Refactr: private AI hosted and operated in Europe
Refactr runs open AI models on hardware you already own, through its Epsilon inference layer. It's a sound approach when you want to keep data in house. Bunker starts from the same confidentiality requirement, but delivers a complete private AI, hosted in Europe or on your own servers, operated by an EU company, with RAG, the interface and the support included.
Refactr makes Epsilon, an inference serving layer that runs open AI models on the graphics cards you own, on premises or in your own private cloud. Refactr installs that layer, monitors it and keeps it running, but it does not build applications and does not touch your documents: it makes the AI engine underneath work. The company is based in Islamabad, Pakistan.
On one point Refactr and Bunker agree: your data stays in an environment you control, with nothing sent to a public API. The difference lies elsewhere. Refactr optimizes hardware you have to own and operate, from an operator located outside the European Union. Bunker hosts and operates a complete private AI in Europe, or on your servers, on an open-source stack, with support and reversibility: the day you want to bring everything back in house, you leave with your code and your data.
A public AI API, Refactr and Bunker
| Criterion | Public AI API (OpenAI, Anthropic) | Refactr (Epsilon) | Bunker |
|---|---|---|---|
| Publisher / operator | US vendor | Refactr (company based in Islamabad, Pakistan) | Bunker (EU company) |
| Operator's jurisdiction | United States (CLOUD Act, FISA 702) | Pakistan (outside the European Union) | European Union (GDPR, outside the CLOUD Act) |
| Where the models run | Vendor's cloud, outside the EU | On your hardware, on premises or in your private cloud | Hosted in Europe by Bunker, or on your servers |
| Must you own and run the GPUs | No | Yes, you provide and operate the hardware | No when managed (hosted in Europe), yes on premises |
| Nature of the offer | API billed per token | Inference layer (Epsilon) placed on your hardware | Complete private AI product (model, RAG, interface, support) |
| Scope around the AI | Model only | Serving layer only (no application, does not touch your documents) | Private AI and sovereign IaaS (VMs, Kubernetes, databases) |
| Models | Vendor's proprietary models | Compatible open models, brought by you | Open-source models operated for you |
| Inference software layer | Proprietary | Epsilon, proprietary (no public repository) | Open-source stack |
| Support | Vendor support | Team based in Pakistan | European support and SLAs |
| Certifications | Vary | None claimed by Refactr | Sovereign hosting in Europe, certification targeted |
| Reversibility | Low (proprietary API) | Depends on a non-EU operator to run it | Open source, re-internalizable |
| Pricing model | Per token or per seat | Setup fee and monthly subscription (14-day trial) | Managed infrastructure flat fee |
Refactr and Epsilon, an inference layer on your hardware
Refactr takes on a real problem: running private AI models is expensive, and much of the bill comes from graphics cards sitting idle between requests. Epsilon, its inference serving layer, aims to get more tokens out of each card. The figures Refactr publishes, on two RX 7900 XTX cards with a 27-billion-parameter Qwen model, report higher throughput and a lower cost per million tokens than llama.cpp and vLLM. On the ground of GPU efficiency, this is serious work.
Refactr's logic is clear: you bring the hardware and a compatible open model, and Refactr installs then maintains the layer that serves it fast and reliably. The company says so itself, it does not build applications and does not touch your documents. Two consequences follow. First, you must own and operate the GPUs, including workstation or gaming cards. Second, the Epsilon layer is proprietary software, with no public repository, operated by a company based in Islamabad, Pakistan.
The data stays with you, but the operator matters
Refactr is right on one substantive point: when the model runs on your servers, your questions, documents and answers do not leave, and no telemetry is sent back to the publisher. We do not claim otherwise. The real question, for a European organization bound by sovereignty requirements, is not only where the data lives, but who operates the AI layer, under what law, and with what recourse.
With Refactr, the entity that deploys, monitors and updates your stack falls under Pakistani law, outside the European Union. With Bunker, you contract with an EU company: a GDPR framework, outside the reach of the CLOUD Act, European support, governance of maintenance access and contractual recourse in Europe. Bunker also adds an option Refactr does not offer: an AI hosted in Europe, on our infrastructure, when you do not want to buy or run GPUs.
What Bunker adds around the AI engine
Refactr serves the model and stops there. Bunker delivers the complete product: the open-source model operated for you, RAG over your own documents, a private chat interface, and support from scoping to production. Around it, you get the sovereign infrastructure, managed sovereign cloud, virtual machines, Kubernetes and databases, from a single European operator.
The economic model differs too. Refactr charges a setup fee then a monthly subscription, on hardware you fund. Bunker starts from a managed infrastructure flat fee, hosted in Europe or on your own servers. And nothing is locked in: the stack is open source, so you can start managed then bring everything back in house the day your size or your strategy calls for it.
Frequently asked questions
What is Refactr and its Epsilon product?
Refactr makes Epsilon, an inference serving layer that runs open AI models on hardware you own, on premises or in your private cloud. Refactr installs, monitors and maintains that layer, but does not build applications and does not touch your documents. The company is based in Islamabad, Pakistan.
Does Refactr host my data?
No, and that is a genuine strength of its approach: the models run on your hardware, and your questions, documents and answers stay there, with no telemetry sent back to Refactr. The difference with Bunker is therefore not about where the data lives, but about the operator: Refactr is a company based in Pakistan, outside the European Union, while Bunker is an EU company, under GDPR and outside the CLOUD Act, with European support.
Is Refactr a European sovereign provider?
No. Refactr is a company based in Islamabad, Pakistan, outside the European Union. The entity that deploys and maintains your AI layer therefore falls under Pakistani law. Bunker is an EU company: the contract, the support and the governance of maintenance access are in Europe, under European law.
Do you need to buy GPU hardware for private AI?
With Refactr, yes: its Epsilon layer optimizes the graphics cards you own and operate yourself. With Bunker, you have a choice: a managed private AI hosted in Europe, with no hardware to buy, or an installation on your own servers if you prefer on premises. In both cases, you keep reversibility.
How does Bunker compare with Refactr for private AI?
Refactr serves the model on your hardware and stops at the inference layer. Bunker delivers the complete product: an open-source model operated for you, RAG over your documents, a private chat interface, support, and the sovereign infrastructure around it. All hosted in Europe or re-internalizable, at a flat fee, under European law.
A sovereign private AI, hosted in Europe
An open-source model operated for you, with RAG, interface and support included, hosted in Europe or on your servers. Let us talk about your private AI project.
Refactr and Epsilon are trademarks of Refactr; OpenAI is a trademark of OpenAI and Anthropic a trademark of Anthropic. These names are cited for objective comparison. Bunker has no commercial relationship with Refactr. Details observed on 29 September 2026 from public information on refactr.net.