Why is Musk willing to rent 220,000 GPUs to 'competitor' Anthropic?

Mirae Asset Securities in its May 8th research report made a core judgment: xAI's decision to fully lease Colossus 1 to Anthropic is not about giving its core assets to a competitor, but rather transforming a heterogeneous cluster with limited training efficiency into a cash-flow asset better suited for inference business.

Heterogeneous Cluster Amplifies Tail Effect

The constraints of Colossus 1 initially stem from the hardware structure. The report states that this data center, located in Memphis, is equipped with over 220,000 NVIDIA GPUs, including approximately 150,000 H100s, 50,000 H200s, and 20,000 GB200s, with three generations of chips mixed within the same cluster, forming a heterogeneous architecture.

This configuration is not unacceptable for inference, but when used for large-scale distributed training, it can amplify efficiency losses. The reason is straightforward: during 10,000-level GPU training, each step of the computation has to wait for all chips to synchronize, with the fastest GB200 also having to wait for the slowest H100 or any problematic node to catch up.

This is the tail effect that the research report emphasizes. The report cites The Information, stating that xAI's recent MFU (Model Floating-point Utilization) is only 11% — MFU refers to the utilization rate of model floating-point operations, which is the proportion of actual computational power achieved compared to the theoretical peak computational power. In contrast, Meta and Google manage to reach above 40%, and this gap reflects not the performance of individual chips but the collaborative efficiency of the entire training system.

The network topology further amplifies this issue. The report suggests that NVIDIA NCCL is traditionally more suited for scales from thousands to tens of thousands of cards; once scaled up to the 100,000 card level, the latency of data transmission along the ring topology increases significantly, and the time GPUs spend waiting for data suppresses overall utilization.

GB200 also brings additional software adaptation pressure. Former xAI Multimodal Pretraining Head Zeeshan Patel stated that the Blackwell GPU's power consumption potential is very aggressive, as the chip has built-in power smoothing mechanisms, while xAI's existing software stack was primarily aimed at Hopper, necessitating the rewriting of the model stack to address the new hardware characteristics.

Rental Cash Flow Rewrites IPO Narrative

In this framework, xAI's choice seems more like an asset reconfiguration. The report notes that xAI migrates its training load to Colossus 2, which is 100% homogeneous in Blackwell, and hands over Colossus 1 to Anthropic, which requires more inference capacity, since inference tasks have lower demands on GPU synchronization and communication, thereby reducing the weaknesses of a hybrid architecture.

For Anthropic, Colossus 1 provides readily available computational power. This single-tenant lease covers over 220,000 GPUs and 300MW of electricity, with plans to go live this month, filling in the short-term capacity gap in its expanding inference needs.

For xAI, the transaction will form a financial fortress before the IPO. The report estimates that Colossus 1, rented out at approximately $2.60 per GPU hour, could generate an annual revenue of $5 billion to $6 billion, close to the annualized net loss of xAI of about $6 billion from the first quarter.