths produce deficits near −$2.0-2.2 trillion because the floor forces spending far above what low revenue would justify. In a bust, the intensity path barely matters — contractual momentum dominates.

Note: Depreciation breakeven and capex payback measure different things. Depreciation breakeven is a P&L metric — when does AI stop dragging reported earnings? CapEx payback is a cash flow metric — when does annual gross profit exceed annual cash spent? Hyperscalers will report “profitable” AI segments (beating depreciation) before they’ve recovered the cash invested.

7.2 Gross Infrastructure Return — 3×3 Matrix

Methodological note: The return metric used throughout this report is Annual Hyperscaler Gross Profit / Cumulative AI CapEx — a gross-return-on-infrastructure measure. This is not ROIC in the institutional sense (Mauboussin defines ROIC as NOPAT ÷ invested capital, including working capital, with taxes and operating expenses deducted). I use gross profit rather than NOPAT because AI-specific operating expenses are not separately disclosed by any hyperscaler — Google Cloud and AWS report segment operating income, but do not isolate the AI-infrastructure portion from general cloud operations. The gross-profit metric overstates true returns by excluding operating expenses (R&D, sales, overhead allocated to AI), but understates them by excluding non-capex infrastructure value (existing datacenter shells, power contracts, land). For a cross-check against actual cloud segment margins, see the table below. Applying the blended ~35% operating margin to the hyperscaler AI GP figures suggests operating income ≈ 35/55 of GP at ~55% GP margin — meaning a “gross infrastructure return” of 8.0% corresponds to an operating-level return of approximately 5%. However, these segment margins cover all cloud operations (AI + non-AI); the AI-specific operating margin is likely lower given higher R&D intensity and earlier-stage scaling, so the true operating return may be closer to 4-5%.

Segment operating income cross-check (2025 actuals, from 10-K filings):

Note: Microsoft does not disclose Azure standalone operating income; Intelligent Cloud includes Azure plus other server products. These segments cover all cloud operations (AI + non-AI); no hyperscaler isolates AI-specific operating costs. Blended operating margin across the three disclosed segments is ~35%.

Throughout this report, “ROIC” refers to this gross infrastructure return metric unless otherwise noted.

Gross Infrastructure Return = Annual Hyperscaler GP / Cumulative AI CapEx (from 2024).

ROIC at 2030

ROIC at 2035

Only A×Low reaches acceptable ROIC (>10%) by 2035 at 12.1%. B×Low, which cleared 10% without the transition floor, now reaches only 9.3%. The transition mechanism compresses ROIC differentiation within Scenario C — all three intensity paths produce nearly identical 2030 ROIC (2.7%) because the floor forces the same cumulative capex regardless of intended efficiency. This is the quantitative expression of the capex trap: in a bust, capital discipline is impossible.

7.3 Breakeven Timelines — 3×3 Matrix

Depreciation Breakeven (Year GP ≥ Depreciation)

CapEx Payback (Year Annual GP ≥ Annual CapEx)

Under blended depreciation (Section 5.2), breakeven advances by ~1 year for Base and Low intensity paths under Scenarios A and B; High intensity is unchanged.

These matrices reveal two patterns. First, intensity dominates both metrics — High intensity never breaks even regardless of revenue (see “Key Insight” in Part VIII). Second, the transition floor introduces a revenue-intensity interaction: in Scenario C, even Low intensity doesn’t achieve depreciation breakeven until ~2035, because the floor forces years of spending that far exceeds demand. C×Low capex payback is delayed to ~2033 — two years later than without the floor.

7.4 NVIDIA Returns — 3×3 Matrix (Declining Market Share)

NVIDIA’s revenue is a derivative of total hyperscaler capex × NVIDIA’s market share. In the 3×3 framework, Total CapEx = AI CapEx / AI share (Section 3.1). With declining share (see rationale below), NVIDIA’s returns depend critically on the intensity dimension — but in the opposite direction from hyperscalers.

NVIDIA Share Erosion Rationale

Three structural forces erode NVIDIA’s dominance:

1. Custom silicon: Google TPUs (v5e/v6, powers Gemini training and inference), Amazon Trainium/Inferentia (Trainium2 shipping 2025, multi-billion run rate), Microsoft Maia (custom AI accelerator, deployed in Azure), Meta MTIA (inference chip for recommendation/ranking). Each chip displaces NVIDIA GPUs for internal workloads.

2. AMD gaining share: MI300X competitive on price/performance for inference. Data center GPU revenue growing >100% YoY. Projected $10-15B by 2026.

3. Inference-dominant workload mix: Training requires cutting-edge GPUs (NVIDIA advantage). Inference is cost-sensitive and amenable to custom/alternative hardware. As inference becomes 70-80% of compute spend, cheaper alternatives gain.

Projected NVIDIA Share: 30% (2024) → 28% (2025) → 27% (2026) → 26% (2027) → 24% (2028) → 23% (2029) → 22% (2030) → 18% (2035). Calibrated to Intel’s historical server CPU share erosion (~4.5pp/year relative decline once AMD achieved competitiveness), with a structural floor reflecting neoclouds (~90% NVIDIA-locked) and training workload lock-in.

Assumptions: NVIDIA gross margin 63% (conservative vs. 71% GAAP reported FY2026; haircut reflects competitive pressure from AMD/custom silicon and potential mix shift). Own capex ~3% of revenue, operating expenses ~5%. FCF = Revenue × 55%.

3×3 NVIDIA Cumulative FCF 2025-2035 ($B)

3×3 NVIDIA Average Annual Cash Yield at $4.45T Market Cap

The transition floor is slightly favorable for NVIDIA: forced continued spending means more GPU purchases during the wind-down period. Scenario C FCF compresses to a narrow $440-464B range because the floor, not intensity, drives spending. NVIDIA benefits from the capex trap.

Probability-weighted expected return: Using revenue scenario probabilities (A: 42.5%, B: 37.5%, C: 20%) and intensity path probabilities (High: 20%, Base: 50%, Low: 30%):

Expected cumulative FCF ≈ $593B → 1.2%/yr average annual cash yield at $4.45T vs 4.3% risk-free rate. (This is cumulative FCF ÷ years ÷ market cap — an average cash yield, not an IRR.)

The NVIDIA-Hyperscaler Paradox

The intensity dimension affects NVIDIA and hyperscalers in exactly opposite directions:

NVIDIA benefits from high intensity — hyperscalers spending aggressively on GPUs. Hyperscalers benefit from low intensity — same revenue from less capex. Every dollar of efficiency that makes cloud AI more profitable is a dollar NVIDIA doesn’t collect. This is not a coordination problem; it is a structural feature of the value chain.

Sensitivity: Share erosion pace (B×Base capex)

In every cell of the 3×3 matrix and every share trajectory, NVIDIA at $4.45T delivers an average annual cash yield below the risk-free rate. The transition floor slightly improves NVIDIA’s yield (more forced GPU purchases during the wind-down) but not enough to change the conclusion.

7.5 AI Software Layer Returns

Anthropic (at $380B valuation, 20x revenue)

• Scenario A actual revenue path: $24B (2026) → $85B (2028) → ~$138B (2030)
• Gross margin: 45% → 65%
• 2030 gross profit: ~$90B
• At 25x GP (mature tech multiple): ~$2,240B valuation
• From $380B: 490% total / ~56%/yr annualized (pre-dilution)
• Post-dilution estimate: 25-35%/yr
• Targets cash-flow positive by 2028 (delayed from 2027; less dilution required)

OpenAI (at $840B valuation, 42x revenue)

• Scenario A actual revenue path: $32B (2026) → $88B (2028) → ~$132B (2030)
• Gross margin: 33% → 55%
• 2030 gross profit: ~$73B
• At 25x GP: ~$1,815B valuation
• From $840B: 116% total / ~21%/yr annualized (pre-dilution)
• Post-dilution estimate: 3-6%/yr ($115B cumulative losses require massive funding)

7.6 Returns by Cost Basis (The Marginal Price Problem)

Market cap is not capital invested. It is the marginal price (last trade) multiplied by shares outstanding. Using the probability-weighted NVIDIA FCF of ~$593B (3×3 declining share model):

This explains why “overvalued” stocks stay overvalued: existing holders are anchored to unrealized gains and have no urgency to sell. The marginal buyer sets the price but represents a minority of ownership.

7.7 What NVIDIA’s Price Implies — Reverse DCF

The 3×3 analysis shows NVIDIA yielding 0.9-1.8%/yr at $4.45T. But what would NVIDIA need to deliver for its current enterprise value (~$4.40T, after subtracting ~$51B net cash) to be fairly valued? This reverses the question: instead of computing returns from the model’s outputs, we solve for the implied inputs.

NVIDIA FY2026 baseline (fiscal year ended January 2026):

• Revenue: $215.9B. Data center: $193.7B (90% of total).
• FCF: $96.6B (44.7% FCF margin). FCF yield at $4.40T EV: 2.2%.
• GAAP gross margin: 71.1%. Operating margin: 60.4%.
• Consensus FY2027 revenue: $360.7B (+67% YoY). FY2028: $467.7B (+30%).

Reverse DCF assumptions: 10% discount rate, 3% terminal growth, 40% terminal FCF margin (below current 44.7% — assumes margin compression from competition), 10-year explicit forecast, terminal value via Gordon growth.

Implied revenue path to justify $4.40T EV:

At a flat 20% CAGR for a decade with 40% FCF margins, NVIDIA’s EV is roughly justified. This is the implied growth rate embedded in the current price.

The critical question: analyst consensus vs. the 3×3 model

Wall Street consensus (69 analysts) projects FY2027 revenue of $360.7B (+67%) and FY2028 of $467.7B (+30%), with a 5-year revenue CAGR of ~26%. Front-loading these growth rates and then decelerating to ~13-15% by year 10 produces ~$1,700B in revenue by FY2036 — implying an EV of ~$6.2T, well above the current $4.40T. Under analyst consensus, NVIDIA appears undervalued.

The 3×3 model reaches the opposite conclusion because it projects NVIDIA’s share of hyperscaler capex declining from 28% to 18% over the decade — calibrated to Intel’s historical server CPU share erosion — as custom silicon (Google TPUs, Amazon Trainium, Microsoft Maia) and AMD competition erode GPU dominance. This share erosion causes NVIDIA’s revenue to grow much more slowly than the AI infrastructure market itself — the model’s probability-weighted cumulative FCF of ~$593B over 11 years corresponds to an average annual FCF of ~$54B, roughly half of NVIDIA’s current annual FCF ($96.6B).

The disagreement is entirely about share erosion:

The market is pricing no meaningful share erosion. The 3×3 model prices substantial erosion. Both are defensible — NVIDIA’s CUDA ecosystem and architecture lead are real advantages, but every hyperscaler is investing billions in custom silicon alternatives. The bet on NVIDIA at $4.45T is not primarily a bet on AI growth; it is a bet that NVIDIA maintains its competitive position for a full decade while growing at historically unprecedented scale.

Historical context: Intel — the last company to dominate a computing paradigm — peaked at ~$250B market cap (2000) and generated $18B peak annual FCF (2020). NVIDIA at $4.45T is priced at 18x Intel’s peak, requiring proportionally larger and more sustained cash flows. No semiconductor company has ever sustained >20% revenue growth past $100B for more than 3 years.

7.8 Shareholder Returns vs. Infrastructure Returns — A Timing Problem

The ROIC figures in Section 7.2 measure whether the AI infrastructure investment earns an adequate return on capital deployed. But ROIC is a point-in-time ratio (2035 GP ÷ cumulative capex); it does not account for when the cash flows occur. For the companies making these investments — and their shareholders — the timing matters.

The AI capex cycle is structurally front-loaded: the largest capital outlays ($488B in 2026, $545B in 2027) arrive years before the revenue base is large enough to generate matching gross profit. This creates a time value of money problem that ROIC does not capture.

NPV of Hyperscaler AI Cash Flows Through 2035 (at 10% cost of capital, $B)

Every cell is negative. Even A×Low — the only scenario with positive 2035 ROIC above 10% — has an NPV of −$1.1 trillion through the projection window, because eight years of losses (2024-2031) are discounted less heavily than four years of gains (2032-2035).

For comparison, the same B×Base scenario that shows 8.0% ROIC at 2035 has an NPV of −$1.5 trillion. An 8% return on invested capital does not mean an 8% return for the entity bearing the cost — it means the annual profit has reached 8% of cumulative spending, while the cumulative cash deficit remains −$2.0 trillion.

Terminal Value Matters — But Only for Low Intensity

The NPV through 2035 does not capture post-2035 cash flows, which could be substantial if the AI business generates strong returns on depreciated infrastructure. But the size of the terminal value depends critically on the steady-state net margin — the gap between annual GP and annual capex at maturity:

At Base intensity, hyperscaler GP roughly equals ongoing capex — the AI business generates near-zero net cash flow at steady state. This is not a failure of the model; it reflects the structural reality that AI infrastructure requires continuous reinvestment proportional to revenue (GPU refresh cycles, datacenter expansion, power infrastructure). Unlike software businesses where margins expand with scale, infrastructure businesses require ongoing capital proportional to the service delivered.

Technology advancement — each hardware generation (Blackwell, Rubin) bringing per-token costs down dramatically — shows up in the model primarily through the intensity dimension, not the margin dimension. The 73% gross margin at 2035 already prices in substantial efficiency gains from today’s 40%. The open question is whether further hardware and algorithmic gains translate to lower capex (the Low intensity path, where each dollar of infrastructure supports more revenue) or just more capacity at the same capex (Base/High, where hyperscalers reinvest savings into additional GPUs). Section 5.3 discusses this tension in detail: per-token inference costs decline 5-10x/year, but per-task token consumption rises (Jevons paradox), and competitive dynamics push hyperscalers to reinvest rather than harvest efficiency. The intensity path — not the margin assumption — is the decisive variable for steady-state cash generation. Even pushing gross margins to 85% (well above any current cloud business) only adds 6.6pp to net margin at Base intensity; moving from Base to Low intensity adds 15pp.

Only the Low intensity path produces terminal values large enough to meaningfully offset the accumulated deficit. Incorporating terminal values:

¹ B×Base terminal is near-zero because steady-state net margin is 0.15% of revenue. ² A×Low terminal assumes margins reach ~80% and intensity stays 0.25x post-2035 — the most optimistic sustainable configuration.

Only A×Low produces positive total NPV — and it requires the best case on both revenue and efficiency dimensions, plus continued margin improvement beyond 2035.

The Distinction for Shareholders

This creates a gap between the infrastructure return and the shareholder return. A project can show an attractive accounting ROIC while destroying present value for the investors funding it. The mechanism is straightforward: $1.4 trillion of the $3.6 trillion cumulative capex (B×Base) is committed in 2024-2027 — before AI revenue has scaled — and the transition floor locks in another $0.6 trillion through 2029. By the time annual GP approaches annual capex (~2034), the discounted cost of the early years has compounded far beyond what the late-arriving gains can offset.

The gap becomes concrete when compared to the AI premium currently embedded in hyperscaler valuations:

¹ Approximate: 25× estimated non-AI free cash flow.

The market prices approximately $3 trillion of AI value into the Big 4 hyperscalers. The model’s NPV of AI cash flows ranges from −$1.5T (B×Base) to +$324B (A×Low). For the implied AI premium to be justified, shareholders need some combination of:

1. A×Low outcome (best-case revenue AND efficiency) — which still only delivers +$324B, roughly one-tenth of the implied premium
2. Significant AI-driven uplift to non-AI businesses — better ad targeting, enterprise productivity gains, search quality improvements — generating additional GP beyond what the AI revenue model captures
3. Intensity declining well below 0.25x — AI infrastructure becoming dramatically cheaper than any historical infrastructure technology

The central point is not that hyperscaler stocks are mispriced — markets reflect expectations, risk preferences, and information that extend beyond any single model. The point is structural: even when the AI infrastructure investment earns a reasonable return on capital (8-12% ROIC by 2035), it does not automatically follow that shareholders of the companies making that investment earn a commensurate return. The front-loaded cost structure, the near-zero steady-state net margin at Base intensity, and the AI premium embedded in current valuations mean that the shareholder return depends on outcomes beyond what the capex-to-revenue model alone can deliver.

7.9 Second-Order Effects — What the Model Doesn’t Capture

The 3×3 model treats AI as a standalone infrastructure business: capex in, cloud revenue out. But for hyperscalers, AI also improves existing non-AI businesses — advertising, enterprise software, e-commerce, search — in ways the model’s revenue line doesn’t capture. These second-order effects are the strongest candidate for closing the gap between the model’s NPV (−$1.5T in B×Base) and the ~$3T AI premium embedded in hyperscaler valuations.

Advertising — The Quantifiable Channel

Meta and Google earned a combined ~$491B in advertising revenue in 2025 (Meta ~$196B⁴, Google ~$295B per 10-K filings). AI demonstrably improves ad performance through both channels identified in the introduction:

¹ 3.5% of Meta’s ~$196B = ~$7B direct; actual incremental revenue higher due to improved auction density and advertiser willingness to pay. ² The $60B figure is total ad revenue flowing through Meta’s end-to-end automated solutions, not incremental revenue attributable to AI. The incremental lift is embedded in the GEM and ROAS figures above. ³ Partial rollout; assumes 20-30% of Search campaigns participate at 19% lift. ⁴ Derived estimate. Meta’s 10-K reports Family of Apps revenue of $198.759B and states 2025 revenue growth was driven almost entirely by advertising, but does not present a single “advertising revenue” line item in the segment table. The ~$196B figure reflects FoA revenue less estimated non-advertising revenue (Payments, Horizon Worlds, WhatsApp Business).

Conservative sizing: A 5-10% AI-driven uplift to combined Google+Meta ad revenue = $25-49B in incremental annual revenue. At ~60% marginal operating margin (incremental ad revenue has near-zero marginal cost), this generates $15-29B in incremental operating profit per year — accruing to Google and Meta, not to the AI infrastructure business.

By 2030-2035, if AI-driven ad improvements compound (better models, more data, broader adoption), a 10-15% uplift on a growing base could reach $60-100B in incremental annual ad revenue, generating $35-60B in operating profit. This is material: $35-60B/year at a 20x multiple supports $700B-1,200B of enterprise value — potentially covering 25-40% of the ~$3T AI premium.

Enterprise Software — Overlap Risk with First-Order Revenue

• Microsoft Copilot: 4.7M paid subscribers at $30/user/month = ~$1.7B run rate. Penetration is 3.3% of 450M Office installed base. If penetration reaches 15-20% by 2030: ~$24-32B annual revenue. At ~85% gross margin, this is $20-27B GP.
• Google Workspace AI: Similar economics, smaller installed base.
• Amazon AWS AI services: Already captured in the model’s AI revenue line (these are “first-order” cloud rental effects).

Double-counting caveat: The model’s AI revenue scenarios (Section 4) define hyperscaler AI revenue broadly, including own AI products such as Copilot, Gemini, and similar offerings. To the extent Copilot and Workspace AI revenue is already embedded in the revenue scenarios, counting it again here overstates the second-order contribution. Advertising uplift is a clean second-order effect (it flows through non-AI revenue lines). Enterprise software is less clean — it may already be partially captured as first-order AI revenue. The summary table below therefore separates advertising effects (cleanly incremental) from enterprise software (potential overlap).

Other Channels — Qualitative Only

• Cloud stickiness: AI workloads create switching costs (data gravity, model training on specific infrastructure). Effect is real but doesn’t directly generate incremental revenue — it reduces churn.
• E-commerce: Amazon’s AI-powered search, recommendations, and supply chain optimization. Material at scale but not separately measured.
• Labor productivity: Internal efficiency gains from AI coding assistants, automated operations, accelerated R&D. Could reduce opex by 5-15% long-term but no hyperscaler has disclosed AI-attributable savings.
• Search quality: Google’s AI Overviews may defend search market share against ChatGPT/Perplexity. Defensive value (avoiding revenue loss) rather than incremental gain.

What This Means for the AI Premium

The model shows A×Low (best case) producing +$324B total NPV from AI infrastructure alone. The ~$3T AI premium implies the market expects something beyond infrastructure returns. Second-order effects — primarily advertising and enterprise software — are the most plausible source:

¹ Enterprise software revenue may already be partially captured in the model’s AI revenue scenarios (see caveat above). The “cleanly incremental” total excludes it to avoid double-counting.

Even with generous second-order estimates, cleanly incremental effects cover roughly one-third to one-half of the embedded AI premium. If enterprise software is fully additive (not captured in model revenue), the range extends to $1,400-2,100B (47-70%). The remainder depends on effects that are currently qualitative (labor productivity, search defense) or on outcomes beyond the model’s most optimistic cell.

This is not an argument that hyperscalers are overvalued — it is an accounting of what needs to be true. The AI premium is rational if: (1) infrastructure returns reach A×Low levels, AND (2) second-order effects on advertising materialize at scale, AND (3) enterprise software, productivity, and search defense contribute meaningfully beyond what the model already captures. All three are plausible. Whether all three are probable simultaneously is the investment question.

---

Part VIII: Summary of Projections

The 3×3 Framework

Returns depend on two independent dimensions: revenue growth (demand) and capital intensity (efficiency). Revenue determines how large the AI market becomes; intensity determines how much infrastructure spending each dollar of revenue requires. The committed 2024-2027 capex ($1,414B) is fixed across all combinations.

Cumulative Hyperscaler AI CapEx through 2035: $2,644B (C×Low) to $5,569B (A×High). Hyperscaler-only (excludes ~$600-750B neocloud capex; see Section 6.3). The transition floor compresses the lower bound — contractual momentum prevents rapid capex cuts even in a bust. Cumulative Hyperscaler GP through 2035: $610B (C) to $2,291B (A) Cumulative deficit (GP − CapEx): Every cell negative. Best: −$1,172B (A×Low). Worst: −$3,278B (A×High). Scenario C deficits cluster near −$2.0-2.2T regardless of intensity — the capex trap dominates.

Revenue Dimension — Current Trajectory

Current trajectory (March 2026) tracks Scenario A. Anthropic at reported $19B annualized revenue run rate (Bloomberg), OpenAI at reported $20B+ annualized revenue (Reuters), both adding >$1B/month. Enterprise adoption accelerating. The revenue side of the equation is performing at or above the most optimistic projections.

Intensity Dimension — The Open Question

The intensity dimension is where the real uncertainty lies. Current signals are mixed:

• Spending acceleration: Hyperscalers guiding $675-715B total capex for 2026, up from $410B actual in 2025 — nearly 75% YoY. This is High intensity behavior — capex scaling far faster than revenue.
• Efficiency gains: Per-token inference costs declining 5-10x/year at fixed capability (see Section 5.3 for decomposition — only ~3x/year from algorithmic gains, rest from hardware and competitive pricing). Custom silicon scaling (TPU v6, Trainium2, Maia). These are Low intensity enablers — but they haven’t yet translated to reduced spending because hyperscalers are reinvesting savings into more capacity (Jevons paradox).
• Amazon shortening GPU useful life from 6 to 5 years (2025) signals faster obsolescence — a High intensity factor.
• GPU utilization reportedly >80% at major cloud providers (vendor-specific claims; no cross-provider audited source) suggests demand justifies current spending — consistent with either Base or High.

The observed 2025-2026 intensity (~3.9x declining to ~3.7x under Scenario A) is consistent with the committed capex schedule. The true test comes in 2028-2030 as spending transitions from committed to intensity-driven. But the transition floor means capex can’t fall more than 30% per year, so even in 2028 the paths are partially constrained by 2027’s committed spending.

The Prisoner’s Dilemma: Current behavior is consistent with the High intensity path — the worst column in the 3×3 matrix for both hyperscalers and NVIDIA. The competitive dynamics explain why: each hyperscaler rationally increases capex to avoid falling behind in AI capabilities. If Microsoft builds more GPU capacity and Google doesn’t, Google risks losing cloud AI customers. If Meta invests and Amazon doesn’t, Amazon risks falling behind in AI product quality. The individually rational decision (spend more) produces a collectively irrational outcome (industry-wide High intensity where nobody earns adequate returns).

This is a classic prisoner’s dilemma. Low intensity — the path where hyperscaler ROIC reaches acceptable levels — requires either:

1. Coordination: Hyperscalers collectively agreeing to harvest efficiency rather than reinvest it. Unlikely given antitrust constraints and competitive incentives.
2. A demand shock: A downturn or “AI winter” that forces capex cuts, as happened in the 2001 fiber optic bust. The cuts would be involuntary, not strategic.
3. Organic demand absorption: AI revenue growing fast enough that the installed base becomes sufficient — hyperscalers choose to slow capex because existing capacity meets demand. This is the Scenario A hope, but even A×High never breaks even.

The fiber optic precedent is instructive. From 1997-2001, telecom companies collectively invested ~$2 trillion building fiber capacity, each afraid of being left behind. The bust forced rationalization — bankruptcies, asset sales at pennies on the dollar, and eventually the surviving companies operated profitably on the overbuild. The key question for AI: does demand materialize fast enough that hyperscalers can choose to slow down (path 3), or does the cycle end with a bust that forces them to (path 2)?

Key Insight: Intensity Dominates Breakeven

The most important finding from the 3×3 analysis: intensity matters more than revenue for determining whether AI infrastructure investment pays off.

Low intensity with strong revenue (A×Low) achieves capex payback by ~2031. High intensity never breaks even regardless of revenue. But the transition floor adds a third dimension: in a bust, contractual momentum makes the intensity path nearly irrelevant — hyperscalers are trapped into spending regardless of intended efficiency.

The NVIDIA-Hyperscaler Paradox

NVIDIA and the hyperscalers sit on opposite sides of the intensity dimension:

• NVIDIA’s best outcome (A×High: 1.8%/yr) is the hyperscalers’ worst within Scenario A (ROIC 7.5%)
• Hyperscalers’ best outcome (A×Low: ROIC 12.1%) gives NVIDIA only 1.1%/yr

At $4.45T, NVIDIA’s average annual cash yield is 0.9-1.8%/yr across all nine cells — below the 4.3% risk-free rate in every combination. Probability-weighted expected yield: 1.2%/yr. The transition floor slightly narrows the range because forced continued spending means more GPU purchases even in adverse scenarios.

Hyperscaler Returns — The Full Picture

Gross Infrastructure Return (GP/cumulative capex; see Section 7.2 for methodology; operating-level returns are approximately 35-40% lower, implying ~4-5% operating return at B×Base):

Only A×Low reaches acceptable returns (>10%). B×Base reaches 8.0% — below most hyperscalers’ cost of capital before opex. Scenario C is trapped at 3.1-3.3% regardless of intensity (capex trap). At a 10% cost of capital, every cell has negative NPV through 2035; only A×Low turns positive after terminal value (+$324B). The ~$3T AI premium embedded in Big 4 hyperscaler market caps requires outcomes beyond infrastructure returns alone (see Section 7.9 for second-order effects). Full details: Sections 7.2-7.3 (ROIC, breakeven), 7.8 (NPV, steady-state economics).

What to Watch — Leading Indicators

Revenue Dimension

Intensity Dimension