Following the AI capex money

1. The bill, and who pays it

Start with the buyers, because their capital spending is the revenue of almost everyone else on the map. Over the latest four quarters Microsoft, Amazon, Alphabet, Meta and Oracle generated $692B of operating cash flow and spent $566B of it on property and equipment. Two years earlier the same five spent $184B. That is +208% in two years, at companies that were already the largest corporate spenders in the world.

The ratio that matters is capex to operating cash flow, because it shows how much room is left. It bottomed at 35% in Q3 2023, when the build-out was just starting, and reached 97% in Q2 2026. Company by company the picture varies. Microsoft and Alphabet still fund their spending comfortably from cash. Amazon and Meta are closer to the line. Oracle has crossed it: $75.7B of capex against $46.9B of operating cash flow, so the difference has to come from its balance sheet or from new borrowing. CoreWeave, the largest of the GPU-only clouds, spent $20.6B while generating $6.9B from operations; its model is to borrow against customer contracts.

The chart below shows where the six buyers’ operating cash went. Most of it flowed into capex; what is left over is free cash flow, which is also what pays for buybacks and dividends. The $54.0B on the left that doesn’t come from operations is the gap Oracle and CoreWeave had to fund some other way.

2. The spending you don’t see in capex

Capex is only the part of the build-out a buyer pays for up front. The rest is leased. Microsoft, Meta, Oracle, Alphabet and CoreWeave each disclose data-center leases they have signed that have not yet started. Because the buildings aren’t ready, those leases aren’t in capex and aren’t on the balance sheet yet. Added up, they come to $1.11T of future payments, more than a full year of the buyers’ capex, committed for terms of up to 20 or 30 years.

Two points matter for an investor. First, a lease is still a fixed cost. When these leases start, rent and depreciation of the leased assets will flow through the income statement whether or not the capacity is used. Second, the landlords are often new: Meta’s two largest projects are held in joint ventures it doesn’t consolidate, backed by residual value guarantees (see the money loop). Meta also reports $454.5 billion of non-cancelable contractual commitments, most of them for third-party cloud capacity. The full size of the build-out is visible only in the notes to the financial statements.

• Microsoft: “As of June 30, 2026, we had additional leases, primarily for datacenters, that had not yet commenced of $329.1 billion, with some arrangements subject to certain contractual conditions being met.” 10-K, July 29, 2026
• Meta: “These lease obligations were approximately $355.73 billion, consisting of data centers, energy-related infrastructure, colocations, and certain network infrastructure, which will commence during the remainder of 2026 through 2036 with lease terms ranging from greater than one year to 30 years.” 10-Q, October 22, 2026
• Oracle: “As of May 31, 2026, we had $260 billion of additional lease commitments, substantially all related to data center arrangements, that are generally expected to commence between the first quarter of fiscal 2027 and fiscal 2029 and for terms of fifteen to nineteen years that were not reflected on our consolidated balance sheet as of May 31, 2026 or in the maturities table above.” 10-K, June 22, 2026
• Alphabet: “As of September 30, 2026, we have entered into leases, primarily related to data centers and energy infrastructure, that have not yet commenced with future lease payments of $130.5 billion that are not yet recorded.” 10-Q, October 29, 2026
• CoreWeave: “The aggregate amount of estimated future undiscounted lease payments associated with such leases is $35.5 billion.” 10-Q, August 12, 2026

3. Where the dollars land, and who keeps them

Follow the dollars down a layer and the question changes from how much to who keeps it. Over the past two years the supplier layers we track added $627B of revenue. The chart splits that gain into what each layer kept as gross profit and what it passed on as cost of revenue: to its own suppliers, its factories and its freight bills.

The pattern is stark. Chip companies kept 72% of their extra revenue and memory makers 93%; between them they took 83% of all the extra gross profit. Server makers kept 14%. That is not because they run their businesses badly. An AI server is mostly GPUs and memory, priced by the companies that make them, so the assembler’s revenue rises with the chip prices while its profit rises only with the work it adds.

For an investor this is the most useful single test of an AI supplier: does its gross margin rise when its revenue does? A rising margin means customers are competing for a scarce product. A flat or falling margin on rising revenue means the company is a conduit. Both can make money, but they deserve very different valuations, and they behave very differently when the cycle turns.

The money keeps going down. Nvidia, the largest recipient, reports $279 billion of supply and capacity commitments as of its latest quarter, up from $119 billion three months earlier: the next layer’s revenue, contracted in advance. Data-center revenue at Nvidia was $278B over its latest four quarters, +90% on the year before.

4. Margins show where the scarcity is

Margins are where scarcity shows up. The chart compares each layer’s gross margin over the latest four quarters with the same period two years earlier. Memory moved the most, from 25% to 75%: high-bandwidth memory, which is stacked next to every AI accelerator, has been sold out, and conventional memory prices rose with it. Chips went from 61% to 66%. Servers went the other way, from 17% to 16%, because a larger share of what they sell is someone else’s silicon.

Memory is the layer to watch most closely. Its margins have historically swung from losses to records and back within a few years, as makers add capacity in good times and cut prices in bad ones. If high-bandwidth memory keeps its premium as new factories come online, the industry has changed. If memory margins fall while AI spending is still rising, it has not. Either way the margin, not the revenue, will tell you first.

5. The depreciation still to come

Capex becomes expense slowly. A server bought today is depreciated over the five or six years its owner expects it to last, so this year’s spending shows up in profits for years to come. In Q2 2026 the five buyers spent 3.9× what they recorded as depreciation and amortization. Their depreciation over the latest four quarters was $165B, +108% in two years, while capex rose +208%. That gap is depreciation still to come.

How fast it arrives depends on the useful lives the companies choose, which are management estimates. Since 2022 every one of these buyers has lengthened the life of its servers at least once, which lowered depreciation and raised reported profits. Amazon is the exception that proves the point: it shortened the life of part of its fleet back to five years in 2025, which its 10-K puts down to “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” The table quotes what each says today. If the newest accelerators turn over faster than their predecessors, which their makers’ annual product cycles suggest is possible, these lives could prove too long, and earnings would take the difference.

Source: each company’s 10-K and 10-Q text on property and equipment, read by our useful-life parser (every reading quotes its sentence; hover a life to see it). D&A is the cash-flow statement’s depreciation and amortization over the latest four quarters.

6. Is the money coming back?

Now the return trip. Revenue at the four big cloud businesses whose results are disclosed (AWS, Microsoft Cloud, Google Cloud and Oracle’s cloud infrastructure) reached $132B in Q2 2026. The three with a longer reported history, AWS, Microsoft Cloud and Google Cloud, took in $126B that quarter against $57.3B in Q1 2023: 2.2× in 13 quarters. Growth is accelerating: AWS grew +28% over its latest four quarters, Google Cloud +76% and Oracle’s infrastructure business +94%.

Over the same period the buyers’ quarterly capex rose 5.0×. The second chart puts both on one scale, indexed to 100 in Q1 2023. The lines tell the central story of the AI build-out: the spending is running ahead of the revenue it is meant to earn. That is normal early in an infrastructure cycle, since the capacity has to exist before it can be sold. It becomes a problem if the gap keeps widening once the capacity is live.

There is better news in profits. Google Cloud’s operating income over the latest four quarters was $31.8B, +198% on the year before. AWS earned $54.7B, +28%. Microsoft’s Intelligent Cloud segment earned $57.0B. Cloud businesses are absorbing the depreciation so far and still widening their profits. That is the strongest evidence in the filings that the spending is paying. The test is whether it holds as the leases and capex above reach the income statement.

7. The backlog: revenue promised, not yet earned

Backlog is the bridge between spending and revenue: contracts customers have signed for services not yet delivered, reported as remaining performance obligations. Several AI companies now report backlogs larger than a year of revenue, and Oracle’s is many years of revenue.

Timing matters as much as size. Oracle says it expects to recognize about 13% of its $664 billion as revenue in the next twelve months and the rest over the following years. Microsoft expects about 30% of its backlog within twelve months. A backlog that turns into revenue slowly says the capacity to serve it has yet to be built, which is why Oracle is spending more than it earns. Backlog is also only as good as the customer: a contract with a company that is itself funded by investors carries that company’s risk.

What to watch in the next filings

• The buyers’ capex-to-cash-flow ratio. Above 100% for the group means the build-out depends on debt markets. Watch Oracle’s and CoreWeave’s borrowing costs as the early signal.
• Cloud segment growth against capex growth. The lines in the index chart need to converge within a few years. If cloud growth slows while capex keeps rising, returns are falling.
• Gross margins at the toll booths. Falling margins at Nvidia or the memory makers while their revenue still grows would be the first sign of supply catching up with demand.
• Useful-life changes. A buyer that shortens the life of its servers is telling you the hardware is aging faster than planned, and its earnings will absorb more depreciation.
• Leases starting. As the leases signed but not yet started begin, watch operating lease cost and finance lease depreciation in the cloud segments’ margins.
• New financing links. Each new investment, guarantee, warrant or capacity buy-back in a filing means a supplier is carrying more of its customer’s cost. We add them to the money loop as they appear.

Each layer has its own page with the company-by-company numbers: AI clouds, servers, chips, memory, networking, power & DCs, foundry & tools, software.