Alphabet is deploying $84 billion into AI infrastructure across three distinct arrangements: a $920 million per month rental agreement with SpaceX for orbital GPU clusters, a $35 billion lease backstop for five Anthropic-designated US data centers, and direct capital expenditure on owned facilities. The SpaceX contract runs $11 billion annually and represents the first large-scale commercial deployment of space-based compute for model training. Anthropic's data center financing carries Google's credit guarantee on lease payments through 2034, with first delivery scheduled for Q2 2026.
The SpaceX arrangement is not satellite internet capacity. It is physical compute infrastructure—GPUs, CPUs, shared memory pools—deployed in low Earth orbit and accessed via dedicated fiber links at sub-15ms latency to Google's terrestrial data centers in Oregon and South Carolina. SpaceX provides the hardware, thermal management, and orbital maintenance. Google operates the software stack and trains models that exceed the thermal limits of ground-based facilities during peak batch runs. The $920 million monthly payment includes hardware amortization, launch cost recovery, and a margin for SpaceX's operational risk. First deployment is late Q3 2025, with four additional clusters by mid-2026.
The Anthropic backstop is structured as a sale-leaseback guarantee. Five data centers currently under construction in Virginia, Texas, and Ohio will be leased to Anthropic under fifteen-year terms. Google guarantees 100% of lease payments if Anthropic cannot meet obligations, effectively converting the facilities into Google-controlled assets without triggering immediate balance sheet recognition. The data centers are purpose-built for frontier model training, with total power capacity of 2.1 gigawatts and liquid cooling systems rated for 1,200-watt chips. Construction financing closed in March 2025 with Blackstone and KKR as co-developers. Google's guarantee converts to equity at a 1.3x multiple if Anthropic defaults, giving Alphabet both compute access and a path to outright ownership.
This is response to the 2027 capacity wall. Current projections show frontier labs exhausting available US data center inventory by Q4 2026 if training runs continue scaling at 4x per 18 months. Google is securing compute before the bidding war begins. The SpaceX deal solves thermal density—orbital facilities radiate heat into vacuum rather than fighting thermodynamics in Arizona. The Anthropic backstop solves domestic permitting delays and power grid connection timelines, which now average 38 months from breaking ground to energization. By pre-committing $84 billion, Google locks access to roughly 18% of projected US AI compute capacity through 2030.
Operators should track three follow-on events. First, watch for SpaceX's S-1 filing, expected late Q2 2025. The Google contract will appear as a named customer commitment and likely reset valuation expectations 15-20% higher than the $350 billion January private round. Second, monitor Anthropic's Claude 4 training start date, projected August 2025. If training begins in the Google-backstopped Virginia facility, it confirms the infrastructure is ahead of schedule and increases the probability Google exercises early conversion rights. Third, observe Alphabet's Q2 2025 earnings call in late July for updates on owned data center capex. Management has signaled $48 billion for 2025 alone, but the SpaceX and Anthropic deals suggest the figure is now closer to $62 billion when off-balance-sheet commitments are included.
The decade-long bet is on compute scarcity, not model superiority. Google is paying $84 billion to ensure it has the infrastructure to train models whether or not it builds the best ones. The SpaceX contract delivers in Q3 2025. The first Anthropic data center energizes in Q2 2026.