Comparison · different layers
These are different layers of the stack: OpenRouter is a routing marketplace across many providers and hundreds of models; AxForge is an EU infrastructure operator serving open models on hardware it owns.
At a glance
| Dimension | AxForge | OpenRouter |
|---|---|---|
| What it is | An EU infrastructure operator: open models served on hardware AxForge owns and operates. | A routing marketplace: one API and one bill across many AI providers. |
| Model menu | A short open-model catalogue — Qwen3.8 27B served now, plus embeddings, image generation and speech; more open models deployable on dedicated systems. | Hundreds of models from many providers, proprietary and open, with routing and fallbacks. |
| EU data residency guarantee | Yes — processing in eu-se-1 (Stockholm) or eu-es-1 (Málaga), pinned per key with x-axforge-region and echoed on the response. | Depends on the provider that serves each request; routing preferences can narrow the pool. |
| Hardware transparency | Named machines — NVIDIA DGX Spark (GB10, 128 GB unified memory) — with published serving details and measured numbers. | Abstracted by design: you target a model, not a machine. |
| Retention policy | Zero prompt retention: processed in memory, not written to disk, not logged, never used to train. | Set by each underlying provider; the marketplace documents provider data policies. |
| Dedicated compute | Yes — dedicated DGX Spark systems as a monthly rental, available now; RTX tiers on the way. | Not the product — it routes requests to providers' capacity. |
| OpenAI compatibility | Yes — chat completions, streaming, tool calls, usage reporting; plus embeddings, images and audio. | Yes — an OpenAI-compatible API is the interface. |
The layers can stack: routing marketplaces sit above infrastructure operators, and AxForge operates the kind of EU capacity such marketplaces route to. This is a comparison of layers, not a contest.
Honest routing
If any of these describe you, a routing marketplace is the right layer:
| Maximum model choice | You want frontier proprietary models and open models side by side, switchable with one line of code. |
|---|---|
| Fallbacks | You want requests rerouted automatically when a provider is down or rate-limited. |
| One bill | You want a single account and invoice across many providers instead of managing each one. |
| Model evaluation | You are still comparing models and do not want to commit to one provider's catalogue yet. |
Honest routing
| Auditable EU processing | You must show where inference ran — down to the named machine in a named EU region, pinned per key and echoed on every response. |
|---|---|
| Zero prompt retention | Retention is a hard requirement, not a preference: prompts and completions processed in memory, never persisted, never trained on. |
| A dedicated system | You want a machine that serves only your traffic — a monthly DGX Spark rental, available now. |
| Operator accountability | You want one operator who owns the hardware, publishes the serving config and answers for the numbers. |
The substance
| Regions | eu-se-1 · Stockholm and eu-es-1 · Málaga, both live. More EU regions in deployment. |
|---|---|
| Data handling | Zero prompt retention — in-memory processing in Sweden, no disk, no logs, no training. Only billing metadata is kept. Full policy: axforge.ai/privacy. |
| API | OpenAI-compatible at https://api.axforge.ai/v1 — TLS 1.3, terminated in the EU. |
| Pricing | Qwen3.8 27B: €0.29 / 1M input · €1.77 / 1M output (launch pricing). Everything else is published too — see the price list. |
| Measured speed | 15.9 tokens/s single-stream decode for Qwen3.8 27B with speculative decoding. |
| Dedicated systems | NVIDIA DGX Spark (GB10, 128 GB unified memory), monthly rental, available now in Sweden. |
Speed measured on our production DGX Spark node, single-stream, 2026-08. We publish only numbers we measured ourselves.
FAQ
They sit at different layers. OpenRouter is a routing marketplace across many providers and hundreds of models. AxForge is an EU infrastructure operator that serves open models on hardware it owns. If your requirement is EU processing you can audit down to a named machine, AxForge is the European alternative; if your requirement is maximum model choice, a marketplace serves that better.
Yes. The layers stack: a routing marketplace sits above infrastructure operators, and AxForge operates the kind of EU capacity such marketplaces route to. Using a marketplace for breadth and a pinned EU operator for regulated traffic is a normal split.
OpenRouter lists hundreds of models from many providers. AxForge serves a short open-model catalogue: Qwen3.8 27B on the shared API today, plus embeddings, image generation and speech models, with more open models deployable on dedicated systems on request.
Yes. Processing happens in eu-se-1 (Stockholm) or
eu-es-1 (Málaga), pinned per API key with the
x-axforge-region header and echoed on every response. TLS terminates
in the EU, and prompts are processed with
zero retention.
Yes. AxForge exposes an OpenAI-compatible API at
https://api.axforge.ai/v1 with chat completions, streaming, tool
calls and usage reporting, plus embeddings, images and audio. Official OpenAI
SDKs work by changing the base URL.
Related