sparda.fr

What It Is

SPARDA is an on-premises AI appliance for enterprises, built around the message: “Your AI. Your infrastructure. Zero exposure.” It is not a typical cloud-based AI SaaS product; instead, it delivers GPU hardware, a hardened Debian runtime environment, and the Phantom Agent interface as a packaged deployment inside the customer’s internal network. The focus is on keeping data from leaving the organization, avoiding external tool calls, and maintaining zero outbound connections. The page shows the device located in Toulouse, France, and positions its air-gapped architecture as the core differentiator.

Core Capabilities and Integrations

SPARDA uses a three-layer architecture. L1 is the hardware appliance, based on NVIDIA RTX GPUs with 16–48GB+ VRAM, available in desktop, 2U, and 4U form factors. L2 is the SPARDA Runtime Environment, which includes Debian 12 LTS, rootless Docker, AppArmor, and signed OTA updates. L3 is Phantom Agent, which can be embedded into Google Workspace and Microsoft 365, with Slack / Teams also listed. AI-200 also supports RAG based on enterprise cloud drives, while AI-500 supports multi-tenancy, SSO/SAML, and high-availability options. However, the page does not disclose specific model names, parameter sizes, context length, APIs, or output-quality benchmarks.

Pricing and Deployment

All pricing is quote-based. AI-100 is aimed at up to 25 seats, with 16GB VRAM and up to 5 concurrent chats. AI-200 targets 25–250 seats, with 24GB VRAM, up to 12 concurrent chats, and RAG included. AI-500 is for 250+ seats, with 48GB+ multi-GPU configurations, unlimited concurrency, and a dedicated solution architect. The page mentions that users can request a demo, which is run on a real AI-100 in the Toulouse lab, and also refers to pilots, but it does not state whether pilots are free.

Pros and Cons

The main strength is its very clear focus on data privacy and compliance: prompts and responses remain inside the VLAN, model updates are inbound-only, zero outbound connections can be verified at the firewall, and the design is built around RGPD, NIS2, immutable logs, and audit materials. This makes it friendly to CISOs and IT departments. The limitations are that procurement costs are not transparent, concurrency is limited on lower-tier configurations, CSPN is still only on the ANSSI roadmap with a target of H1 2027, and there is no information about Chinese-language support.

Who It’s For and Access from China

SPARDA is best suited to regulated European industries, groups whose data cannot leave their jurisdiction, and organizations that want to introduce a private AI assistant into M365/GWS quickly. It is less suitable for teams looking for low cost, instant online signup, clearly documented Chinese-language capabilities, or an open API ecosystem. The page provides no evidence regarding access from China, so this remains unknown; payment methods are also not disclosed. If you only need local experimentation or lightweight private RAG, alternatives worth comparing include Ollama, LM Studio, AnythingLLM, PrivateGPT, or a local deployment of Dify.