GPUBeacon
Language
NVIDIA

Cloud GPU · verified 2026-09-05

NVIDIA B200

Blackwell180 GB

Blackwell for people who will actually use FP4 and a 1.8 TB/s fabric. 180–192 GB of HBM3e, ~1 kW, often a waitlist. Not a default H100 replacement.

Author Valentin FourierLast modified

VRAM180 GB
From /GPU/hr$3.75
ArchitectureBlackwell
VendorNVIDIA

In brief

What to know before you rent

Catalog listings

B200 Pricing and Availability

Published on-demand rates first. Spot and waitlists are labeled. Confirm the live rate before you provision.

ProvidersRegionBilling typeInterconnect/GPU/hr
Packet.aiIn stockUS-WestOn-demandNVLink$3.75Open provider
KoyebLimitedEU-WestOn-demandNVLink$4.10Open provider
VerdaIn stockEU-NorthOn-demandNVLink$4.25Open provider
RunPodLimitedUS-EastOn-demandNVLink$4.40Open provider
LambdaWaitlistUS-WestOn-demandNVLink$4.90Open provider
CoreWeaveWaitlistUS-EastOn-demandInfiniBand NDR$5.50Open provider

Our reading

Who this GPU is for

I put B200 on a shortlist when the run is large enough that Hopper memory and NVLink 4 are the bottleneck, and when the row is actually in stock. I do not put it on a demo that needs to boot tomorrow. Quota and liquid cooling are not footnotes.

Rent it if Frontier-ish training, long-context serving that wants FP4, teams with a CUDA 12.8+ stack and a budget that survives 1 kW hours.

Skip it if Anyone who needed an H100 this week. Fine-tunes that fit 80 GB. Shops without a Blackwell software path.

How we compare hours

Datasheet

Technical specifications

NVIDIA B200 · Per GPU · SXM

Compute · dense

FP4
9,000 TFLOPS~18,000 with sparsity, SKU-dependent
FP8
4,500 TFLOPS~9,000 with sparsity
FP16 / BF16
2,250 TFLOPS~4,500 with sparsity
FP32
75 TFLOPS
FP64
37 TFLOPS
Precision support
FP4, FP6, FP8, FP16, BF16, TF32, FP32, FP64, INT8

Memory

Capacity
180–192 GB HBM3eCloud listings often 180 GB; SXM typically 192 GB
Bandwidth
~8,000 GB/s
ECC
Yes

Silicon

Architecture
Blackwell
Process
TSMC 4NP
Package
Dual-die GPU

Fabric and host

GPU interconnect
NVLink 5 · 1.8 TB/s
Host interface
PCIe 5.0 x16

Power

Board power
1,000 WSXM class, liquid cooling in dense racks

Platform

Partitioning
MIG, up to 7 instances

Blackwell numbers move with SKU and NVIDIA’s sparse-vs-dense footnotes. Treat this table as the SXM-class envelope clouds sell, then confirm the live row. GB200 NVL72 is a rack, not a GPU hour.

Trade-offs

Pros and cons of B200

What this GPU does well, and when the hour is the wrong buy.

Strengths

  • HBM3e at ~8 TB/s is a different memory regime than Hopper
  • NVLink 5 at 1.8 TB/s is why 8-GPU jobs stop looking like a network problem
  • Native FP4 is the inference lever H100 does not have
  • When it is in stock, it is the SKU the rest of the market is waiting on

Limits

  • Stock is the product. A datasheet is not a reservation
  • 1,000 W and liquid cooling. Cheap hosts will thermal-throttle you into a worse hour
  • FP4 only helps if the runtime path is real
  • Price vs H100/H200 has to clear a high bar. FLOPs/watt on a slide is not that bar

Stock is the spec that matters

B200 marketing is ahead of B200 inventory. I will not plan a launch on a press deck. If the catalog says waitlist, it is not on-demand. It is a newsletter.

When a specialist cloud actually lists an hour, compare it to H200 on memory and to H100 on cost. Blackwell has to win a real job, not a keynote.

FP4 is not a free 2×

The slide that shows 9 PFLOPS dense FP4 assumes a stack that quantizes and still hits quality. If you are in BF16 because the eval dropped, you bought a B200 to run Hopper math at Blackwell rent.

I only push people here when they already know the precision they can keep. Curiosity is cheaper on an H100.

How I actually rent it

  1. 01

    Confirm it exists

    In-stock on-demand. Not a quote, not a 2026 delivery slot.

  2. 02

    Confirm the bytes

    180 vs 192 GB, NVLink 5, GPU count. Dual-die does not save a bad node.

  3. 03

    Have an H200 fallback

    If Blackwell slips, Hopper with 141 GB is the adult plan.

FAQ

B200 questions

Who has the cheapest B200?+

As of September 5, 2026, the cheapest verified in-stock on-demand B200 tracked by GPUBeacon is $3.75 per GPU per hour from Packet.ai.

How much does the B200 cost per month?+

As of September 5, 2026, at 720 hours per month, one B200 costs an estimated $3,114 at the median on-demand price in this sample. Cheapest verified in stock: $2,700 per month on-demand.

Where can I rent an B200?+

GPUBeacon currently tracks B200 configs from 6 providers. The cheapest verified in-stock on-demand configs come from Packet.ai and Verda. See the full price comparison above for every provider and config.

Are B200 prices going up or down?+

GPUBeacon publishes a catalog sample, not a 90-day price index. As of September 5, 2026, the median on-demand B200 in this sample is $4.33 per GPU per hour. The cheapest verified in-stock hour is $3.75.

Why choose the B200?+

Frontier-ish training, long-context serving that wants FP4, teams with a CUDA 12.8+ stack and a budget that survives 1 kW hours.

When is the B200 not a good fit?+

Anyone who needed an H100 this week. Fine-tunes that fit 80 GB. Shops without a Blackwell software path.

Is B200 180 GB or 192 GB?+

NVIDIA SXM is typically 192 GB HBM3e. Cloud catalogs on this site often list 180 GB. That gap is a SKU, not a rounding error. Read the row.

Should I skip H100 for B200?+

Only if you need FP4, the extra HBM, or NVLink 5, and you can buy it. Otherwise H100 is still the hour that exists.

Keep comparing

Alternatives to NVIDIA B200

Same job, different envelope. Compare VRAM, bandwidth and the hour before you lock a badge.

NVIDIA B300

288 GB HBM3e · similar bandwidth to the B200

From$7.50/ GPU / hr

Read the verdict

NVIDIA H200

141 GB HBM3e · 40% less bandwidth than the B200

From$3.00/ GPU / hr

Compare

NVIDIA H100

80 GB HBM3 · 58% less bandwidth than the B200

From$1.73/ GPU / hr

Read the verdict

AMD MI300X

192 GB HBM3 · 34% less bandwidth than the B200

From$1.99/ GPU / hr

Compare

Browse all GPUs

Sources

Where these numbers come from

TFLOPS and TDP are from the vendor datasheet. Hourly prices are from the GPUBeacon catalog sample. Open the originals, then confirm the live SKU with the provider before you rent.