Cloud Giants Shift Gears on Capex: Revenue Accelerating, Debt Financing Share Rising
Alina Collins
Hyperscaler capex growth is shifting from operating-cash-flow coverage to debt and lease financing. Goldman Sachs estimates direct debt will rise from 33% to 35% of capex by 2027, while $1.2 trillion in lease obligations — over half not yet started — signal that balance-sheet strain is accelerating alongside revenue.
Cloud revenue is still accelerating — so where is the money coming from?
Google's quarterly capex doubled year-on-year to $44.9 billion; full-year guidance was raised by $15 billion to $195–205 billion. Cloud backlog hit $514 billion, up 385% year-on-year, expected to convert to revenue over roughly two years.
Amazon AWS grew 36.7% year-on-year in Q2 — the fastest pace in 18 quarters. Full-year cash capex was raised to about $220 billion; management said even that will not meet all demand in 2026–2027.
This means → demand signals remain strong, but investment has outgrown what operating cash flow alone can cover. The real new variable this cycle is the shift in funding structure.
What do Microsoft's and Meta's books reveal?
Microsoft's full-year cloud revenue topped $214 billion; nearly 90% came from customers outside frontier-model companies. Cloud gross margin fell to 65%, dragged by rising Azure share and AI infrastructure spending.
Meta posted Q2 revenue of $60.8 billion, up 28% year-on-year; Advantage+ annualised above $75 billion. Yet total expenses rose 55% to $42 billion, leaving free cash flow at just $784 million.
In plain terms = "revenue delivery" and "cash-flow pressure" can happen at the same time. Meta is the clearest case: it is earning more but spending faster, with almost no cash cushion left.
Why have custom chips gone from cost-saving tools to customer-locking weapons?
Google uses its TPU — a custom AI chip — to underpin the $514 billion cloud backlog and lock in long-term supply. Amazon deploys Trainium — its custom training chip — to win multi-year, multi-gigawatt commitments from Anthropic and OpenAI.
This means → custom silicon is no longer just an internal cost play; it is the product hyperscalers use to bind marquee customers to long-term compute contracts. Broadcom remains a key TPU supplier; UBS sees the new v8t's chiplet design — assembling chips from smaller modular tiles — as reducing scale-up risk.
Memory and advanced packaging are also being locked in early: Nvidia and SK Group announced a partnership exceeding $500 billion; Broadcom and Samsung signed an MOU exceeding $200 billion — but both figures span multiple years and should not be read as confirmed orders.
What risk sits inside $1.2 trillion in lease obligations?
The five hyperscalers disclosed $1.2 trillion in lease obligations; $725 billion has not yet commenced. These obligations are not fully captured in conventional leverage ratios today, but will convert to fixed payments once facilities are delivered.
This means → if rating agencies pull these commitments into adjusted debt metrics early, credit indicators could tighten before capex actually converts to cloud revenue.
Goldman Sachs estimates the four lower-leveraged hyperscalers carry total leverage of just 0.6×–1.3×; under a conservative scenario, incremental debt capacity reaches $1.3–1.4 trillion. The headroom is ample — but financing costs are not guaranteed to stay stable.
Can this AI-cloud investment cycle continue?
Goldman Sachs' summary of management commentary shows the investment-to-monetisation lag may stretch from three months to two years. Direct debt financing as a share of capex is projected to rise from 33% in 2026 to 35% in 2027.
In plain terms = the cycle's durability hinges on two lines staying apart: cloud revenue keeps accelerating while financing costs hold steady, and the cycle extends; but if capex revisions keep climbing, free cash flow turns negative, and credit spreads widen in tandem, valuations will de-rate before demand does.
This reflects a pivotal shift in the market's core AI-cloud debate — from "is the demand real?" to "which arrives first: the spending peak or the revenue catch-up?"
Content is for reference only, not financial advice.