Private AI Infrastructure
Your dataForged in black
A GPU dedicated to your vault, running open-source models. Your documents are processed there, never by a third-party AI service. Destroy the vault and its GPU instance and disk are deleted.
document
merger-agreement-draft-v3.pdf
142 pages · 3.2 MB · processed in 4.2s
Classification
CONFIDENTIAL
Entities detected
47 personal-data references
Risk assessment
HIGH — contractual exposure identified
Key clauses
Indemnity, change of control, non-compete
Processing
Runs on your dedicated instance
Architecture
Forged for privacy
Six principles behind every vault.
Dedicated GPU
A GPU dedicated to your vault. Its compute is never split with anyone else.
Self-Destructing Vaults
Destroying a vault deletes its GPU instance and disk, and its saved chat history.
Open Source Models
Fully auditable LLMs. No black boxes.
Full Audit Trail
Vault, document and specialist events are logged with timestamps.
Dedicated Processing
Your documents are analysed on the GPU dedicated to your vault.
No Third-Party AI
No third-party AI service ever sees your documents.
Process
The forging
Four stages. From raw silicon to obsidian glass.
Provision
Click 'New Vault'. Your dedicated GPU usually spins up in minutes, occasionally up to 20.
Raw material extracted from the earth. Your instance, carved from silicon.
Provision
Click 'New Vault'. Your dedicated GPU usually spins up in minutes, occasionally up to 20.
Raw material extracted from the earth. Your instance, carved from silicon.
Upload
Drag and drop documents. They pass through our servers to your instance and are processed there.
Ore meets fire. Documents enter the crucible.
Upload
Drag and drop documents. They pass through our servers to your instance and are processed there.
Ore meets fire. Documents enter the crucible.
Analyze
Chat with documents. Extract data. Generate reports.
White-hot intelligence shaped under pressure. Every insight hammered into form.
Analyze
Chat with documents. Extract data. Generate reports.
White-hot intelligence shaped under pressure. Every insight hammered into form.
Destroy
When done, destroy the vault. Its GPU instance and disk are deleted.
The glass cools and breaks. Its saved chat history is deleted too.
Destroy
When done, destroy the vault. Its GPU instance and disk are deleted.
The glass cools and breaks. Its saved chat history is deleted too.
Specialist on demand
When AI isn't enough
Nothing goes to a specialist until you send a brief to one approved by our team. One request, one bill, no marketplace to browse.
Counsel review on the clause the AI flagged.
A CPA's eyes on a number the model wasn't sure of.
A clinical reviewer for the chart you can't sign off.
Available inside every vault
Are you a specialist? Apply to the panelPricing
Pay as you go
No subscriptions. No commitments. GPU time billed by the minute.
FLASH
Qwen 3 4B
Fast extraction and quick analysis. Ideal for straightforward document review.
TURBO
Qwen 3.5 35B
Deep reasoning across complex documents. Nuanced analysis and multi-step queries.
ULTRA
GLM-5 744B
Our largest model, a 744B-parameter open-source MoE. For the hardest analysis.
Start your first vault
Add credit and begin. No contracts, no commitments.
FAQ
Questions
Everything about Obsidian's security and privacy architecture.
General AI chat services run your documents through the provider's own models and systems. Obsidian runs open-source models on a GPU dedicated to your vault. No third-party AI service sees your documents, and they are never used for model training. When you destroy the vault, its GPU instance and disk are deleted.
Your documents are uploaded over TLS 1.3 to Obsidian, and our servers relay them to your vault's GPU instance without storing them. That last hop is not encrypted. The GPU is rented from independent hosts on a third-party GPU marketplace; it is dedicated to your vault, though the host machine (CPU, disk, network) may be shared with other renters. Documents are indexed and processed there by open-source models. We never use your content for training.
Destroying a vault deletes its GPU instance and disk at the GPU provider, including the uploaded documents on it. This can't be undone. If chat saving was on (the default), the vault's saved questions and replies are deleted from our database at the same time.
We run a curated lineup of open-source models — Qwen 3 (4B) on the FLASH tier, Qwen 3.5 (35B MoE) on TURBO, and GLM-5 (744B MoE) on ULTRA — served via vLLM, an open-source inference engine. The model weights are public and run on your dedicated GPU instance, so no document is sent to a third-party AI API. You choose the model tier that fits your workload, from fast extraction to deep reasoning.
No. Obsidian is built for confidential work, but we don't offer a Business Associate Agreement and we haven't been certified against HIPAA. Each vault runs on a GPU dedicated to it, and you can destroy the vault when you're done.
Obsidian uses simple pay-as-you-go pricing with no subscriptions or commitments. GPU time starts at $4.95/hr for the FLASH tier (4B model) and goes up to $99.99/hr for the ULTRA tier (744B MoE frontier model). You only pay for the minutes you use. Add credit to your balance and start analyzing — no commitments.
Nothing goes to a specialist unless you request one. You write a brief, flag the files it's about, optionally auto-redact emails, US-format phone numbers, SSNs and card numbers in the brief, and confirm with a typed phrase. The specialist receives the brief and the names of the flagged files, not their contents. The request is recorded in your audit log.
Still have questions?
support@obsidian.expert