AI data center power: hyperscaler capex analysis

Somewhere between the model demo and the earnings call, AI strategy became a utility interconnection problem wearing a fleece vest. CreditSights, part of Fitch Solutions, estimates the top five hyperscalers will spend about $602 billion in 2026, with about 75 percent aimed at AI related infrastructure. That works out to roughly $450 billion chasing not just chips, but buildings, power, cooling, and the deeply glamorous art of getting electrons to show up on time. The GPU is still the celebrity, obviously, but the grid is now the agent, venue, and bouncer. This is the part of AI scaling that does not fit neatly on a benchmark chart. Capability, latency, regional availability, reliability, and margins increasingly depend on whether infrastructure can be financed, permitted, connected, and energized fast enough. Your agentic workflow may have impeccable reasoning traces, but if the data center is waiting in a grid queue, it is basically a very expensive todo list.

What broke, according to CreditSights and Enverus

CreditSights frames 2026 hyperscaler capex as a spending wave centered on AI infrastructure, with the top five hyperscalers expected to commit about $602 billion and about three quarters allocated to AI related needs. Enverus Intelligence Research contributor Carson Kearl describes a similar investor concern around disclosed 2026 capital spending plans from GOOGL, AMZN, META, and MSFT, totaling roughly $695 billion to $725 billion, up from prior high end expectations of about $670 billion. Enverus also noted split market reactions after April 29 earnings, with GOOGL supported by Google Cloud strength while META saw investor pullback after raising 2026 capex guidance. That split matters because investors are no longer treating AI capex as a magical bonfire into which money disappears and shareholder value emerges wearing sunglasses. The question is whether capital turns into usable compute, then into revenue, then into product leverage. In other words, the new KPI is not just how many accelerators you bought, it is how quickly they become rentable, reliable, margin positive inference.

The queue moved from chips to electricity,

according to Enki AI Enki AI says the main constraint on the AI infrastructure buildout has shifted from semiconductor availability to the physical limits of the global power grid. Its report points to grid access delays, a 7 GW U.S. AI data center capacity crisis, and project cancellations as signs that power delivery has become a bottleneck rather than background plumbing. That is a cruel plot twist for anyone who thought the hard part ended when procurement secured the accelerators. For builders, this changes how AI products should be planned. A model roadmap that assumes infinite regional capacity is now about as realistic as a refrigerator that also does interpretive dance. Latency sensitive features, enterprise data residency requirements, and high volume inference all become geography problems when power availability dictates where capacity can actually land.

Time to power is the new deployment metric,

according to The Economy and MMCG Invest The Economy Research argues that time to power is a decisive economic variable for AI infrastructure, meaning how quickly capital, land, servers, and electricity become working compute. Using a 100 MW AI data center model, it says a one year delay can hurt lifecycle value more than large changes in power or tax variables. Translation for the ML crowd: a delayed interconnection can be worse than a worse electricity rate, which is the infrastructure version of losing the race before your model even imports torch. MMCG Invest’s hyperscale data center overview shows why the pressure is not theoretical. It estimates U.S. hyperscale data center revenue at approximately $111 billion in 2025, with growth forecast to about $165 billion by 2030. MMCG also says AI and ML workloads roughly doubled compute requirements for major hyperscalers in 2024 alone, adding pressure to already expanding cloud demand.

Policy is now part of the stack,

according to PwC and Brookings PwC, working with Oxford Economics, modeled data center capital expenditure across 46 countries and territories and five regions, projecting $31.6 trillion in global data center capex through 2050 in its central scenario, with a plausible upside near $50 trillion. PwC also argues this cycle resets every four to six years, which is a polite way of saying the depreciation treadmill has joined the gym and hired a trainer. If hardware refresh cycles keep pulling infrastructure forward, power planning becomes continuous strategy, not a one time build decision. Brookings has framed global energy demand as part of the AI regulatory landscape, which is exactly where it belongs. AI policy that only talks about model behavior while ignoring electricity demand is like regulating aviation by inspecting tray tables. For readers building AI products, the practical takeaway is simple: ask vendors where capacity lives, how resilient it is, and whether your workloads can move when regional power economics shift. The next round of AI competition will still feature models, papers, benchmarks, and demos with suspiciously perfect prompts. But the quiet winners may be the teams that understand power contracts, siting, latency, and utilization as deeply as context windows. The next model leaderboard may need a column for extension cords.