Google's $920M Monthly SpaceX Compute Deal Explained

The Deal in One Paragraph

Google has committed to paying SpaceX approximately $920 million per month for access to 110,000 NVIDIA GPUs, with the contract running through mid-2029. The deal was disclosed in an SEC filing tied to SpaceX's upcoming IPO, per TechCrunch's reporting on June 5, 2026. At that monthly rate, the total contract value lands somewhere north of $33 billion over roughly three years — making it one of the largest AI compute commitments ever publicly disclosed.

This is Google. The company that builds its own TPUs, runs one of the largest cloud platforms on Earth, and just shipped Gemini 3.5. And it's writing SpaceX a check approaching a billion dollars a month for someone else's GPUs.

That tells you everything about where the AI compute market is right now.

What's Actually in the Filing

Based on reporting from TechCrunch and The Verge, here are the key terms:

• Monthly commitment: ~$920 million
• GPU count: 110,000 NVIDIA GPUs
• Duration: Through mid-2029 (~3 years)
• Estimated total value: $33+ billion
• Disclosure: SEC filing ahead of SpaceX's IPO

The filing surfaced because SpaceX is preparing for what could be one of the largest tech IPOs in years. Major customer contracts like this get disclosed to give investors a picture of SpaceX's revenue streams — and this one paints a very specific picture: SpaceX isn't just a launch company anymore. It's an AI compute provider with customers that include two of the most important AI companies on the planet.

Google Is SpaceX's Second Major AI Customer

If this deal sounds familiar, it should. In early May 2026, Anthropic signed a lease for SpaceX's entire Colossus 1 supercluster — 220,000 NVIDIA GPUs drawing 300MW of power. That deal, which we covered when it broke, gave Anthropic access to what was at the time xAI's flagship training rig, built for Grok.

Here's how the two deals compare based on public disclosures:

Term	Anthropic-SpaceX	Google-SpaceX
GPU count	220,000	110,000
Disclosed monthly cost	Not publicly disclosed	~$920M
Cluster	Colossus 1	Not specified in reports
Duration	Not fully disclosed	Through mid-2029
Disclosed via	Announcement/reporting	SEC filing (SpaceX IPO)

The Google deal is for roughly half the GPUs that Anthropic secured, but Google is the first customer where we have a hard monthly dollar figure attached. That $920M/month number is significant — it puts a concrete price tag on large-scale GPU leasing that the industry can actually benchmark against.

Why Google Needs Outside GPUs

This is the question worth sitting with. Google operates one of the largest computing infrastructures in the world. Google Cloud is a top-three cloud provider. Google DeepMind designs its own TPUs — custom silicon purpose-built for AI workloads. The Trillium TPU (v6e) is already in production. Why go to SpaceX for NVIDIA hardware?

A few factors make this less contradictory than it sounds:

Demand is outrunning everyone's capacity. Training frontier models, running inference at scale for Gemini across Search, Android, Workspace, and Cloud — the compute appetite is genuinely insatiable right now. Even companies that manufacture their own chips can't build capacity fast enough. Google said as much during its I/O 2026 keynote when it acknowledged that Gemini demand was straining internal resources.

TPUs and NVIDIA GPUs serve different roles. Google's TPUs are optimized for specific workloads within Google's ecosystem. But a lot of AI research, third-party model support, and certain training configurations still run better on NVIDIA's CUDA stack. Leasing NVIDIA GPUs from SpaceX doesn't replace TPU production — it supplements it for workloads where NVIDIA hardware is the better fit.

Speed of deployment matters. Building new data centers and manufacturing new TPU generations takes years. Leasing existing GPU clusters from SpaceX gets capacity online in months, not years. When the competitive window for frontier AI models is measured in quarters, that time advantage has real strategic value.

SpaceX as AI Landlord: The IPO Angle

The timing of this disclosure isn't accidental. SpaceX filed this contract as part of its IPO preparation, and the reason is obvious: it transforms SpaceX's revenue narrative.

Starlink already made SpaceX a recurring-revenue business. But AI compute leasing adds a second, entirely different revenue stream — one that ties SpaceX to the fastest-growing sector in tech. Having Google and Anthropic as anchor tenants makes SpaceX's compute business look less like a side project and more like a serious infrastructure play.

Back the math out: $920 million per month from Google alone is $11 billion per year. Add the Anthropic deal (terms not fully public, but likely comparable in scale given the larger GPU count), and SpaceX could be pulling in $20+ billion annually from AI compute leasing alone. For context, SpaceX's launch revenue was estimated at roughly $9 billion in 2025. AI compute may already be the bigger business.

My read: the SpaceX IPO prospectus is going to tell a story about three businesses — launch, Starlink, and AI compute — and the AI compute segment might be the one that gets the highest revenue multiple from Wall Street. Recurring, contracted, inflation-resistant revenue from the two companies most likely to define the next decade of AI? That's exactly what public-market investors pay up for.

The Broader AI Compute Land Grab

Google's SpaceX deal doesn't exist in isolation. It's part of a wave of massive compute commitments that have defined 2026:

• Anthropic-SpaceX: 220K GPUs via Colossus 1 lease
• Anthropic-Akamai: $1.8B, 7-year cloud deal
• IREN-NVIDIA: $3.4B infrastructure partnership
• NVIDIA-Corning: $3.2B optics supply deal
• Google-SpaceX: ~$33B total (this deal)

The pattern is clear: every major AI company is locking up compute capacity years in advance, and they're willing to pay enormous premiums to do it. The companies that don't secure GPU access now risk being compute-constrained when the next generation of models needs to train.

What's newer is who is providing the compute. A year ago, the AI infrastructure conversation was dominated by the hyperscalers — AWS, Azure, GCP — and the big colo providers like Equinix and Digital Realty. Now SpaceX, a rocket company, is one of the largest GPU lessors in the world. Akamai, traditionally a CDN company, is signing multi-billion-dollar AI deals. The compute supply chain is diversifying fast because demand has simply overwhelmed the incumbents' ability to build.

What We Don't Know

A few open questions that the SEC filing and current reporting don't answer:

• Which GPUs specifically? The reporting says NVIDIA, but doesn't specify whether these are H100s, H200s, B200s, or GB200s. The generation matters a lot — it's the difference between last-gen and cutting-edge performance per watt.
• Where are the clusters located? SpaceX's Colossus 1 is in Memphis, Tennessee. It's not clear whether Google's allocation is in the same facility, a different SpaceX data center, or spread across multiple sites.
• What's the workload split? Is Google using these for Gemini training, inference, Google Cloud customer workloads, or some combination? The use case affects how to read the deal's strategic significance.
• Are there renewal or expansion options? The deal runs through mid-2029, but contracts this size usually include options. Whether Google can extend or scale up is an important detail for the long-term picture.

These gaps will likely get filled as SpaceX's IPO filing becomes more detailed in the coming weeks.

The Honest Take

Three things stand out about this deal:

First, compute scarcity is real, not marketing. When Google — a company that literally manufactures its own AI chips — goes outside for 110K GPUs, the "there aren't enough chips" narrative isn't hype. It's operational reality. Every major AI lab is scrambling for capacity, and even the best-resourced players can't build fast enough internally.

Second, SpaceX just became a critical node in the AI ecosystem. With Google and Anthropic as customers, SpaceX's compute infrastructure is now upstream of two of the most important AI models in the world (Gemini and Claude). That's a concentration of dependency that deserves attention — both from investors pricing the IPO and from policymakers thinking about AI supply chain resilience.

Third, the unit economics of AI compute are staggering. $920 million per month for 110K GPUs works out to roughly $8,400 per GPU per month. That's a premium price, and Google is paying it willingly. It suggests that the revenue Google expects to generate from these GPUs — through Gemini, Cloud AI services, and internal applications — far exceeds even that astronomical cost. The companies buying GPU access at these prices clearly see returns that justify the spend.

For everyone watching the AI infrastructure race, this deal is confirmation that 2026 is the year compute access became the defining competitive advantage — not model architecture, not data quality, not talent. Compute. And the companies willing to write the biggest checks are the ones positioning to win.