Hyperscalers will dominate capacity
Analysts expect hyperscalers to control about two‑thirds of global data‑center capacity by 2031, concentrating where large compute and AI projects will run. The report also flags rising power‑market stress and explicit energy commitments by big AI tenants as factors shaping where and how capacity is built. That concentration makes selective cloud use an economic and control decision, not just a latency tradeoff. (ciodive.com)
The internet’s biggest cloud companies are on track to control 67% of all data-center capacity by 2031, up from 48% at the end of 2025. The same forecast says enterprise-owned server rooms fall to 19%, down from 56% in 2018. (srgresearch.com) That shift means more of the world’s computing will happen inside buildings run by Amazon Web Services, Microsoft, Google, and a small set of peers, not inside a company’s own office park. Synergy says hyperscale operators already ran 1,360 large data centers by the end of the fourth quarter of 2025. (srgresearch.com) The reason is not ordinary web hosting. Since late 2022, artificial intelligence spending has pushed quarterly hyperscaler capital spending up almost 180%, reaching $142 billion in the third quarter of 2025. (srgresearch.com) These facilities are also getting bigger, so the count alone understates the change. Synergy says total operational hyperscale capacity has increased more than fourfold since early 2018 as the average size of each new site keeps rising. (srgresearch.com) The buildout is splitting into two models at once. About 60% of hyperscale capacity now sits in data centers the cloud companies built and own themselves, while the rest sits in leased facilities run by landlords. (srgresearch.com) That matters because the cloud is no longer just rented software in someone else’s warehouse. For the biggest artificial intelligence jobs, the real choice is whether to buy access to a landlord’s shelves or to chase scarce land, transformers, and power connections yourself. (jll.com) Power is the bottleneck underneath all of this. JLL says nearly 100 gigawatts of new data-center capacity will be added between 2026 and 2030, and that pace will require energy innovations to ease grid constraints. (jll.com) The International Energy Agency says electricity demand is rising partly because of artificial intelligence and data centers, and a new estimate reported on April 10 says global data-center power demand could reach 945 terawatt-hours by 2030. That is roughly the scale of Japan’s current total electricity use. (iea.org) (spglobal.com) Big tenants are responding by locking in energy directly, not just buying server time. Amazon Web Services signed a 17-year agreement for 1,920 megawatts from the Susquehanna nuclear plant in Pennsylvania, with full delivery expected by 2032. (datacenterdynamics.com) Google is doing the same in a different way. In February 2026, Clearway said Google signed three 20-year power purchase agreements totaling 1.17 gigawatts to support data-center operations, adding to Google’s earlier 1.5-gigawatt clean-energy buying push tied to its round-the-clock carbon-free power goal. (esgdive.com) (cloud.google.com) So the old argument for keeping workloads on-premises because the cloud was “someone else’s computer” is getting weaker. If two-thirds of capacity ends up concentrated with hyperscalers, then using cloud selectively becomes a decision about price leverage, power access, and operational control, not just speed to a nearby user. (ciodive.com)
