Aimed at banks, public bodies and startups in southeast Europe that must keep AI workloads inside Serbia yet want current Blackwell GPUs, with a new account activated within 24 hours of payment. Look elsewhere for instant self-serve signup, several regions or a published uptime percentage, because inference starts with a sales form.
What changed
Orion AI Factory is a GPU cloud run by the Serbian telecom Orion telekom from its Belgrade data center, launched on 19 January 2026. It rents NVIDIA DGX B200 capacity by the hour, runs bare-metal training clusters and sells token-priced inference, with data stored and processed only in Serbia.
Frequently Asked Questions
What does Orion AI Factory actually cost?
Orion lists four hourly packages excluding VAT: Entry at €3.00, Pro at €5.50, Max at €8.90 and Enterprise at €9.78 with premium support and NVIDIA AI Enterprise software. Hosted inference is metered per million tokens, from €0.50 in and €1.50 out on DeepSeek V4 Flash to €1.00 in and €4.00 out on the GLM and Kimi models. Fees are prepaid in Serbian dinars, and a terabyte of extra storage is €50 a month.
Is Orion AI Factory free to use?
No free tier exists. Orion offers a guided free trial with limited GPU resources, an AI workspace, an inference test environment and onboarding help, but its team reviews every request before granting access. The request form asks you to pick a package from Entry to Enterprise and a workload of develop, train or inference.
What should you use instead of Orion AI Factory?
[Crusoe](/hub/services/crusoe) suits teams that want openly published dollar rates for Blackwell GPUs. For Kubernetes and Slurm across 59 availability zones, look at [CoreWeave](/hub/services/coreweave) instead, and for a marketplace of mixed GPUs in many countries with no KYC, try [Lium](/hub/services/lium). Stay with Orion when a Serbian operator and Serbian data residency decide the purchase.
How does Orion AI Factory compare to Crusoe in 2026?
Both rent Blackwell GPUs by the hour and state that they charge nothing for data egress, but Orion sits in Belgrade only while Crusoe's HokAI-listed regions are in the US, Norway and Canada. Crusoe's pricing page lists its single-GPU Blackwell instance at $8.50 an hour, and Orion's full-GPU package works out to about $9.99 at today's rate, roughly $1.50 more per GPU-hour. Take Crusoe for managed Kubernetes or serverless fine-tuning on a wider footprint, and Orion when residency or a Serbian contract is the requirement.
How long does it take to get going with Orion AI Factory?
Orion says access and an activation code arrive within 24 hours of payment. Before that, you pick a package, a duration and any extra storage, enter company details with tax identifiers and a domain email, confirm the order with a six-digit emailed code, then pay by card or a pro forma bank-transfer invoice. Inference accounts take longer to begin, since a sales form and a service agreement come before any API key.
Top Alternatives
- Crusoe Cloud: Choose Crusoe for openly published dollar rates and managed Kubernetes in US and Norwegian sites; choose Orion when your data has to stay in Serbia.
- CoreWeave: CoreWeave wins on Kubernetes and Slurm at multi-region scale; Orion wins when one Belgrade site under a Serbian contract is what compliance asks for.
- Lium: Lium suits a cheap, no-KYC marketplace of mixed GPUs; Orion suits buyers who need a named operator with ISO-listed facilities and direct engineer access.