Crusoe vs Scaleway

Both providers use per-hour billing, minimizing short-job overhead compared to per-second models like AWS. Crusoe differentiates with spot instances, offering deep discounts (often 50-80% off on-demand) for interruptible workloads, ideal for fault-tolerant training. No reserved instances are prominently featured, but per-hour granularity suits variable usage. Scaleway sticks to straightforward per-hour on-demand pricing across GPU instances (e.g., A100, H100 via Nabu), lacking public spot markets but potentially bundling with integrated services for volume discounts. Implications: Spot availability favors Crusoe for spiky, preemptible ML experiments, reducing costs for non-urgent jobs but risking interruptions. Scaleway's predictability benefits steady inference/production, though higher base rates may apply without spots. For multi-hour runs, both align well; sub-hour tasks incur full-hour charges, pressuring efficiency in both.

Crusoe delivers superior value for small experiments and large training runs via spots—e.g., H100 clusters at ~$2-4/hour spot vs. $8-12 on-demand, unbeatable for budget-constrained prototyping or batch inference (savings amplify over days). Less ideal for production inference needing reliability. Scaleway shines in production inference and fine-tuning with Nabu’s scaled H100s, offering integrated storage/networking that cuts setup overhead (effective value at ~$5-10/hour equivalents). For small experiments, per-hour billing without spots reduces edge over Crusoe. Large training favors Scaleway’s EU sovereignty if compliance trumps cost; otherwise, Crusoe’s green spots win. Overall, Crusoe for cost-sensitive batch (30-50% better value), Scaleway for sovereign scale.

Crusoe emphasizes bare-metal GPU servers (A100, H100) with high-speed InfiniBand networking for multi-node scaling, vertically integrated for low-latency energy delivery. Limited to U.S. regions (e.g., Denver, San Antonio); supports Kubernetes via bring-your-own but focuses on simplicity. Storage via block/object with S3 compatibility. Scaleway offers virtualized/bare-metal GPUs in Nabu (thousands of H100s), EU-focused (France, Netherlands) with 400Gbps interconnects, native Kubernetes (Kapsule), and rich storage (Block, Object, Rclone-compatible). Scaleway edges in managed services and regions.

Crusoe GPUs deliver hyperscaler-comparable performance, excelling in multi-GPU batch scaling (e.g., 8x H100 pods) with sustainable uptime; spots may vary availability. Scaleway’s Nabu shines for massive parallelism (10k+ H100s), strong NVLink/InfiniBand for training throughput, potentially lower EU latency. Both handle DGX-like configs; Crusoe reports 95%+ utilization in green data centers, Scaleway benchmarks competitive for inference. Limited public benchmarks note Crusoe’s edge in cost-per-FLOP for spots, Scaleway in sovereign scale.