Silo Server.
Sovereign compute,
shipped as a product.

Law firms, RIAs, medical practices, and accounting firms hold data they cannot lawfully or commercially place in a public cloud: case files, health records, portfolio allocations. The market forces them to choose between easy and private. Cloud AI is easy but not private. Enterprise on-prem hardware is private but not easy. Local-first tooling is a kit, not a product. Nobody sells private and turnkey as a single ordered object.

Silo Server closes that gap by turning private infrastructure into a consumer product. Each unit ships from the factory pre-configured with the full operational stack: local language models, retrieval over the firm's own documents, ingestion, access control, and the vertical template for the buyer's industry. The customer configures the machine online the way they would configure a Mac, and it arrives ready to run. Nothing phones home. The only network the machine knows is the one inside the building.

Capacity is physical. When the firm outgrows one unit, it orders another and clicks it onto the stack. The modules align magnetically, lock mechanically, and join a private fabric automatically. No rack, no integrator, no migration project. Scaling infrastructure becomes a purchasing decision instead of an engineering one.

This is not a cold start. Silo Systems has deployed this class of infrastructure by hand for consulting clients in regulated legal and financial services since 2019, under the internal designation Obox. The consulting practice proved the demand, funded the R&D, and supplies the design partners. Silo Server is the productization of work clients already pay for, aimed at the thousands of firms one consultant cannot reach.

The buyer is a managing partner or practice principal at a 1–50 seat regulated firm. They already spend on compliance, IT, and per-seat software. They have watched competitors adopt AI and they have read their own bar opinions, HIPAA guidance, and client engagement letters. Their constraint is not budget. It is that every AI product offered to them requires surrendering custody of the one asset they are professionally obligated to protect.

Why now, specifically. Three curves crossed between 2024 and 2026. Open-weight models reached genuine professional utility at the 30–70B scale. Unified-memory edge silicon made 70B-class local inference possible in a shoebox instead of a rack. And compliance pressure on cloud AI hardened from theoretical to explicit, in bar ethics opinions, insurer questionnaires, and client outside-counsel guidelines. The window where a local appliance is both feasible and differentiated is open now.

The concept document left the compute platform open between Apple Silicon, NVIDIA, and x86. This revision closes it: AMD Strix Halo (Ryzen AI Max 300 series) across every tier, differentiated by unified memory capacity alone.

The reasoning is recorded in full at D/01, but the shape of it: Apple does not sell M-series silicon to third parties, so an Apple-based appliance cannot legally exist at scale. NVIDIA Jetson Thor is a strong inference module but binds the entire software estate to CUDA on ARM at a materially higher cost per gigabyte of unified memory. Strix Halo delivers up to 128GB of unified LPDDR5X, an XDNA NPU plus a 40-CU RDNA GPU, full x86 compatibility, and, decisively, a supply chain already proven open to small integrators. One architecture means one NixOS image, one thermal design, one test jig, and one spares bin across the entire product line. For a build-to-order operation, fleet homogeneity is not a preference. It is the support model.

Manufacturing posture · integrate, don't fabricate

Version one is an integration business, not a fabrication business (D/01, manufacturing clause). Mainboards are sourced as Strix Halo OEM modules from the same supply chain that already serves small system builders. The chassis is extruded and CNC-finished at a job shop in lots of 50–200, anodized to the burnt-copper specification. Assembly, flashing, burn-in, and QA happen in-house in California, which is both a cost decision and a brand asset: every unit ships with a serialized certificate of provenance, assembled and attested where the company lives. Contract manufacturing is milestone-gated at a sustained 500 units per year.

Regulatory engineering: internal power is a pre-certified PSU module, which confines compliance work to FCC Part 15 Class B emissions testing and ETL listing of the assembly. Budgeted at $30k and eight weeks inside Phase 2 of the roadmap (§X).

The signature interaction: lift a new module onto the stack, feel the magnets seat it, close two latches, and watch the fabric absorb the capacity. No cables between modules, no configuration, no integrator. This section specifies how that experience is engineered without a single dishonest claim underneath it.

Three physics problems had to be resolved. First, retention: magnets alone cannot safely carry a 2.9 kg cantilevered module against a shear pull, so neodymium cone pairs handle alignment and self-seating within ±1.5 mm, and dual tool-free cam latches carry the structural load at a rated 25 kg shear (D/04). Second, signal integrity: the concept's pogo-pin interconnect cannot carry modern serial lanes reliably at over 10 GT/s across a separable interface, so the electrical join is a blind-mate high-speed connector of the ExaMAX class, recessed behind the magnetic registration so it can never be force-mated. Third, the pooling claim itself, resolved below.

Every Silo Server runs the Silo Engine: an immutable, declaratively configured operating stack whose single governing rule is inherited from every product this company has shipped. Zero bytes egress by default. No telemetry, no accounts, no external APIs, no phone-home. The build system enforces it; the front lamp displays it.

The architecture extends the factory discipline already proven across the Silo app line: edit the manifest, re-emit everywhere. One file per machine, silo.server.json, declares the tier, the vertical template, the model pack, the role map, retention policy, and access control sources. From that one file the factory emits the NixOS configuration, the service topology, the compliance defaults, and the customer's provenance certificate. There is no hand-configured machine anywhere in the fleet, which is the only way a small team services a thousand appliances.

Compliance marketing usually asserts conclusions. This product asserts mechanisms, and lets the buyer's counsel draw the conclusion, which is exactly how regulated buyers evaluate. Every claim below names the physical or cryptographic fact that produces it.

Pricing anchors to the buyer's alternative, not our BOM. A 20-seat firm's cloud AI bill runs $7,200 a year forever, with custody surrendered; an integrator-built private stack starts near $60,000. Silo Server prices between those poles, with hardware margin at appliance norms and recurring revenue attached to every unit (D/11).

Every unit moves through one pipeline with four hard gates. Nothing ships around a gate. The pipeline is designed for a founder-led operation today and a contract manufacturer tomorrow, because every step is executed against the manifest and recorded to the ledger, the process itself is the training material for the first ops hire.