# install
$ pip install lyceum-cli
# launch an A100
$ lyceum launch --gpu a100
VM ready in 4.2s. ssh user@a100-eu-west-19f3.
# or an L40S
$ lyceum launch --gpu l40s
Dedicated Training
A100 80GB and L40S 48GB. Available now. Per-second billing.
Dedicated GPU VMs for fine-tuning and production inference. Full root SSH. EU and US data centers.
GDPR by default EU + US data centers Per-second billing
Why teams move to Lyceum for A100 and L40S
Quota walls and waitlists
Hyperscalers gate A100 capacity behind quota requests, regional unavailability, and reservation queues. Lyceum has supply you can launch today.
Hourly minimums and egress fees
Hyperscalers bill in 1-hour blocks, charge for cross-region egress, and add networking fees. Lyceum bills per second with no egress fees.
Black-box pricing
Hyperscalers publish list prices that do not match the actual bill. Lyceum’s price is the price: $1.59/hr A100, $1.19/hr L40S, billed per second.
Provision in seconds, not days
1
Install the CLI
One command, any platform.
$ pip install lyceum-cli
2
Launch a VM
A100 or L40S, EU or US, by name.
$ lyceum launch --gpu a100
3
SSH in
Full root, any framework. Per-second billing starts when the VM is ready.
$ ssh user@<provisioned-host>
Cards in this offer
NVIDIA A100 80GB
$1.59/hr- GPU memory
- 80 GB HBM2e
- System RAM
- 120 GB
- vCPU
- 16
- FP16
- 312 TFLOPS
- TF32
- 156 TFLOPS
- Price
- $1.59/hr on-demand
- Billing
- Per-second, no minimums
- Regions
- EU West (Netherlands), EU Central (Germany), EU North (Finland), US
- Best for
- Fine-tuning (LoRA, QLoRA, full SFT), distributed training up to ~70B params, mid-scale inference
NVIDIA L40S 48GB
$1.19/hr- GPU memory
- 48 GB GDDR6
- System RAM
- 128 GB
- vCPU
- 12
- FP16
- 362 TFLOPS
- FP8
- 1466 TFLOPS
- Price
- $1.19/hr on-demand
- Billing
- Per-second, no minimums
- Regions
- EU West (Netherlands), EU Central (Germany), EU North (Finland), US
- Best for
- Production inference, batch inference, vLLM deployments, smaller fine-tunes
What you will actually pay
| Card | Lyceum (per-second) | Typical hyperscaler (per-hour, list) |
|---|---|---|
| A100 80GB | $1.59/hr | ~$3.06 to $12.29/hr depending on provider and region |
| L40S 48GB | $1.19/hr | ~$2.50 to $5.00/hr depending on provider and region |
Hyperscaler rates above are publicly listed prices as of 2026-04 and vary by provider, region, and commitment. Per-second billing on Lyceum means partial hours are not rounded up.
Fine-tuning on A100
- LoRA, QLoRA, full SFT
- Frameworks: PyTorch, JAX, TensorFlow, custom CUDA
- Multi-GPU support: up to 8x A100 in a single VM
- Per-second billing means iterative experiments do not pay the rounding-up tax
Production inference on L40S
- vLLM, TGI, Triton, custom serving stacks
- FP8 inference (L40S native FP8 path)
- Cold-start to first token: 30s typical
- Scale-to-zero on inactive replicas
How A100 + L40S on Lyceum compares
| Lyceum A100 / L40S | Hyperscaler GPU VM | GPU rental marketplace | Self-hosted | |
|---|---|---|---|---|
| Availability today | Yes | Quota-gated | Variable | n/a (capex) |
| Billing | Per-second | Per-hour | Per-second (varies) | n/a |
| EU data residency | Yes | Region-dependent | Provider-dependent | Yes |
| Full root SSH | Yes | Yes | Yes | Yes |
| Egress fees | None | Yes | Provider-dependent | n/a |
| Hardware ownership | Lyceum-owned | Provider-owned | Marketplace (variable) | Yours |
| Stability | Lyceum SLA | Provider SLA | Highly variable | Operator-dependent |
Frequently asked questions
What is the difference between A100 80GB and A100 40GB?
This page is A100 80GB only. The 40GB card is older and not in this offer.
Can I get multiple GPUs on a single VM?
Yes. Up to 8x A100 80GB or 8x L40S 48GB in one VM. Larger configurations are available via quote.
How does per-second billing work?
The meter starts when the VM is provisioned and stops when you terminate it. Partial seconds are billed pro-rata. No minimum hold time.
Where are the data centers?
EU West (Netherlands), EU Central (Germany), EU North (Finland), and US.
Is there a free trial or credit?
Lyceum CLI signup includes a small starter credit. Contact sales for larger evaluation budgets.
Can I bring my own VPN or firewall?
Yes. Each VM is a dedicated machine with full root, configure as you would any Linux box.
What about Serverless Inference?
If you want pay-per-token API access to pre-hosted models (Llama, Mistral, etc.) without managing a VM, see /products/inference/serverless/.