Hyperscaler Cash and Debt: Can Operations Pay for It, How Much Is Borrowed, and Can They Afford It

This is the key chart of the board and the check for map station 4. CapEx ÷ operating cash flow, latest quarter Amazon 178.9%, Google 121.4%, Meta 106.5%, Microsoft 77.9%. Above 100% the business does not generate enough cash to pay for capex, and the gap is bonds or cash reserves. What station 4 tests is not how high the ratio gets but whether operating cash flow grows faster than capex by the end of 2027 and turns this line around.

Left: operating cash flow minus capex. Amazon $-33B and Google $-9B are both negative, Meta $-2B is near zero, and only Microsoft $13B is clearly positive. Right: cash plus short-term investments minus borrowings, excluding lease liabilities; Microsoft is down from $55B to $29B, Meta has only $2B left, Google still has a $140B cushion. Station 5, free cash flow turning positive, is about when the left chart gets back above zero.

Left: free cash flow minus buybacks minus dividends, trailing four quarters. The gap to free cash flow is the burden of shareholder returns; Microsoft's roughly $42B a year of returns absorbs all its free cash flow, leaving headroom at zero. Google and Meta look better only because they cut buybacks; Amazon has deteriorated to $-46B. Right: shrinking buybacks are the leading sign of a funding squeeze, appearing before bond issuance; Google went from $15B a quarter to zero.

Left: quarterly capex as each company reports it, Microsoft $45B, Amazon $76B, Google $51B, Meta $35B, all at records; Microsoft and Meta include finance leases, Google is cash basis, Amazon is property and equipment plus finance leases. Right: cash-basis actuals for the four plus Oracle against analyst estimates (dashed), with actuals running ahead of estimates and the estimate line still chasing; by late 2027 the five together are pencilled in above $300B a quarter.

Left: capex plus new operating leases, adding back the compute that sits off the balance sheet, a step above reported capex; some quarters are undisclosed, so lines break. Right: new operating leases as a share of all-in investment; a rising share means compute spending is leaving the capex line. The recent decline is because cash purchases of GPUs and servers inflated the denominator while new operating leases are mostly real estate; once the $800B of leases signed but not started in the next section begin, this line turns back up.

Left: leases signed but not yet started, Microsoft $365B, Meta $356B, Amazon $137B, Google $131B, the most forward-looking capacity signal, every dollar of which becomes rent over the next few years. Middle: the stock of purchase and contract commitments, holding GPUs, take-or-pay power contracts and cloud capacity, Google $802B, Meta $455B; Google is on a new combined definition from 1Q26. Right: operating lease liabilities already running, debt in another form, not counted in the long-term debt chart.

Left: annual capex estimates for listed neoclouds plus SpaceX; SpaceX alone at $187B in 2027 exceeds the five listed neoclouds combined, and its 2Q26 actual was $18.4B with 86% of it AI compute. Right: NVDA revenue (calendarized) as a ratio of the big five's combined capex, stable between 0.45 and 0.54 over the long run because NVDA sells beyond the five and the five buy more than chips; a sharp drop means in-house chips are displacing NVDA faster, a sharp rise means buyers outside the five are ramping, and 0.54 in 1Q26 is the latter.

Left: gross bond issuance by quarter. Amazon $53B and Oracle $27B in 1Q26 and Meta $30B the quarter before are the three spikes that mark the turning point in how this capex cycle is funded; before that the big four barely issued. Right: long-term debt outstanding, with Oracle $135B rising steepest and Amazon at $173B; lease liabilities are excluded and belong alongside the operating lease chart in the previous group.

Left: debt as a share of total assets, where only Oracle 55.0% stands out, because its expansion is almost entirely bond-funded on a small asset base; the big four are diluted by their huge assets no matter how much they issue and always look safe. Right: the rating agencies watch debt ÷ EBITDA, where borrowed cash and finished data centers both count in the denominator. Oracle is at 4.9×, cut by S&P to BBB− with a negative outlook in July 2026, one notch above junk; the big four are all below 0.5×, Microsoft 0.2×, Meta 0.8×.

Left: order-book coverage, peak book divided by final size, latest Google in August 2026 at 4.6×, still above the 1.5× warning line; coverage is held down by deal size, and the 2026 US investment-grade average is about 4×, so the absolute book matters too. Right: 40-year new-issue spreads, like for like; the slope between one issuer's deals is how fast its cost is worsening, and the vertical gap between issuers is the credit tiering. Oracle at 195 basis points has worsened fastest, Google 130 and Meta 147 have added 35 basis points in six months, Amazon 125 is flattest.

The formula is capex estimates plus buybacks and dividends (held at the trailing four-quarter level) minus estimated operating cash flow, a lower bound on net new external funding: no refinancing of maturities, no pre-funding or cash buffer, and no draw on existing net cash, so actual issuance is usually larger. 2026 uses first-half actuals plus a second-half extrapolation, $20B for the four plus Oracle; 2027 is $106B in the base case (operating cash flow growth continues) and $223B in the conservative case (growth halves), or $44B after netting what was over-raised in 2026.

One row per supplier, every number from the related-party and concentration notes in 10-K and 10-Q filings. Microsoft is the most transparent: 65 to 70% of its AI revenue comes from one customer, OpenAI, with its own Copilot and the rest only about 30%. This section answers the core question about circular deals: if the model companies' money was invested by the cloud providers, and the cloud providers' revenue is paid by the model companies, how much of the growth on the books is the same dollar going around once.

Spread over SOFR on the secured tranche, one bar per deal, each verified. CoreWeave's GPU-backed spread tightened from 975 basis points in 2023 to 225, then went back to 450 to 550 in 2026: lenders' confidence in GPU residual value turned cautious in 2026, which says more than any commentary. Read three layers: the secured spread, the amortization structure, and the same issuer's unsecured tranche as a comparison.