The number behind the spending sprint
Estimates for 2026 capital expenditure across Microsoft, Alphabet, Amazon, Meta, and Oracle cluster between 660 billion and 700 billion dollars, roughly 60 percent above what the same group spent in 2025. Broken out by company, projections put Amazon at about 200 billion dollars, Alphabet at 175 to 185 billion dollars, Meta at 115 to 135 billion dollars, Microsoft above 120 billion dollars, and Oracle at 50 billion dollars. About 75 percent of that total, roughly 450 billion dollars, goes specifically to AI-related spending such as servers, GPUs, and data center shells rather than general-purpose infrastructure.
That is a genuinely new order of magnitude for the industry, and it is being funded largely out of operating cash flow and debt rather than patient long-term capital. Nvidia's own results show where a meaningful share of that spending lands: fourth quarter fiscal 2026 revenue reached 68.1 billion dollars, and full fiscal year revenue hit 215.9 billion dollars, with 193.7 billion dollars of that attributed to data center products. Jensen Huang has said he expects roughly 500 billion dollars in combined Blackwell and Rubin platform revenue through the end of 2026 alone.
The cash flow squeeze this creates
Spending at this scale is starting to show up in the hardest number to disguise: free cash flow. Amazon is projected to turn free cash flow negative in 2026 for the first time in years, a direct consequence of capex outpacing operating cash generation. Pivotal Research projects an even sharper move at Alphabet, forecasting free cash flow falling almost 90 percent, to roughly 8.2 billion dollars from 73.3 billion dollars in 2025. Both companies remain highly profitable by any normal measure, and the math still shows real strain between what the AI buildout costs today and what it generates in revenue today.
Microsoft and Meta face a version of the same pressure, even where public free cash flow projections are less dramatic. Meta's capex guidance sits at 115 to 135 billion dollars for 2026, and the company has layered on multi-gigawatt power commitments, including a January 2026 set of agreements with three nuclear companies for more than 6 gigawatts and an earlier 20-year, 1.1 gigawatt deal with Constellation Energy in Illinois starting in 2027. Every one of those commitments is a long-dated cash obligation sitting on top of an already stretched capex budget.
Why Rubin and Blackwell scarcity makes this worse
Part of why this capex is so cash-intensive is that the underlying chips remain scarce even at these spending levels. Nvidia's Blackwell platform is reportedly sold out through mid-2026, and the Vera Rubin platform, launched at CES 2026 and reaffirmed at GTC in March 2026, is targeting volume production only in the second half of the year. Jensen Huang has pointed to roughly 1 trillion dollars in orders through 2027, which means hyperscalers are committing capital years ahead of delivery, locking in spend against capacity they cannot yet use.
That timing mismatch, paying now for chips that arrive later, is a core reason free cash flow is cracking even as capex climbs. Hyperscalers are effectively prepaying for a multi-year supply chain while simultaneously building the power, land, and shells those chips will eventually sit in, and both halves of that spend hit the balance sheet well before the corresponding AI revenue materializes. For a CTO watching a vendor's quarterly numbers, this is the detail that matters more than the headline capex figure: a provider can be spending aggressively and still be cash-constrained in the exact window when your workload is scheduled to land.
Power is compounding the cash problem
The International Energy Agency projects data center electricity use reaching roughly 1,100 terawatt-hours in 2026, comparable to Japan's entire annual consumption, and the Uptime Institute projects global AI-specific data center load reaching about 10 gigawatts by the end of the year. Because grid interconnection remains slow almost everywhere, hyperscalers are increasingly building their own on-site generation, including natural gas turbines, microgrids, and batteries, a pattern operators have started calling energy islands on their own earnings calls.
Offtake agreements between data center operators and small modular reactor projects have grown from roughly 25 gigawatts at the end of 2024 to about 45 gigawatts now, nearly doubling in about eighteen months. Every one of those agreements represents additional long-term capital commitment layered on top of the chip and shell spending already straining free cash flow, and none of it produces revenue until the underlying power, and the AI capacity it supports, actually comes online.
What this means for your contract negotiations
None of this means the hyperscalers are in financial trouble. Companies generating tens of billions in quarterly revenue can absorb negative or sharply reduced free cash flow for a period without existential risk. What it does mean is that the vendors you depend on for cloud and GPU capacity are under real internal pressure to protect margin and manage cash, and that pressure eventually shows up in your contract terms, whether through price increases, slower capacity delivery, or less generous negotiating room on renewals.
The practical response is to negotiate as if that pressure is already affecting you, because it likely will soon. Push for multi-year price locks before the next renewal cycle rather than after, insist on specific delivery SLAs with financial remedies rather than best-efforts language, and build contingency into your capacity planning for the possibility that a strained vendor delays a committed deployment. A vendor's quarterly earnings call is no longer background reading for your infrastructure team, it is a leading indicator for your own budget.
