B200 GPU pricing spans $3.75 to $27.04 per hour for identical silicon. Here is what drives the 7x spread across clouds, and how to pick a tier without overpaying.
NVIDIA B200 cloud pricing on packet.ai starts at $3.75/GPU-hr on Dynamic capacity, against a 2026 market that spans $3.75 to $27.04 per GPU-hour for the same silicon.
Key takeaways
If you searched for B200 GPU pricing in 2026, you have probably noticed the numbers make no sense at first glance. The same NVIDIA B200 rents for under $4 an hour on one cloud and over $27 on another. This guide breaks down what a B200 actually costs across providers, what the hardware gives you, and how packet.ai's Dedicated, Dynamic, and bare metal tiers compare, so you can match the right price to your workload instead of overpaying for someone else's SLA.
packet.ai prices the B200 across two capacity tiers and two billing terms. Dynamic capacity is $3.75/GPU-hr whether you bill hourly or monthly. Dedicated capacity is $5.90/GPU-hr on hourly billing and drops to $5.60/GPU-hr on monthly terms. You can rent a B200 cluster on packet.ai and be running a vLLM workload the same day.
B200 Dynamic capacity on packet.ai starts at $3.75/GPU-hr, 46% below Lambda's $6.99/hr single-GPU rate and 74% below AWS p6 at roughly $14.24/hr per GPU.
At 730 hours a month, that pricing translates to about $2,728 per GPU per month on Dynamic, $4,307 on monthly Dedicated, and $4,248 on hourly Dedicated. The same always-on GPU on AWS p6 runs to roughly $10,395 a month before egress and storage.
The B200 rental market in mid-2026 is wide open. Pricing trackers put the spread between the cheapest and most expensive public B200 listings at roughly 7x, from $3.75/hr at the low end to $27.04/hr on Azure's managed instances. Hyperscaler rates bundle enterprise SLAs, managed networking, and support. Neoclouds like packet.ai, Lambda, and Nebius price closer to the silicon.
As of mid-2026, published B200 cloud rates span $3.75 to $27.04 per GPU-hour, a 7x spread for the same hardware.
Competitor rates from published pricing pages and market trackers, June to July 2026. Rates change frequently, verify with each provider.
Prices have not been standing still either. Silicon Data's SDB200RT rental index rose 24.4% between January and late March 2026, driven by HBM3e cost increases and a run of frontier model launches soaking up spare capacity. The direction of travel matters if you are deciding between locking a monthly rate now or staying on hourly billing.
The NVIDIA B200 GPU is the flagship of the Blackwell generation, a dual-die design with 208 billion transistors. For inference workloads the headline number is memory: 192 GB of HBM3e per GPU. That is enough to hold a 70B-parameter model in FP16 on a single card, which removes tensor parallelism overhead entirely for that model class.
Against the previous generation, the practical jump is in inference throughput. NVIDIA cites up to 4x the inference performance of the H100, driven by higher memory bandwidth and FP4 Tensor Cores that halve bytes-per-parameter versus FP8. If your workload fits comfortably in 141 GB, the H200 SXM on packet.ai at $2.49/hr remains the better cost-per-token option for many serving setups. The B200 earns its premium when you need the extra memory headroom, longer context at high batch sizes, or FP4 serving.
The $2.15/hr gap between Dynamic and hourly Dedicated is not a discount for worse hardware. It is a scheduling model choice. Dynamic capacity gives you the same B200 silicon at $3.75/GPU-hr on capacity that packet.ai schedules dynamically. Dedicated capacity at $5.90/GPU-hr hourly, or $5.60/GPU-hr monthly, pins the hardware to you.
✓ Dynamic fits
✗ Go Dedicated instead
A useful rule of thumb: if an interruption costs you more than the roughly $1,580 per GPU per month you save on Dynamic (at 730 hours), pay for Dedicated. If your job checkpoints and resumes cleanly, Dynamic is the cheapest B200 hour you will find on packet.ai. If you only need tokens rather than raw GPUs, the packet.ai Token Factory skips the tier decision entirely.
For teams that want the full HGX B200 node with no virtualization layer, packet.ai offers B200 bare metal servers on monthly terms. Bare metal gives you the whole box: all eight GPUs, NVLink fabric, host CPUs, and NVMe, with direct hardware access for custom kernels, NCCL tuning, and container stacks you control end to end.
packet.ai B200 bare metal is priced at $5.60/GPU-hr on monthly terms during the current Independence Day offer.
⚡ Limited offer
B200 bare metal is available monthly-only at promotional pricing. Full node configs and terms are on the packet.ai B200 bare metal Independence Day offer page.
Bare metal makes sense when virtualization overhead, noisy neighbors, or driver-level control are actual constraints for you, typically sustained pretraining, large-scale RLHF, or inference fleets where you have already squeezed the software stack. If you would rather have orchestration handled, browse available clusters on the standard tiers instead.
Where you should rent a B200 in 2026 comes down to three questions. First, do you need one GPU or a full node? Lambda and packet.ai both offer single-GPU B200 access, while AWS p6 and GCP's A4 series only sell 8-GPU nodes. Second, does your workload tolerate interruption? If yes, dynamic and spot tiers cut your rate substantially. Third, do you actually need hyperscaler SLAs and integrations, or are you paying a 3x to 4x premium for a logo?
For most AI teams the practical answer in 2026 is a neocloud. The silicon is identical, provisioning is faster, and the rate difference compounds: a single always-on B200 at $3.75/hr on packet.ai versus $14.24/hr on AWS is a difference of roughly $7,550 per GPU per month. Hyperscalers earn their premium when you are deeply embedded in their data and networking stack and egress would dominate your bill anyway.
B200 supply also loosened through 2026 as TSMC 4NP yields improved and more providers stocked inventory, which is why on-demand access now exists at all. It was waitlist-only at most providers a year ago. If B200 economics still do not fit, the same decision framework applies one rung down: deploy H200 on packet.ai from $2.49/hr, or packet.ai H100 pricing starts at $2.50/hr for workloads that fit in 80 GB.
Last reviewed: July 9, 2026. Pricing verified against provider listings at time of writing. Ready to run Blackwell? Browse available clusters on packet.ai →
Same models. Same API. Fraction of the cost. Start free — no credit card required.
Start Building →
