Start Building →
Pricing

Pay for compute.
Not commitments.

Three plans, seven NVIDIA SKUs, one transparent rate card. Same peak performance & VRAM as Dedicated, often at half the price.

Dynamic· Shared, same peak performance

Multi-tenant GPU with scheduler-enforced isolation. Same peak performance & VRAM as Dedicated. Hourly billing.

Dedicated· Whole card, exclusively yours

Single-tenant GPU committed to your account. Zero scheduler interference. Predictable performance. 99.99% SLA.

Clusters· Multi-node, flexible terms

Multiple nodes at wholesale pricing, around 30% below retail. Custom storage, named TAM, dedicated SLAs.

Billing
GPU
Region
VRAM
Any Showing 8 of 8 GPUs
Neocloud comparison

Same silicon.
Lower bill.

Side-by-side starting rates on the same NVIDIA silicon. packet.ai delivers the same peak performance and VRAM through smart scheduling, often at a fraction of what other neoclouds charge.

GPU
packet.ai
starts from
RunPod
starts from
Vast.ai
starts from
Lambda Labs
starts from
ShadeFarm
starts from
RTX 4090
24 GB GDDR6X
$0.39~30% avg
$0.69
$0.35
/
$0.60
RTX 5090
32 GB GDDR7
/
$0.99
$0.41
/
$0.65
RTX 6000 Pro
96 GB GDDR7
/
$2.09
$1.00
/
$2.19
L40S
48 GB GDDR6
$0.92~9% avg
$0.86
$0.47
/
$1.70
A100 80GB
80 GB HBM2e
$1.43~23% avg
$1.49
$0.77
$1.99
$3.28
B200
192 GB HBM3e
$5.25~11% avg
$5.89
$4.34
$6.99
$6.52
All prices in USD per GPU-hour, published starting rates. / = SKU not offered by that provider.
Which plan should you choose?

Match the plan to the workload.

Mix and match per project. Most customers run Dynamic for dev and Dedicated for production on the same account, then move to Clusters when they outgrow single-node training.

Workload
Recommended
Why
Production inference (LLM API)
Dedicated
Predictable p99 latency · zero scheduler interference
Fine-tuning under 100 GPU-hours
Dynamic
Hourly billing · spin up in under 5 min · no commit
Evaluation / batch inference
Dynamic
Bursty workload, Dynamic is 30-40% cheaper
Distributed training · 64+ GPUs
Clusters
Requires InfiniBand fabric · multi-node NVLink topology
Frontier model pre-training
Clusters
1,000+ GPU pool · custom storage tiering · named TAM
Always-on RAG / agent compute
Dedicated
Single-tenant card · committed monthly rate
Research notebooks · interactive dev
Dynamic
Hourly billing · spin down between sessions
Pricing FAQ

Questions teams ask before signing.

Real answers from our solutions team. For anything not here, reach help@packet.ai.

What is the difference between Dynamic and Dedicated?
Dedicated gives you the whole GPU card, exclusively yours. Single-tenant, predictable p99, ideal for production inference and regulated workloads. Dynamic is shared infrastructure with the same peak performance and VRAM; the packet.ai scheduler co-locates workloads that stress different GPU dimensions so latency stays within ±2-5% of dedicated, at roughly half the price.
How does packet.ai compare to RunPod, Vast.ai, Lambda Labs, and ShadeFarm?
On the same NVIDIA SKU, packet.ai starting rates are competitive with or below every major neocloud. An NVIDIA A100 80 GB is $1.43/GPU-hr on packet.ai vs $0.77-$1.99/GPU-hr across the others. The B200 lands at $3.75 vs $4.34-$6.99. See the comparison table above for the full breakdown.
Do you offer monthly billing?
Yes. Monthly commits are up to 20% off the on-demand hourly rate. 6-month and 12-month terms include additional discounts (talk to sales). All commits include rollover credits for unused hours.
How fast can I deploy a GPU?
Dynamic-tier launches in under 5 minutes from API call to SSH-ready. Dedicated single-GPU launches in 5-10 minutes. Multi-node Cluster deployments take 2-6 weeks for InfiniBand cabling and validation.
What about cluster pricing?
Clusters use InfiniBand-connected multi-node fabric (up to 1,024+ GPUs). Pricing depends on fabric topology, node count, storage tier, and term. Typical 64-node B200 clusters land at $2.80-$3.20/GPU-hr on 12-month terms. Talk to clusters team.
Are there hidden fees? Ingress, egress, storage?
No platform fees, no ingress. Egress is $0.04/GB (vs $0.09/GB on AWS). NVMe local storage included up to 2 TB per node. Object storage is $0.018/GB-month hot and $0.004/GB-month cold (S3-compatible).
Which NVIDIA GPUs are available?
Published today: RTX 4090, RTX 5090, RTX 6000 Pro, L40S, A100 80GB, and B200. H100 SXM and H200 are launching soon — join the notify list. New SKUs roll in monthly.
Which regions do you serve?
Capacity is live in the United States and Europe, California, Virginia, Texas, Oregon, Frankfurt, Amsterdam, Paris, London, Dublin. Use the region search above the rate card to filter. APAC capacity is rolling out in Q3.

Launch in under 5 minutes.
Or talk to a human.

Most teams ship their first inference workload before their AWS quote comes back.

Start building →Get a wholesale quote
No credit card · usage-based billing · cancel anytime