Dwarkesh Podcast · Dylan Patel
Episode explainer

Who pays for the AI buildout?

Dylan Patel expects AI datacenter spending to reach trillions of dollars a year. A big part of that money has to be borrowed, and borrowing at that scale pushes interest rates up for everyone else. That is the episode's core argument, and it ends at crashed stock valuations and defaulting countries. I could follow the argument. The finance concepts underneath it kept losing me, so this page explains those concepts first, with models you can play with, and then goes through the argument itself.

Podcast: Dwarkesh Podcast Guest: Dylan Patel, SemiAnalysis Transcript ↗ YouTube ↗

§1Everything is priced per megawatt

AI capacity is measured in power, because you can buy more chips much faster than you can get more electricity supply built. A large AI datacenter draws a gigawatt — 1,000 megawatts, roughly a nuclear reactor's output. So the industry's unit economics are dollars per megawatt per year. Two numbers from the episode anchor everything else. Running a megawatt of compute costs $10–15M a year, and the frontier labs now earn around $50M a year in revenue from that same megawatt, heading (in Dylan's projection) toward $100M. Drag the market price and watch who can still afford to buy compute.

what a megawatt of AI capacity rents for per year
$0$30M$60M$90M$120M

This is the mechanism behind "the labs will outbid everyone". Anyone can profit from a megawatt at $10–15M — the episode's example is renting a GB300 rack, serving open-weight models, and selling tokens on OpenRouter. So the price is already rising. But a lab earning $50M per megawatt can keep paying after everyone else has dropped out, which is how two companies end up taking half of all new compute.

§2Two ways to raise money

A company that needs more cash than it earns can sell ownership or borrow. The labs ran on the first option for years, and the buildout now mostly runs on the second.

Equity venture capital, shares

Investors give you money in exchange for a piece of the company.
  • Never has to be repaid. No ongoing bill.
  • Investors only make money if the company becomes worth more, so they accept losses for years.
  • The cost is permanent, because the slice you sold is gone.
This funded the labs while they lost money on every token — "venture-funded losses", as the episode puts it.

Debt loans, bonds, credit

Lenders give you money that must be repaid, plus interest, on a schedule.
  • Interest is a cash bill that arrives every year whether business is good or bad.
  • Lenders get no upside, so all they care about is whether you can pay. Predictable revenue is what makes debt cheap.
  • You keep full ownership.
This funds the buildout now. Signed multi-year compute contracts are exactly the predictable revenue lenders want.
Meta borrowed at 5–6% recently; the episode says a lab would pay 20% for marginal capacity

You need $100B for datacenters. How do you raise it?

borrowed raised by selling ownership
Annual interest bill
Ownership sold
Ownership-sold assumes a $500B company valuation, roughly the scale the frontier labs raise at. The valuation is an example, not a number from the episode.

The episode's headline numbers: about $11 trillion of AI capex from 2024 through 2029 in SemiAnalysis's modeling, of which roughly $6T can be paid from company cash flows and roughly $5T has to be borrowed. Five trillion dollars of new borrowing is the input to everything in the next two sections.

§3The interest rate is a price, and AI is bidding it up

An interest rate is the price of borrowing money. Like any price, it is set by supply and demand: the world has a pool of savings looking for a return, and borrowers compete for it. A company's rate is the government's rate plus a credit spread for its riskiness (rates move in basis points: 1bp = 0.01%, so "250bps" means 2.5 percentage points). Now add a new borrower who can turn $1 of capacity into $4 of revenue. It will happily outbid everyone for the savings pool — and once lenders can get 8% from Meta, nobody else gets to borrow at 6% anymore. That effect is called crowding out. Slide the AI credit demand and watch it happen.

the episode's modeled number is $5T
Market borrowing rate
Change
The rate curve and the affordability thresholds are invented for the demo and only show the mechanism. The anchors from the episode are real. Meta recently borrowed at 5–6%, would rationally pay 8%, and a 250bps rise repricing everyone else's credit is the scenario discussed.

§4What higher rates do to governments

Governments are the biggest borrowers of all, and they borrow short: much of US federal debt rolls over within about five years, so a rise in market rates becomes the government's actual interest bill quickly. The episode's arithmetic, reproduced here with round numbers, is that US federal tax revenue is about $5T a year, debt is about $30T, and interest already eats about 20% of tax revenue. Raise rates and watch that share.

applied to the whole stock once the debt rolls over

Where a year of tax revenue goes:

interest on the debt everything else the government does
Interest bill
Share of tax revenue

The episode's view is that the US comes out fine, because the datacenters are on US soil and can be taxed. The countries in trouble are the ones that get the higher interest rates with none of the AI revenue — heavily indebted, low tax base, short-duration debt. Pakistan and Nigeria are the episode's examples. The precedent cited (via economist Basil Halperin) is the Volcker shock: when the Fed pushed rates up hard in the early 1980s, some forty countries, mostly in Latin America, defaulted over the decade.

§5Why higher rates crush stock prices

A stock is a claim on a company's future cash flows. Money that arrives in twenty years is worth less than money today, so each future year gets shrunk by a discount rate before you add them up — and that rate tracks interest rates, because the alternative to holding the stock is lending the money out. The bars below are thirty years of a steady company's cash flows ($10B a year, growing 2%) after discounting. Raise the rate and watch the far-out years disappear. Those far years are where most of the value sits.

tied to market interest rates plus a risk premium
Value of all 30 years
vs value at 5%

This is why the episode says a rate rise craters exactly the "safe" stocks — the Johnson & Johnsons and railway companies whose entire appeal is steady cash flows for thirty years. Dylan pushes it further. If AI demand is real enough to blow up interest rates, then everything should trade at two or three times earnings, AI suppliers included.

§6Why supply can't just catch up

If compute is this profitable, why not simply build more? The episode runs the numbers on the deepest layer: about $6B of chip-fab equipment can produce roughly a gigawatt of AI chips every year, and a gigawatt generates around $100B of revenue a year at the end of the chain. The fab keeps producing, each year's gigawatt keeps earning, so the cumulative revenue stacks up fast.

Fab equipment capex
$6B
Cumulative end revenue
Multiple
Each color band is one year's gigawatt of chips, earning $100B every year after it ships. Halve everything for the middlemen — datacenters, power, installation, the model itself — and the multiple is still around 100x.

The episode's answer to why capitalism doesn't instantly close a 100x gap is the bullwhip effect. The demand signal takes years to travel from "labs want more compute" down to the firms that grind mirrors for lithography machines. Zeiss plans on the order of 100 EUV tools' worth of optics a year by 2030 and only recently raised that plan. Meanwhile the shortages produce odd trades — the episode mentions people buying gas turbines purely to resell them, and estimates an EUV machine bought for $400M today would flip for over a billion. Dylan's own answer to "why not just pay Zeiss $10B to expand?" is that you would have to do it for every company in the chain at once, and the world is capital constrained — which is section §2 and §3's story again.

§7The argument, from megawatts to defaults

The steps below follow the conversation in order. Numbers are Dylan's estimates unless marked otherwise.

The labs turned profitable on inference. A year ago every token was sold below cost. Now a megawatt costing $10–15M generates up to $50M for Anthropic — and that margin funds training.
They are buying compute faster than anyone. OpenAI and Anthropic each went from about 2GW to over 5GW in a year, taking ~30% of all new compute — 40–50% next year on contracts already signed.
Because they monetize best, they outbid everyone. Compute that rented for $10–15M/MW is moving to $25–40M (SpaceX is already selling there), and the labs can keep paying past $50M.
The buildout costs more than anyone earns. Capex is over $1T this year and headed past $2T in 2028 — $11T cumulative through 2029 in SemiAnalysis's model. Lab revenue is hundreds of billions. The gap has to be financed.
About $5T of it has to be debt. Hyperscalers already raise debt for capex (Meta at 5–6%). The marginal borrower would rationally pay 8% — the returns still clear it.
Everyone else's borrowing reprices too. A 250bps rise hits mortgages, consumer-goods companies, telecoms, banks. That is crowding out. The lenders they used to rely on now have a better-paying customer.
Government debt service balloons. The US already spends ~20% of tax revenue on interest; +5% on rolled-over debt plus continued borrowing pushes it past half. The US can tax datacenters to cope. Pakistan and Nigeria cannot.
Non-AI equities get repriced. Higher rates mean higher discount rates, and stable-cash-flow stocks lose the most (§5). Dwarkesh's summary is that the S&P survives while the Berkshire-type stock craters.
The precedent is the Volcker shock. The last time rates jumped this way, ~40 countries defaulted in a decade. The episode expects a rerun for AI-less debtor nations.
Supply can't catch up quickly. The 100x gap between fab capex and end revenue (§6) persists because the supply chain reacts on a multi-year lag. Mirror makers and turbine factories take years to expand, and expanding them all at once needs exactly the capital that's scarce.
Meanwhile the labs point compute inward. Dylan's non-consensus call: inference share falls over time, because a megawatt spent on AI research raises all future earnings — worth more than $100M of token revenue. Labs sell tokens mainly to fund bigger training fleets.
Everything points at centralization. Economies of scale in training, compute scarcity, best-model-builds-next-model. Neither host finds a force pushing the other way, and regulation and politics are the only brakes they can name. They leave open whether a decentralized future can take these economics seriously.

§8Test yourself

Twelve questions — half on the finance concepts, half on the episode's argument.

§9The vocabulary, in one place

Capex / opexCapital expenditure buys long-lived assets (buildings, chips, power plants); operating expenditure is the ongoing cost of running things. The $11T figure is capex.
Megawatt / gigawattUnits of power. 1GW = 1,000MW ≈ one nuclear reactor. AI capacity is quoted in power because electricity is the binding constraint.
$ per megawattThe industry's unit economics: annual cost (~$10–15M) or revenue (labs: ~$50M and climbing) per megawatt of capacity.
Gross marginRevenue minus the direct cost of delivering it. "Negative gross margin" means every sale loses money — the labs' position until roughly a year ago.
EquityMoney raised by selling ownership. No repayment. Investors profit only if the company's value grows. Venture capital is early-stage equity.
Debt / creditBorrowed money with a schedule of interest and repayment. Cheaper than equity when revenue is predictable, and the bill arrives regardless of how business goes.
Interest rateThe price of borrowing, set by supply and demand for the world's savings. A borrower with better returns bids it up for everyone.
Basis point0.01 percentage points. "Rates rose 250bps" = from 5.5% to 8%.
Credit spreadWhat a borrower pays above the government rate, priced to its risk. Banks suffer when spreads jump, because their funding reprices faster than their assets.
Crowding outWhen one sector's borrowing raises rates enough to push other borrowers — home buyers, poorer governments — out of the market.
Discount rate / DCFDiscounted cash flow: value an asset by shrinking each future year's cash by (1+r)^years. Higher r destroys far-future value first — that's why stable stocks fall hardest.
Sovereign defaultA government failing to pay its debt. Risk rises with high debt, low tax revenue, and short-duration debt that reprices fast — the episode's worry for AI-less countries.
Bullwhip effectDemand signals amplify and lag as they travel up a supply chain. Why mirror grinders in Germany expand years after labs want more compute.
Rule of 70Divide 70 by a growth rate to get the doubling time. 3% growth doubles an economy in ~23 years; 70% growth, in one.