Who pays for the machines? The balance-sheet mechanics of a $750bn build-out

AI & ROBOTICS INVESTING — EXPLAINED SO YOU CAN ACTUALLY ACT ON IT

Deep Dive · Issue 18 · 10 July 2026 · R. Lauritsen

Who pays for the machines? The balance-sheet mechanics of a $750bn build-out

This week's Signals made a claim: the AI trade is turning into a financing story, and the week proved it — SK Hynix handed $29bn of equity on Thursday, Oracle cut to BBB− the same afternoon, Fed minutes showing participants split 9–8 over the price of every borrowed dollar. The issue named the channels. What it deliberately didn't do is the working: the balance-sheet numbers behind the labels, the hybrid cases that don't fit, what each channel actually costs, and what happened the last three times a funding window slammed shut. That's this piece.

The bill, properly itemised

The scale first: consensus estimates for 2026 hyperscaler capex crossed $750bn this week, per Bloomberg's tally of analyst models. Capex-to-sales is the headline ratio — it's the one S&P cited in Oracle's downgrade — but on its own it's a spending measure, not a risk measure. What matters is spending relative to the cash a business generates and the balance sheet behind it. Here's the fuller picture for the major builders:

Methodology: capex/sales ratios are 2026 analyst-consensus estimates against estimated revenue; company guidance is used where given (Alphabet, Meta) and marked where it isn't. “n/m” = not meaningful. These are estimates that will move with each earnings cycle — treat the column ordering, not the decimals, as the finding.

Read down the last column and the week reorganises itself. The downgrade didn't happen to the biggest spender; it happened to the biggest spender without an internal funding source. A company spending 25 cents of every revenue dollar on construction is investing. A company spending 86 cents without the cash flow to cover it is borrowing its future — and this week the rating agencies started saying so out loud.

What each channel costs

The four channels aren't interchangeable — they carry very different price tags, and the gap between those price tags is the whole thesis.

Customer capital — from commitments to true pre-payments. This channel is a spectrum, and the economics differ along it. A true pre-payment — cash before the asset exists, like the floor-priced memory contracts from earlier this year — is roughly free capital; the supplier even earns float. A purchase commitment, like Apple's $30bn Broadcom agreement, guarantees demand but not early cash: volume risk transfers, while price, timing and renegotiation risk usually stay with the supplier. Both beat borrowing. Only one is free.

Own cash flow — the opportunity cost of patience. No market can shut this channel; only shareholders can, by demanding the cash back. That's why Meta's Model API launch matters beyond the product: revenue against the capex line converts a patience problem into a P&L answer. That's our interpretation of the move rather than Meta's stated rationale — the 29 July call is where it gets tested.

Public equity — priced by enthusiasm. As open as it has been at any point in this boom: a seven-times-covered book measures depth of appetite, not cost — equity's real bill arrives later, as dilution and the returns new shareholders expect. What the oversubscription does establish is that the window is wide open today. The catch is reflexivity — the channel is only open while the crowd believes, and the crowd can read a broken issue price too.

Debt — the channel priced by non-believers. Equity is priced by outsiders too — but by outsiders who believe. The bond market holds no AI conviction at all: coupons, covenants and refinancing windows answer to inflation prints and a Fed whose June minutes put participants' preferences at 9–8. That's the fragility: this channel's price can rise for reasons that have nothing to do with whether the data centres fill up.

The hybrids — where the map gets honest

Assigning one channel per company flatters the analysis. Most real balance sheets blend, so the classification rule we use in the map below is the marginal dollar: whichever channel funds the next data centre, not the last one. Four cases show why the rule matters.

Amazon funds from enormous operating cash flow and has flagged possible equity and debt raises — the marginal dollar is drifting toward the capital markets even while the average dollar stays internal. CoreWeave is debt-funded, but its debt is secured on contracted revenue from customers — meaning its true exposure is a blend of its own coupon and its customers' commitment. SK Hynix raised equity this week while also collecting customer pre-payments; the equity funds the long-dated plants, the pre-payments fund the near-term lines. Meta self-funds — but its backstop commitments to neocloud suppliers (the $15bn Nebius tranche from Issue 17) quietly make it a funding source for someone else's balance sheet. When you classify a holding, ask which dollar you're actually exposed to. And one more distinction worth carrying: company-level versus project-level funding. A company can borrow centrally while a specific site runs on customer deposits or state incentives — follow the project you're actually underwriting.

The map, by name

Nvidia sits in the strongest row and deserves its own caveat, because channel-independence is not risk-independence. It doesn't need equity, debt or anyone's patience — net cash, customers pre-paying for allocation, roughly 74% share of AI chips per The Information. But its order book is concentrated in a handful of hyperscalers, some of whom fund their purchases through the fragile rows of this table. If the debt row seizes, Nvidia doesn't feel it in its balance sheet; it feels it in bookings, two quarters later. Add export policy — which this year has already shown it can switch markets off — and the safest name in the table still carries transmission risk from the weakest.

What happens when a funding window shuts

This isn't the first infrastructure boom to migrate from an engineering story to a financing story, and the history has an uncomfortable pattern: the funding channel, not the demand story, decided how the ending felt.

Telecom fibre, 1999–2002. The build-out was real and demand for bandwidth never stopped growing. But the carriers built on high-yield debt, and when that window shut in 2000, carrier capex fell by more than half within two years, on industry tallies — and the suppliers (Lucent, Nortel, the toll collectors of their day) fell harder than the borrowers' customers ever imagined. Pre-payments and vendor financing turned out to be exposure, not protection.

US shale, 2014–16. An equity-funded boom this time: when the secondary-offering window closed in mid-2015, drilling capex roughly halved within a year, per industry data, regardless of what any individual company's geology promised. The equity channel closed on the whole cohort at once — the exact reflexivity risk SKHY's $149 line now carries.

Cloud capex digestion, 2018–19. The counter-example. Hyperscalers paused capex growth for several quarters — and almost nothing broke, because the build was funded from operating cash flow. No forced sellers, no refinancing cliff, just a pause. Same demand story as the other two; utterly different ending.

The limits of the analogy are worth stating — demand disappointments played a part in each episode too; the funding channel set the speed, not the direction. And today's five biggest builders are the most cash-generative companies in history, which 1999's carriers were not. That's precisely why the framework worries about the edges — the Oracles and CoreWeaves — rather than the centre. In 2000 the edge was the whole sector. In 2026 it's a cohort. But edges are where repricings start.

Three scenarios, one template

Each scenario below carries the same structure — trigger, first-order effect, second-order effect, most exposed, and what would invalidate it. We deliberately assign no probabilities; the tripwires exist so the market can tell us.

Scenario A — the money stays cheap. Trigger: soft CPI (14 July), the Fed's hold camp keeps its majority, SKHY holds $149. First order: credit spreads stay tight; the levered names keep their bounce. Second order: the Applications signal downgrade gets reversed within a month and the best trade was buying the fear. Most exposed: anyone who de-risked into the scare. Weakened by: a hot CPI or a poorly received bond deal — either puts the thesis straight back on the table.

Scenario B — the price of money turns. Trigger: a hot CPI recruits the Fed's next hiking vote, or Oracle's next benchmark bond prices wide. First order: the debt row of the map reprices — Oracle, CoreWeave, the levered tail. Second order: the market starts paying up for self-funding balance sheets; dispersion inside “AI” exceeds the sector's move itself; supplier bookings from the levered cohort soften with a lag. Most exposed: debt-funded builders, then their suppliers. Invalidated by: two consecutive soft inflation prints — two, because a single print is noise in this regime. The faster tell is investment-grade spreads and the two-year yield, which aggregate the market's verdict in real time.

Scenario C — the equity window cracks first. Trigger: SKHY breaks $149 inside a fortnight. First order: the raise calendar behind it — Korean secondaries, neocloud follow-ons, the rumoured trillion-dollar lab listings — reprices or postpones. Second order: capacity plans that assumed cheap equity quietly shrink; the supply wave the memory bears feared arrives smaller and later, which is perversely bullish for incumbents' pricing. Most exposed: recent issuers and anyone needing to raise within twelve months. Invalidated by: SKHY holding issue price through end-July on real volume.

The five-question stress test

The weekly issue asked you to label each AI holding by funding channel. This is the sharper tool: five questions, one point each, answerable from any annual report or free financial site. It's a heuristic, not a credit model — five blunt questions that catch most of what matters.

1. Does trailing-year operating cash flow cover guided capex? (Yes = 1)

2. Is interest coverage — operating profit over trailing-year interest expense — above roughly 4x? (Yes = 1. Why 4x: a common investment-grade comfort line; below about 2x, the debt is steering the business.)

3. Could cash on hand retire the debt maturing in the next 24 months? (Yes = 1. Why 24 months: long enough to ride out a shut refinancing window without new borrowing.)

4. Are its customer commitments contractual pre-payments or take-or-pay — rather than cancellable orders? (Pre-paid = 1)

5. Would a meaningful rise in its borrowing costs — say 150 basis points, roughly what past credit scares added to weaker issuers' spreads — leave the build plan intact? (Yes = 1. Mostly fixed-rate debt softens the blow; floating-rate or near-term refinancing takes it immediately.)

Scoring: 4–5 builds through a squeeze. 2–3 is fine until the window narrows — watch its bonds, not its stock. 0–1 means you don't own an AI bet; you own a rates bet wearing an AI badge. For calibration: Microsoft scores 5 of 5 on public figures. CoreWeave, on its own filings — capex above revenue, interest consuming about half of EBITDA, roughly $25bn of debt — scores 0 or 1, and the difference is question four: its backlog is contracted and long-dated, but public filings don't establish how much is pre-paid or take-or-pay rather than commitments. Award the point only if your own reading of the filings does. Either way, the spread between it and Microsoft — two names the market files under the same trade — is the entire point of this piece.

The bottom line

$750bn doesn't get spent; it gets funded — and the four channels that fund it are diverging in price for the first time since the boom began. Equity is euphoric. Contracts are gold. Cash flow is quietly winning. Debt is one CPI print from expensive. Run the five questions on what you own, and you'll know before the bond market tells you.

The headline said record listing. The rating action said last rung. Both were about the same machines.

Sources for figures in this piece: S&P Global (Oracle rating action), company guidance (Alphabet, Meta capex ranges), Morgan Stanley (Amazon FCF model — a bank estimate), CoreWeave public filings (debt and interest burden), The Information (Nvidia share), and press reports on the SK Hynix ADR pricing. Where a number is a consensus estimate rather than a reported figure, it is marked as such above.