SpaceX rents out xAI's GPUs to Reflection AI in $6.3bn pact

The News

Reflection AI, an open-source artificial intelligence lab founded in 2024 by two former Google DeepMind researchers, has agreed to pay SpaceX $150 million every month for access to high-end computing power. The arrangement runs from 1 July 2026 through 2029 and is worth up to $6.3 billion.

In return, Reflection gets immediate use of Nvidia's latest GB300 chips and supporting hardware housed at Colossus 2, the sprawling data centre near Memphis, Tennessee, that was originally built by Elon Musk's xAI for its own model training. xAI is now part of SpaceX, which has begun renting out the facility's capacity as its internal AI ambitions cooled.

The contract is not locked in for the long haul. Either side can walk away with 90 days' notice once the first three months have elapsed, an unusually quick exit for a deal this size.

Reflection, which has raised $2 billion to pursue an open-weight strategy and publishes the trained parameters of its models, framed the move as a vote for openness. "Recent events highlight how important open source is to the AI ecosystem," a company spokesperson said, pointing to nations and enterprises wary of leaning entirely on closed systems.

Why It Matters

The headline number looks enormous until you set it beside its peers. At the same Colossus 2 complex, Anthropic is paying $1.25 billion a month and Google $920 million a month, with all three contracts running through July 2029. Reflection's spend, in that company, is the budget option.

That hierarchy is the real story. Compute, not algorithms, has become the binding constraint on who gets to build frontier models, and the gap between a $150 million tenant and a $1.25 billion one maps almost directly onto how much intelligence each can train. The last comparable scramble was the 2023 rush for Nvidia H100s, when start-ups signed away equity just to jump the queue.

It also marks a quiet inversion. xAI, which built Colossus to own its destiny, has become a landlord. Owning silicon proved less attractive than collecting rent on it, a lesson for every would-be compute champion before it pours concrete.

Indian Angle

For India, the deal is a mirror held up to an uncomfortable gap. The IndiaAI Mission has been assembling subsidised GPU capacity, more than 38,000 graphics processors pooled for start-ups and researchers, precisely so that local builders need not sign billion-dollar leases abroad. Reflection's $150 million monthly outlay is roughly the entire compute budget many Indian AI start-ups would hope to spend across several years.

The open-weight thread matters even more here. Indian model builders such as Sarvam, picked to develop a sovereign foundation model, and Ola's Krutrim lean heavily on open weights precisely because training from scratch at frontier scale is out of reach. A well-funded open lab renting world-class compute strengthens the public-weights ecosystem that Indian firms quietly depend on, lowering the cost of the base models they fine-tune.

There is a regulatory prompt too. MeitY's push for sovereign compute and the data-localisation instincts baked into the DPDP framework both assume India will eventually host such capacity at home. Until Krutrim's and Reliance's domestic data centres mature, Indian developers will keep paying dollar-denominated rates for compute that sits overseas, and a weaker rupee makes every GB300 hour dearer.

FAQ

How big is this deal compared with rivals?

Reflection AI's $150 million a month is the smallest of three known tenancies at Colossus 2. Anthropic pays $1.25 billion a month and Google $920 million a month. All three contracts run through July 2029, underlining how compute spend now separates the frontier pack from everyone else.

What does open-weight mean for Indian start-ups?

Open-weight models have their trained parameters released publicly, so Indian firms like Sarvam and Krutrim can fine-tune them rather than train from scratch. A better-resourced open lab effectively subsidises the base layer that cost-constrained Indian builders rely on.

Can either party cancel the contract?

Yes. After the first three months, either Reflection AI or SpaceX may terminate with 90 days' notice, an unusually flexible clause for a deal worth up to $6.3 billion.

When does the arrangement begin?

Payments start on 1 July 2026 and the agreement runs through 2029, giving Reflection immediate access to Nvidia GB300 chips at the Memphis facility.

Where can I read the original announcement?

The full reporting is at TechCrunch, linked in the attribution below.

This story was reported by TechCrunch. Read the full original coverage at TechCrunch.