SpaceX just signed its third major GPU rental deal. This time it’s $6.3 billion with Reflection AI for access to GB300 chips. Stack that on top of existing contracts with Anthropic and Google, and SpaceX is now pulling in an estimated $2.32 billion per month from renting compute. That’s $28 billion annualized. For context, that’s roughly twice what CoreWeave makes in a year.
If you run any kind of service business that touches AI, this matters. The company building rockets is now a critical player in the infrastructure layer that determines what your AI tools cost, how fast they run, and who controls access. The compute supply chain just got a new giant.
What happened
- Reflection AI signed a $6.3 billion compute deal with SpaceX for immediate access to GB300 chips. Payments of $150 million per month start July 1, 2026, and run through 2029.
- This is SpaceX’s third major GPU deal. It joins existing contracts with Anthropic (Colossus 1 and Colossus 2) and Google.
- Anthropic’s deal is the biggest. They took all of Colossus 1 at $1.25 billion per month, covering roughly 325,000 total chips.
- Jamin Ball from Clouded Judgement tallied the numbers and noted that SpaceX is charging above $10/hour for Blackwell GPUs, which he called a very high rate.
- The deals reportedly include 90-day out clauses, giving customers some flexibility.
The numbers
- $2.32 billion/month in total GPU rental revenue across three deals
- $28 billion/year annualized, roughly 2x CoreWeave’s current revenue
- $1.25 billion/month from Anthropic alone (Colossus 1)
- $150 million/month from the new Reflection AI deal
- >$10/hour implied rate for Blackwell GPUs
- 325,000+ chips in the Colossus 1 cluster
- CoreWeave holds a $60 billion valuation one year after IPO
5 things service business operators should know about SpaceX’s GPU play
- Compute is becoming a brokerage market. SpaceX isn’t building AI models. It’s renting raw GPU power to the companies that do. This creates a new layer between hardware makers (NVIDIA) and model builders (Anthropic, Google). That layer sets prices you’ll eventually pay.
- $10/hour for Blackwells is expensive. Jamin Ball flagged this as a high rate. If you’re buying AI services from companies paying these prices, those costs get passed downstream to you. Every dollar per GPU-hour shows up in your API bill or your vendor’s pricing.
- The 90-day out clauses matter. These aren’t locked-in-forever contracts. Big AI labs can shift compute providers relatively quickly. That means pricing pressure could increase as more neoclouds compete for the same customers.
- CoreWeave has a real competitor now. CoreWeave has been the poster child for GPU-as-a-service. SpaceX doing twice their revenue changes the competitive landscape. More competition generally means better pricing for buyers further down the chain.
- Open source models are driving demand. Reflection AI is using SpaceX compute to train open source models. Combined with GLM-5.2’s breakout week and Baseten’s $1.5 billion raise, there’s a clear trend: open models are getting good enough that companies want dedicated compute to fine-tune and run them.
The hot take
SpaceX becoming a $28B/year compute landlord is the clearest sign yet that AI’s real bottleneck isn’t intelligence, it’s electricity and chips. We keep talking about model breakthroughs. But the companies printing money right now are the ones with power connections and GPU clusters. SpaceX has both (thanks to its existing infrastructure footprint), and it’s exploiting that advantage hard. The model layer is getting commoditized. The infrastructure layer is where the pricing power lives. If you’re planning your AI strategy around which model is smartest, you’re optimizing the wrong variable.
The Agency OS play
This week, pull up every AI vendor contract and API bill you’ve signed in the last six months. Look at what you’re actually paying per token, per API call, or per seat. Then ask your vendor a simple question: what’s your underlying compute cost, and who’s your provider? You don’t need them to open their books. You just need to know if they’re on CoreWeave, AWS, SpaceX-backed infrastructure, or something else. That tells you how exposed your pricing is to a single compute provider’s rate changes.
If you’re running any AI workloads directly (fine-tuning models, running local inference, or building internal tools), benchmark the open source alternatives right now. GLM-5.2 is showing up as genuinely capable for agent-style work at a fraction of the cost of closed models. Baseten is serving it at over 280 tokens per second. That means you might be able to swap expensive API calls for self-hosted inference on cheaper compute, especially for repetitive, high-volume tasks like document processing, lead scoring, or intake automation.
Finally, start treating compute procurement like you treat any other major cost center. Get quotes from at least two providers. Ask about commitment terms, out clauses, and price locks. The $10/hour Blackwell rate SpaceX is charging is a useful benchmark. If your provider can’t explain how their pricing compares to that number, you’re flying blind on one of the fastest-growing line items in your business.