deadrisk

Simulation — not financial advice

The AI build-out, 2026–2035

Demand is the world's work, counted in the tokens an AI would need to do it. Supply is what the fleet of AI chips can serve, built year by year and worn out over time. Pick a world, move the inputs, and watch who keeps the margin.

Read what version 1 says: who the AI build pays

Load a world

Base
Boom / bust
Bigger worlds
Credit
Balanced…

    What the filings say

    the model's 2026 is set against these · rebuilt 2026-10-15
    • Top-5 hyperscaler capex, a quarter $182B 2026Q2 · a year before: $97B · $726B a year at this pace
    • …as a share of their operating cash 97% 2026Q2 · last four quarters: 82%
    • Top-5 hyperscaler debt $487B 2026Q2 · a year before: $245B
    • Nvidia revenue, a quarter $96B 2026Q2 · a year before: $47B
    • Micron revenue, a quarter $54B 2026Q3 · a year before: $11B
    • SK Hynix revenue, a quarter ₩79.3tn 2026Q2 · a year before: ₩22.2tn
    • TSMC revenue, a quarter NT$1270B 2026Q2 · a year before: NT$934B
    • Contracted backlog: MSFT, ORCL, GOOGL, CRWV $1.95T 2026Q2 · a year before: $651B · all customers, not only AI
    • OpenAI + Anthropic revenue, annualised $105B 2026-08 · Epoch AI estimate
    • 10-year Treasury yield 5.19% 2026-10-14
    • H100 rent, per GPU-hour $2.65 2026-10-11 · Ornn index
    Inputs
    Demand
    Adoption follows an S-curve from 3.3% in 2026: slow, then fast, then levelling off near the ceiling.
    Agents take longer jobs and check their own work, so each task uses more tokens.
    Smarter models and leaner agent loops pull the other way over time.
    The growth in knowledge work with no help from AI prices.
    The % more work demanded for each 1% fall in the cost of a year of AI work, capped at 4× the base. At 0.25, halving that cost brings about 19% more work. The digital economy grew about 6.5% a year in real terms against 1.8% for GDP; spending on computing hardware has been close to unit-elastic, software above 1.
    Work done on capacity this model doesn't count: Chinese-hosted models selling cheap tokens worldwide, open models on buyers' own hardware, phones and laptops.

    Supply
    Gigawatts of AI capacity already ordered. About 50% a year takes the fleet from ~40 GW to ~80–90 GW in 2028.
    Builders size each year's build so last year's demand, grown at last year's rate, would use 80% of capacity, up to +60% a year. Because they extrapolate, they overbuild when growth slows. Projects under way still finish (30% of last year's build), and labs keep buying each new generation for the frontier (6% of the fleet a year), so the build tapers rather than stopping dead. Untick to hold the committed rate to 2030 and then this rate:
    Tokens per watt of each year's new chips over the last year's. Chips wear out over years four to nine, about six on average; a chip is scrapped early when the power and upkeep to run it cost more per token than a new chip costs all in.
    Serving software and model tricks that speed up every chip in the fleet.
    Each year both gains shrink by this share. Chip physics and the 4-bit arithmetic floor point to a slowdown.
    Google TPU, Amazon Trainium, Meta, Microsoft, OpenAI–Broadcom. They save owners money (SemiAnalysis puts TPU cost ~44% below GB200 for Google itself) but draw on the same wafers, packaging, HBM and grid, so they add no supply.
    Each layer delivers only so many new GW a year (forecast to 2030: HBM 23→50, CoWoS packaging 26→75, wafers 50→120, power equipment and grid 22→50). The tightest binds. When builders want more than a layer can make, its makers add capacity that arrives a year later (packaging, HBM) or two (wafers, grid), up to four times the forecast. Chips that replace retired ones go into sites already wired, so only growth needs new grid. Above capacity suppliers charge up to 30% more per GW (memory 60%); far below it, prices and margins fall.
    0: hyperscalers build to keep the market about 80% busy. 1: they hold back until it is 95% busy, to lift prices. −1: they chase demand and build until it is only 65% busy. Neoclouds (about a quarter of the fleet today, borrowed money, ~3 points dearer) build whenever the market runs tighter than 85%, if lenders fund them.
    AI labs and prices
    The share of the fleet's full cost (cost of capital included) that compute rents recover when supply and demand balance. About 0.4 today: the fleet earned ~$175B in 2026 against ~$400B needed. Rents sink toward running costs in a glut and rise in a shortage, up to three times full cost — but never past what the labs can pay.
    Buyers pay for the best models' work off what it is worth. In 2026 that works out at about 20% of what a person doing the work costs, solved from what buyers paid (about $150B, an estimate: OpenAI and Anthropic ran from $30B to $105B a year through 2026, per Epoch). It rises as agents take whole jobs, or falls as rivals catch up.
    About half of AI work in 2026. The rest is commodity work — last year's capability, open models — sold at about its compute rent (below it in 2026, while labs buy share). Capability spreads, so the share falls.
    Labs pay compute rent out of up to 80% of their revenue, and the rest from investors: OpenAI's $30B+ bridge, Anthropic's listing, DeepSeek's $12–15B, Broadcom's $42B note to Anthropic. Rounds grow with revenue while it grows fast and shrink when it slows or credit tightens. What the labs cannot pay, the fleet does not earn.
    Capital
    The 10-year Treasury yield, above 5.2% in October 2026. AI's cost of capital is this plus 3.5 points: a blend of hyperscaler bonds, project debt and equity. Neoclouds pay 3 points more.
    New debt raised each year by governments, companies and households worldwide: about $13–15T (IIF). Grows about 5% a year.
    Governments take about half today. The US deficit was $2.0T (6.2% of GDP) in FY2026; CBO projects 6.7% by 2036.
    AI takes about 7–8% of private net borrowing today; the dot-com tech peak was 14% of US bond sales. Lenders stretch this ceiling by up to 80% when the fleet is sold out and prices are rising, and cut it by up to 40% when it sits idle.
    Basis points added to AI's cost of capital for each point of private borrowing AI takes above 5%. Each year's chips lock in that year's rate.
    Share of capex paid by new stock, vendor financing and customer prepayments.
    Operating cash the big platforms can put into AI before borrowing, growing 10% a year, plus whatever AI itself earns above its 2026 cash margin. In 2026Q2 the top five spent 97% of their operating cash on capex.
    Power and cost
    The fleet cannot grow past this. All data centres use about 1.5% today; the IEA expects about 3% by 2030.
    About 30,000 TWh in 2026, growing about 3.5% a year (IEA).
    Starts at $70 per MWh for data-centre supply.
    Chips, systems, buildings and power hookup for one gigawatt. Estimates run $38–47B for 2025 builds (McKinsey, SemiAnalysis, Barclays); Nvidia says $50–60B.
    Nvidia's revenue per GW went from about $18B (Hopper) to $25B (Blackwell) to $40B (Rubin). Tokens per GW still rise faster, so cost per token keeps falling.
    Bigger worlds
    Domestic, care, logistics, factory and construction work: $2.5T of US wages a screen can't reach (TEXXR's occupation table), times the world multiplier.
    About $2.7T of world military spending (SIPRI). Served first in a shortage and pays half again per token.
    About $3T of world R&D. New value rather than replaced wages, and pays twice the price per token.
    Sovereign AI: GW a year built whatever the returns, first in line for power and parts, paid for with government borrowing.
    Wages AI replaces, less re-employment, cut consumption; savings come back as lower prices (spent) or profits (mostly saved). The levy phases in as consumption falls below where it would have been.
    A credit event
    2026 means no failure. The chill cuts AI's borrowing ceiling and fades with a two-year half-life; spreads jump ~200bp; neoclouds can't raise money for three years; lenders dump the collateral GPUs for about two years. A guarantee mostly undoes the chill. A government purchase moves the failed fleet onto the state's books.

    Tokens a year

    Log scale, quadrillions (Q) of tokens. Usable capacity is 65% of running every chip flat out. On a log scale a gap that looks small can be large: the label at the right gives it.

    Usable capacityDemand

    Who gets paid

    A year, in dollars. Chip and system makers' revenue from the build; the labs' margin, what buyers pay them less the fleet's rent; and the fleet owners' return over their cost of capital, below the zero line when they fall short. The essay walks through it.

    Chip and system makers: revenueLabs: margin after rentFleet owners: return over cost of capital

    Demand as a share of capacity

    2026 is set at 77%: tight, as GPU rents show. Below 50% is a glut; above 100%, price rations demand.

    Balanced, 70–95%

    The fleet: built, worn out, running

    GW of AI capacity installed each year (up) and retired (down), with the fleet running at year end.

    BuiltRetiredRunning

    Capex: the filings, then the model

    Solid: the top five hyperscalers' capex from their filings, the running year annualised. Line: the model's spending by every builder. The top five are most of it, not all.

    Filed, top fiveTop five, Nvidia's outlookModel, all builders

    AI in the world economy

    Log scale, each as a share of world GDP: about $125T in 2026, growing 4.5% a year, plus the capex and the change in consumption the model adds. Wages AI replaced is the work AI does, valued at what people were paid for it; the other line is what buyers pay the labs for it.

    What buyers pay for AIAI capexWages AI replaced

    Revenue against what the build needs

    Log scale. What buyers pay the labs for AI; what the fleet earns in compute rent, paid by the labs from that revenue and from investors; and what the fleet needs to earn its cost of capital plus power and running costs. The gap between the first two is the labs' margin. Rings: the two biggest labs' revenue run rate, part of what buyers pay.

    Buyers pay the labsFleet earnsFleet needsOpenAI + Anthropic, annualised (Epoch)

    Growth each year

    Where the magenta line runs above the blue, demand is gaining on capacity.

    Usable capacityDemand

    What a year of AI work costs

    Break-even price for one worker-year of AI output, against the average knowledge worker's cost.

    AI worker-year, break-evenHuman worker-year

    AI's share of world electricity

    The fleet's draw (IT load × 1.2 for cooling, at 60% average load) against world generation. Dashed: the ceiling you set.

    The economics at a glance

    How the model works

    • Demand = total intelligence demand × AI's share. Total demand starts at 4.6 sextillion tokens a year: about 1 billion knowledge workers, weighted by the tokens an AI needs per job (about 0.5 billion a year for routine service work, about 25 billion for expert work). It grows with the work demanded and the tokens per unit of work.
    • AI's share starts at 3.3%, set so 2026 demand matches the roughly 150 quadrillion tokens served worldwide as a full-year average. That total is scaled up from Google's 3.2 quadrillion a month, assuming Google serves 25–30% of all tokens.
    • Supply is built by chip generation. The 2026 fleet of about 40 GW mixes chips installed in 2023–26, half of it this year. Each year's new chips are more efficient than the last; software gains apply to every chip. Chips wear out over a spread of years, so capex does not echo the build of six years before.
    • The money runs buyers, labs, fleet. Buyers pay labs for AI work: frontier work off what it is worth, commodity work at about its compute rent. Labs pay the fleet's rent from that revenue and from investors. The rent follows the fleet's cost and how busy it is, and is capped by what the labs can pay. What buyers pay beyond the rent is the labs' margin.
    • Suppliers and lenders answer the market. Packaging, memory, wafer and grid makers add capacity when their order books overflow, after a lead time. Lenders lend more against a sold-out fleet and less against an idle one, and builders spend what AI earns.
    • Usable capacity is 65% of the theoretical maximum. On that basis 2026 runs at 77%, which matches rising GPU rents and sold-out older chips.
    • Builders move halfway each year from last year's build to the one they want, since sites, power and chip orders take about two years. Near the power ceiling the grid fills gradually: each year builders get at most half the room left.
    • The happy ending needs demand to outrun the fleet. The owners earn their cost of capital only in a shortage, and software gains of more than about 60% a year end it by moving the margin to the labs.
    • Every input is an estimate. The model's 2026 is set against the filed figures above; later years are scenarios, not forecasts.

    What version 1 says: Who the AI build pays. Built for The Token Ceiling and The same chip, three businesses.

    Show the numbers as a table