CoreWeave vs TensorDock

Both providers use per-second billing with spot and on-demand options, minimizing waste for variable workloads compared to per-hour models like AWS. CoreWeave's pricing emphasizes predictable on-demand rates for H100/A100 GPUs (e.g., ~$2-4/hour on-demand, lower spots), with reserved options for long-term commitments. Spot instances provide deep discounts but compete in auctions, suiting bursty patterns. TensorDock differentiates via marketplace aggregation, boasting extremely low spot prices (often <$1/hour for A100s, up to 80% below competitors), stabilized post-Voltage Park acquisition for better inventory. No explicit reserved instances noted, focusing on opportunistic bidding. Implications: Short, interruptible jobs favor TensorDock's aggressive spots; sustained training benefits CoreWeave's reliability, avoiding restart overheads. High-utilization users (>70%) see better economics on CoreWeave due to scale efficiencies, while sporadic experiments thrive on TensorDock's lows.

TensorDock delivers superior value for small experiments and fine-tuning, where spot prices enable 3-5x savings over on-demand, ideal for <24-hour runs with tolerance for preemption. CoreWeave offers better value for large training runs (e.g., multi-day LLM jobs), as InfiniBand and Kubernetes reduce checkpointing costs from interruptions, potentially halving effective TCO despite higher base rates. For production inference, CoreWeave's consistent availability justifies premiums for SLAs. Batch inference leans TensorDock for cost if fault-tolerant, but CoreWeave wins for tight deadlines. Overall, TensorDock maximizes value at <50% utilization; CoreWeave at >70%, factoring reliability into ML engineer productivity.

CoreWeave employs a Kubernetes-native architecture on bare-metal clusters with InfiniBand for RDMA-scale networking, NVMe storage, and managed storage options like Rook Ceph. Supports multi-tenant isolation at hyperscale. TensorDock's marketplace model brokers virtualized/bare-metal GPUs from varied data centers, with less emphasis on unified orchestration—Kubernetes support uncertain, relying on user-managed setups. Networking/storage more heterogeneous, post-acquisition improvements noted but details sparse.

CoreWeave delivers top-tier multi-GPU scaling via InfiniBand (up to 400Gbps+), excelling in NCCL collectives for LLM training with minimal overhead. High H100/A100 availability for enterprises. TensorDock offers competitive single-node perf at low cost, but spot preemption and mixed interconnects (Ethernet common) hinder large-scale efficiency. Availability improved via stabilization, yet less proven for 100+ GPU jobs; user reports note good for mid-size but variable latencies.