Rent a B300, the Blackwell Ultra ceiling. 288 GB HBM3e at 8,000 GB/s, FP4 tensor cores, 5th-gen NVLink at 1.8 TB/s. The extra 96 GB over B200 unlocks 1T+ parameter pretraining without aggressive sharding, DeepSeek-V3 671B MoE serves at ~5,200 tok/s aggregated with full KV cache headroom, Llama-3 70B FP4 at ~7,500 tok/s with 256-request concurrency. Available on Clore.ai bare-metal 8-GPU NVLink-Switch pods. Billed per-minute, paid in BTC or USDT/USDC.
B300s are what you rent when memory matters. 288 GB HBM3e per card holds 1T+ MoE active weights with margin. FP4 tensor cores, 8 TB/s HBM keeps the SMs saturated; 5th-gen NVLink lets eight cards train as one.
B300 raises the per-card memory ceiling to 288 GB HBM3e, 50% more than B200. This is the unlock for 1T+ parameter pretraining and serving 671B MoE with full KV cache resident on one node. Blackwell Ultra tensor cores keep B200's FP4/FP8 throughput while giving you the memory headroom that B200 couldn't sustain.
TensorRT-LLM, vLLM, SGLang — all tuned for Blackwell Ultra. 288 GB VRAM serves 70B FP4 with massive KV-cache headroom, or DeepSeek-V3 671B MoE from an 8-GPU pod.
288 GB HBM3e + FlashAttention-3 means 256k-token contexts run without offload, and 671B MoE fits with full KV cache resident on a single 8-GPU pod. Ideal for long-doc RAG and agent loops with tool use.
Blackwell Ultra is the architecture trillion-parameter foundation models are being pretrained on. Specs from Nvidia's SXM5 datasheet; pricing reflects the lowest live spot floor.
// prices are spot-market lows · refreshed every 60 s
Every server is priced by its host. These are the live floors across the marketplace — you'll see hundreds of variants once you're in.
No sales call. No quota request. No three-week procurement. The first four commands are all you need.
Filter the marketplace by B300 288GB, country, GPU count, reliability score, network speed.
Choose a Docker image — PyTorch, vLLM, ComfyUI, Blender — or paste your own.
You get a public endpoint, an SSH key, and Jupyter on port 8888 in under 90 s.
Per-minute billing rounds to the second. Stop the instance and the meter stops with it.
B300 is Blackwell Ultra — same FP4/FP8 tensor cores as B200 but with 288 GB HBM3e (50% more memory) and higher sustained throughput. The extra VRAM unlocks 1T+ parameter pretraining and 671B MoE serving with full KV cache headroom, where B200 had to swap experts. CLORE.AI lists B300 as supply ramps in 2026; available on bare-metal 8-GPU NVLink-Switch pods.
Single-card, no - 70B pretraining needs an 8-GPU node minimum. CLORE.AI lists 8x H100, 8x H200, and 8x B200 pods with NVLink fabric for exactly this. A100 80GB pods run 70B FSDP training but at lower throughput than Hopper-class. For multi-week training, contact host operators for reserved-instance terms - listed in the marketplace under 'Reserved'.
A100 80GB has no FP8 - peak is BF16/TF32. H100 introduces FP8 with TransformerEngine and roughly 4x the BF16 training throughput at 2x the rental price - so ~2x perf-per-dollar on FP8-eligible workloads. H200 matches H100 compute but adds 141 GB HBM3e. B200 doubles H100 FP8 again with 192 GB HBM3e. Pick by VRAM and bandwidth ceiling, not just sticker FLOPS.
8-GPU H100 SXM, H200 SXM, and B200 nodes ship with NVSwitch fabric - 900 GB/s peer bandwidth on H100/H200, 1.8 TB/s 5th-gen NVLink on B200. PCIe variants (H100 PCIe, A100 PCIe) have NVLink Bridge in pairs only. Multi-node fabric (NVLink-Switch across racks) is available on B200 hyperscale pods - filter by 'NVSwitch' in the marketplace.
Yes. Multi-GPU listings expose all cards in a single rental as a coherent node with NVSwitch (where present), shared NVMe scratch, and InfiniBand or 100 GbE fabric for multi-node training. The standard PyTorch torchrun, DeepSpeed, and Megatron-LM launchers run unmodified. Filter the marketplace by GPU count to find 8x A100, 8x H100, 8x H200 nodes.
V100 (HBM2, 900 GB/s) -> A100 40GB (HBM2e, 1,555 GB/s) -> A100 80GB (HBM2e, 1,935 GB/s) -> H100 (HBM3, 3,350 GB/s) -> H200 (HBM3e, 4,800 GB/s) -> B200 (HBM3e, 8,000 GB/s). Each generation roughly doubles bandwidth or VRAM; KV-cache-bound serving and bandwidth-bound training scale almost linearly with this number.
288 GB HBM3e + Blackwell Ultra FP4 — the trillion-parameter inference card. Pretrains 1T+ models and serves 671B MoE without expert swap.
288 GB per card carries 1T+ parameter active weights without B200's expert-swap overhead. Frontier-lab unlock for trillion-scale pretraining.
Read the guide →288 GB per card serves 671B MoE with full KV cache headroom — no off-card expert spill, ~37% higher sustained throughput than B200.
Read the guide →Doubles B200 70B concurrency at the same per-request latency budget. The serving-throughput ceiling in 2026.
Read the guide →Side-by-side specs across the datacenter tier. Click any row to see that GPU.
Step-by-step guides verified on CLORE.AI hardware. Pick a workload, copy the docker image, ship in minutes.
Per-minute payouts in BTC, USDT or USDC. No listing fee, no contracts, withdraw any time.
Texas and EU bare-metal hosts are accepting workloads right now. Sign up, top up your wallet, and the next hour is yours.