Follow the AI Capex Dollar

Everyone quotes the headline spend number. Far fewer trace where it actually goes. So I mapped all $407 billion of the five hyperscalers' latest-year capex — from the balance sheets that wrote the checks, through the spending buckets, down to where it finally settles as someone's gross profit. The map answers two questions the headline can't: who keeps the money, and how much longer it stays self-funded.

The number that gets the headlines is the spend itself — $700 billion in 2026, more than $1 trillion penciled in for 2027. It's a genuinely staggering figure, and it's also where most of the analysis stops. But a capex dollar doesn't vanish into "AI." It leaves a hyperscaler's balance sheet, lands in a specific spending bucket, gets paid to a specific supplier, and ends its journey as gross profit on someone's income statement or as the cost of physical inputs — concrete, copper, wafers, electricity. Map that whole journey and the spending stops being an abstraction. It becomes a set of very concrete claims about who is actually getting paid, and how the people writing the checks are paying for it.

This piece does that mapping for the five companies that dominate the build — Amazon, Microsoft, Alphabet, Meta, and Oracle — using their latest reported fiscal year. The funding source and the total are hard numbers from SEC filings: cash capex plus finance-lease additions, what I'll call economic capex, because a leased GPU cluster is every bit as real a commitment as a bought one. The downstream splits — how much of each bucket goes to which supplier layer — are modeled estimates, not disclosed bills of materials, and I'll flag that plainly wherever it matters. The point isn't false precision. It's the shape of the thing.

The map, by year

Before tracing any single dollar, look at how the five companies split their spend across the six destination buckets — and how that split scales over three years. Each grid below is one scenario: companies down the side, buckets across the top, every cell the modeled dollars ($B) flowing from that company into that bucket. Darker means more money. The three grids share one color scale, so the story isn't just where the money goes in any one year — it's watching the whole board go dark as the build roughly triples.

Cells are modeled company-to-bucket allocations in $B; shared color scale across all three years. Latest-reported totals are SEC cash capex plus finance-lease additions; 2026E/2027E are FT/Goldman-style capex scenarios. Destination splits are modeled, not company-disclosed.

Two things jump off the board. First, the Accelerators & AI servers column is the darkest in every year, for every company — it's the single most-funded destination on the map, and it only deepens with time. Second, the grids don't just get darker uniformly; they get darker fastest in the compute and power columns, which is the spending mix tilting harder toward AI as the build scales. Oracle's row stays the lightest in absolute dollars but climbs the most in percentage terms — a tell I'll come back to on the funding side.

The rest of this piece is two readings of that board. First: the money concentrates, violently, at the silicon layer once you follow each bucket one step further down to its suppliers. Second: the way it's being paid for is quietly changing character — and that second shift is the one I think the market is under-weighting.

Reading one: where the money lands

Start with the six buckets every capex dollar flows into. This is the first fork in the road — before any of it reaches a named supplier, the hyperscalers are implicitly deciding how much of the build is compute versus the physical plant that houses and powers it.

Accelerators and AI servers take roughly 37 cents of every dollar — about $149B, more than double the next bucket. That alone tells you the build is, at its core, a compute build: the data centers, the power, the cooling, and the construction are all in service of cramming as many accelerators as possible into them. The three "physical plant" buckets — construction, power/cooling, and networking — together come to about $163B, roughly on par with the compute bucket. That's the part of the story the power and electrical names have been riding, and it's real. But it's not where the margin pools.

The bucket fractures at the supplier layer

Follow that $149B accelerator bucket one step further down and it splits across the supply layers — but unevenly. The bulk of it, around $80B, flows straight to NVIDIA and AMD. The memory makers (SK hynix, Samsung, Micron) take roughly $24B of HBM and DRAM. Custom-silicon designers — Broadcom, Marvell, and the hyperscalers' own internal TPU and Trainium programs — take about $15B. The server integrators (Dell, HPE, Supermicro, the ODMs) take around $19B, most of which is pass-through component cost they don't actually keep.

And that distinction — what a supplier receives versus what it keeps — is the whole game. A server integrator booking $19B of revenue at a low-teens gross margin keeps a sliver. NVIDIA booking $80B at a ~71% gross margin keeps most of it. So the more revealing chart isn't who gets paid. It's who keeps it.

This is the picture in one bar. NVIDIA and AMD's accelerator layer captures about $55B of gross profit — and it isn't a close race. That single layer keeps more gross profit than the entire power-and-grid chain, the construction-and-EPC chain, and the networking-systems chain combined. Vertiv, Eaton, GE Vernova, Constellation, the EPC contractors, Arista, Cisco — add up everything those layers retain and you're still short of what flows into two chip designers' margin line.

It's worth being precise about why this happens, because it's mechanical, not magical. The accelerator layer sits at the one point in the chain with no substitute and no competition worth the name. Concrete has a dozen suppliers; megawatts have a hundred; a frontier training cluster has, functionally, one-and-a-half. Scarcity at a bottleneck is what lets a supplier price at 70% gross margin while the contractors pouring the foundation underneath it work for single digits. Every layer in this map is doing real work. Only one of them is doing it from a position where the customer has nowhere else to go.

The mirror image is just as instructive. The construction, power, and electrical layers move enormous dollars — they have to; you can't run a gigawatt data center on enthusiasm — but they keep a thin slice of each one. EPC and site work runs single-digit to low-teens gross margins. The OEM/ODM server integrators are worse, often high-single-digit. These are good businesses riding a real volume wave, and the wave is large enough that even a thin margin on a huge number is meaningful. But they are volume stories, not margin stories, and the map makes the difference impossible to miss.

Reading two: the funding is about to change character

So far this is a story about a pie and who eats the biggest slice. The second reading is more interesting, and I think more under-appreciated, because it's about the left edge of the map — the funding sources — rather than the right.

Here's the thing almost nobody emphasizes: the current build is overwhelmingly self-funded. Of the $407B of economic capex, roughly $378B came straight out of operating cash flow, and only about $29B from finance leases — and most of that is one company (Microsoft). On the latest-year numbers, this is a ~93%-cash-funded buildout. It's a flex of balance-sheet strength, not a credit story. These businesses are generating so much cash that they can spend at this scale without meaningfully touching the debt markets.

But that's the current snapshot. Now run it forward on the spending scenarios the sell-side is actually using — roughly $725B of Big Four capex in 2026, and a Goldman-style ~$1.1T hyperscaler total by 2027 — while holding each company's operating cash flow flat at its latest reported level. That second assumption is the load-bearing one, and it's deliberately conservative: it assumes these companies keep generating cash at today's pace, no better. Under that frame, the arithmetic is unforgiving. If the build roughly triples while the internal cash engine stays the same size, the gap has to be borrowed.

The debt-funded share goes from 7% to about 26% to nearly 48% across the three periods. Said plainly: on these assumptions, by 2027 roughly half of the entire hyperscaler build is being financed rather than paid for out of pocket. That is a different kind of build. A self-funded capex boom is an earnings-quality debate — are these investments earning their cost of capital, are the depreciation schedules honest. A half-debt-funded capex boom is a credit story, and credit stories have a different failure mode: they don't depend only on whether the AI demand shows up, they depend on whether the financing stays cheap and available while everyone waits to find out.

The flat-cash-flow assumption is the conservative anchor, not the only case — so the chart lets you relax it. Grow each company's cash generation alongside the build and the debt share eases, but it's stubborn: even at a healthy +10%/year, the 2027 debt-funded share only falls to about 37%. It takes sustained +20%/year cash-flow growth — roughly the group's recent pace, and a strong assumption to hold for two years against a tripling capex bill — to keep 2027 down near a quarter financed (~24%). In other words, the funding shift isn't an artifact of pessimism about cash flow. It shows up even when you're generous, because the capex line is simply growing faster than any plausible cash engine behind it.

Oracle is the canary

The aggregate hides the company that gets there first. Oracle has the smallest cash engine of the five — roughly $21B of operating cash flow against a capex ambition that the scenarios put near $55B in FY2026 and as high as ~$95B by FY2027. It simply cannot self-fund the way Amazon or Alphabet can. On the same flat-cash-flow framing, Oracle's debt-funded share runs from roughly 14% today to about 63% in 2026 to nearly 78% by 2027. It is, by a wide margin, the most leveraged version of this trade — spending like a hyperscaler on a software company's cash generation, and borrowing the difference.

That's not a prediction that anything breaks. Plenty of great businesses have been built on borrowed money against a real demand wave. It's an observation about where the risk has moved. For the cash-rich four, the AI-capex question is still mostly "is this a good use of your own money." For Oracle, and increasingly for the marginal dollar across the whole group, it's becoming "what happens to this build if credit gets more expensive before the revenue arrives."

How the build could go — three ways to hold the funding picture

What the map is really telling you

Two findings, and they pull in different directions, which is exactly why both are worth holding at once. The concentration finding is a story about strength: the AI build routes its richest margin stream through a tiny number of suppliers who can price like monopolists because, at the bottleneck, they nearly are. That's a durable, structural advantage, and it's priced as one.

The funding finding is a story about a changing risk profile: a buildout that has so far been a demonstration of corporate cash-generation strength is, on the spending scenarios everyone's using, about to become substantially debt-financed — and the company least able to self-fund is leaning into it the hardest. Neither finding is a forecast that anything goes wrong. Both are just what falls out of taking the headline number seriously enough to ask where each dollar actually goes, and where each dollar actually comes from.

The headline asks how big the AI build is. The map asks two better questions: who keeps the money, and who's borrowing to spend it. The answers aren't the same companies.

The interesting thing to watch from here isn't the spend number — that's going up, everyone knows it. It's the two ratios underneath it. The share of capex flowing into accelerators tells you whether the toll booth is holding or starting to crack. And the share of capex funded by debt tells you whether this is still a story about earnings, or quietly becoming a story about credit. Both are knowable, both move slowly enough to track, and both will tell you more than the next trillion-dollar headline.

Figures and assumptions: company funding sources and totals are from latest reported SEC filings (cash capex plus finance-lease and operating-lease asset additions; company facts via SEC XBRL). Destination-bucket and supplier-layer splits, and the gross-profit capture by layer, are Hammockistan modeled estimates — not company-disclosed bills of materials — built by applying reported segment/company gross margins to modeled revenue allocations. 2026 capex uses FT/Tom's Hardware coverage of roughly $725B across Amazon, Microsoft, Alphabet, and Meta, plus Oracle's ~$55.7B FY2026 capex from post-earnings coverage; 2027 uses a Goldman-style ~$1.1T hyperscaler scenario with Oracle fixed near its ~$95B FY2027 ceiling. The funding-crossover math holds each company's latest reported operating cash flow flat as cash-funded capacity and shows the remaining capex gap as debt/lease funding; equity issuance is modeled at zero. These are illustrative scenarios meant to show the shape and direction of the funding shift, not point forecasts. Underlying data are from the AI-capex value-chain model (June 2026).