What $3 Trillion Has to Earn — A Token-Level Feasibility Model

Builds-on: five-years-of-capex-already-signed, compute-as-collateral-the-residual-value-wrap Related: token-cost-velocity-2023-2026, ai-bubble-collapse-token-price-direction, the-mask-that-eats-what-it-hides, cyclical-20-and-the-ai-capex-mask, ai-token-economics-and-open-source-competition, the-efficiency-counterthesis, anthropic-unit-economics-and-the-power-user-loss, ai-circular-financing-and-banking-exposure-audit Led-to: ai-capex-watchlist-check-august-20-2026

five-years-of-capex-already-signed established that nine companies have contracted roughly $3T of AI infrastructure over five years against ~$600B of annual capex. This doc asks the obvious follow-up and answers it with a model rather than a vibe: what does that have to earn, in tokens, from real customers, and is there any adoption path that gets there?

The model lives in workspaces/ai-capex-mask/model/ — tokenomics.py, demand.py, unit_econ.py. Inputs are documented inline so the assumptions can be argued with individually. Everything below is reproducible by re-running them.

The short version: the required number is about $1.17 trillion per year of end-user AI revenue by 2031, against roughly $110B today. Consumer and enterprise seats, pushed to optimistic 2031 levels, produce about a quarter of it. The remaining three quarters has to come out of payroll, because there is no other pool that size. And the unit economics that make it work are undone by the price decline that the same competition produces.

Step 1 — The mortgage payment

The framing of "$3T is like a mortgage" is exactly right, so treat it as one. Level-payment amortization on each tranche at its own useful life.

Sensitivity is mild: 7% WACC gives $552B, 10% gives $616B. The five-year IT life dominates, which is the whole depreciation argument from cyclical-20-and-the-ai-capex-mask showing up as a cash number instead of an accounting one. Use four years instead of five and the charge rises to ~$680B.

Add operating cost. Netting out ~$350B of the purchase commitments as prepaid energy rather than iron, $3T buys about 76 GW of facility capacity at $35M/MW all-in. At $900K/MW/year — the benchmark for a 1GW AI hyperscaler facility, covering power, cooling, staff and maintenance — that is $68B/year.

Total cost to serve: ~$652B per year. Capital charge is 90% of it. Power and everything else is 10%.

That split is worth pausing on, because it contradicts the dominant public framing. This is not an energy story on a cost basis. It is a silicon-amortization story with an energy garnish. The political fight over electricity (the-mask-that-eats-what-it-hides) is real and it matters to households, but it is a rounding error against the depreciation.

Step 2 — The number that should not be a coincidence

Two independent cross-checks put the installed base at about 37 million GPU-equivalents: $1.8T of IT spend at ~$48.6K per GPU (GB200 NVL72 at ~$3.5M for 72), and 76 GW de-rated by PUE at ~1.8 kW per GPU. They agree to within 1%.

Divide the cost by the machines:

Now put that next to two numbers already in this thread:

• Nvidia's take-or-pay backstop strike, per SemiAnalysis's illustrative six-year schedule: $2.36/hr
• H100 one-year contract rental, March 2026: $2.35/hr

Three numbers derived from completely different directions — a bottom-up amortization of the WSJ commitment tally, a vendor's negotiated floor, and a spot contract market — land on top of each other.

That is not a coincidence, and it has a specific meaning. Nvidia's backstop is struck at the system's breakeven, not below it. A conservative floor would sit under the cost line with room to spare. This one sits on it. Which is exactly consistent with SemiAnalysis's finding that a neocloud's IRR is 25.4% in the good case and zero or slightly negative if the backstop activates — the guarantee is calibrated to make the lender whole and wipe everything above the debt. It also means the market is currently clearing compute at cost, with the entire return to the buildout resting on volume growth that has not happened yet.

Step 3 — From compute cost to end-user revenue

$652B is what the compute layer must collect. End users must pay more than that, because the model layer sitting in between carries training, R&D, safety, sales and its own margin. Marking up 1.5x/1.8x/2.2x:

Against today: OpenAI at a $40B run rate (Bloomberg, Aug 13) and Anthropic at roughly $70B per trackers — call the frontier labs ~$110B combined, generously, and noting that a meaningful slice of it is recycled capex rather than external demand. Microsoft's own disclosure is the cleanest evidence of that: OpenAI accounted for $24.1B, about 70% of Microsoft's FY26 AI revenue, and OpenAI-related obligations are 45% of the $625B commercial cloud backlog.

$110B → $1,170B. 10.6x in five years. A 60% CAGR, sustained, for five consecutive years, on external customers.

On your 70% assumption: it is correct for Microsoft and should not be generalized to the system. Google serves Gemini to itself, Meta serves itself, AWS serves Anthropic plus a real external book. But it is correct for the merchant compute layer specifically, and the merchant layer is where the debt in compute-as-collateral-the-residual-value-wrap sits. Lab concentration is a credit problem even where it is not a demand problem.

Step 4 — Bottom-up: who actually pays?

This is the part that answers "what is adoption like out there in the real world." The honest reference points first:

• Pew (Feb 2026, n=5,119): 49% of US adults have used an AI chatbot. 24% use one daily, 12% several times a day, 4% "almost constantly." Under 50, usage is 63%; over 65, it is 23%. 38% of employed adults use one for work.
• Census BTOS: only 17–20% of US businesses report using AI to produce goods or services, with 20–23% expecting to within six months. Large-firm adoption has plateaued; small firms are now adopting faster.
• OpenAI: 900M weekly actives, 50M paying subscribers, 9M paying business users. A 5.6% conversion rate.

That last number is the load-bearing one. Consumer AI is adopted and not paid for. Half of American adults have touched it; one in eighteen users pays anything.

Building 2031 forward from there, with deliberately generous assumptions:

Even the aggressive consumer case — 3.5B users, 18% conversion, $20/mo, plus $25/yr ad ARPU on the free base — only reaches $223B. And 55% seat penetration at $50/month only reaches $363B. Push both to the aggressive end and you still land near $590B, half the target.

Three quarters of the required revenue cannot come from people using chatbots or from software seats. It has to come from work being done.

Correction — the seat line was wrong, and the seats/agentic split was worse

(Added 2026-08-19, after the a16z enterprise survey and the Ramp AI Index. This revision cuts against the conclusions above and is the more important of the two corrections in this doc.)

The $35/seat/month above is a hand-set number — $420 per employee per year. It is not what anyone measures. The 2026 benchmarks:

So the seat assumption was 3–5x low. But the arithmetic error is the smaller problem.

The structural error: this doc treats "enterprise seats" ($185B) and "agentic work" ($881B) as two different markets, and they are one market at different intensities. The Ramp distribution runs 680x from the top 1% to everyone else, and a16z's Global 2000 panel shows LLM budgets going $7M → $11.6M (+65%) with ~75% expected next year. The median firm buys something shaped like a SaaS seat; the top 1% buys something priced against a salary. There is no new market that has to appear on schedule for the buildout to work. There is a distribution that has to shift right — and the right-hand tail already exists, in production, today. The payroll conclusion below is not a 2031 hypothetical. It is an observed behaviour at the frontier of adoption.

But the naive multiply is also wrong, and badly. Applying $2,068 across 1.1B global knowledge workers gives numbers larger than the entire global software market, which is a reductio. Calibrating instead: US employment of ~135M at the $2,068 mean is $279B; adding the rest of the world at lower intensity gives roughly $419B of gross global enterprise AI spend in 2026. Only part of that reaches the compute layer — the rest is AI-embedded SaaS, consultants and internal build — and a ~30% pass-through squares with the two frontier labs' ~$110B combined.

The required rate is below both the observed rate and the expected rate. That is the single most encouraging number in this thread, and it means the "10.6x, 60% CAGR" framing earlier in this doc is too harsh — it measured growth against the frontier labs' revenue rather than against total enterprise AI spend, which is the correct base and is roughly four times larger.

Two caveats that cut back the other way, and they are not small. Ramp's panel is expense-platform customers, which skews heavily toward startups and tech companies and is not a sample of global employment. And the mean is dragged by that 680x tail, so it is not a typical firm — the median firm spends under $200 and is not on a path to $5,580.

What survives: the payroll conclusion below, restated more precisely. $2.79T of gross enterprise AI spend is 6.6% of global knowledge-worker compensation, and $5,580 per AI-equipped employee per year is 5.9% of a $95k salary. The money still has to come from the labor line. What changes is that the diffusion path to it is visible in current data rather than hypothetical, and the required growth rate is slower than what enterprises are currently doing.

Step 5 — Which pool is big enough

The residual is 63% of every dollar spent on enterprise software on earth. It is not going to be carved out of software budgets — that would require AI to become, by itself, roughly two-thirds the size of the entire packaged software industry within five years, without displacing the rest of it.

Against payroll, the same number is 1.8–2.5%. That is a modest-sounding figure, and it is the only pool with the capacity.

The $3T pencils only if AI revenue is booked against the labor line, not the software line.

This is the load-bearing conditional of the entire buildout, and it is rarely stated this plainly. Every hyperscaler capex defense that routes through "productivity gains" and "TAM expansion" is, when you do the arithmetic, an argument that AI captures 2%+ of global white-collar compensation within five years. That may happen. But it is a claim about labor substitution at civilizational scale, on a five-year clock, and it should be evaluated as that claim rather than as a software forecast. It also has a political failure mode that a software forecast doesn't: capturing 2% of payroll means visibly displacing a lot of people, in an environment that has already produced the-mask-that-eats-what-it-hides's twenty-three states legislating on data center cost allocation.

Step 6 — The throughput-price coupling

(Distinct from the scissors in ai-bubble-collapse-token-price-direction, which is about how prices bifurcate after a collapse — commodity down, frontier up. This is a constraint that binds before one: whether any price-and-throughput combination clears the mortgage at all.)

Now the part that makes it harder than a demand-growth problem. Required revenue per GPU-hour is fixed at ~$2.25. Revenue per GPU-hour is throughput times price. And throughput and price move in opposite directions, driven by the same variable — model size.

Market, August 2026 (BenchLM, 142 models): median proprietary blended $2.63/Mtok, median open-weight blended $0.53/Mtok. The real floor is lower still — ai-bubble-collapse-token-price-direction has DeepSeek V4 Flash at $0.14/$0.28, which makes the open-weight column below optimistic. The three-year tier-by-tier price history behind all of this is in token-cost-velocity-2023-2026.

Line those up and the required prices sit directly on top of the market prices, with no headroom. Frontier reasoning needs $2.08 against a $2.63 median. Mid-size needs $0.52 against a $0.53 open-weight median. The system is priced at cost across every tier simultaneously.

Expressed as utilization — what fraction of the year the machines must be sold to clear the mortgage:

flowchart TB
  A[Required: 2.25 dollars<br/>per GPU-hour] --> B{What fills the GPUs?}
  B -->|frontier tokens| C[High price per token<br/>2.63 per Mtok]
  B -->|cheap tokens| D[High throughput<br/>4000 tok/s/GPU]
  C --> E[Low throughput<br/>300 tok/s/GPU]
  D --> F[Low price per token<br/>0.53 per Mtok]
  E --> G[Needs 79% utilization<br/>on paid frontier demand]
  F --> H[Needs volume the<br/>installed base cannot emit]
  G --> I[Clears only if paid demand<br/>grows 10x in 5 years]
  H --> I
  J[Open-weight substitution<br/>at frontier quality] --> K[Frontier tier collapses<br/>to 393% utilization]

The kill shot is not weak demand. It is open-weight substitution at frontier quality. If the tokens that people are willing to pay $2.63 for can be served by a model priced at $0.53, the frontier tier's required utilization goes to 393% and no amount of adoption fixes it. That is ai-token-economics-and-open-source-competition and the-efficiency-counterthesis arriving as a hard arithmetic constraint rather than a competitive worry.

Step 7 — The treadmill

Industry-wide LLM prices fell roughly 80% from 2025 to 2026. Assume that decelerates sharply. Volume growth needed just to hold revenue flat:

Stack that on the 10.6x needed for growth. At a 35% annual decline — less than half the current rate — the industry needs roughly 90x today's paid token volume by 2031. At 50%, it needs 340x.

For scale: Google processes 3.2 quadrillion tokens per month across every surface it owns, up 7x year over year, and monetizes almost none of it directly. That is 38 quadrillion a year. The requirement here is 400–1,200 quadrillion paid tokens a year — 10–30x Google's entire current throughput, all of it billed.

And the installed base can physically emit that only in the cheap-token regime: 37M GPUs at frontier-reasoning throughput produce ~313 quadrillion tokens/yr; at mid-size MoE throughput, ~1,250 quadrillion. The volume requirement and the price requirement point at different halves of the product line. You need frontier prices on mid-size economics.

What this actually says

The model does not say the buildout is impossible. It says something more specific and more useful.

1. The infrastructure clears at today's prices and plausible utilization. At $2.63/Mtok a mid-size MoE needs 20% utilization. Nobody is going bankrupt on unit economics this year, and anyone arguing from cost-per-token alone will keep being wrong for a while.
2. The gap is entirely a volume-of-paid-demand problem, and the required volume is 10.6x on revenue and up to 90x on tokens. That requires the agentic/automation thesis to be right, at scale, on schedule.
3. The required revenue can only come out of payroll. $881B is 63% of global enterprise software and 2.5% of global knowledge-worker compensation. Only one of those is a viable denominator.
4. Price decline is the mechanism that turns a growth story into a shortfall, and it is generated by the same competition that drives adoption. This is the genuinely uncomfortable structural feature: the thing that expands the market destroys the revenue per unit faster than it expands the units.
5. Consumer AI is not the answer and is not close. 900M weekly users converting at 5.6% produces about $100B at optimistic 2031 assumptions — under 10% of the requirement. The public's mental model of the AI economy is the part of it that pays for almost none of the mortgage.

Held against the rest of the thread: five-years-of-capex-already-signed says the spending cannot be turned off, and this doc says the revenue required to justify it needs a 60% five-year CAGR against a falling price. Those two facts together are the whole thesis. The capex prints into GDP either way. What varies is whether it earns, and the earning has to show up as displaced labor cost inside five years or it does not show up at all.

Correction: the model prices everything as sold tokens, and that is wrong for two firms

The steelman for the buildout is not a token forecast, and building the model around tokens understates the case. A large share of AI value is captured internally, never crossing a price list. Meta sells no tokens at all. Google sells some, but its largest AI payoff is defending a search franchise, not billing for inference. For those two, "required $/Mtok" is the wrong question entirely.

Run Meta on its own numbers. 2026 ad revenue of $243.5B, growing 24.1% — that is $47.3B of year-over-year growth, and eMarketer has Meta overtaking Google in global digital ad revenue for the first time on the strength of Advantage+ automation. Meta's trailing-four-quarter capex of $75.7B carries an annual capital charge of about $14.7B.

Meta's own disclosures put a 6% lift in landing-page-view conversions on Lattice/GEM and 1.6% on the Adaptive Ranking Model, with GEM claimed as "4x more efficient" — all vendor-stated and unaudited, which is the right caveat to keep. But even at the conservative end, Meta's AI capex roughly pays for itself out of ad ranking alone, with no token revenue, no enterprise seats, and no agentic thesis required. Google's equivalent: defending 5% of $239.5B of search ad revenue from chat substitution is worth $12B/yr, before counting Cloud.

This does not rescue the aggregate. It relocates the problem, and the relocation is the finding:

The internal-substitution justification is strongest exactly where the debt isn't, and absent exactly where it is. Meta and Google fund from cash flow and capture value internally. Oracle and the neoclouds have no internal business to capture value in, are the most leveraged, and sit downstream of the residual-value wrap in compute-as-collateral-the-residual-value-wrap. So the $1.17T token requirement in this doc should be read as falling almost entirely on the merchant layer, not on the nine firms evenly. That makes the aggregate number somewhat too pessimistic and the distribution considerably worse.

Two structural notes that complete the picture:

The individually-rational, collectively-irrational shape. Each firm's best response to the others building is to build. Sell-side estimates for five US companies' 2026 capex have risen $173B since January, to $697B. No participant can unilaterally stop without conceding the franchise, which is why the stated justification is almost always asymmetry — the risk of underinvesting exceeds the risk of overinvesting — rather than a return calculation. That is a defensible position for a firm with $150B of operating cash flow and a coherent one for the industry to hold collectively while the aggregate fails to pencil. It is not a forecast, and it should not be read as one.

The cash constraint explains the financing. AI capex has gone from 33% of hyperscaler cash flow from operations in 2023 to roughly 93% in 2026, with Amazon already over 100%. The SPVs, the leases-not-yet-commenced, and the residual-value guarantees in five-years-of-capex-already-signed are not clever tax structuring. They are what happens when the capex race runs past internal funding capacity and the participants still cannot stop. The financial engineering is downstream of the arithmetic in this doc, not independent of it.

What would move this materially

• The blended realized price, not the list price. The single most important unknown in the model and the least disclosed. If somebody publishes an actual revenue-per-token figure for a major lab, that pins the whole thing. The CME rental futures curve from compute-as-collateral-the-residual-value-wrap is a partial proxy on the cost side; there is no equivalent on the revenue side.
• Paid conversion at OpenAI moving off 5.6%. Every point of conversion on 900M weekly users is worth roughly $2B/yr at current pricing. Getting to 12% is $65B. This is the cheapest lever anyone has and it has not moved.
• Spend per employee, tracked as a distribution rather than a mean. The Ramp AI Index and a16z's Global 2000 panel are the two live series. What matters is not the average — it is how fast mass moves from the sub-$200 median toward the $2,800+ decile, because that migration is the demand curve.
• Enterprise seat pricing holding above ~$30/month. Seat pricing is where competitive pressure hits first and where Microsoft's bundling incentive is to give it away.
• Any credible measurement of AI revenue displacing payroll rather than software budgets. BLS or Census data showing headcount reduction attributed to AI in non-tech sectors would be the first real evidence for the only sufficient pool. Census BTOS at 17–20% production usage says that has not started.
• An open-weight model at genuine frontier reasoning parity. The 393% cell in the utilization table. This is the fastest path from "tight but workable" to "structurally impossible," and it is not a demand event.

Model inputs and their softness

Stated plainly, worst to best:

The result is most sensitive to the markup and the realized price, which are the two least observable inputs. That is a real limitation and it cuts both ways. What survives it is the structure: the required revenue is an order of magnitude above today's, the only pool large enough is payroll, and price decline compounds against volume growth. Those conclusions hold across the full plausible range of the soft inputs.

Sources

• OpenAI's Revenue Run Rate Tops $40 Billion Ahead of IPO — Bloomberg
• Microsoft Disclosures Suggest OpenAI Sales Account For Around 70% Of FY26 AI Revenue — Ed Zitron
• Americans and AI 2026: Chatbots, Smart Devices and Views on Impact — Pew Research Center
• Large Firms With at Least 20 Employees Biggest AI Users — US Census Bureau BTOS
• Monitoring AI Adoption in the U.S. Economy — Federal Reserve FEDS Notes
• Google processes over 3.2 quadrillion tokens monthly — Crypto Briefing
• LLM Pricing Statistics 2026: Cost Per Million Tokens — BenchLM
• LLM API Pricing — $ per Million Tokens by Model (2026) — Silicon Analysts
• Nvidia GPU Debt Backstop Unleashes the AI Project Trinity — SemiAnalysis
• AI Data Center Cost per MW: 2026 Benchmarks by Tier — Axis Intelligence
• Data centre operating cost structures: traditional cloud versus AI — ITK Research
• Gartner Forecasts Worldwide IT Spending to Grow 14.2% in 2026, Totaling $6.37 Trillion
• 1 Million Tokens Per Second: Qwen 3.5 27B on GKE with B200 GPUs — Google Cloud
• The Hidden Cost Driver in Agentic Coding Sessions in 2026 — Vantage
• ChatGPT Statistics 2026 — Nerdynav