Amazon Web Services is the default cloud for most companies, and its GPU instances (P4, P5, P6, G6e) sit inside the largest service catalogue in the industry: SageMaker, S3, IAM, VPC, and everything else your platform already uses. For an enterprise that lives on AWS, keeping GPU workloads there has obvious appeal.
packet.ai is a GPU cloud and only a GPU cloud. It does not have 200 services. What it has is the same NVIDIA silicon at a fraction of the hourly price, no quota approval, no 8-GPU minimum on the high-end cards and egress at less than half of AWS's rate. This comparison is for the team deciding whether the GPU line item should stay inside the AWS bill.
Platform overview: packet.ai vs AWS
What AWS offers
AWS GPU pricing is per instance, not per GPU, and the high-end cards come in 8-GPU shapes only. Normalised per GPU in us-east-1 on-demand: p6-b200.48xlarge (8× B200) about $113.93/hr, or roughly $14.24 per GPU-hour; p5.48xlarge (8× H100) about $55.04/hr, or $6.88 per GPU-hour; p4d.24xlarge (8× A100 40GB) around $4.10 per GPU-hour; g6e.xlarge (1× L40S) about $1.86/hr. Capacity Blocks and Savings Plans reduce these substantially with commitment.
- Per-second billing on Linux, 60-second minimum; Savings Plans and Capacity Blocks for reserved rates
- Service quotas gate P-series access; new accounts typically need to request an increase and wait
- Egress $0.09/GB after the free tier; inter-region and cross-AZ transfer charged separately
- SageMaker, Bedrock, EKS and the rest of the platform are a real integration advantage
What packet.ai delivers
The first difference is not price, it is access. AWS gates P-series instances behind service quotas: a new account typically has to file a quota increase request and wait, sometimes days, before it can launch a single B200 or H100. packet.ai has no quota system. You sign up and deploy.
The second difference is the unit. AWS sells B200 and H100 only in 8-GPU shapes, so a single-card workload still pays for the full node's hourly rate, unless the workload is scaled to actually use all eight. Normalised per GPU, that puts AWS's B200 at roughly $14.24/hr and its H100 at roughly $6.88/hr, both far above what a comparable single card costs elsewhere. packet.ai sells B200 at $3.75/hr Dynamic, a 74% gap.
- Dynamic: scheduler-enforced multi-tenant, one card at a time, no quota. RTX 6000 Pro at $0.66/hr, B200 at $3.75/hr versus AWS's ~$14.24/hr per-GPU equivalent.
- Dedicated: a whole card, single-tenant, 99% SLA. RTX 4090 at $0.39/hr, L40S at $0.92/hr, A100 80GB at $1.43/hr versus AWS's ~$4.10/hr per-GPU A100 rate, B200 at $6.99/hr.
What AWS has that packet.ai does not: SageMaker, Bedrock, EKS and the rest of a platform most enterprises already have contracts and compliance sign-off for. If your workload lives inside an existing AWS environment and needs to talk to S3, IAM and VPC without leaving it, that integration is worth real money. If you just need a GPU without the quota wait or the per-node markup, packet.ai is built for exactly that.
Billing is hourly with monthly commits up to 20% off. There are no platform fees and no ingress charges; egress is $0.04/GB. Up to 2 TB of local NVMe per node is included. Dynamic instances are SSH-ready in under five minutes, Dedicated in five to ten. Capacity is live in the US (California, Virginia, Texas, Oregon) and Europe (Frankfurt, Amsterdam, Paris, London, Dublin). H100 SXM, H200 and RTX 5090 are on the notify list.
Against AWS, packet.ai's case is arithmetic. A single B200 is $3.75/hr Dynamic against roughly $14.24 per GPU-hour on p6, and you can rent one instead of eight. An A100 80GB is $1.43/hr against about $4.10 per GPU-hour on p4d, which only comes in 40GB. Egress is $0.04/GB against $0.09/GB. There is no quota request; Dynamic capacity is SSH-ready in under five minutes.
packet.ai vs AWS at a glance
Published starting rates in USD per GPU-hour, on-demand unless noted. Percentages compare the first price in each cell.
| Category | packet.aiYOU | AWS |
|---|---|---|
| B200 (180–192GB) | $3.75/hr Dynamic · $6.99/hr Dedicated | ~$14.24/hr per GPU (p6-b200, 8× only) |
| A100 80GB | $1.43/hr Dedicated | ~$4.10/hr per GPU (p4d, 40GB, 8× only); 80GB p4de higher |
| L40S (48GB) | $0.92/hr Dedicated | ~$1.86/hr (g6e.xlarge, 1×) |
| RTX 6000 Pro (96GB) | $0.66/hr Dynamic | Not offered |
| RTX 4090 (24GB) | $0.39/hr Dedicated | Not offered (no consumer GPUs) |
| H100 (80GB) | Launching soon (notify list) | ~$6.88/hr per GPU (p5, 8× only) |
| Smallest high-end unit | 1 GPU | 8 GPUs for A100/H100/B200 |
| Access | Self-serve, no quota approval | Service quota increase usually required for P-series |
| Billing | Hourly; monthly commits up to 20% off | Per second (60s min); Savings Plans; Capacity Blocks |
| Egress | $0.04/GB, no ingress fee−56% | $0.09/GB after free tier; inter-AZ/region extra |
| Adjacent services | Object storage, Token Factory (waitlist), Pixel Factory | SageMaker, Bedrock, EKS, S3, IAM and 200+ others |
| Compliance | Dedicated tier, single-tenant, 99% SLA | SOC 1/2/3, ISO 27001, HIPAA, FedRAMP, PCI DSS and more |
Detailed comparison
Price per GPU-hour
The gap is not subtle. On the B200, AWS's p6 instance works out to roughly $14.24 per GPU-hour on-demand; packet.ai's Dynamic B200 is $3.75/hr and Dedicated is $6.99/hr. That is a 74% saving at Dynamic and 51% at Dedicated, before you account for having to rent eight B200s on AWS when you need one.
The A100 story is similar (about $4.10 versus $1.43 per GPU-hour, and AWS's p4d is the 40GB part), and even the L40S, where AWS does sell a single card, is roughly double packet.ai's price. AWS Savings Plans and Capacity Blocks narrow these gaps for committed 24/7 workloads, but they are commitments, and they still do not get below packet.ai's on-demand rate on Blackwell.
Quotas, minimums and time to GPU
New AWS accounts start with a P-series vCPU quota of zero. Getting eight H100s means filing a Service Quotas request and waiting for approval, sometimes hours, sometimes days. Then you rent all eight, because that is the instance shape.
On packet.ai you pick a card, pick Dynamic or Dedicated, and have SSH within minutes. One card, no ticket. For a fine-tune that needs a single B200 for a weekend, the difference is between doing the work and doing paperwork.
Egress and the rest of the bill
AWS charges $0.09/GB for internet egress after the free allowance, plus inter-AZ and inter-region transfer. Serving model weights, datasets or generated images out of AWS is a line item people underestimate. packet.ai charges $0.04/GB with no ingress fee and includes up to 2 TB of local NVMe per node.
Where AWS earns its keep is everything around the GPU. If your data is already in S3, your auth is IAM, and your MLOps is SageMaker, the integration tax of moving compute elsewhere is real. Many teams end up hybrid: control plane and data on AWS, GPU-hours on a neocloud.
Compliance and enterprise controls
AWS has the longest compliance list in the industry: SOC 1/2/3, ISO 27001, HIPAA, PCI DSS, FedRAMP and dozens more, plus IAM, GuardDuty, CloudTrail and mature audit tooling. Regulated enterprises choose it partly for that. packet.ai's Dedicated tier gives you single-tenant hardware in known US and EU data centres with a 99% SLA; check its current certifications against your requirements before moving regulated workloads.
Who should choose which
Choose packet.ai if
- You need 1 to 8 high-end GPUs - not a forced multiple of 8 on a full node
- You want B200 or A100 at a fraction of AWS P-series per-GPU pricing
- You move a lot of data out and want egress at under half of AWS rates
- You want capacity SSH-ready in minutes without a quota request or service limit increase
Choose AWS if
- Your data, identity and MLOps already live on AWS and the switching cost is high
- You need FedRAMP, HIPAA or other compliance certifications on the compute itself
- You are committing to 24/7 usage for a year or more and can use Savings Plans or Capacity Blocks
- You need AWS-native managed services like SageMaker or Bedrock more than the raw GPU
Conclusion
AWS is a platform; packet.ai is a GPU. If you are buying the platform, the GPU premium is the price of admission. If you are buying GPU-hours, packet.ai delivers the same B200 at roughly a quarter of AWS's per-GPU on-demand price, one card at a time, without a quota ticket. Most teams we see do both: AWS for the system of record, packet.ai for the compute.
Same silicon. Smarter economics.
Every rate on this page is a published starting price you can deploy against today. Dynamic launches in under five minutes; no credit card to start.
Additional resources
- packet.ai GPU pricing: the full Dynamic, Dedicated and Clusters rate card.
- packet.ai documentation: API, CLI and platform guides.
- How intelligent scheduling works: why Dynamic keeps peak performance at a lower price.
- Clusters and wholesale quotes: multi-node InfiniBand capacity at ~30% below retail.
- Rent GPUs on packet.ai: specs, pricing and workload fit for every card.