Blackwell Ultra capacity, reserved before it exists.
Orion Factory is bringing NVIDIA GB300 NVL72 racks online in our European data center from January 2027 - reservable from a single tray to multiple racks. Capacity is allocated from a priority queue: join now and lock your position before the first racks are committed.
You're in the queue
We've logged your request. Expect a confirmation of your position within 48 hours - allocation review begins Q4 2026.
Size your cluster
Drag to your GPU count - from a single tray to a multi-rack NVLink-connected fleet. We allocate at any scale; no full-rack minimum.
January allocation - Rack 01
Built for what you're building
Frontier training
Train trillion-parameter and MoE models on a unified 72-GPU NVLink domain per rack - 37 TB of fast memory keeps checkpoints, batches and context where they belong.
AI reasoning
Test-time scaling can demand 100× the compute of one-shot inference. Blackwell Ultra's 2× attention acceleration and FP4 efficiency are built for exactly this.
High-throughput inference
Hold large models and long KV caches entirely in GPU memory for low-latency serving at maximum tokens per second - and per megawatt.
How allocation works
Reserve your position
Join the waiting list with your cluster size and target start. No payment, no commitment - your position in the queue is timestamped.
Allocation review
Our engineers walk through your workload, configuration and terms - with transparent reserved-capacity pricing, confirmed in a signed allocation agreement.
Deployment
Racks come online in our European data center. Confirmed customers get white-glove onboarding and direct engineering support.
Why Orion Factory
Telecom-grade infrastructure operated by Orion telekom, delivered as an AI factory built with NVIDIA.
Any scale, no full-rack minimum
Most providers allocate GB300 only as complete racks. We reserve from a single 4-GPU tray to multi-rack clusters - grow into your capacity, don't overbuy it.
NVIDIA Partner Network
We deploy GB300 NVL72 as an NPN partner - reference architecture, validated networking and direct access to NVIDIA engineering.
European data residency
Your models and data stay in Europe, in facilities we own and operate - with low latency to EU markets and full jurisdictional clarity.
Carrier backbone included
As a telecom operator we bring our own fiber and peering. Multi-terabit connectivity to your cluster is part of the offer, not an add-on.
Reserved means reserved
Capacity is allocated once, in queue order. We would rather run a real waiting list than oversell racks that don't exist yet.
Prefer to talk first?
Skip the queue form - our infrastructure engineers will walk you through configurations, networking and timelines directly.
Message received
One of our infrastructure engineers will get back to you within one business day.