288GB
VRAM
HBM3e memory
AIVory · GPU Marketplace
Live spot pricingNVIDIA B300 — the most memory you can fit on one card.
288 GB of HBM3e on a single accelerator. The B300 Blackwell Ultra pushes memory bandwidth to 12 TB/s and packs enough VRAM to hold a 400B dense model in FP16 without sharding. At $8.19/hr on spot, it's the only practical way to deploy the largest open-weight models on a single GPU.
At a glance
B300 specifications.
Key hardware specs that determine what workloads this GPU handles.
12.0 TB/s
Memory Bandwidth
peak throughput
1200W
TDP
thermal design power
Blackwell
Architecture
NVIDIA GPU architecture
Spot pricing
B300: live hourly rates.
Every provider offering this GPU on the spot market, sorted cheapest first.
Prices in USD per GPU-hour · spot instances · sorted cheapest first
Recommended models
AI models that run well on B300.
Tested model-GPU pairings with notes on why each is a good fit.
Llama 4 Maverick
400B MoE model fits comfortably in FP16 with room for batched KV caches. No tensor parallelism needed — single-card deployment eliminates inter-GPU latency.
View model pricing →
DeepSeek V3.2
671B MoE at FP8 precision sits well within the 288 GB envelope. The 12 TB/s bandwidth keeps expert dispatch under 2ms even at high batch sizes.
View model pricing →
Qwen 3 235B
Full FP16 deployment on one card with 50+ GB headroom for KV cache, context batching, and LoRA adapters running concurrently.
View model pricing →
Use cases
What the B300 is built for.
-
Single-GPU deployment of 400B+ parameter models
The B300 eliminates the biggest operational headache of frontier model deployment: multi-GPU coordination. Models like Llama 4 Maverick (400B MoE) that previously required 4-8 GPU clusters now run on one card at $8.19/hr. Fewer GPUs means simpler infrastructure, no tensor parallelism overhead, and predictable latency.
-
Research experimentation with full-precision frontier models
Academic and research teams running ablation studies on 200B+ models need fast iteration cycles. The B300 lets you load the full model in FP16, modify architecture components, and re-run experiments without juggling multi-node orchestration — cutting experiment turnaround from days to hours.
-
Massive-context serving for retrieval-augmented generation
RAG pipelines processing 500K+ token contexts generate KV caches that exceed 100 GB alone. The B300's 288 GB VRAM holds both the model weights and enormous KV caches simultaneously, enabling true long-context retrieval without the latency penalty of offloading to CPU memory.
FAQ
Common questions.
B300 vs B200 — when does the extra memory matter?
The B300 adds 108 GB over the B200 (288 vs 180 GB) and costs $8.19/hr vs $6.48/hr — a 26% premium for 60% more VRAM. Choose the B300 when your model exceeds 180 GB in target precision, when you need massive KV cache headroom for long-context serving, or when you want to avoid multi-GPU setups entirely for 200-400B models. For models under 180 GB, the B200 is the better value.
Is the B300 available in multi-GPU configurations?
Yes. B300 NVLink clusters can scale to 8-GPU nodes with 2.3 TB of aggregate VRAM and inter-GPU bandwidth exceeding 1.8 TB/s. These configurations are designed for pre-training runs on trillion-parameter models. However, the B300's primary advantage is that it pushes the single-GPU frontier so far that most models don't need multi-GPU at all.
What models are too large even for a B300?
Dense models above ~400B parameters in FP16 (requiring 800+ GB) still need multi-GPU setups. Similarly, training runs on 200B+ models with Adam optimizer states (which triple the memory footprint) exceed 288 GB. For these workloads, a B300 cluster is the answer, not a single card.
How does spot pricing work for such an expensive GPU?
B300 spot instances follow the same model as cheaper GPUs — you pay the hourly rate with per-second billing, accept the risk of preemption, and save 60-80% versus on-demand pricing. The interruption probability is typically low (under 5%) because B300 supply is still ramping and demand is concentrated in reserved contracts. Spot is a genuine bargain at this tier.
Need 8, 16 or 32 GPUs?
Pick the GPU, the count and how long. See every live offer, priced per GPU hour and for the whole job.
Do you run GPUs or an inference API?
Send a test key and we benchmark your endpoint. You keep the numbers either way.
Rent a B300. Right now.
Spot pricing, per-second billing, no commitment.
Browse the live marketplace, pick your GPU, deploy in one click. Credits from $10.