Bernstein: Cloud Giants' AI Capex Approaching Cash Flow Ceiling, Financing Pressure Has Arrived
0xBroomberg
Bernstein's latest tracker shows the four major hyperscalers will spend $623 billion in 2026 capex against roughly $635 billion in operating cash flow. This means → self-funding can no longer sustain the AI buildout, and the financing window is now open.
Where is the money going, and why is it running short?
In 2026, Microsoft, Amazon, Google, and Oracle are projected to spend a combined $623 billion in capex. Operating cash flow for the same period: roughly $635 billion — nearly a wash.
This means → most hyperscalers can no longer fund expansion from profits alone; they must borrow.
Add Meta, and the five-company total approaches $760 billion. Microsoft alone raised its 2026 capex guidance to $190 billion, of which roughly $25 billion reflects component price increases.
Who has already started borrowing?
Amazon, Google, and Oracle have all launched financing rounds. Oracle moved the most aggressively: roughly $48 billion raised via debt and equity in FY2026, with plans for another ~$40 billion in FY2027.
Bernstein's view: if 2027 capex stays strong while cash-flow growth lags, every hyperscaler except Microsoft will need the capital markets.
In plain terms = the bottleneck has shifted. It used to be "can't buy enough GPUs." Now it's "can't get enough power to the data center." Whoever connects compute to the grid fastest and cheapest wins.
How wide is the performance gap across cloud businesses?
Google Cloud was the standout: Q1 revenue grew 63% year-on-year, topping $20 billion for the first time. Sequential revenue additions of ~$2.4 billion exceeded AWS's ~$2.0 billion. Backlog nearly doubled quarter-on-quarter to roughly $462 billion, with over half expected to convert within 24 months.
AWS grew 28%, reaching an annualized run rate of ~$150 billion — its fastest growth in 15 quarters. Backlog hit $364 billion, not yet including the $100 billion+ Anthropic contract. Azure grew 39% at constant currency; AI annualized recurring revenue crossed $37 billion.
Oracle OCI surged 93%; total contract backlog reached $638 billion, up 363% year-on-year. Alibaba Cloud revenue rose 38.2%; AI-related product revenue posted triple-digit growth for the 11th straight quarter, now 30% of external cloud revenue.
The four major hyperscalers' combined backlog has hit $2 trillion, growing 176% year-on-year. Bernstein notes, however, that operating cash-flow growth remains volatile with no clear acceleration yet.
Is AI actually dragging margins?
Bernstein argues AI will pressure IaaS/PaaS gross margins at least through the infrastructure-build phase. GPU server useful life and compute supply-demand balance are the two key unknowns.
AWS operating margin: ~38%, up 270 bps quarter-on-quarter, but management flagged Q2 headwinds from equity-compensation vesting and rising depreciation. Google Cloud operating margin rose to ~33%, a sharp improvement year-on-year, though the Wiz acquisition will create a low-single-digit margin drag in 2026.
Oracle's new CFO named return on invested capital (ROIC — how much profit each dollar of investment generates) as the core metric; OCI AI data-center ROIC is projected above 20%. Alibaba Cloud adjusted EBITA margin stood at 9.1%; management guides to low double digits over the next few quarters, with a long-term 20% target.
CoreWeave — high-growth star or hidden risk?
CoreWeave (CRWV) posted Q1 revenue of $2.1 billion, up 112% year-on-year. Backlog expanded to $99.4 billion (~4× year-on-year), with over $40 billion in new bookings in a single quarter.
Yet adjusted operating margin was only ~0.5%. This reflects a timing mismatch: lease, power, and depreciation costs hit the moment a facility is delivered, while revenue recognition waits roughly 6–8 weeks until deployment is complete.
On financing: total debt exceeds $25 billion, collateralized mostly by GPU hardware and customer contracts. This quarter saw $8.5 billion in investment-grade issuance at an effective rate of ~7%, backed by Meta's take-or-pay contract — a long-term commitment to pay whether or not capacity is used.
Bernstein views this deal as a proof of concept, not a repeatable benchmark — very few hyperscale customers carry an Aa2+ credit rating. The company targets a steady-state margin of 25%–30%, but the path depends heavily on execution. Bernstein rates CoreWeave underperform, target price $67.
How does Bernstein rate these companies?
Outperform: Microsoft (target $646), Amazon ($315), Oracle ($325), Alibaba ($180 / HK$176).
Market perform: Google (target $390). Underperform: CoreWeave ($67).
This means → Bernstein believes AI-driven cloud demand is real, and surging backlogs plus accelerating AI revenue penetration offer medium-term support. But investors need to see revenue growth pick up further over the next few quarters before they can confirm this unprecedented infrastructure bet will actually pay off.
Content is for reference only, not financial advice.