New NVIDIA B300 · 288 GB HBM3e servers — from $4,019/mo per GPU

See the B300
Pay in crypto Deploy a GPUDeploy

10 GPU models · 1 to 8 GPUs per server

Dedicated GPU servers,
priced by the month.

From the GeForce RTX 5080 to the NVIDIA B300 and the AMD MI355X. Every server is dedicated to you, with root access and one flat monthly price about 30% under the market median.

Configure a server Size my model
  • 10GPU models
  • 16–288 GBmemory per GPU
  • From $199per month
  • −30%vs market median

Three families, ten GPUs.

Filter by family. Each card opens the full specifications, every server size and the market check for that GPU.

  • NVIDIA · Blackwell Ultra

    B300

    New

    288 GB HBM3e

    Bandwidth
    8 TB/s
    Link
    NVLink 5
    Server
    1–8 GPUs

    Frontier training, 400B+ inference, long context

    from $4,019/mo

    $5,745 market −30%

  • NVIDIA · Blackwell

    B200

    180 GB HBM3e

    Bandwidth
    8 TB/s
    Link
    NVLink 5
    Server
    1–8 GPUs

    Large-scale training and high-throughput inference

    from $3,439/mo

    $4,920 market −30%

  • AMD · CDNA 4

    MI355X

    Best value

    288 GB HBM3E

    Bandwidth
    8 TB/s
    Link
    Infinity Fabric
    Server
    1–8 GPUs

    Memory-hungry inference, ROCm training, FP64 science

    from $1,409/mo

    $2,022 market −30%

  • NVIDIA · Hopper

    H200

    141 GB HBM3e

    Bandwidth
    4.8 TB/s
    Link
    NVLink 4
    Server
    1–8 GPUs

    70B-class models on one GPU, fine-tuning, serving

    from $2,279/mo

    $3,259 market −30%

  • NVIDIA · Hopper

    H100

    80 GB HBM3

    Bandwidth
    3.35 TB/s
    Link
    NVLink 4
    Server
    1–8 GPUs

    The proven workhorse for training and inference

    from $1,779/mo

    $2,548 market −30%

  • NVIDIA · Blackwell

    RTX PRO 6000

    96 GB GDDR7 ECC

    Bandwidth
    1.6 TB/s
    Link
    PCIe Gen5 x16
    Server
    1–8 GPUs

    96 GB per card for inference, fine-tuning, rendering

    from $959/mo

    $1,380 market −31%

  • NVIDIA · Ada Lovelace

    L40S

    48 GB GDDR6 ECC

    Bandwidth
    864 GB/s
    Link
    PCIe Gen4 x16
    Server
    1–8 GPUs

    Cost-efficient inference, diffusion, video pipelines

    from $789/mo

    $1,139 market −31%

  • NVIDIA · Blackwell

    RTX 5090

    Hot

    32 GB GDDR7

    Bandwidth
    1.8 TB/s
    Link
    PCIe Gen5 x16
    Server
    1–4 GPUs

    Image & video generation, 3D rendering, dev boxes

    from $329/mo

    $481 market −32%

  • NVIDIA · Ada Lovelace

    RTX 4090

    24 GB GDDR6X

    Bandwidth
    1 TB/s
    Link
    PCIe Gen4 x16
    Server
    1–4 GPUs

    Budget inference, Stable Diffusion, rendering

    from $269/mo

    $387 market −30%

  • NVIDIA · Blackwell

    RTX 5080

    16 GB GDDR7

    Bandwidth
    960 GB/s
    Link
    PCIe Gen5 x16
    Server
    1–4 GPUs

    Entry GPU for small models, CI, development, video encoding

    from $199/mo

    $299 market −33%

Every GPU, every key number.

Sort by price, memory, bandwidth or price per GB of GPU memory. Prices are per GPU per month; a server of n GPUs costs n times as much.

Specifications and monthly price per GPU for every GPU model. Column headers can sort the table.
Interconnect Precisions GPUs / server Market median
NVIDIA B300Blackwell Ultra 288 GBHBM3e 8 TB/s NVLink 5 · 1.8 TB/s FP4 · FP8 · BF16 1–8 $4,019 $13.95 $5,745 −30%
NVIDIA B200Blackwell 180 GBHBM3e 8 TB/s NVLink 5 · 1.8 TB/s FP4 · FP8 · BF16 1–8 $3,439 $19.11 $4,920 −30%
AMD MI355XCDNA 4 288 GBHBM3E 8 TB/s Infinity Fabric · 1.07 TB/s FP4 · FP6 · FP8 1–8 $1,409 $4.89 $2,022 −30%
NVIDIA H200Hopper 141 GBHBM3e 4.8 TB/s NVLink 4 · 900 GB/s FP8 · BF16 · FP64 1–8 $2,279 $16.16 $3,259 −30%
NVIDIA H100Hopper 80 GBHBM3 3.35 TB/s NVLink 4 · 900 GB/s FP8 · BF16 · FP64 1–8 $1,779 $22.24 $2,548 −30%
NVIDIA RTX PRO 6000Blackwell 96 GBGDDR7 ECC 1.6 TB/s PCIe Gen5 x16 FP4 · FP8 · BF16 1–8 $959 $9.99 $1,380 −31%
NVIDIA L40SAda Lovelace 48 GBGDDR6 ECC 864 GB/s PCIe Gen4 x16 FP8 · BF16 1–8 $789 $16.44 $1,139 −31%
NVIDIA RTX 5090Blackwell 32 GBGDDR7 1.8 TB/s PCIe Gen5 x16 FP4 · FP8 · BF16 1–4 $329 $10.28 $481 −32%
NVIDIA RTX 4090Ada Lovelace 24 GBGDDR6X 1 TB/s PCIe Gen4 x16 FP8 · BF16 1–4 $269 $11.21 $387 −30%
NVIDIA RTX 5080Blackwell 16 GBGDDR7 960 GB/s PCIe Gen5 x16 FP4 · FP8 · BF16 1–4 $199 $12.44 $299 −33%

$ / GB = monthly price per GB of GPU memory. Market median = median on-demand list price per GPU from 40 GPU clouds, 23 Sep 2026 × 730 hours. Full market comparison

Memory first, then speed.

Three questions settle most choices. The matrix shows where each GPU fits best.

  1. Does it fit?

    Weights, KV cache and activations must fit in GPU memory. 16 GB on the RTX 5080, up to 288 GB per GPU on the B300 and MI355X, and 2.3 TB in an 8-GPU server. Size a model

  2. How fast must it run?

    Token and image generation are mostly limited by memory bandwidth: up to 8 TB/s on HBM flagships, 1.6 to 1.8 TB/s on the RTX PRO 6000 and RTX 5090, under 1 TB/s on the L40S and RTX 5080.

  3. Will the GPUs work together?

    Sharded training exchanges data between GPUs at every step: NVLink (900 GB/s–1.8 TB/s) and Infinity Fabric are built for it. PCIe cards suit independent jobs and inference.

Where each GPU fits best, by workload
GPU LLM trainingFine-tuningInference & servingImage & video3D rendering & VFXResearch & HPC
B300 Recommended Good fit Good fit — — —
B200 Recommended Good fit Good fit — — —
MI355X Recommended — Recommended — — Recommended
H200 — Recommended Recommended — — Recommended
H100 Good fit Recommended Good fit — — Recommended
RTX PRO 6000 — Recommended Good fit Recommended Recommended —
L40S — — Recommended Recommended Good fit —
RTX 5090 — — Good fit Recommended Recommended —
RTX 4090 — — Good fit Good fit Recommended —
RTX 5080 — — Good fit Good fit — —

RecommendedGood fit

What every server comes with.

  • Dedicated GPUsNo time-slicing and no other tenant on your cards.
  • Root SSH accessInstall your drivers, CUDA or ROCm, containers and frameworks.
  • DDR5 RAM and local NVMevCPU, RAM and NVMe scale with the number of GPUs.
  • Same price per GPUAt every server size, from 1 to 8 GPUs.
  • Month to monthPrepaid monthly, renewed when you choose. No long-term contract.
  • Pay in cryptoBTC, ETH, USDT, USDC or LTC.

Choosing hardware.

Billing and payments are covered in the full FAQ.

Which GPU should I choose?

Start from memory: the model, its context and the batch must fit in GPU memory. Then look at memory bandwidth, which sets how fast tokens or images are generated, and at the interconnect if you train on several GPUs. The sizing helper lists every configuration that fits a given model, cheapest first.

What is the difference between HBM GPUs and PCIe cards?

The B300, B200, MI355X, H200 and H100 use stacked HBM memory (3.35 to 8 TB/s) and a high-speed link between the GPUs of a server: NVLink or Infinity Fabric. The RTX PRO 6000, L40S and GeForce cards use GDDR memory and talk to each other over PCIe. HBM platforms are faster for training and very large models; PCIe cards cost much less for serving, generation and rendering.

Are GeForce cards suitable for production?

They are fast and cost-efficient for image and video generation, rendering, development and smaller models. They have no ECC memory and no NVLink, and servers take up to four cards. For long-running services that need ECC memory, choose the RTX PRO 6000 or the L40S.

Can I switch to another GPU later?

Yes: order the new configuration, move your data, and let the old server’s term end. Each server keeps its own monthly term; moving data between servers is your responsibility.

Why is there no A100 or older GPU?

The range focuses on current generations: Blackwell, Blackwell Ultra, Hopper, CDNA 4 and the latest GeForce and RTX PRO cards. The L40S and RTX 4090 (Ada Lovelace) stay in the range as budget options.

Found your GPU? Build the server.

Pick 1 to 8 GPUs and see the exact monthly price, next to the market median.

Configure a server Price list