The sovereignty tier: running local inference for banking and government briefs
Hardware, latency and the honest tradeoffs of running the whole stack inside your own network — including the one requirement people try to skip.
Some matters cannot leave the building. A regulator-facing brief, a bank's customer file, a ministry's procurement papers — for these the question is not which model is best but whether the text crosses a network boundary at all.
LexCore Sovereign is the same product as the managed cloud, installed on hardware you own or rent. Documents, matters, the database and user accounts stay inside your network. There is no phone-home and no remote access, by design — which also means no vendor-side telemetry to reassure you it is healthy. That is the trade.
What you receive, and what you provide
- Container images from a private registry with a read-only token for your organisation. Not source code.
- A licence file — your signed entitlement, carrying seat count and expiry.
- An install package — compose file, interactive installer, deployment runbook.
| Minimum | Recommended | |
|---|---|---|
| CPU / RAM | 8 vCPU / 32 GB | 16 vCPU / 64 GB |
| Disk | 200 GB SSD | 500 GB SSD |
| GPU | None (demo only) | 24 GB VRAM; 48 GB for the larger models |
| OS | Ubuntu 22.04 LTS + Docker Engine 24+ |
The latency conversation, honestly
A local model on a 24 GB card will not match a frontier model on a provider's fleet, on either speed or headroom. What it does instead is remove an entire class of question from the procurement review, and for the matters that qualify, that is worth more than the difference in tokens per second. The failure is choosing the sovereign lane for everything and then watching review times push people back to email.
Set the lane per matter
Model routing exists precisely so a firm does not have to make one choice for all work. Route confidential matters local and let routine agreements use the faster lanes. Deciding the routing rule in advance, by matter type, is what makes the policy hold under deadline pressure — a rule that has to be re-argued per upload is not a rule.
The install path, hardware sizing and the licensing model are documented in full under Deployment & Sovereignty.
Keep reading
NDPA, NDPR and cross-border legal documents: what the regulation actually requires
A reading of the Nigerian data-protection position that matters to legal ops teams — which instrument is the primary statute, what a cross-border transfer is, and where an AI review sits in it.
Per-seat pricing is broken for legal AI, and here is the math
Why charging per lawyer makes firms hoard licences and kills adoption, and what a credit model has to get right to be better rather than merely different.