
GPU infrastructure from training to serving
Start by the hour on GPU Cloud today. When you scale, Cluster Build designs, deploys and operates a dedicated cluster for you.
Six things that stall an AI infrastructure decision
These are the questions reviewers raise most often, with how PacketStream answers each one.
We can't get GPUs when we need them
GPU Cloud creates instances in the Portal as soon as capacity is available. For periods you cannot miss, reserve capacity for one month or longer in advance.
OptionsRenting GPUs alone doesn't make training fast
Node-to-node traffic runs on InfiniBand NDR/XDR. NVMe-based parallel storage is provided on the same fabric.
ComponentsOur data can't leave a set location
When data residency applies, we build a closed-network cluster inside that region. Encryption keys can be managed by your team.
SecurityWe don't have people to run a cluster
The Cluster Build operations contract covers incident response and part replacement. Alert thresholds are agreed with you at handover.
ProcessWe can't forecast the cost
On-demand prices per GPU hour are published. For steady use, compare reserved rates and a Cluster Build quote side by side.
PricingIt may not fit the tools we already use
Metrics export to Prometheus, Grafana or Datadog. Object Storage speaks the S3 API, so your SDKs and CLI keep working.
Integration
Three ways to start
All three use the same GPU lineup, so you can start small and move up without changing your stack.
GPU Cloud
On-demand, by the hour
Best whenExperiments, PoCs, short training runs and inference with uneven demand
- Lead time
- Immediate to a few days
- Billing
- Hourly from creation to deletion, no commitment
- Tenancy
- Instances in a shared cluster
- Network and topology
- Standard configuration
- Location
- Choose from available regions
GPU Cloud
Reserved, one month or longer
Best whenTraining on a fixed schedule and always-on inference
- Lead time
- A few days for large or region-specific requests
- Billing
- Rate agreed by term and volume; longer and larger costs less
- Tenancy
- Capacity held for you first
- Network and topology
- Standard configuration
- Location
- Choose from available regions
Cluster Build
Dedicated cluster, designed and operated
Best whenLarge training running for months, regulated data, or a cluster in your own data center
- Lead time
- Weeks from stock, months for a new order
- Billing
- Build plus operations contract
- Tenancy
- A cluster dedicated to you
- Network and topology
- Designed to your requirements
- Location
- A chosen region or your site
Everything around the GPU, in one design
Training slows down at the network and storage as often as at the GPU. We design all six layers together.
- GPU
B300 · B200 · GB300 · H200 SXM · RTX PRO 6000
Choose Blackwell or Hopper by workload. GB300 and RTX PRO 6000 are reserved only, and RTX PRO 5500 is coming soon.
- Interconnect
InfiniBand NDR / XDR
Carries node-to-node traffic. Topology is set by cluster size and communication pattern.
- Storage
PFS + Object Storage
NVMe-based parallel file storage with a POSIX mount, and S3-compatible Object Storage.
- Network
Own backbone, AS139648
Data moves in and out over the PacketStream backbone and direct ISP connections.
- Observability
Prometheus · Grafana · Datadog
GPU, node and fabric metrics export to the monitoring stack you already run.
- Handover
Bare metal or Kubernetes
We hand over bare metal with the OS and drivers, or a cluster with Kubernetes and a scheduler set up.
Configurations by workload
A starting point for the first conversation. The final design follows your model, data and schedule.
| Typical work | GPU | Network and storage | Suggested option | |
|---|---|---|---|---|
| Model training | Foundation model pre-training, fine-tuning and alignment, multi-node distributed training | B300 for the most memory per node, B200 as the Blackwell default, GB300 by NVL rack | InfiniBand interconnect, PFS storage | Reserved or Cluster Build |
| Inference serving | LLM serving, long-context processing, batch inference | H200 SXM with steady supply, RTX PRO 6000 for inference and rendering | Traffic out over the backbone and direct ISP connections | On-demand or reserved |
| Data processing | Preprocessing and embedding, vector indexing, RAG pipelines | Sized to the batch; on-demand capacity is enough to start | PFS and Object Storage | On-demand |
| Regulated environments | Data residency requirements, closed networks, audit logs | Chosen in design to fit the site's power and cooling | Closed network in the region, customer-managed keys | Cluster Build |
What changes after you start
GPU infrastructure changes how long you wait, what you pay and who operates it.
Immediate
Time · GPU Cloud instances when capacity is available
Less waiting for GPUs
On-demand instances start as soon as capacity is available. A Cluster Build from stock takes weeks, not months.
Hourly
Cost · on-demand billing, no commitment
Pay for what you use, less as you commit
On-demand bills by the hour. Reserved rates fall with term and volume, and for months of steady use Cluster Build reaches a lower total cost.
Included
Reliability · part replacement in the operations contract
Training keeps running through failures
When a GPU fails, the node is isolated and work moves to another node. Operations contracts include part replacement.
From first call to operation
GPU Cloud skips this process: create instances in the Portal and start.
- 1
Requirements
We go through your model, data size, schedule and regulatory conditions together.
- You provide
- Workload, target schedule, data location rules
- You receive
- Requirements summary
- Estimated duration
- 1 week
- 2
Design and quote
We size GPUs and nodes, the interconnect, PFS capacity and bandwidth, and power and cooling. Pick only the parts you need.
- You provide
- Review of the design, site power details if hosting on site
- You receive
- Configuration and quote
- Estimated duration
- 2 weeks
- 3
Order and build
Weeks if covered by stock, months for a new order. New power or a new site adds time.
- You provide
- Contract, access and network details
- You receive
- Installed and tested cluster
- Estimated duration
- 4 to 6 weeks
- 4
Handover and operation
We hand over bare metal or Kubernetes, agree alert thresholds with you, then operate including part replacement.
- You provide
- Alert contacts and thresholds
- You receive
- Operations contract and monitoring
- Duration
- Ongoing
We reply within 24 business hours of your inquiry.
Security and data control
Write down who can reach the cluster and where data goes before anything is built.
- Network boundary
- Network boundaries and access paths are stated in the design.
- Encryption keys
- Keys can be configured so your team manages them directly.
- Logs
- Access logs and change history are provided.
- Data residency
- Under residency rules, the cluster runs as a closed network inside the region.
- Tenancy
- Cluster Build is dedicated to one customer.
- Review documents
- Documents for regulated-industry reviews are scoped before the contract.
Ask sales for the current certification list.
Works with what you already run
No new console to learn. Each layer connects to a standard you likely use today.
Monitoring
Prometheus · Grafana · Datadog
Export GPU, node and fabric metrics.
Orchestration
Kubernetes
Handed over with Kubernetes and a scheduler, or as bare metal.
Object storage
S3 API
Keep your AWS SDK, CLI and tools; change only the endpoint.
File storage
POSIX
PFS mounts like a local file system on every GPU node.
Public cloud
Cloud Connect
Private links to AWS Direct Connect, Azure ExpressRoute, Google Cloud Interconnect and Oracle FastConnect.
Your data center
Colocation
Place a Cluster Build in your chosen data center after a site power check.
What it costs
On-demand prices are public. Reserved capacity and Cluster Build are quoted to your term and volume.
| Item | Detail | Price |
|---|---|---|
| GPU Cloud on-demandPer GPU, hourly | NVIDIA B300288GB · Blackwell | $7.38/hour |
| NVIDIA B200192GB · Blackwell | $5.88/hour | |
| NVIDIA H200 SXM141GB · Hopper | $4.38/hour | |
| Reserved and dedicatedQuoted | GPU Cloud reservedAll models, one month minimum; GB300 and RTX PRO 6000 are reserved only | Quote |
| Cluster BuildBuild plus operations contract | Quote | |
| StorageBilled separately | PFS Cluster StorageNVMe parallel file system | $0.20/GiB/month |
- Prices are per GPU.
- Storage and data transfer are not included in the hourly price.
Questions before you decide
Can we start small and move to a dedicated cluster later?
Yes. GPU Cloud and Cluster Build use the same GPU lineup. Validate on-demand first, then move to reserved or Cluster Build once usage is clear.
How long does a Cluster Build take?
Weeks if stock covers it, months if we order new hardware. Power upgrades or a new site add time.
Can the cluster sit in our own data center?
Yes. We check the site's power conditions first, then design for it. PacketStream data centers already have power, cooling and circuits ready.
What happens when a GPU fails?
The node is isolated and work moves to another node. In Cluster Build, part replacement is part of the operations contract and spare quantity is agreed in design.
Are storage and data transfer included?
No. Both are billed separately from the hourly GPU price. PFS is $0.20 per GiB per month.
Can we hand over only part of the work?
Yes. Choose from GPU and node selection, interconnect, storage, power and cooling, handover and operations.
Related products
GPU Server
GPU Cloud and Cluster Build lineup, prices and calculator.
Open GPU ServerObject Storage
S3-compatible storage for datasets and outputs.
Open Object StorageCloud Connect
Private links to AWS, Azure, Google Cloud and Oracle.
Open Cloud Connect
Tell us about your workload
Share your model, data size and schedule. We reply within 24 business hours with a configuration and next steps.
