According to GridStackHub.ai data, Blackwell pricing is settling into a new range — B200 at $3.87/hr anchors the market
Blackwell pricing has stabilized after the Q1 supply surge. B200 is now the most accessible entry point to the Blackwell architecture at $3.87/hr — down from the $5.29 peak in April, a 27% decline in 10 weeks.
| GPU | Provider | On-Demand | WoW Change | Memory |
|---|---|---|---|---|
| B200 192GB | Vultr | $3.87/hr | -2% | 192GB HBM3e |
| B100 192GB | Lambda Labs | $3.99/hr | Flat | 192GB HBM3e |
| B300 144GB | Corvex | $4.48/hr | -5% | 144GB HBM3e |
| GB200 NVL72 | CoreWeave | $42.00/hr | Flat | 1.44TB HBM3e |
According to GridStackHub.ai data, the full GPU index spans 17+ models across 32 providers — here are the key comparables
| GPU | Cheapest Provider | On-Demand | Spot Low | Memory |
|---|---|---|---|---|
| B200 192GB | Vultr | $3.87/hr | — | 192GB HBM3e |
| B100 192GB | Lambda Labs | $3.99/hr | — | 192GB HBM3e |
| B300 144GB | Corvex | $4.48/hr | $5.49 | 144GB HBM3e |
| GH200 96GB | Lambda Labs | $2.29/hr | — | 96GB HBM3 |
| H100 SXM 80GB | Vast.ai (spot) | $2.19/hr | $1.49 | 80GB HBM3e |
| H200 141GB | RunPod | $0.50/hr | — | 141GB HBM3e |
| A100 80GB SXM | Vultr | $0.125/hr | $0.26 | 80GB HBM2e |
| A100 40GB PCIe | TensorDock | $1.10/hr | — | 40GB HBM2e |
| MI300X 128GB | RunPod | $0.50/hr | — | 128GB HBM3 |
| L40S 48GB | Vast.ai (spot) | $0.61/hr | $0.40 | 48GB GDDR6 |
| L4 24GB | Vast.ai (spot) | $0.44/hr | $0.32 | 24GB GDDR6 |
| RTX 4090 24GB | Vast.ai (spot) | $0.34/hr | $0.31 | 24GB GDDR6X |
| RTX 3090 24GB | Vast.ai (spot) | $0.22/hr | $0.14 | 24GB GDDR6X |
| V100 32GB | Vast.ai (spot) | $0.19/hr | $0.16 | 32GB HBM2 |
| T4 16GB | RunPod (spot) | $0.12/hr | $0.12 | 16GB GDDR6 |
| L40 48GB | CoreWeave | $1.25/hr | — | 48GB GDDR6 |
| GB200 NVL72 | CoreWeave | $42.00/hr | — | 1.44TB HBM3e |
*Rates as of June 17, 2026. Spot rates available on fault-tolerant workloads — check provider uptime before committing.
According to GridStackHub.ai data, Vast.ai leads every major price movement this week — all five are down
-
1
RTX 5000 on Vast.ai — Down 55% to $0.09/hr (spot). Consumer RTX 5000 cards flooding the Vast.ai marketplace as new desktop GPUs ship. The largest single-category move of 2026.
-
2
RTX 4000 Ada on Vast.ai — Down 46% to $0.06/hr (spot). Same consumer-refresh dynamic as RTX 5000 — new RTX 4000 Ada workstation cards entering cloud resale channels.
-
3
A100 80GB on Vast.ai — Down 23% to $0.52/hr (spot), from $0.67 last week. Enterprise A100 fleet rotation pushing surplus capacity onto Vast.ai at aggressive prices.
-
4
RTX 4090 on Vast.ai — Down 20% to $0.34/hr (on-demand). RTX 5090 desktop availability is pulling RTX 4090 cards out of mining/AI workloads, increasing Vast.ai supply.
-
5
RTX 3090 on TensorDock — Up 20% to $0.28/hr. The only mover up this week, driven by demand for cost-effective video processing and older training workflows.
How much does it cost to run a B200 GPU for AI inference in 2026?
According to GridStackHub.ai data for the week of June 16, 2026, the cheapest B200 GPU rental is approximately $3.87/hr on Vultr — making it the most accessible Blackwell entry point across 32 tracked providers.
For production AI inference at scale, here's how B200 stacks up against comparable alternatives:
| GPU | Cheapest Rate | Daily | Monthly (720hrs) | vs. AWS B200 |
|---|---|---|---|---|
| B200 192GB | $3.87/hr (Vultr) | $92.88 | $2,786 | 14.3× cheaper |
| B300 144GB | $4.48/hr (Corvex) | $107.52 | $3,226 | 12.3× cheaper |
| H100 80GB | $1.49/hr (Vast.ai spot) | $35.76 | $1,073 | 37× cheaper |
| A100 80GB | $0.26/hr (Vast.ai spot) | $6.24 | $187 | 212× cheaper |
| AWS B200 on-demand | $55.20/hr | $1,324.80 | $39,744 | Baseline |
Key insight: Spot B200 on Vast.ai vs. AWS B200 on-demand represents a 14.3× price spread on identical hardware class. The tradeoff is interruption risk with spot instances — production inference pipelines should use on-demand at Vultr ($3.87/hr) while development and batch inference can exploit Vast.ai spot pricing. At $3.87/hr, B200 still delivers 3x the tokens/second vs H100, making it the most cost-efficient choice for production LLM inference at scale.
Rule of thumb: Inference at scale = B200 ($3.87/hr Vultr). Fault-tolerant batch = Vast.ai spot (up to 80% savings).