Four companies are on pace to spend more on AI infrastructure in 2026 than most countries spend running their governments.
Microsoft, Amazon, Alphabet, and Meta disclosed combined capital expenditure guidance for 2026 that analysts have pegged near 650 billion dollars, according to a Morgan Stanley research note circulated after the companies' Q4 2025 earnings calls. That number gets repeated in every headline. What almost nobody breaks down is where it goes once it leaves the balance sheet.
For a buyer trying to plan a 2026 GPU contract or a storage migration, the total is close to useless on its own. What matters is the split: how much becomes GPUs you can actually rent, how much goes into buildings and power infrastructure that takes years to pay back, and how much is quietly exposed to a depreciation problem that could reshape pricing before the year is out.
Microsoft guided to capital expenditure growth it described as continuing at a similar pace to fiscal 2025 during its October 2025 earnings call, with quarterly spend already running above 34 billion dollars. Meta raised its full year 2025 capital expenditure outlook to a range of 70 to 72 billion dollars and told investors that 2026 dollar growth would be notably larger, without giving a hard ceiling. Alphabet lifted 2025 guidance to roughly 91 to 93 billion dollars and signaled a significant increase again for 2026. Amazon has not published a single 2026 figure but ran at an annualized pace above 125 billion dollars through the back half of 2025 on AWS infrastructure alone.
Add those trajectories together and independent estimates land in the 550 to 650 billion dollar range for calendar 2026 across the four hyperscalers, before counting Oracle, CoreWeave, and the sovereign compute programs covered elsewhere on this site. The headline number is directionally real. The problem is that a capex dollar buys wildly different things depending on where in the stack it lands, and none of the four companies break that out cleanly in their filings.
| Company | 2025 Capex | 2026 Signal |
|---|---|---|
| Microsoft | ~$88B (FY2025) | Growth continuing at similar pace |
| Amazon (AWS infra) | ~$125B annualized (H2 2025) | No hard figure, elevated pace expected |
| Alphabet | $91 to $93B | Significant increase guided |
| Meta | $70 to $72B | Notably larger dollar growth guided |
Industry analysts who model hyperscaler capex, including teams at Morgan Stanley and TD Cowen, generally split AI related capital spend into four buckets: silicon and servers, data center shells, power and cooling infrastructure, and networking plus storage. The silicon bucket, meaning GPUs, accelerators, and the servers that hold them, typically accounts for roughly half of total AI capex. The remaining half is split unevenly across land, buildings, power generation and transmission commitments, and the networking and storage fabric that makes the GPUs usable at scale.
That second half is the part buyers underestimate. A rack of GPUs sitting in a data center with insufficient power allocation or inadequate east to west bandwidth to storage does not compute anything faster than a rack that was never purchased. Sunny Smith, founder and CTO of Massed Compute, made this point directly on the DataStorage.com Podcast: storage has to sit adjacent to the GPU with real bandwidth, because moving a petabyte between data centers at the point of use is prohibitively expensive. Capex spent on GPUs without matching capex spent on storage and power is capacity that cannot be turned on.
This is also why the power and cooling line has grown faster than the GPU line in percentage terms over the past two years. Meta, Microsoft, and Amazon have all signed multi gigawatt power purchase agreements since 2024, some running into the 2030s, because the constraint on deploying new GPU capacity is no longer chip supply from NVIDIA. It is finding a substation with spare capacity.
Every hyperscaler assumes a useful life for its GPU fleet when it calculates depreciation, and that assumption drives the effective price they can charge renters. Microsoft and Google have both used a six year useful life assumption for server hardware including GPUs in recent filings. Independent hardware analysts, including coverage from Tom's Hardware and The Register, have questioned whether a Hopper generation GPU purchased in 2024 will still be commercially rentable at a competitive price point six years later given the pace of Blackwell and Vera Rubin generation transitions.
If the real useful life is closer to four years than six, the depreciation expense understates the true cost of the fleet today and overstates it later, which means today's GPU rental pricing is arguably subsidized by an optimistic accounting assumption. That gap does not stay hidden forever. It shows up as either a future write down, a future price increase, or both. Buyers signing multi year GPU reservations in 2026 are effectively betting on which way that correction goes.
None of the four hyperscalers report storage capex as a standalone figure, but it is not a footnote. Training checkpoints for frontier scale models now run into hundreds of terabytes per run, and inference fleets serving agentic workloads write and read context at a volume that dwarfs the traffic patterns storage systems were built for five years ago. NVMe pricing has roughly tripled since early 2025 according to component pricing trackers cited by Blocks and Files, and scarcity is expected to persist into 2027, which means the storage portion of every GPU cluster buildout costs more per unit today than it did eighteen months ago even before accounting for the added capacity.
This is the same dynamic that makes egress free storage relevant to a conversation that is nominally about GPUs. Enterprises moving workloads between hyperscalers, or between a hyperscaler and a neocloud, to chase available GPU capacity carry their data with them. Egress fees on a multi terabyte move can erase the savings that motivated the switch in the first place, which is why zero egress providers such as Backblaze and Wasabi keep coming up in procurement conversations that start out purely about compute.
The bull case for this level of spending assumes enterprise AI deployment accelerates sharply from where it sits today. Most enterprises are still in pilot phases. Talent scarcity in building production grade AI infrastructure, and the production readiness cycle itself, have kept actual enterprise GPU consumption well below the consumer chatbot boom of 2023 and 2024. If that gap closes, demand could run several times higher than current utilization and the capex looks conservative in hindsight.
If it does not close on the timeline hyperscalers are betting on, the industry has a supply problem instead of a shortage. CoreWeave's stock volatility following news of Meta's own capacity build out, and Nebius facing the same pressure, are early signals that investors are already pricing in the possibility that neocloud capacity outpaces near term demand. Buyers in that scenario gain negotiating leverage on contract length and pricing, which is a meaningfully different posture than the shortage driven market of 2023 through 2025.
Three practical takeaways follow from the breakdown above. First, treat headline capex figures as a demand signal, not a supply guarantee. Money committed to power and land does not become usable GPU capacity for eighteen to thirty six months. Second, ask any provider, hyperscaler or neocloud, what depreciation schedule they are using on the hardware you would be renting. A provider running an aggressive useful life assumption is more exposed to a future repricing event, and that risk gets passed to long term contract holders. Third, model storage and egress costs into any capacity decision from day one rather than treating them as an afterthought once the GPU contract is signed. The compute number gets the headline. The storage and power numbers determine whether the compute is actually usable at the price you were quoted.
The 650 billion dollar figure is not fiction, but it is not a supply guarantee either. Half of it buys GPUs. The other half decides whether those GPUs ever turn on.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
