The AI Compute Buildout of 2026: Who's Spending What, and How It's Financed

The 2026 AI compute buildout is the largest private infrastructure program in history, and its financing structure — not its gigawatts — is what operators should be watching. Here's the map, from disclosed numbers. The Spend, By the Numbers - Meta guided 2026 capital expenditure to roughly $125–145 billion, and disclosed $237.67 billion of non-cancelable contractual commitments as of March 2026, mostly tied to third-party cloud capacity. It's building a 1GW AI-optimized data center in Alberta and signed for up to 6 gigawatts of AMD Instinct GPUs in a deliberately diversified silicon strategy. - Microsoft reported quarterly capex of $37.5 billion (Q2 FY2026) — about two-thirds on short-lived assets, primarily GPUs and CPUs — added another gigawatt of capacity in a single quarter, and says it will double its footprint in two years. - Alphabet put Q1 2026 capex at $35.7 billion, overwhelmingly on technical infrastructure. - Oracle disclosed FY2026 capital funding that included $43 billion of debt and $5 billion of equity — plus a remarkable structure: $75 billion of customer-prepaid or customer-supplied GPU hardware inside large AI contracts. Its remaining performance obligations jumped $85 billion in one quarter to $638 billion. - CoreWeave passed 1 GW of active power with over 3.5 GW contracted, and reported a $99.4 billion revenue backlog on $2.1 billion of quarterly revenue. The Financing Layer Is the Story Three structures dominate, each with different risk: 1. Cash-flow-funded (Meta, Microsoft, Alphabet). Hyperscalers fund from operations, but the strain shows: Microsoft's quarterly free cash flow compressed to $5.9 billion against $37.5 billion capex; Amazon's trailing-twelve-month free cash flow fell to $1.2 billion on a $59 billion year-over-year increase in property and equipment purchases. 2. Debt-financed neoclouds (CoreWeave, Oracle). CoreWeave carries roughly $25 billion of debt (including an $8.5 billion non-recourse facility at SOFR+2.25%) against contracted backlog. Even mid-size players borrow: Mistral raised $830 million in debt from a seven-bank consortium for its Paris-area data center. 3. Customer-prepaid capacity (Oracle's $75B, the OpenAI multi-cloud deals). Buyers commit years ahead — OpenAI's $100 billion, 8-year AWS expansion for ~2 GW of Trainium; Anthropic's Amazon agreement securing up to 5 GW — turning compute procurement into something resembling power-purchase agreements. Concentration Cuts Both Ways The buildout's dependency graph is tight. Microsoft disclosed that about 45% of its commercial remaining performance obligation comes from OpenAI — one customer. CoreWeave's backlog leans on a handful of anchors (Meta committed up to ~$21 billion through 2032; Jane Street signed $6 billion). If frontier-lab revenue growth stalls, the exposure propagates up the chain to lenders within quarters, not years. What Operators Should Watch 1. Backlog-to-revenue conversion rates at Oracle and CoreWeave — the honest indicator of whether contracted gigawatts become paid gigawatts. 2. Custom silicon share — Amazon's chip business passed a $20 billion run rate; Broadcom's AI semiconductor revenue hit $10.8 billion in a quarter on Google TPU and Meta custom-accelerator work. Every custom-silicon gigawatt is margin leaving NVIDIA's column. 3. Power, not chips, as the binding constraint — deals are now denominated in gigawatts (SK's 2 GW Vera Rubin factory, Meta's 6 GW AMD agreement). Siting and grid interconnect timelines gate everything.