SpaceX Just Became an AI Infrastructure Company

SpaceX will supply open-source AI startup Reflection AI with computing power from its Colossus 2 data center, a deal worth up to $6.3 billion if it runs its full course through 2029. Reflection will pay $150 million a month starting July 1, drawing on Nvidia GB300 chips housed in the same facility.

Either company can walk away with 90 days' notice after the first three months, a clause that matters given how young this arrangement is. Reflection was last valued at $25 billion and has already taken an $800 million investment from Nvidia, which puts Nvidia in the odd position of being both an investor in Reflection and, through SpaceX's data center, one of its indirect suppliers.

This is not SpaceX's first compute deal this year. Colossus 2 has now signed similar arrangements with Anthropic, Google, and Cursor, turning Elon Musk's rocket company into a landlord for the AI industry's most compute-hungry tenants.

Why a rocket company is renting out chips

SpaceX's core business does not obviously require a data center full of Nvidia's newest silicon. What it does have is capital, land, and power infrastructure built for a different kind of demanding workload, and a growing list of AI companies willing to pay for guaranteed access to scarce GB300 capacity rather than wait in line at a traditional cloud provider.

For Reflection, the calculation is straightforward: open-source model training is compute-bound, and locking in $150 million a month of dedicated capacity is cheaper than competing for spot capacity against better-funded labs. The 90-day exit clause suggests neither side is fully confident the arrangement survives at that price if better options appear.

The bigger pattern is what this says about who ends up supplying AI infrastructure. Traditional cloud providers, chipmakers, and now aerospace companies are converging on the same business: selling guaranteed compute to whoever can pay for it, regardless of what industry they started in.

Sources: CNBC · Data Center Dynamics · MLQ News

Comments

Post a Comment

Popular posts from this blog

Claude Sonnet 5's Price Cut Is the Real Story, Not the Benchmarks

Anthropic's Claude Sonnet 5 launched on June 30 with the kind of benchmark chart every model release ships with these days: 63.2% on SWE-bench Pro, up from Sonnet 4.6's 58.1%, closing in on flagship Opus 4.8's 69.2%. On Humanity's Last Exam, it hits 57.4% with tool use, essentially matching Opus 4.8's 57.9%. On GDPval-AA v2, a knowledge-work benchmark, Sonnet 5 actually edges past Opus 4.8, 1,618 to 1,615. Those numbers are real and worth noting. But they're not the interesting part. The interesting part is the price tag attached to them. Sonnet 5 is shipping at an introductory $2 per million input tokens and $10 per million output tokens through August 31, dropping to a standard $3/$15 after that. For a model landing within a few points of a frontier-tier system on agentic coding and reasoning benchmarks, that's a meaningfully different cost structure than what "near-frontier performance" has historically demanded. Why this matters more than th...
Read more