SpaceX has rented its Colossus 1 data center to Anthropic after facing latency problems and incompatible hardware while attempting to train its Grok AI models. The data center’s limitations forced SpaceX to shift from using the facility in-house to generating revenue through rentals.
- Latency and mixed chip generations caused training delays at Colossus 1.
- Anthropic pays $1.25B monthly to lease the underutilized SpaceX facility.
- SpaceX’s Grok AI model struggled to meet expectations amid infrastructure limits.
What happened
SpaceX initially built the Colossus 1 data center with the intention of training its advanced AI models, including Grok, across multiple campuses. However, the company encountered significant latency issues when trying to connect Colossus 1 in Memphis with two other data centers over 10 miles apart. The aging network infrastructure was unable to provide the ultra-fast interconnectivity required for distributed AI model training.
Besides networking challenges, the facility suffered from hardware mismatches. Colossus 1 contains a mix of Nvidia chip generations, including older accelerators alongside Hopper and Blackwell processors. This heterogeneous hardware setup slowed overall cluster performance because workloads had to synchronize at the speed of the slowest components. As a result, SpaceX decided to lease Colossus 1 to Anthropic rather than letting the facility remain underutilized.
Why it matters
The leasing of Colossus 1 highlights the technical difficulties even well-funded companies face in building optimal AI training infrastructure. While SpaceX built the data center rapidly, the compromise on uniform hardware and network capacity has limited its ability to run cutting-edge models efficiently. This impacts the prospects of Grok, which has shown declining user engagement and slower adoption compared to competitors.
Moreover, SpaceX has turned its initial infrastructure investment into a significant revenue stream from external renters such as Anthropic and Google, collectively earning roughly $2.17 billion monthly. This shift positions SpaceX as an AI compute landlord rather than solely a direct AI services provider, altering the company's AI strategy and financial narrative.
What to watch next
SpaceX retains the option to reclaim Colossus 1 capacity if its internal AI needs grow stronger, with a lease agreement allowing cancellation within 90 days. However, recent trends in Grok’s performance suggest that reclaiming the facility may not be urgent in the near term. Monitoring Grok’s user engagement and paid conversion rates will be key indicators of whether SpaceX intensifies its internal AI development.
Industry observers should also watch how SpaceX balances efforts between continuing AI infrastructure rentals and developing proprietary AI products. The evolving arrangement could set a precedent for other technology firms building costly AI compute centers, potentially accelerating the monetization of large-scale AI infrastructure through third-party leasing.