The leaked call indicates the company achieves approximately 85% inference margins, generating a sixfold profit on these operations. It also states the company secured two Atlas Super Clusters, totaling 16,000 950DT GPUs from Huawei, and plans to port its TileLang software to Huawei's ecosystem.
High inference margins suggest significant profitability for AI model providers, while the Huawei GPU procurement highlights a strategic shift to non-NVIDIA compute for large-scale inference.