ISSUE #002 · JUNE 2026
|
Serverless inference is live. The GPUs are back.
Pay-per-token access to open-source models on EU-sovereign GPUs. Plus: H100 capacity is back, and a hardware drop is on the way.
|
SERVERLESS INFERENCE IS LIVE
Until now, running open-source models meant either managing your own GPUs or sending your tokens to a US provider. Serverless Inference is live.
Pay-per-token API access to 30+ open-source models, with almost all hosted within the EU. No infrastructure to manage, no idle costs, and it scales from zero to thousands of requests per second. You pay only for the tokens you process.
It’s fully OpenAI-SDK compatible: swap the base URL and your existing code works out of the box:
|
from openai import OpenAI
client = OpenAI(
base_url="https://api.lyceum.technology/v1",
api_key="your-api-key"
)
response = client.chat.completions.create(
model="meta-llama/Llama-3.3-70B-Instruct",
messages=[{"role": "user", "content": "Hello!"}]
)
|
Models available at launch include Llama 3.3 70B, DeepSeek V3.2, Qwen3 235B, Kimi K2, Nemotron Ultra 253B, and GPT-OSS 120B, plus 30+ more with new ones added continuously.
|
GPU CAPACITY UPDATE
GPU availability has been tight over the last few months. If you tried to start a run and couldn’t get the hardware you needed, you weren’t imagining it, and we know how much that hurts when you’re trying to ship.
That’s changed. We’ve brought a lot more capacity online across every GPU type, from L40S and A100 up to H200, B200, and B300, so you can start a run whenever you need one, on whatever hardware your workload calls for.
Single GPUs up to multi-GPU configs, provisioned in seconds, per-second billing, zero minimum commitment.
Spin up a GPU
|
COMING SOON
Beyond on-demand, we’ve got larger reserved-capacity blocks landing soon: dedicated nodes on commitment, for teams that need guaranteed capacity for a training run, a product launch, or steady production load.
Here’s what’s coming:
| Nodes |
GPU |
Commitment |
Lead time |
| 8 |
B300 |
12 months |
12 weeks |
| 2 |
H100 |
Up to 6 months |
— |
| 32–128 |
RTX PRO 6000 |
3 months–3 years |
— |
| 5 |
H200 |
Up to 3 months |
— |
We want to give the people on this list first access before these go more widely. If any of these blocks fit what you’re planning, reply or grab a slot below and we’ll walk you through pricing and timing.
I’m interested, let’s talk capacity
|
|
Not sure what GPU shape your workload needs?
Book a 30-minute validated-infrastructure assessment with our GPU solutions engineer. You walk away with a specific recommendation, not a pitch.
Book your assessment
|
|
|
Lyceum Technology, Berlin, Germany
Unsubscribe
|
|
