Scaleway vs Vultr

Both Scaleway and Vultr employ per-hour billing for GPU instances, minimizing costs for variable workloads compared to monthly commitments. Scaleway's Nabu AI Supercomputer prices H100s at around €3.50/hour and A100s at €1.50/hour (as of late 2024), with no per-second granularity but flexible scaling via autoscaling groups. Vultr mirrors this with hourly rates—H100s ~$2.50/hour, A100s ~$1.00/hour—offering per-second billing for non-GPU compute, which extends to some GPU configs for ultra-short bursts. Neither heavily promotes reserved instances for GPUs, but Vultr provides spot instances at 50-90% discounts for interruptible workloads, absent in Scaleway. This favors Vultr for spiky usage like experiments, while Scaleway's flat hourly model suits predictable long runs. Implications: short experiments (<1h) cost more on both without minimums; long training benefits from hourly resets, but Vultr's spots reduce risk for fault-tolerant jobs.

Scaleway delivers superior value for sustained EU-based training runs, where H100 pricing undercuts Vultr by 20-30% when sovereignty justifies premium, ideal for large LLM pretraining (e.g., 8xH100 clusters at ~€28/hour total). Vultr shines for small experiments and batch inference via spot pricing, slashing costs by up to 90% for interruptible jobs, and global regions avoid data transfer fees. For production inference, Vultr's multi-region autoscaling offers better ROI for high-traffic apps, while Scaleway's integrated Nabu storage reduces egress costs for EU inference. Budget-conscious fine-tuning favors Vultr's cheaper A100s and RTX options for prototyping. Overall, Vultr provides broader value for variable, global workloads; Scaleway for committed, sovereignty-bound runs—calculate TCO using regional spot availability for decisions.

Scaleway focuses on dedicated bare-metal-like Nabu GPU instances with KVM virtualization, InfiniBand/RoCE networking up to 400Gbps, and elastic block/object storage optimized for AI (e.g., 1PB scalable). Kubernetes is fully managed via Kapsule. Vultr blends virtualized Cloud GPUs with bare metal options, 10-100Gbps networking, block storage up to 10TB/volume, and global object storage. Both support managed K8s, but Vultr's 32+ regions enable multi-cloud federation, while Scaleway prioritizes dense EU clusters.

Scaleway's Nabu delivers top-tier H100 performance with NVLink 900GB/s interconnects for multi-GPU training, strong in sustained FLOPS for EU workloads; limited regions may bottleneck global scaling. Vultr GPUs match specs (A100 80GB, H100 SXM), with reliable multi-node scaling via Slurm/K8s, but variable inter-region bandwidth. Both offer high availability (>99.9%), with Vultr edging in provisioning speed (minutes) due to scale; Scaleway noted for lower jitter in EU benchmarks.