Amazon Web Services plans to spend $200 billion on capital expenditure in 2026, the majority of it on data centers, according to CEO Andy Jassy on the February 2026 earnings call. That figure lands on top of $96.5 billion spent in 2025, an 81% jump that pushed AWS capital intensity to 75% of its own revenue. The company now runs more than 900 facilities across 50 countries, a number revealed in November 2025 by a Bloomberg and SourceMaterial investigation of leaked internal documents. Previous industry estimates had put the count between 350 and 450.
The scale has a simple explanation. Training and deploying frontier AI models needs power and silicon at a level that general-purpose cloud never required. AWS's Project Rainier campus in Indiana, announced at $11 billion in April 2024, occupies 1,200 acres near New Carlisle and is designed for 2.2 gigawatts across 30 buildings, each over 200,000 square feet. Seven buildings were operational within seven months of construction starting in September 2024, and a second $15 billion expansion across northern Indiana brings the state's total AWS commitment to roughly $26 billion and 4.6 gigawatts. That single state's footprint would exceed the entire current U.S. capacity of Google Cloud.
Why this matters now
The nut of it: electricity, not capital, is now the binding constraint on AWS's expansion. PJM Interconnection projects 32 gigawatts of peak load growth between 2024 and 2030, with data centers responsible for 94% of it. Its December 2025 capacity auction fell 6,625 megawatts short, and capacity prices rose from $28.92 to $333.44 per megawatt-day. Interconnection queues now stretch eight years. Free cash flow at Amazon compressed to $7.7 billion in 2025 from $32.9 billion in 2024, and whether demand will justify $200 billion of 2026 spend against roughly $142 billion in annualized AWS revenue remains unanswered.
This is a story about edges as much as centers. AWS counts 220-plus rented edge locations among its 900-plus facilities, plus 700-plus CloudFront points of presence and 43 local zones. The edge layer is what puts compute near users and devices, and it is the fastest-growing part of the network by site count even if the megawatts concentrate in Virginia and Indiana. The company maintains 39 geographic regions and 120-plus availability zones, each region holding at least three physically separate data centers.
The power problem nobody can engineer around
Northern Virginia's US-East-1 region alone holds 163 facilities drawing about 2,754 megawatts, with $91.5 billion invested since 2006. Dominion Energy provides grid power, and AWS paid $542.9 million in Virginia property taxes in 2024. Contracted data center power in the state reached 40 gigawatts, up 88% in six months as of December 2024. Lead times for large power transformers now average 128 weeks, and costs are up 77% since 2019, so securing grid interconnection has become as strategically important as securing land.
The buildout has a human cost that the capital figures skip. A Rolling Stone investigation in November 2025 linked AWS groundwater withdrawals in eastern Oregon to worsening nitrate contamination and a cluster of rare cancers in nearby communities. In the high desert around Boardman and Umatilla, the US-West-2 region runs 46 data centers with an estimated 1,000 to 1,500 megawatts of capacity, backed by $39.2 billion invested since 2011. Data Center Watch counted $18 billion of projects blocked and $46 billion delayed by community resistance in 2025.
What competitors are doing
Microsoft Azure runs roughly 400 data centers across 70-plus regions but holds only 1.2 gigawatts of active U.S. IT capacity, about half of AWS. Google Cloud runs 508 megawatts across 42 regions, roughly a quarter of AWS's American footprint, though its power usage effectiveness of 1.09 leads the industry. AWS holds 29 to 31% of global cloud infrastructure spending against Azure's 20% and Google's 13%, per Synergy Research Group data from late 2025.
The four largest hyperscalers spent $416 billion on infrastructure in 2025 and are projected to approach $645 billion in 2026. Amazon's $200 billion guidance for 2026 is the largest single line. Hyperscale facilities cost roughly $10 to 12 million per megawatt in 2025, while AI-optimized facilities run $20 million or more. The math implies that every new gigawatt of AI capacity costs more than an entire mid-sized utility's annual budget.
Edge sites and the latency race
The 220-plus rented edge locations and 43 local zones exist for one reason: latency. Financial trading, industrial IoT, and real-time AI inference all lose value when a request round-trips to Ashburn, Virginia. AWS's 700-plus CloudFront points of presence handle content delivery, but the local zones put actual compute close to population centers. Nearly 20 million kilometers of fiber optic cabling stitch the network together.
This is where the competitive fight will land. Azure and Google are adding regions, but AWS's head start in edge presence is measured in hundreds of sites. Whether that lead holds depends on whether the power grid cooperates. Eight-year interconnection queues mean a site planned today may not energize until 2033, and by then the AI workload mix may have shifted again.
The bet and its downside
It is not clear this will hold. AWS is spending like a utility while pricing like a software company, and the gap between $200 billion of 2026 capital expenditure and $142 billion of annualized revenue has to close through either higher prices or higher utilization. Both are plausible. Neither is guaranteed. The infrastructure being poured into Indiana cornfields and Virginia suburbs will define the physical substrate of artificial intelligence for decades, and the communities absorbing it have not all agreed to the terms.
For more on the AWS buildout, see the MMCG Invest analysis. For related coverage on this site, see our cloud and edge computing archive.
AWS's next earnings call will be the first real read on whether 2026 demand matches the $200 billion plan. Watch capacity prices in PJM and the interconnection queue length in Virginia; those two numbers tell you more about the cloud's future than any product announcement ever could today, and they are worth tracking weekly.