Microsoft disclosed $14.9 billion in Q1 AI infrastructure commitments across three continents. Google parent Alphabet allocated $13.2 billion to the same purpose in the same quarter. Oracle signed $22 billion in multi-year GPU cluster and cooling contracts with three suppliers, none of them Nvidia. The three firms alone spent more on AI infrastructure in ninety days than the global semiconductor industry spent on M&A in all of 2025.
The capex is structural, not speculative. Microsoft's commitments include 740 megawatts of contracted power capacity in Texas and 1.2 gigawatts under negotiation in Sweden, both tied to multi-year take-or-pay agreements with utilities. Google is pre-leasing 18 million square feet of data center space across six U.S. markets, with occupancy beginning Q3 2026. Oracle's GPU orders are split between two Asian foundries and one U.S. assembler, with delivery schedules extending into 2028. The deals are denominated in fixed prices, not floating estimates, and most include liquidated damages clauses if suppliers miss delivery windows.
The spending reflects a shift in how AI product economics work. Training a frontier model cost $100 million in 2023. Inference—running that model for millions of users—costs $2 billion annually at scale, and the cost curve is moving in the wrong direction. Token processing has not gotten cheaper per unit. It has gotten more expensive as models grow more capable and users demand lower latency. Hyperscalers are internalizing that reality and locking in supply before prices rise further or capacity disappears. The Oracle contracts, in particular, sidestep Nvidia's H100 and H200 chips entirely, opting instead for custom silicon from two suppliers under NDA. That is not a hedge. That is a belief that Nvidia's roadmap is too slow or too expensive for Oracle's margin requirements.
Allocators should watch three things. First, whether Amazon Web Services discloses comparable capex in its April earnings call. AWS has been silent on Q1 infrastructure commitments, which is unusual given its historical pattern of matching or exceeding Microsoft's spend within sixty days. Second, whether any of the three firms begin selling equity or issuing debt specifically to fund AI infrastructure. SpaceX just borrowed $20 billion against projected AI revenue. If a hyperscaler does the same, it signals that internal cash flow is no longer sufficient to fund the build-out at current velocity. Third, whether power utilities in Texas, Virginia, or Northern Europe begin filing rate increase requests with regulators. The contracted megawatts are real. Someone is paying for the grid upgrades, and if it is ratepayers rather than tech firms, the political cost will follow.
The Oracle contracts are denominated in Korean won and Japanese yen, not dollars. That is the forward-looking fact.