The AI Capex Ledger: Who Pays, Who Earns, and What the Bond Market Is Missing

GeometricInvestor

Jun 12, 2026

The AI debate is stuck on the wrong question.

The question is not whether AI is a bubble. Nor is it whether NVIDIA is expensive, whether Michael Burry is early again, or whether productivity gains will eventually lower inflation. The better question is an accounting one: who needs to earn what return for the AI capex cycle to make sense?

At the bottom of the stack, GPU, HBM, networking, power, and cooling suppliers are already earning. The capex is real. The checks have cleared. That was the first phase of the trade.

The harder question sits one layer above. Hyperscalers and neoclouds are converting capital into compute. Compute becomes tokens. Tokens must become revenue. Revenue must become gross profit after depreciation, power, financing, and model costs. And finally, the buyers of those tokens must earn a return high enough to keep spending.

Only then does AI become a true macro productivity shock rather than a capital-spending boom with better branding.

So this is not another bubble essay — that genre is already a landfill with charts. It is not a valuation note on any single stock, and it is not a referendum on Michael Burry’s track record. It is also not a claim that AI is inflationary, or that it is disinflationary; it will turn out to be both, through different channels, on different clocks. The claim is structural: AI is a chain of required returns. A stack of linked ledgers, each of which must clear its own hurdle rate for the layer below it to stay funded. The market has priced the first ledger with enthusiasm. The macro consequences — for growth, for inflation, and above all for the long end of the bond market — depend on the ledgers it has barely begun to audit.

Every AI headline — a GPU order, a neocloud financing, an enterprise pilot, a Treasury selloff — is an entry on one of those ledgers, posted against one of those hurdles. What follows is the map.

The Four Ledgers of AI

Why does the bubble framing fail? Because it compresses four different balance sheets into one word.

A bubble verdict treats AI as a single asset with a single price. But the AI cycle is a stack of four linked ledgers, with different owners, different time horizons, and different required returns:

The infrastructure ledger. NVIDIA, HBM and memory suppliers, foundries, advanced packaging, networking, power equipment, and cooling. They earn today because hyperscalers and neoclouds are spending today. Their proof is quarterly and financial.

The hyperscaler and neocloud ledger. Cloud incumbents and GPU-rental specialists must convert installed compute into rented or sold capacity at utilization and gross margin sufficient to cover depreciation, power, financing, and the risk that the assets age out before they pay back.

The token buyer ledger. Enterprises and consumers must receive more value from tokens than the tokens cost — labor savings, revenue lift, faster software, automation, fewer errors. If the buyer’s return is negative, every ledger below it is being funded by a future that will not arrive.

The macro ledger. If token usage raises economy-wide productivity, trend growth and the neutral real rate can rise. If AI capex strains power, copper, grid equipment, and capital markets faster than it raises productivity, the same buildout shows up as bottleneck inflation and term premium instead. Either way, the bond market is in this trade whether it wants to be or not.

The geometry matters, so make it concrete. Each ledger is the revenue line of the one below it. NVIDIA’s revenue is hyperscaler capex. Hyperscaler revenue is enterprise token spend. Enterprise token spend is justified only by corporate operating leverage. And corporate operating leverage becomes macro productivity only if it is broad enough to move the aggregate data.

That is why AI cannot be judged by the bottom ledger alone. The bottom ledger can look spectacular while the top ledger is still unproven. And it is why the stack carries a peculiar funding asymmetry right now. The bottom has been paid in cash. The middle has paid in capital. The top has, so far, paid mostly in expectations.

That asymmetry is the entire macro question.

The Capex Is Real. That Was Phase One.

For the bottom ledger, the proof requirements are old-fashioned and near-term. Are orders real? Are margins holding? Are lead times tight? Are ASPs firm? Are customers taking delivery? Are inventories clean?

On those tests, the first ledger has already closed its books. NVIDIA’s most recent quarter — Q1 of fiscal 2027 — reported data-center compute revenue of $60.4bn and data-center networking revenue of $14.8bn, the latter up 199% year over year.1 Memory and HBM remain structurally tight. Power and cooling sit in the queue-constrained part of the cycle, where the binding question is delivery slots, not demand.

So the first-layer trade was simple: capex was real, supply was tight, suppliers earned. That part is not the debate anymore.

The relevant question about the bottom ledger is different: how long can the layer above keep paying it? Supplier earnings are not self-justifying. They are a derivative of someone else’s capital budget, and capital budgets are a derivative of someone else’s expected return. Which is why the cycle is best read not as one trade but as a ladder of hurdles.

The Required Return Ladder

Every layer’s revenue is the layer above’s cost. So each layer must generate a return that lets the next layer keep paying. Read from the bottom of the stack up, the ladder looks like this:

The middle rung deserves its arithmetic, because it anchors everything above and below it. Compute is short-lived capital. If the economic life of a deployed AI dollar is on the order of four to six years, depreciation alone consumes roughly 17–25 cents of it per year. Power, cooling, and operations take their share on top. Financing is no longer free. And the capital is supposed to earn a return, not merely amortize itself. Stack those, and annual monetizable AI revenue — revenue someone actually pays for AI capability, as opposed to internal usage and bundled giveaways — probably needs to approach something like 0.35–0.60x deployed AI capex over time for the system to earn its keep. Call the central hurdle 0.5x. The band is there to be argued with: stretch asset lives toward eight years and the hurdle slides toward 0.3x; shorten them toward three, or let power and financing costs climb, and it pushes past 0.6x.

Now scale it. If cumulative AI capex eventually approaches $3 trillion, the system needs roughly $1 trillion to $1.5 trillion of annual monetizable AI revenue — the bottom of the band up through the central 0.5x hurdle; the full 0.6x top would demand $1.8 trillion — to fully validate the capital cycle. That band is an estimate, not divine scripture delivered from Mount Spreadsheet. Its job is not to be right to the decimal. Its job is to give the debate a number — because a debate with a number can be settled by evidence, and a debate without one just gets louder. Whatever today’s monetizable AI revenue actually is — definitions vary widely enough to be their own argument — nobody serious puts it near a trillion dollars a year. The gap between here and the hurdle is not a footnote to the AI cycle. It is the AI cycle. The exact number matters less than the direction of travel: the ratio of monetizable AI revenue to cumulative AI capex has to rise materially — and soon — for the supplier-led capex boom to become a self-funding productivity cycle.

But the 0.5x hurdle is only the middle layer. It tells us what hyperscalers and neoclouds need to earn on deployed compute. It does not tell us whether the enterprise buyer earns a return on tokens. That is the next and more important hurdle. If a hyperscaler earns a 40% gross margin selling tokens, but the buyer gets only 70 cents of value for every dollar spent, the system can grow for a while but cannot compound. Eventually the buyer stops paying, and the hyperscaler’s return collapses back into the supplier’s order book.

The ladder, in other words, is only as strong as its weakest hurdle. The next three sections climb it.

The Token Ledger

What does a hyperscaler actually have to prove?

Not that demand exists. Demand observably exists. What the hyperscaler must prove is that capital converted into compute converts into gross profit. The chain runs through five conversions — capital into compute, compute into tokens, tokens into revenue, revenue into gross profit, gross profit into an acceptable return on the deployed base — and each conversion has a loss factor:

GPU and server depreciation,

power,

cooling,

financing,

networking,

maintenance,

model and inference costs,

obsolescence risk,

utilization risk.

Here a fence is needed, because the most quoted statistic in this debate is the wrong one. Revenue per token is not the metric. Falling revenue per token is routinely cited as evidence the economics are deteriorating. On its own it proves nothing. If inference costs fall faster than prices, and volume elasticity is high, falling revenue per token is exactly what a successful cost-curve collapse looks like from the inside. Cheaper tokens can mean more gross profit dollars, not fewer. Usage alone proves nothing either — usage is a necessary condition for the economics, not evidence of them.

The metrics that actually settle the token ledger are:

total token gross profit dollars,

gross profit per watt,

utilization-adjusted compute margin,

AI revenue as a share of cumulative AI capex,

depreciation-adjusted return on deployed compute.

None of these are disclosed cleanly today. That is worth pausing on. The decisive metrics of the largest capital-spending cycle in technology history are currently being estimated by outsiders from fragments — an accounting vacuum that both bulls and bears fill with temperament.

The Buyer’s Ledger

Why would anyone keep buying a trillion dollars of tokens a year?

This is the least discussed layer and probably the most important. Enterprises do not buy tokens because Sam Altman needs a data center. They buy tokens if — and only for as long as — the tokens generate a return. That return must show up somewhere concrete:

fewer labor hours,

higher output per worker,

faster software development,

customer-service automation,

lower SG&A,

higher conversion,

faster product cycles,

better retrieval and fewer errors,

lower external service spend.

The arithmetic is brutally simple. If an enterprise spends $1 on AI tokens and gets $1.30 of value, the cycle can compound: budgets grow, deployment widens, and token spend becomes a durable expense line, like cloud or electricity. If it gets $0.70 of value and keeps spending because the CEO wants to sound modern on earnings calls, that is not productivity. That is shareholder-funded theater — and CFOs eventually close theaters.

The current evidence is mixed in exactly the way you would expect a few years into a general-purpose technology: visible, measurable wins in narrow domains — software development, support automation, document-heavy workflows — and a long tail of pilots that have not yet escaped the innovation budget. The macro question is whether token spend graduates to the operating budget. Innovation budgets are sentiment. Operating budgets are returns.

And if buyers do earn their return, where does it show up first? Not in GDP. One ledger up.

The Corporate-Profit Ledger

The rosy AI case is not simply higher GDP. It is higher corporate operating leverage.

If enterprises earn positive returns on token spend, the first visible macro-financial evidence should be corporate margins. A firm using AI well should be able to grow revenue faster than headcount, reduce SG&A intensity, shorten software and product cycles, automate support, improve conversion, or cut external service spend. That is how token ROI becomes corporat

[truncated for AI cost control]