Allod — on-premises AI, leased
Small task-specific models, trained every night on your own accepted and corrected work, running on hardware we deliver and operate on your floor. The training data never leaves the building. No query touches a closed model provider. One fixed monthly fee.
Book a deployment call See the training loop
The stalemate
The thing that would work
On a narrow, repeated workflow, a small model trained on that workflow's own history beats a general frontier model called through an API. Not marginally. It has seen your exceptions, your house style, your precedent, and the specific ways your people correct a wrong answer. It also runs at a fraction of the energy and a fraction of the cost, on hardware that fits in one rack.
Why you cannot build it
Training it means putting proprietary workflow data — credit files, charts, case records — into a model provider's pipeline. Your policy forbids it. In most cases your regulator forbids it. And the provider is a company that may compete with you in three years, whose retention terms you do not control and cannot audit from the outside.
So the institutions with the most valuable proprietary data are the ones least able to train on it. That is a stalemate, and it is not a capability problem. It is an architecture problem, and it has an architectural answer: put the hardware, the data, and the weights in the same building, and make sure that building is yours.
What we install
Space and power are the two binding long-term constraints on AI. Compute gets cheaper every year; megawatts and square metres do not. You already have both, which is the part of this that is genuinely hard, and it is already paid for.
| Item | Specification | Supplied by |
|---|---|---|
| Compute node | 8 accelerators, ~640 GB accelerated memory, NVMe local store | Allod |
| Enclosure | One 42U rack. Locking front and rear doors, tamper-evident seals. | Allod |
| Network | Isolated VLAN. One outbound-only path. No inbound route exists. | Allod |
| Installation | Delivery, racking, power-on, burn-in, workflow bring-up on site. | Allod |
| Operation | Monitoring, nightly training, gating, hot-swap, spares, on-site service. | Allod |
| Floor space | One rack footprint plus service clearance. Slab or raised floor. | Customer |
| Power | ≈10 kW continuous, 208 V three-phase, A+B feeds to the rack. | Customer |
| Cooling | ≈34,000 BTU/h removed at the rack. Standard hot/cold aisle is fine. | Customer |
Final configuration is sized to the workload after a one-day survey. Accelerator count and memory are the two figures that move; the rack, the power envelope, and the boundary do not.
From signed order to first workflow in production is eight to twelve weeks, and most of that is hardware lead time and your change-control calendar, not us. Install itself is two days on site. We do not need a network drop into your production environment, a VPN into our office, or an account on any of your systems.
The loop
Every workflow we deploy is instrumented from day one. The model gets better because your people are already correcting it, and those corrections are the most valuable training signal in your organization. Today they are thrown away.
The workflow runs in production against real cases. Every request and every response is written to a local store on the node.
Each output is instrumented for outcome: accepted as-is, rejected outright, or corrected by the person doing the work. The correction is the signal — it is the label a frontier provider cannot buy.
At 02:00 local, a LoRA adapter trains on that day's accepted, rejected, and corrected pairs. It runs on the same node, on idle capacity, on your power.
The candidate adapter is scored against your evaluation set — your cases, your rubric, your acceptance threshold. An adapter that does not beat the incumbent is discarded. There is no override.
A passing adapter is hot-swapped into the serving path without restarting the model. Every version is retained on the node; reverting to any prior adapter is a single command.
The week-over-week delta on your eval set is recorded and reported. If the number stops moving, we say so — that is a finding about the workflow, not something to hide.
Back to 01. The cycle repeats every night. No station in it is reachable from outside your building, and no data in it leaves.
Measured
Credit operations — covenant exception triage. Both systems saw the same cases and were scored against the same frozen evaluation set, held by the customer and never used for training.
Placeholder figures — benchmark not yet published
| Measure | Frontier API | Allod, week 1 | Allod, week 8 |
|---|---|---|---|
| Agreement with adjudicator | 00.0% | 00.0% | 00.0% |
| Material breaches missed | 00 | 00 | 00 |
| Officer corrections per 100 memos | 00 | 00 | 00 |
| p95 latency | 0,000 ms | 000 ms | 000 ms |
| Cost per 1,000 memos | $00.00 | $0.00 | $0.00 |
Method
Ownership
Not a license to use them. Not access for the term. Ownership, perpetual, including after you stop paying us.
Security posture
Data residency
Prompts, retrieved context, outputs, corrections, and training data are written to local storage on the node and never leave it. There is no replication target outside the building.
Update channel
Outbound-only. The node pulls signed artifacts from a fixed endpoint and verifies them against a key pinned at install. Nothing outside can open a connection inward.
Telemetry
Metrics only — utilization, temperature, error counts, adapter version, gate pass/fail. Never prompt content, never output content, never a record identifier. The schema is fixed and published.
Air-gapped mode
Supported. The update channel is disabled entirely and artifacts arrive on physical media against a published hash. Roughly a quarter of prospects choose this.
Read the security architecture
Written to be forwarded to third-party risk without editing.
Where this lands first
Covenant exception triage
A credit officer reads a borrower's quarterly financials against the covenant schedule, decides whether a breach is technical or material, and drafts the exception memo. The judgment is in the bank's own precedent — thousands of prior exceptions, none of them public, none of them in any frontier model's training set.
Prior authorization packet assembly
Assembling the clinical justification a specific payer accepts for a specific procedure code, from the chart. What each payer accepts is learned from that hospital's own approvals and denials. Every input is PHI and none of it can be sent anywhere.
Records request redaction review
First-pass exemption marking on a records release, applying the agency's own long-standing reading of each exemption. The corpus is the agency's prior releases and withholding logs. Much of it is unclassified but not public, and it cannot go to a commercial API.
What it costs
The lease covers the hardware, the platform, the nightly training, the gating, the spares, and on-site service. There is no separate license, no per-seat charge, no per-token charge, and no professional services line for the install. A workflow that runs ten times as often next quarter does not cost more.
You provide space and power. Those are real costs and we will not pretend otherwise — roughly 10 kW continuous per node, on your meter.
A typical single-node deployment is a mid-five-figure monthly line item. The range is driven by node configuration, number of workflows, and service level, and we will give you the number on the first call rather than after three.
Objections
The lease ends and we collect the hardware. You keep the weights — base model, every adapter you trained, every checkpoint — exported to disk in safetensors format before the rack leaves the building, along with the serving container and the eval harness. They run on any comparable hardware without us. Nothing phones home and nothing expires. The thing we take back is steel.
One outbound-only channel. The node polls a fixed endpoint on a schedule you set, over a path your firewall allows in one direction. It pulls signed artifacts — platform updates, base model updates, security patches — verifies the signature against a public key pinned at install, and stages them for your approval. Nothing outside your building can initiate a connection inward, because there is no inbound route to initiate it on. If you prefer, we disable the channel entirely and deliver updates on physical media against a published hash. That is air-gapped mode and roughly a quarter of prospects want it.
It is our hardware and our problem. We hold spares for every field-replaceable part in the node and swap on a next-business-day target, faster where the site warrants it. The lease covers it; there is no separate maintenance line. A failed accelerator degrades throughput, it does not stop the workflow — the model keeps serving on the remaining cards. Adapters and training data live on the node's local store, which is mirrored, and the weights are yours in either case.
Buying the hardware is the easy quarter of the problem. The hard parts are the training loop and the gate: instrumenting a live workflow so accepted, rejected, and corrected outputs are captured cleanly, turning that into a nightly LoRA run that does not degrade the model, and building an evaluation set honest enough that you can let a new adapter into production on its say-so. That is a standing team you would have to hire, and most of that team's work has nothing to do with your business. We lease you the hardware and that team as one line item. You can also just buy GPUs from us in effect — take the hardware, decline the loop — but then you have bought a rack, and the rack is not the product.
One node, one workflow. That is the unit. We will not sell a half-node or a shared node, because the entire argument is that the hardware is yours and nothing else runs on it. One node comfortably carries several workflows once the first is in production, and most customers add the second and third within two quarters. If you want to test the loop before committing a rack, we run a six-week paid pilot on a node we install and then remove — same architecture, same boundary, same gate.
Book a deployment call
Thirty minutes, an engineer on the line, no deck. If your workflow is wrong for this — too varied, too low-volume, no correction signal — we will say so on that call rather than sell you a rack.