Google rents 110,000 GPUs from Elon Musk, which is a thing that happened

Google has agreed to pay SpaceX $920 million a month for access to roughly 110,000 NVIDIA GPUs hosted in the Colossus cluster in Memphis, from October 2026 through June 2029. The deal was disclosed in a SpaceX SEC filing on 5 June ahead of its IPO. Google calls it "bridge capacity" for Gemini Enterprise. The structural story is that Google, which designs and operates its own AI silicon, is renting someone else's chips, in a building owned by a direct competitor, for thirty-two months, at roughly $30 billion all-in.

What was actually filed. The Cloud Service Agreement, disclosed as part of SpaceX's pre-IPO paperwork, specifies approximately 110,000 NVIDIA GPUs plus associated CPUs, memory, and supporting hardware, with Google's access ramping before full delivery and fees discounted during the ramp.¹ If SpaceX misses the 30 September 2026 capacity milestone, Google can terminate after a short grace period or accept proportionally reduced capacity at a proportionally reduced price. Both sides hold a 90-day termination right after 31 December 2026.¹ So the headline 32-month, $30 billion number is the ceiling. The floor is closer to fifteen months and an exit clause.

Why this is the interesting part. Google owns TPUs. Tensor Processing Units (Google's proprietary AI accelerator chips, designed to reduce dependence on NVIDIA). The TPU program is on its fifth and sixth generations and is, by any sensible measure, the most credible non-NVIDIA AI silicon in production. Google's strategic rationale for building TPUs in the first place was to not be in the position it is now in. And yet here it is, paying market or above-market NVIDIA rates to a competitor's data centre because, per its own spokesperson, Gemini Enterprise demand outran the build.²

Inference economics, in the open. The frame predicts that as model use shifts from training to serving, the binding constraint moves from one-off compute spend to continuously available GPU-hours, and that this constraint will bite even vertically integrated players. This deal is that prediction in primary-document form. Google is not short on capital, not short on silicon expertise, and not short on data centres in aggregate. It is short on NVIDIA-compatible inference capacity in the specific window between now and whenever its own builds come online, and that window is apparently large enough to be worth $30 billion to close.

The "bridge" framing deserves a moment. Bridges are short. Thirty-two months is not short. A 32-month bridge is a building. If Google's own capacity were arriving in, say, Q2 2027, you would write a twelve-month deal with an extension option. Writing a deal through mid-2029 with an early-exit window says one of two things: either Google's internal build program has slipped further than it would like to disclose, or Gemini Enterprise consumption is growing fast enough that even the post-build capacity will need topping up. Both readings are bullish on inference demand and bearish on hyperscaler capex discipline. They are also not mutually exclusive.

The CUDA gloss. One reasonable counter-reading: Google is not capacity-constrained, it is software-constrained. Enterprise customers running Gemini Enterprise want to bring their own CUDA-based pipelines, CUDA being NVIDIA's proprietary software stack that most AI workloads are written against, and TPUs do not run CUDA. On this reading the lease is not about silicon shortage but about meeting customers where their code already lives. This is plausible and probably partially true. It does not change the structural conclusion much: in either case, Google is paying a competitor to host workloads it cannot run on its own fleet, and the constraint is binding enough to be worth nine-figures-monthly.

The third compute landlord. Add the Anthropic contract, $1.25 billion a month, also hosted at Colossus, disclosed separately in the S-1, and SpaceX walks into its IPO with roughly $2.17 billion a month, or about $26 billion a year, in contracted AI compute revenue from two blue-chip counterparties on multi-year terms.² That is not speculative GPU upside. That is a utility revenue line, the kind public-market investors price generously. The structural read is that there are now effectively four AI compute landlords of consequence: AWS, Azure, Google Cloud, and Colossus. The fourth is owned by a private rocket company whose founder runs the lab that competes with its two largest tenants.

Which brings us to the slightly strange part. Anthropic and Google are both now operationally dependent on infrastructure whose ultimate beneficial owner has a direct competitive interest in their products underperforming. The commercial contracts presumably contain the usual SLAs, audit rights, and data-isolation language. SLAs being service-level agreements, the contractual performance guarantees vendors offer customers. But counterparty risk in compute is not only about uptime. It is about prioritisation, about future renewal terms, about whose workloads get the newer racks when the cluster upgrades. None of that is easily contracted around when the landlord also runs a model lab.

I would expect the SpaceX S-1 to address this in a risk factor. I would expect Google's next 10-Q to address it less. The interesting disclosure question is whether Google flags concentration risk to a single non-hyperscaler provider, and whether the language acknowledges the competitive posture of that provider's affiliates. Filings tend to be honest about the first and shy about the second.

What this is a case of. It is a case of inference economics overrunning hyperscaler capex planning by enough to force the most vertically integrated AI infrastructure player on earth to lease from an outsider. It is a case of compute becoming a tradable utility with a small number of large landlords, one of whom is structurally conflicted. And it is a case of "agent platform demand" being concrete enough to put $30 billion of contracted spend behind, which is the most material evidence yet that agentic workloads are not a 2027 story but a current capacity emergency.

What to watch. Whether Microsoft or AWS files a comparable external-lease agreement in the next two quarters — the Google deal is only a one-off if no one follows. The September 2026 capacity milestone and whether SpaceX hits it. The risk-factor language in the final SpaceX S-1 on customer-affiliate competition. And the renewal behaviour at the 90-day termination window in early 2027, which will reveal whether "bridge" was the right word or whether Colossus has just become permanent Google infrastructure under a different name.

Footnotes

¶