FinBlade Deployment Options

Deployment topology

Three ways to run FinBlade

The product is identical in all three. What changes is where the boundary of your network sits — and therefore what crosses it. Read each diagram top to bottom: the higher the dashed line, the more of the system runs on FinBlade’s side.

Your site — your network, your storage
FinBlade-managed cloud
The boundary, and what crosses it
A

On-prem

Air-gapped

Everything runs inside your fence line. The model, the agents, the admin panel and every document sit on hardware you own. The system is built to keep working with the internet cable pulled out.

YOUR SITE · AIR-GAPPED NETWORK Enterprise users Workstations on the site LAN FinBlade Stack App · Agents · Admin panel · Audit log Documents, embeddings and logs on your disks LLM runtime Open-weights model on your GPU servers Inference never leaves the rack NOTHING CROSSES No outbound calls, no telemetry, no model callbacks PUBLIC INTERNET — NOT CONNECTED

What you get

  • +Works on an isolated plant network, in a classified area, or during an outage
  • +No cross-border data transfer to argue through legal
  • +You hold the only copy of prompts, documents and audit trail

What it asks of you

  • −GPU servers to buy, rack, cool and keep patched
  • −Model and skill updates arrive as offline packages you schedule
  • −The longest path from decision to first user
B

Hybrid

Middle ground

Your data stays home; only the thinking is rented. The stack, the storage and the audit log run on your servers. A prompt travels out to a FinBlade-managed model and the answer comes back — nothing is written on the far side.

YOUR SITE · YOUR SERVERS Enterprise users Workstations on the site LAN FinBlade Stack + your data App · Agents · Admin panel · Audit log Documents, embeddings and reports on your disks Retrieval and redaction run before anything leaves Every record the system generates is written here PROMPT OUT · ANSWER BACK Encrypted in transit · no bulk export FINBLADE-MANAGED CLOUD LLM inference GPU capacity sized and patched by FinBlade Stateless — prompts not stored, not trained on

What you get

  • +No GPUs to own; the model tier improves without a project
  • +Documents, embeddings and the audit trail never leave your disks
  • +One firewall rule to review instead of a data-migration plan

What it asks of you

  • −A controlled outbound path to the FinBlade region
  • −Prompt text — not your repository — is processed off site
  • −Application servers still run under your ops team
C

Cloud

Private tenant

Nothing to install, and a tenant that is only yours. Stack, model and storage run in FinBlade’s cloud inside an isolated tenant with its own admin panel — your users, your roles, your data, never a shared pool.

YOUR SITE Enterprise users Browser only — signed in with your directory No servers, no GPUs, nothing to patch on site HTTPS · SSO FROM YOUR DIRECTORY Session traffic only · you upload documents once FINBLADE CLOUD PRIVATE TENANT · YOUR USERS ONLY FinBlade Stack App · Agents · An admin panel you control LLM inference, dedicated to your tenant Your data store — isolated, exportable, yours to delete

What you get

  • +Fastest to stand up — no hardware to procure
  • +Your own admin panel, roles and audit log, not a shared workspace
  • +Contractors and multi-site teams reach it without VPN plumbing

What it asks of you

  • −Your documents are stored in the FinBlade cloud, not on your disks
  • −Sign-off on residency and retention terms
  • −Availability follows the site’s link to the internet

Side by side

Same product, same skills library, same outputs — different custody of the parts.

  A · On-prem B · Hybrid C · Cloud
Internet required None — runs air-gapped One outbound HTTPS path to the FinBlade region Standard HTTPS from user browsers
Where documents live Your storage, your network Your storage, your network Isolated tenant storage in the FinBlade cloud
Where prompts are processed Your GPU servers FinBlade GPUs — stateless, not retained FinBlade GPUs — dedicated to your tenant
Records the system creates Written on site Written on site Written in your tenant, exportable on demand
Hardware you run GPU servers and application servers Application servers only None
Model & skill updates Offline packages you schedule and test Managed by FinBlade, on your release window Managed by FinBlade, on your release window
Admin & audit Your admin panel, your logs Your admin panel, your logs Private admin panel for your users; logs exportable
Rollout effort Highest — hardware, network, acceptance testing Moderate — servers plus one firewall change Lowest — accounts and directory integration
Typical fit Gov or classified enterprise clients with isolated or classified plants; strict no-egress rules Clients with a controlled corporate network Pilots, distributed contractors, teams that need to start now

These diagrams show topology and custody of each component, not a network design. Ports, regions, retention periods and isolation guarantees should be confirmed against the FinBlade deployment agreement before they are quoted into a client’s security review.