The NVIDIA B300 is one of the most powerful data center GPUs you can rent in 2026, built on the Blackwell Ultra architecture with 288 GB of memory per GPU. Demand is high and supply is still ramping, which keeps pricing volatile and access limited.
The B300 ships only inside HGX and DGX systems bought through NVIDIA's partner network, so buying is rarely an option. The practical question is what it costs to rent, and whether you need this much GPU at all.
This guide covers NVIDIA B300 cloud pricing for September 2026, including on-demand rates across providers. It also breaks down specs, compares the chip against the B200, H100, and H200, and shows which LLMs fit on a single card.
Key Takeaways
- B300 premiums are the steepest. On-demand rates start near $7.10/GPU-hr on neoclouds and reach $17.80 on hyperscalers.
- Its value is memory. 288 GB per card lets a single B300 hold very large models without sharding across multiple GPUs.
- Choose B300 only when you must. Very large model training or serving a model that overflows 180 GB.
- Thunder Compute roadmap. We don't offer B300 nodes. You can launch an H100 80GB at $3.20/hr, or 2x H100 for $6.40, with one-click VS Code, per-minute billing, and persistent storage.
The pricing information in this guide is reviewed weekly.
NVIDIA B300 Pricing
The B300 isn't sold as a standalone card, and NVIDIA doesn't publish list prices, so cloud rental is how the chip actually gets used. On-demand rates currently start near $7.10/GPU-hr on neoclouds and climb to $17.80 on hyperscalers.
| Provider | GPU / Instance | On-Demand $/GPU-hr | Notes |
|---|---|---|---|
| Modal | B300 | $7.10 | |
| Nebius | B300 | $7.85 | |
| Runpod | B300 | $7.89 | |
| Oracle Cloud | OCI B300 | $15.00 | |
| AWS | p6-b300.48xlarge | $17.80 | Normalized from an 8-GPU node. |
| Last reviewed on September 11, 2026. | |||
Methodology: Why You Can Trust These Numbers
- On-demand only. The table excludes capacity reservations, reserved instances, and spot instances.
- Same silicon. Every row is a 288 GB NVIDIA B300 (SXM6).
- Public price lists only. Figures come straight from each provider's pricing page or API in September 2026.
- US regions, USD. Node prices are normalized to a per-GPU figure where a provider only sells full 8-GPU nodes.
B300 Cost Benchmark
The 10-hour column shows what a single B300 costs for a typical training or inference session, which makes the gap between neoclouds and hyperscalers concrete.
| Provider | On-Demand $/GPU-hr | 10-Hour Cost |
|---|---|---|
| Modal | $7.10 | $71.00 |
| Nebius | $7.85 | $78.50 |
| Runpod | $7.89 | $78.90 |
| Oracle Cloud | $15.00 | $150.00 |
| AWS | $17.80 | $178.00 |
Bottom line: 10 hours on a hyperscaler B300 costs far more than 22 hours of an H100 80GB on Thunder Compute at $3.20/hr. Unless your model truly needs 288 GB per card, available H100 capacity may be the more practical path.
B300 Hardware Price: Buy vs Rent
After comparing cloud rates, it helps to understand what a B300 system actually costs to own. NVIDIA does not publish list prices, and the B300 is sold into integrated systems rather than at retail. The figures below reflect reseller quotes and market estimates tracked through 2026.
| Configuration | Price Range1 | Notes |
|---|---|---|
| Single B300 (SXM), effective street price | $50,000-$60,000 | Single-unit reseller quotes. |
| 8-GPU HGX B300 system | $550,000-$650,0001 | Integrated platform with NVLink fabric, NICs, and CPUs. |
| 1 An integrated 8-GPU system is roughly eight GPUs plus $120,000-$180,000 of CPU, networking, chassis, and integration. | ||
Cloud rental is the right and often only choice for most teams: no capital tied up, no cooling or power engineering, no depreciation, and the flexibility to switch hardware as workloads change.
NVIDIA B300 Specifications
The B300 is the Blackwell Ultra refresh of the B200, adding memory and power on the same SXM-only, liquid-cooled footprint. Its headline number is 288 GB of HBM3e per GPU, enough to hold very large models on a single card.
| Spec | NVIDIA B300 |
|---|---|
| Architecture | Blackwell Ultra |
| Tensor Core Generation | 5th Gen (native FP4) |
| VRAM | 288 GB HBM3e |
| Memory Bandwidth | Up to 8,000 GB/s1 |
| NVLink | 5th Gen, 1.8 TB/s per GPU |
| Form Factor | SXM6 |
| TDP | 1,400 W |
| Cooling | Liquid cooling required |
| PCIe Version | None; datacenter systems only |
| 1 Peak theoretical figure. Real-world bandwidth varies with hardware and software bottlenecks. | |
B300 vs B200: Which Blackwell GPU Should You Rent?
The B300 leads on memory, but the gap only matters for specific workloads. It carries 288 GB versus the B200's 180 GB, draws 1,400 W versus 1,000 W, and rents for about a dollar more per GPU-hour at every provider that lists both.
Choose the B300 when a single model needs the extra memory. If you are serving a 130B-parameter model that would otherwise span multiple cards, one B300 can simplify the deployment and lower cost per query. For anything that fits comfortably in 180 GB, the B200 is the better value.
See current rates and specs in our NVIDIA B200 pricing guide.
B300 vs H100 and H200
The B300's advantage is memory capacity and native FP4, not a free performance win across every workload. For most fine-tuning, inference, and training that fits inside 141 GB, the older Hopper cards provide sufficient compute at a fraction of the price.
| Feature | NVIDIA H100 | NVIDIA H200 | NVIDIA B300 |
|---|---|---|---|
| Architecture | Hopper | Hopper | Blackwell Ultra |
| GPU Memory | 80 GB HBM3 | 141 GB HBM3e | 288 GB HBM3e |
| Memory Bandwidth | 3.35 TB/s | 4.8 TB/s | 8 TB/s |
| FP4 Support | No | No | Yes (5th-gen Tensor Cores) |
| On-Demand Cloud Price | $3.20 - $10.98/hr | $3.99 - $10.85/hr | $7.10 - $17.80/hr |
Compare current H200 rates in our NVIDIA H200 pricing guide.
What LLMs Fit on a B300?
The B300's 288 GB of VRAM lets a single card hold models that overflow a B200 or H200.
| Model | Parameters | Precision | VRAM (weights only)1 | Single B300 (288 GB)? |
|---|---|---|---|---|
| Llama 3 70B | 70B | FP16 | ~140 GB | Yes, with context headroom |
| Qwen 3 72B | 72B | FP16 | ~144 GB | Yes |
| Dense 130B-class model | 130B | FP16 | ~260 GB | Yes, single card |
| DeepSeek R1 | 671B | FP8 | ~670 GB | No |
| 1 Weights only. Add 20-30% for KV cache, activations, and framework overhead at moderate context. 2 MoE models must keep all expert weights in VRAM even though only a fraction activate per token. |
||||
The B300's extra 108 GB over the B200 makes a difference for dense models in the 100B-150B range: a model that overflows the B200 in FP16 may fit on a single B300, removing the need for tensor parallelism and the interconnect overhead. An 8-GPU B300 node provides 2,304 GB of pooled VRAM (8 x 288 GB), enough to serve some of the largest open-weight models.
AWS P6-B300 Instances
AWS offers the B300 through its EC2 P6 instance family in an 8-GPU configuration. As is typical for AWS, egress fees apply on top of the hourly rate after the first free 100GB, which makes the effective cost even higher than the listed price.
| SKU | GPUs | Hourly Price | Price Per-GPU |
|---|---|---|---|
| p6-b300.48xlarge | 8 x B300 | $142.42 | $17.80 |
Why NVIDIA B300 Cloud Pricing Is So High
Three factors keep B300 rates high:
- Low supply. Blackwell Ultra parts are in heavy demand for large-scale AI training, and production has not caught up.
- Restricted access. Much of the B300 capacity is committed to large enterprise customers and long-term contracts, so on-demand supply for smaller teams is thin.
- Infrastructure requirements. The B300 only runs in liquid-cooled, high-power systems with fast interconnects, and that specialized hardware raises the cost of every GPU-hour a provider sells.
Last Thoughts on NVIDIA B300 Pricing
The B300 justifies its premium only for very large model training and memory-bound inference with the budget to match. Availability stays constrained while the Blackwell Ultra rollout matures.
Explore the full cloud GPU market landscape in our AI GPU rental market trends analysis.
FAQ
What Is the NVIDIA B300 Cloud Price in 2026?
On-demand B300 pricing currently ranges from about $7.10-$17.80 per GPU-hour across providers, with neoclouds at the low end and hyperscalers at the high end. Spot rates can drop lower but are interruptible.
How Much Does an NVIDIA B300 Cost to Buy?
NVIDIA does not publish list prices. A single B300 runs roughly $50,000-$60,000 through resellers, and a full 8-GPU HGX B300 system runs about $550,000-$650,000.
How Much VRAM Does the NVIDIA B300 Have?
The B300 has 288 GB of HBM3e VRAM per GPU with up to 8,000 GB/s of bandwidth.
What Is the Difference Between the B300 and B200?
The B300 (Blackwell Ultra) carries 288 GB of VRAM versus the B200's 180 GB and delivers higher throughput. Both are SXM-only datacenter modules; the B300 draws more power at 1,400 W.
Can You Buy a Single NVIDIA B300?
Not practically. The B300 is a 1,400 W SXM6 module with no PCIe version, so it only runs in purpose-built liquid-cooled servers. Renting from a cloud provider is the realistic way to access one.
What Is the AWS B300 Price?
AWS prices the B300 at $17.80/GPU-hr on the p6-b300.48xlarge ($142.42/hr for the full 8-GPU node), normalized from an 8-GPU node.