01
B200
192 GB HBM3e
Frontier-grade Blackwell silicon for the largest training and inference runs.
- Performance10 PFLOPS
- Bandwidth8 TB/s
From
$5.49/GPU·hr
Limited
GPU Cloud
The AI developer
cloud, on demand.
Pods for building. Clusters for scaling. Reserved capacity for shipping. B200, B200 SXM, H100, A100, and L40S — billed by the second, NVLink-ready, in 12 regions.
<90sProvisioning
12Regions
99.99%Uptime SLA
1sBilling increment
The lineup
Every NVIDIA class, ready to deploy
From Ada Lovelace inference at $1.49 / GPU·hr to Blackwell B200 for frontier training — every node is NVLink-capable, NVMe-backed, and billed by the second.
Deploy
02
B200 SXM
141 GB HBM3e
Big-memory Hopper for retrieval-heavy training and long-context inference.
- Performance3,958 TFLOPS
- Bandwidth4.8 TB/s
From
$3.99/GPU·hr
Limited
03
H100 SXM
80 GB HBM3
The flagship workhorse — balanced for pretraining, fine-tuning, and serving.
- Performance3,958 TFLOPS
- Bandwidth3.35 TB/s
From
$2.99/GPU·hr
In stock
04
H100 NVL
94 GB HBM3
PCIe-form H100 paired for inference and mid-scale fine-tuning.
- Performance3,958 TFLOPS
- Bandwidth3.9 TB/s
From
$2.59/GPU·hr
In stock
05
A100 SXM
80 GB HBM2e
Proven Ampere class for general-purpose AI and HPC workloads.
- Performance312 TFLOPS bf16
- Bandwidth2 TB/s
From
$1.89/GPU·hr
In stock
06
L40S
48 GB GDDR6
Cost-efficient Ada Lovelace for inference, embeddings, and diffusion.
- Performance733 TFLOPS
- Bandwidth864 GB/s
From
$1.49/GPU·hr
In stock
Reserved capacity available · 1mo+ commitments
Talk to salesHow it works
From zero to inference in four steps
No replatforming. No lock-in. No hyperscaler tax. Bring your container, your framework, your code — we handle the rest.
01
Spin up
Pick a GPU, pick a region, pick a base image. Pod is ready with SSH and a public IP in under 90 seconds.
02
Build
Train, fine-tune, or batch-process. Your containers, your framework, your weights — persistent NVMe volumes follow your jobs.
03
Deploy
Push to production with templated inference endpoints, blue-green rollouts, and built-in TLS + auto-scaling.
04
Scale
Single pod to thousand-GPU clusters across 12 regions. Reserve capacity for predictable load, burst on demand for spikes.
Workloads
Picked for the work you actually do
Not sure which GPU to start with? Match the workload to the silicon — and start with a single node, scale to a cluster when the job demands it.
01
LLM training & pretraining
Frontier model training, full fine-tunes, and large-scale pretraining on multi-node clusters with NVLink fabric.
Recommended
B200B200 SXM
B200B200 SXM02
Fine-tuning & adaptation
LoRA, QLoRA, full-parameter tunes, and RLHF on 7B → 70B+ models with shared NVMe checkpoint volumes.
RecommendedH100 SXMH100 NVL
H100 SXMH100 NVL03
Production inference
High-throughput LLM serving with vLLM, TGI, or TensorRT-LLM. Sub-100ms first-token latency at scale.
RecommendedL40SH100 NVLB200 SXM
L40SH100 NVLB200 SXM04
Diffusion, vision & multimodal
SDXL, FLUX, SVD, vision encoders, and embedding pipelines on memory-flexible Ada and Hopper class GPUs.
RecommendedL40SA100 SXM
L40SA100 SXMPer-second billing
Pay only for time the GPU is up. Stop a pod, billing stops within the second.
NVLink + InfiniBand fabric
900 GB/s GPU-to-GPU bandwidth on SXM nodes, 3.2 Tbps east-west on reserved clusters.
Persistent NVMe volumes
Checkpoint, dataset, and weight volumes that survive pod restarts and follow your jobs.
Single pod to 1,000+ GPUs
Start with a single accelerator, scale into reserved multi-node clusters when you need to.
Need bigger?
Reserve a cluster.
Multi-node H100, H200, and B200 clusters with NVIDIA NVLink fabric, dedicated capacity, and committed pricing. From a single 8-GPU node to thousand-GPU training runs — we handle the rest.
Multi-node NVLink fabric
8× SXM per node · 900 GB/s GPU-to-GPU bandwidth
Reserved pricing
Up to 60% off on-demand · 1-mo to 3-yr commitment
Dedicated support
Priority access, 24/7 coverage, and SLA-backed reliability
Rapid provisioning
Clusters ready in hours, not days
Custom networking
Tailored VPC, routing, and isolation to fit your architecture
99.99% uptime SLA
Enterprise-grade reliability you can build on
Ready to build?
Talk to our infrastructure team and get a custom quote.
Cluster configuration
Review & requestB200 SXM · 8-node cluster
NVLink fabric · Redundant power · PCIe 5.0
GPUs
64×B200GPU memory
12 TB (192 GB / GPU)Interconnect
NVLink Switch SystemvCPUs / node
96 vCPUsNetworking
400 Gbps · RDMATerm
12-month reserved
Secure·Private·Enterprise ready
FAQ
Common questions
01How fast can I get a GPU?
On-demand pods provision in under 90 seconds for available capacity (H100, A100, L40S). Limited-stock SKUs like B200 and B200 typically provision in a few minutes. Reserved multi-node clusters are provisioned in under 24 hours after sales handoff.
02What's included in the per-hour price?
The GPU(s), bundled vCPU and system RAM, the container disk, public IP, ingress and egress bandwidth (with a generous monthly allowance), and persistent NVMe storage up to a quota. Egress beyond the allowance and reserved capacity are billed separately.
03Can I run multi-node training jobs?
Yes. SXM nodes are NVLink-fabric-connected within a chassis (900 GB/s GPU-to-GPU). Reserved clusters add InfiniBand or RoCE east-west networking (up to 3.2 Tbps) across nodes — purpose-built for distributed training with FSDP, DeepSpeed, or NVIDIA's Megatron-LM stack.
04How does reserved pricing work?
Commit to capacity for 1 month to 3 years and save up to 60% off the on-demand rate. Reserved capacity is dedicated, region-pinned, and SLA-backed. Talk to sales for a quote tailored to your training schedule.
05Do you support spot pricing for interruptible jobs?
Yes. Spot pods are roughly 50–70% cheaper than on-demand and are interruptible with 60 seconds of notice. Best fit for fault-tolerant training with checkpointing, hyperparameter sweeps, and batch inference.
06Which frameworks are pre-installed?
CUDA 12.4, PyTorch 2.x, JAX, TensorFlow, vLLM, TGI, TensorRT-LLM, and common scientific stacks ship in base images. You can also bring your own Docker image from any public or private registry.
Ready when you are
The AI developer cloud
No replatforming. No lock-in. No hyperscaler tax. Pick a GPU, pick a region, and you're running.