·       The partnership lets Google resell NVIDIA CUDA capacity to enterprises while protecting its TPU supply for internal models.

·       Economics deliver Google strong margins; SpaceX locks in high-margin revenue just before IPO, shifting its valuation narrative.

·       Both parties benefit from symbiotic co-opetition in the AI buildout, with flexible termination after December 2026.

On June 5, 2026, SpaceX filed paperwork with regulators disclosing a major cloud computing deal with Google. The contract amounts to over $30 billion across three years and gives Google access to a massive AI computing capacity. However, there is lot to dig into from the why to the economics and the timing.

The Big Deal:

There is insatiable demand for the companies that are seeing quadrillion tokens being consumed per month, and Google is one of the top three. Anthropic being another one. The demand from Google’s AI products, particularly its Gemini models and AI-powered search, is outpacing what it can build internally.

Though the demand stems beyond Gemini on Google’s massive heterogeneous infrastructure, consisting of Google TPUs to NVIDIA GPUs, from the growing Google Cloud customers who are looking to get on the AI bandwagon. We believe Google isn’t buying this capacity purely for internal use. A large portion will be resold to enterprise customers through Google Cloud.

Google doesn’t have enough compute capacity of its own, and it can’t build more fast enough, even though the AI ASIC server shipments driven by Google are going to triple by the end of next year. (Read here: AI Server Compute ASIC Shipments to Triple by 2027 as Custom Silicon Enters Hyper-Growth Phase)

This, we believe, is the premise of Google and others going out looking for infrastructure to not lose current demand to competitors. Thus, the practical benefit for Google to partner here with SpaceX is to get access to a fully operational, powered compute infrastructure, and Google avoids the enormous cost and complexity of building out its own data center power infrastructure for this capacity. This deal lets Google effectively skip that construction queue. Rather than waiting 18 to 24 months, it gets access to 110,000 GPUs more or less immediately.

There’s also a chip ecosystem angle. Google’s own internal AI work runs heavily on its proprietary TPU chips. But enterprise customers overwhelmingly prefer NVIDIA’s CUDA ecosystem. By leasing NVIDIA capacity from SpaceX, Google Cloud can sell what enterprise clients actually want without diverting its TPU supply away from its own model development and inferencing.

Google isn’t the first major AI company to rely on SpaceX’s infrastructure. We believe Anthropic has also made a significant deal with SpaceX for access to the same xAI Colossus data center footprint. Both companies are, in effect, paying to use a potential direct competitor’s infrastructure, which is not making money yet from its own AI offerings because there’s simply no other option at this scale right now.

This has quietly turned SpaceX into a central infrastructure provider for the AI industry.

The Economics:

According to the deal note, the core of the deal runs from October 2026 through June 2029, which is exactly 33 months. During that period, Google pays SpaceX $920 million per month for access to roughly 110,000 NVIDIA GPUs along with supporting hardware: CPUs, memory, and high-speed networking. We don’t exactly know the mix of these GPUs, but I believe it should be Blackwells B200/B300 and expanding to Rubin. There’s also a ramp-up period before October with slightly lower fees, which pushes the total contract value to around $31.5 billion over three years.

Breaking that down per GPU, Google is paying about $8,363 per GPU per month, or roughly $11.45 per hour, assuming the machines run around the clock. That’s a premium, but it reflects the fact that this compute is available immediately, at scale, and fully set up, taking care of the opportunity costs.

Google’s wholesale cost from SpaceX works out to around $11.45 per GPU per hour. Top-tier NVIDIA chips on major cloud platforms currently go for roughly $5 to $15.00 per hour, depending upon instance types, with newer architectures like the Blackwell or Reubin LPX for Agentic AI type applications could fetch from $15 to $50 per hour or a million tokens of a trillion-parameter model based on potential tiering.

Google will layer its own software, support, and service guarantees on top of the SpaceX infrastructure, then charge enterprise customers somewhere between $18 and $70 per GPU per hour, depending on contract length, type of AI workload instances and SLAs. Even at $18 per hour, that’s roughly $6 per hour per GPU in gross margin, which is not bad on 110,000 GPUs running continuously. That equates to a minimum of $350-$450 million per month in just gross margins or $5 Billion per year at base price.

The Timing & Timeouts:

The timing of this deal looks deliberate. The filing landed alongside an amendment to SpaceX’s S-1 registration statement, just days before its expected public listing.

A $30 billion guaranteed revenue contract with Google could change how investors think about SpaceX’s valuation. It could change from a potential next trillion-dollar company, but with high CAPEX and long payback cycles, to a high-margin AI infrastructure business with locked-in cash flows. This could change the conversation on how much the company’s stock is worth.

This deliberate timing is genius to fuel SpaceX’s stock pricing and IPO.

However, some of the disclaimers for “timeouts” in the deal are also worth paying attention to:

• SpaceX must have the full 110,000-GPU cluster operational and accessible by September 30, 2026. If it misses that date, Google can walk away after a short grace period or reduce its fees proportionally. This protects Google against delays in chip supply or power infrastructure.

• The contract explicitly states that Google retains ownership of all its content, AI models, and data processed on the hardware. Given that SpaceX and xAI are closely linked, Google must have made it clear on a firewall on its proprietary data or model training information to remain private from xAI’s competing Grok models.

• After December 31, 2026, either party can terminate the agreement with 90 days’ notice. For Google, this is a good exit if its infrastructure is ready or its find a cheaper alternative infrastructure or something goes south with SpaceX. For SpaceX, the deal has already served its main purpose by that point of anchoring the IPO narrative with a high-profile, long-term revenue commitment. So even if Google exists in some months post its IPO and business becomes stable, it’s a win-win and easy to comply with Google’s terms.

The Symbiotic Co-opetition:

To wrap up, the deal is another example of symbiotic coopetition in this great AI buildout. Some might see it as circular, but I think it is more tangible and somewhat concentric with a sweet spot in between. And, what it definitely does is help drum up the AI ecosystem valuations and maintain the momentum.

Google gets the compute it needs now without waiting years for additional new buildout. SpaceX gets a blue-chip customer and a compelling revenue story right before going public. Win-Win.

Whether the arrangement lasts beyond 2026 is another question entirely. I believe none of them cares.