Hyperscalers commit $300–500B capex
- Social posts and industry commentary estimate hyperscalers will spend roughly $300–500 billion on AI infrastructure through 2026–27, massively expanding data‑center footprints. (x.com) - That buildout is already driving 30–50% cost increases for data‑center projects and creating localized power scarcity and grid bottlenecks in several regions. (x.com) (x.com) - Observers flag infrastructure as now about 71% of AI deployment costs and say energy capacity is the key operational constraint. (x.com)
Data centers are becoming the real story in AI. Not the chatbots. Not the demos. The buildings, chips, substations, and power contracts underneath them. Over the last few weeks, the clearest signal yet came from the biggest buyers themselves: Meta lifted 2026 capital spending to $125 billion to $145 billion, Alphabet’s Q1 2026 capex hit $35.7 billion with most of it going into technical infrastructure, and Microsoft kept talking about margin pressure from AI infrastructure buildout. (fool.com) So the headline estimate — hyperscalers committing roughly $300 billion to $500 billion of capex across the next year or two — is not some wild social-media extrapolation. Basically, it is what you get when you annualize and stack the public spending plans of Meta, Alphabet, Microsoft, and Amazon, then assume those plans stay elevated as AI demand keeps rising. Meta alone is guiding to as much as $145 billion in 2026. Alphabet has already pointed investors to a 2026 capex range of $180 billion to $190 billion. Amazon has not put out a neat full-year AI capex number in the same way, but AWS demand and broader infrastructure spending remain central to the story. (fool.com) Why are they spending like this? Because AI has shifted from a software race to a capacity race. Training frontier models still matters, but inference — actually serving models to users and apps all day — is turning into the bigger long-run load. That means more GPUs, denser racks, faster networking, more cooling, and much larger campuses. Deloitte’s 2025 infrastructure work lays it out pretty bluntly: the biggest hyperscale sites under construction or in planning are moving toward 2 gigawatts, and some early-stage campuses point even higher. (deloitte.com) The catch is that money is not the only bottleneck. Power is. In Virginia, the most important U.S. data-center market, commercial electricity sales jumped by nearly 30 million megawatt-hours from 2019 to 2025, with data centers a major driver. PJM now expects the Dominion zone in Virginia to see the biggest absolute increase in summer peak demand from 2026 through 2030, largely because of data-center load. That is the shape of the problem everywhere else too — not “can they afford servers,” but “can they get enough electricity to the site on time?” (eia.gov) That is also why project costs are rising. When everyone wants the same transformers, turbines, switchgear, land parcels, and utility interconnections at once, the whole stack gets more expensive and slower. Deloitte’s survey found grid stress was the top infrastructure challenge, and it flagged interconnection waits that can stretch to seven years in some cases. So even if chip supply improves, the physical buildout can still jam. (deloitte.com) The broader backdrop is even bigger. EPRI now projects U.S. data centers could consume 9% to 17% of national electricity by 2030. McKinsey’s global estimate is starker still — $6.7 trillion in data-center investment by 2030 to keep up with compute demand. (powering-intelligence.epri.com) Bottom line — hyperscalers are no longer just buying AI. They are rebuilding chunks of the industrial economy around it. The winners will not only be the best model companies. They will be the firms that can secure chips, land, cooling, and especially power before everyone else does.
