The AI buildout has stopped being a question of whether the chips exist and become a question of who is financing them, on what terms, and against what return. In the first half of 2026 the largest builders raised their spending guidance even as the same spending drove free cash flow to zero or below. The frontier the numbers now describe is not capacity. It is the balance sheet.
Capex crossed into negative free cash flow
The scale is no longer abstract. Alphabet raised its 2026 capital-expenditure guidance to $195-205 billion, with Sundar Pichai tying the increase to accelerated capacity delivery and the start of TPU shipments to customer datacentres in Q2 (per abc.xyz). The bill is visible in cash flow: Alphabet reported Q2 2026 capex of $44.9 billion and negative $5.9 billion in free cash flow, with the vast majority of spend going to AI-supporting infrastructure (per abc.xyz).
It is not alone. Amazon's trailing-twelve-month free cash flow fell to just $1.2 billion, driven by a $59.3 billion year-over-year jump in property and equipment purchases it attributed mainly to AI (per Amazon IR). Meta guided 2026 capex to $115-135 billion (per Meta IR) and disclosed $237.67 billion of non-cancelable contractual commitments, most tied to cloud capacity and technical infrastructure (per an SEC filing). Microsoft's fiscal Q2 capex was $37.5 billion, roughly two-thirds of it short-lived GPU and CPU assets — the kind that depreciate fast — against $5.9 billion of free cash flow (per Microsoft). At the foundry, TSMC guided 2026 capital spending to the high end of its $52-56 billion range to meet AI and HPC demand (per TSMC IR).
The financing is getting structured
When operating cash flow no longer covers the build, the money has to come from somewhere, and the sourcing is turning creative. Oracle's FY2026 capital funding included $43 billion of debt and $5 billion of equity, while customer-prepaid or customer-supplied GPU hardware inside large AI contracts totalled $75 billion — a sign that buyers, not just Oracle, are now fronting the silicon (per Oracle IR). The neocloud model is more levered still: CoreWeave's Q1 disclosure showed $7.5 billion of current debt and $17.3 billion of non-current debt against a $99.4 billion revenue backlog (per an SEC filing) — a backlog anchored by commitments such as Meta's pledge to pay CoreWeave up to about $21 billion for capacity through 2032.
These arrangements increasingly loop the same names as supplier, customer, and financier. OpenAI expanded its agreement with AWS by $100 billion over eight years, committing to consume roughly 2 gigawatts of Trainium capacity (per Amazon). The counterparties are underwriting each other's demand, which flatters backlog and remaining-performance-obligation figures — Oracle's RPO rose $85 billion in a single quarter to $638 billion (per Oracle IR) — while concentrating risk if any one node's revenue disappoints.
Who is actually earning on it
For now, the returns are concentrated at the silicon layer. NVIDIA reported fiscal Q1 2027 data-center revenue of $75.2 billion at a 74.9% gross margin (per NVIDIA IR) — the buildout's capex is somebody else's income statement. On the model side, revenue is climbing fast but from a base that still trails the spend: Anthropic said run-rate revenue surpassed $30 billion, up from about $9 billion at end-2025, then crossed $47 billion by May, raising $65 billion in Series H at a $965 billion valuation (per Anthropic). Valuations and run-rates are compounding on the promise that enterprise demand converts.
The ROI evidence is thinner than the spend
That conversion is where the case gets soft. The hardest demand-side numbers are vendor-published and anecdotal: a Microsoft-cited study of SimCorp One customers claimed 134% ROI over three years and 10 hours saved per person per week (per blogs.microsoft.com). Encouraging, but a single sponsored study is not a sector return curve. More telling is the posture underneath the spend: BCG found that 94% of organisations say they will keep investing in AI even if it does not deliver returns in 2026 (per BCG). That is conviction — and conviction, not measured payback, is currently underwriting a good share of the buildout.
The tension resolves into one line. Supply-side commitment is now measured in hundreds of billions and locked in through 2029-2032; demand-side proof is measured in quarterly run-rates and pilot studies. The builders are betting the second catches the first before the debt, depreciation, and prepayment terms come due.
What to watch: the first quarter in which a hyperscaler trims capex guidance rather than raising it — or a marquee prepay or backlog commitment gets restructured. Either would signal the financing story, not the compute story, has started to bind.
