Accelerating Mixture of Experts Training With Rail-Optimized InfiniBand Networking in Crusoe Cloud
AI Engineer World's Fair 2024 · 17:45
AI cloud and data center infrastructure
Crusoe builds AI data centers and operates Crusoe Cloud for model training, inference, and high-performance computing. Its customers include AI startups, enterprises, and hyperscalers. Crusoe Managed Inference lets developers deploy open-weight or custom models through managed API endpoints, while Crusoe Intelligence Foundry provides model discovery and API-key creation without requiring users to manage GPU clusters. Cloud users include Cursor, robotics company 1X, and Fireworks.
Founded in 2018, Crusoe initially converted wasted natural gas into electricity for computing. Co-founders Chase Lochmiller and Cully Cavness remain CEO and president/chief strategy officer, respectively. Its engineering extends beyond physical infrastructure: MemoryAlloy separates inference KV caches from individual model processes and shares cached context across machines. A cache-aware gateway routes requests using available context and estimated time to first token, reducing repeated prompt computation for workloads such as long-document questions and multi-turn chat.
Crusoe divested its Digital Flare Mitigation business in 2025 to focus on AI infrastructure. It co-develops power and construction and manufactures electrical components in-house. In June 2026, the company reported 4.9 GW of contracted AI infrastructure across data-center projects and cloud capacity. Its Oracle campus in Abilene had two operational buildings and six under construction.
AI Engineer World's Fair 2024 · 17:45
Affiliations reflect their AIE appearances, not necessarily current employment.
The talk describes conventional VPC networking alongside a separate rail-optimized InfiniBand fabric designed for high-bandwidth, low-latency communication between GPUs.
Affiliations reflect each recorded session, not necessarily current employment.