Request access
Private Preview

B300Blackwell Ultra

NVIDIA B300

NVIDIA Blackwell Ultra with 270 GB of HBM3E per GPU and NVLink 5 at 1.8 TB/s per GPU.

On-demand
$10.45 /GPU·hr
Spot
From $1.09 /GPU·hr
Reserved
Request a quote
Request this GPU

HBM3E per GPU

270 GB

NVIDIA reference figure

§01 Specifications Reviewed 2026-10-10
DWG 01HGX B300 8-GPU baseboard Source · NVIDIA HGX reference architecture
NVIDIA HGX B300 8-GPU baseboard, dimetric shop drawing
  1. 1. SXM module · 8 per baseboard, four per side
  2. 2. Heatsink · one per module; this one is drawn lifted
  3. 3. B300 GPU · one per module, under the heatsink
  4. 4. NVLink Switch chip · 2 per baseboard, in the board centre
  5. 5. Baseboard · joins the 8 GPUs in one NVLink domain
  6. 6. ConnectX-8 SuperNIC · 8 on the baseboard, one per GPU

Point at a part or a row. The other one lights.

B300 specifications.

TAB 01B300 specifications Source · NVIDIA Reviewed 2026-10-10
B300 specifications
Parameter Value Scope Source
Memory
Memory 270 GB HBM3E Per GPU [2]
Memory bandwidth 7.7 TB/s Per GPU [2]
Memory on one board 2,160 GB · 8 × 270 GB Per HGX board [2]
Compute, peak · dense / sparse
FP4 Tensor Core 14 / 18 PFLOPS Per GPU [2]
FP8 Tensor Core 4.5 / 9 PFLOPS Per GPU [2]
BF16 Tensor Core Not published Per GPU
Interconnect
NVLink 5 1.8 TB/s Per GPU [3]
NVLink domain 8 GPUs per HGX B300 board · 14.4 TB/s aggregate Per HGX board [1]
Host interface PCIe Gen6 · 256 GB/s Per GPU [2]
Power and form factor
Power Up to 1,100 W, configurable Per GPU [2]
Form factor SXM module on the HGX B300 8-GPU baseboard [2]
Board 8 GPUs per HGX B300 board Per HGX board [1]
  1. [1] NVIDIA HGX platform
  2. [2] NVIDIA Blackwell Ultra datasheet (GB300 NVL72, HGX B300)
  3. [3] NVIDIA Blackwell Architecture Technical Brief v2.1
Reviewed 2026-10-10

Specifications are NVIDIA reference figures.

Best for

  • Reasoning and test-time-scaling inference
  • Large-model training
  • Video generation

Look elsewhere for

  • FP64 HPC
  • INT8 workloads

Notes on the figures

  • Memory: NVIDIA lists 270 GB per GPU (2.1 TB per 8-GPU board).
  • FP4 dense: the datasheet rounds to 14 PFLOPS; the HGX B300 system figure (108 PFLOPS dense for 8 GPUs) gives 13.5.
  • Blackwell Ultra trades away FP64 and INT8 throughput.
§02 Price and terms USD per GPU-hour

On-demand, spot or reserved.

Metered by the second, billed hourly. Reserved rates are quoted per term.

  • On-demand

    $10.45 /GPU·hr

    Per GPU-second
    $0.002903
    Per 8 GPUs, per hour
    $83.60
    Per GPU-month, 730 h
    $7,628.50

    Preview list price. It applies when your account opens.

  • Spot

    From $1.09 /GPU·hr

    Work that checkpoints and can wait. Follow the current spot price, or set a maximum hourly price. If the spot price rises above your max, the workload may stop.

    "From" is the lowest spot price for that GPU. The spot price moves with supply and demand. The spot price always stays below the on-demand rate for the same GPU.

    Spot price moves with supply and demand. Follow it, or set a max price per GPU-hour: you pay the spot price, and nodes stop if it rises above your max. A max price limits what you pay. It does not reserve capacity. To guarantee capacity, reserve it. How spot works

  • Reserved

    Request a quote

    Capacity you need on a date, guaranteed. 1 month, 6 months or 12+ months.

    Reservations

Ways to pay Prepaid credit, pay-as-you-go and Enterprise. Each one covers on-demand, spot and reserved capacity. How billing works

EQ 01B300 terms Source · Fantasti price list Reviewed 2026-10-10
Terms for NVIDIA B300, USD per GPU-hour. The on-demand rate is the preview list price of $10.45. Reserved rates without a published figure are quoted.
TermCommitmentCapacityRate
On-demandNone · Not heldNoneNot held$10.45 /GPU·hr
1 monthTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
6 monthsTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
12+ monthsTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
Terms for NVIDIA B200, USD per GPU-hour. The on-demand rate is the preview list price of $9.35. Reserved rates without a published figure are quoted.
TermCommitmentCapacityRate
On-demandNone · Not heldNoneNot held$9.35 /GPU·hr
1 monthTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
6 monthsTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
12+ monthsTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
Terms for NVIDIA H200 SXM, USD per GPU-hour. The on-demand rate is the preview list price of $5.94. Reserved rates without a published figure are quoted.
TermCommitmentCapacityRate
On-demandNone · Not heldNoneNot held$5.94 /GPU·hr
1 monthTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
6 monthsTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
12+ monthsTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
Terms for NVIDIA H100 SXM5, USD per GPU-hour. The on-demand rate is the preview list price of $5.40. Reserved rates without a published figure are quoted.
TermCommitmentCapacityRate
On-demandNone · Not heldNoneNot held$5.40 /GPU·hr
1 monthTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
6 monthsTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
12+ monthsTake-or-pay · Held for the term, on one fabricTake-or-payHeld for the term, on one fabricRequest a quote
Terms for NVIDIA RTX PRO 6000 Blackwell Server Edition, USD per GPU-hour. The on-demand rate is the preview list price of $2.16. Reserved rates without a published figure are quoted.
TermCommitmentCapacityRate
On-demandNone · Not heldNoneNot held$2.16 /GPU·hr
1 monthTake-or-pay · Held for the termTake-or-payHeld for the termRequest a quote
6 monthsTake-or-pay · Held for the termTake-or-payHeld for the termRequest a quote
12+ monthsTake-or-pay · Held for the termTake-or-payHeld for the termRequest a quote
Terms for NVIDIA L40S, USD per GPU-hour. The on-demand rate is the preview list price of $1.86. Reserved rates without a published figure are quoted.
TermCommitmentCapacityRate
On-demandNone · Not heldNoneNot held$1.86 /GPU·hr
1 monthTake-or-pay · Held for the termTake-or-payHeld for the termRequest a quote
6 monthsTake-or-pay · Held for the termTake-or-payHeld for the termRequest a quote
12+ monthsTake-or-pay · Held for the termTake-or-payHeld for the termRequest a quote

List price for the base shape: 1 GPU, 8 vCPU, 32 GiB. Larger shapes are quoted.

Reserved rates are quoted per term. Reservations are tied to a GPU type, a region and one InfiniBand fabric. Take-or-pay under a signed order form. No early cancellation without penalty (Terms §3).

§03 Node shape Platform configuration

Runs as instances and as cluster nodes.

HGX B300: 8 GPUs and 8 ConnectX-8 SuperNICs on the baseboard, 2,160 GB of HBM3E per node.

B300 machine shapes. Source · Fantasti platform configuration.

B300 runs as a 1-GPU instance (24 vCPU · 346 GiB) or as an 8-GPU node (192 vCPU · 2,768 GiB). 8-GPU nodes join one InfiniBand fabric at 800 Gb/s per GPU. Host network adapter: 400 Gb/s per machine. Connections: 8 GPUs to One InfiniBand fabric, 8 links: 800 Gb/s per GPU; More 8-GPU nodes to One InfiniBand fabric; 1 GPU to Host network adapter.

  • 1 GPUInstance
    vCPU
    24
    RAM
    346 GiB
    1 GPU
  • Host network adapter
    rating
    400 Gb/s
    scope
    per machine
  • 8 GPUsNode
    HGX B300 8-GPU baseboard, line drawing
    vCPU
    192
    RAM
    2,768 GiB
    CPU
    Intel Xeon 6776P
    NVMe
    6 × 3.84 TB
    8 GPUs · NVLink 5
  • One InfiniBand fabricShared by the cluster
    per GPU
    800 Gb/s

    Every node of one cluster sits on the same fabric. A cluster is placed onone fabric and stays there.

  • More 8-GPU nodesSame shape
    HGX B300 8-GPU baseboard, line drawing
    HGX B300 8-GPU baseboard, line drawing
    HGX B300 8-GPU baseboard, line drawing
  • 1 GPUInstance
    vCPU
    24
    RAM
    346 GiB
    1 GPU
  • Host network adapter
    rating
    400 Gb/s
    scope
    per machine
  • 8 GPUsNode
    HGX B300 8-GPU baseboard, line drawing
    vCPU
    192
    RAM
    2,768 GiB
    CPU
    Intel Xeon 6776P
    NVMe
    6 × 3.84 TB
    8 GPUs · NVLink 5
  • One InfiniBand fabric
    per GPU
    800 Gb/s
  • More 8-GPU nodesSame shape
  • 1 GPUInstance
    vCPU
    24
    RAM
    346 GiB
    1 GPU
  • Host network adapter
    rating
    400 Gb/s
    scope
    per machine
  • 8 GPUsNode
    HGX B300 8-GPU baseboard, line drawing
    vCPU
    192
    RAM
    2,768 GiB
    CPU
    Intel Xeon 6776P
    NVMe
    6 × 3.84 TB
    8 GPUs · NVLink 5
  • One InfiniBand fabric
    per GPU
    800 Gb/s
  • More 8-GPU nodesSame shape

B300 listing.

CPU, RAM, local NVMe and the fabric vary by pool and are reported at placement. GPUs and InfiniBand adapters are passed through whole to one machine.

Status
Private Preview
Runs as
Single instances and GPU Clusters
Capacity
On-demand · Spot · Reserved
Metering
Metered by the second, billed hourly

GPU Clusters

TAB 02B300 node shape Source · Fantasti platform configuration Reviewed 2026-10-10
B300 node shape
ParameterValueScope
1 GPU24 vCPU · 346 GiBPer machine
8 GPUs192 vCPU · 2,768 GiBPer machine
Host CPUIntel Xeon 6776PPer 8-GPU node
InfiniBand800 Gb/sPer GPU
InfiniBandNot publishedPer 8-GPU node
Host network adapter400 Gb/sPer machine
Local NVMe6 × 3.84 TBPer 8-GPU node
Local NVMe, priceIncludedIn the node price

NVIDIA B300 on Fantasti: fantasti.ai/gpus/b300 · Request capacity: fantasti.ai/contact

Sheet
01 / 01
Title
B300 datasheet
Figures
NVIDIA reference
Reviewed
2026-10-10
§04 Datasheets 8 listings
§05 Request Private Preview

Request B300 capacity.

Tell us the count, the term and where it should run. We reply to a reviewed request, usually within one business day. A reserved term is quoted.

Request this GPU GPU catalog

Next