Technical Architecture

This is the one page on this site written for technical readers. What the hardware and software actually are — modules, memory, storage, throughput, power draw, the agent runtime, the data boundaries — is set out here in full. Nothing on the rest of the site depends on your reading it.

If a term below is unfamiliar, the glossary at the bottom defines every one of them in plain English.

The three systems, side by side

Three builds, one software image family. They differ in the compute module, memory, and storage; the assistant, the pairing flow, and the data boundaries are identical across all three.

Digital Twin Pro Digital Twin Pro Studio Digital Twin Pro Plus
Compute module NVIDIA Jetson Orin Nano Super NVIDIA DGX Spark (GB10 Grace Blackwell) NVIDIA Jetson Orin NX 16GB
Memory 8GB LPDDR5, 102 GB/s 128GB unified LPDDR5x, 273 GB/s 16GB LPDDR5, 102 GB/s
Storage 1TB NVMe 4TB self-encrypting NVMe 1TB NVMe
AI performance Up to 67 sparse INT8 TOPS (33 dense, Super Mode) Up to 1 PFLOP sparse FP4 (theoretical, per NVIDIA) Up to 157 sparse INT8 TOPS (Super Mode)
Power draw 7W–25W, configurable modes Per NVIDIA DGX Spark specifications 10W–40W, configurable modes
Local model class Small local open models; external providers optional, with your own key Up to 200B parameters (NVIDIA capacity statement); two linked systems up to 405B Larger small-class models and more concurrent work than the standard build
Availability Standard, in production Built to order Limited hand-built run, by invitation

Full specifications — Digital Twin Pro

Assembled in Miami, FL, USA

Platform figures are NVIDIA's published specifications for the Jetson Orin Nano Super. Configuration rows describe what ships in every unit.

NVIDIA Jetson Orin Nano Super platform
Module NVIDIA Jetson Orin Nano Super
AI performance Up to 67 sparse INT8 TOPS in Super Mode (33 dense) — NVIDIA-published figure
GPU NVIDIA Ampere architecture — 1,024 CUDA cores, 32 Tensor Cores
Memory 8GB 128-bit LPDDR5
Memory bandwidth 102 GB/s
Power 7W–25W, configurable power modes
Digital Twin Pro configuration
Storage 1TB NVMe SSD, installed and configured
Network Gigabit Ethernet and wireless
Enclosure ROBOCO aluminum cooling case — black or white
Software NVIDIA JetPack 7.2 · your assistant preinstalled · Nemotron and other leading open models, pinned and validated per release
Local and external models Answers are produced in the box. You can optionally connect Claude, GPT, or Gemini using your own provider accounts, governed by those providers' terms.
Setup Phone-guided pairing — no monitor, keyboard, flashing, or command line required
Best suited to One owner, everyday work, private memory, always-on at low power
Included Setup assistance, documentation, factory-reset image, and covered fixes for the supported software version · 1-year limited hardware warranty · optional one-time services (Life Upload, setup, refresh, migration)
Order terms Built to order — confirmed in writing · all sales final

Full specifications — Digital Twin Pro Studio

Assembled in Miami, FL, USA

Platform figures are NVIDIA's published specifications for the DGX Spark. Configuration rows describe the Digital Twin Pro build.

NVIDIA DGX Spark platform
Module NVIDIA DGX Spark — GB10 Grace Blackwell Superchip
AI performance Up to 1 PFLOP sparse FP4 — theoretical peak, per NVIDIA
Memory 128GB unified LPDDR5x at 273 GB/s
Storage 4TB self-encrypting NVMe
Network 10GbE, ConnectX-7 (200Gbps), Wi-Fi 7
Power Per NVIDIA DGX Spark specifications
Digital Twin Pro Studio configuration
Model capacity Up to 200B parameters (NVIDIA capacity statement); two linked systems up to 405B
Software DGX OS (tested version) · Digital Twin Pro configuration · your assistant · validated open models
Included services 90-minute onboarding · one integration · 30 days of deployment support · encrypted recovery image
Best suited to Large local models, development work, document intelligence, team-ready deployment
Not intended for concurrent use beyond the limits we document, or frontier cloud-model equivalence on every task
Warranty 1-year limited Digital Twin Pro service warranty, subject to hardware warranty terms
Order terms Built to order — confirmed in writing · all sales final

Full specifications — Digital Twin Pro Plus

Assembled in Miami, FL, USA

A limited hand-built run on the Jetson Orin NX 16GB, offered by invitation. Platform figures are NVIDIA's published specifications for that module.

NVIDIA Jetson Orin NX 16GB platform
Module NVIDIA Jetson Orin NX 16GB on a Digital Twin Pro-validated carrier board
AI performance Up to 157 sparse INT8 TOPS in Super Mode — NVIDIA-published figure
GPU NVIDIA Ampere architecture — 1,024 CUDA cores, 32 Tensor Cores
Memory 16GB 128-bit LPDDR5
Memory bandwidth 102 GB/s
Power 10W–40W, configurable power modes
Digital Twin Pro Plus configuration
Storage 1TB NVMe SSD, installed and configured
Enclosure Precision aluminum cooling case, black or white; thermal solution selected and validated per unit
Software NVIDIA JetPack 7.2 · your assistant preinstalled · Nemotron and other leading open models, pinned and validated per release
Setup Phone-guided pairing, plus a personal installation session with the founder
Included Setup assistance, documentation, factory-reset image, and covered fixes for the supported software version · 1-year limited hardware warranty backed by a dedicated batch reserve · optional one-time services
Order terms Built to order — confirmed in writing · all sales final

The software stack

  • System image — built on NVIDIA JetPack 7.2 (Digital Twin Pro and Digital Twin Pro Plus) or the NVIDIA DGX software stack (Studio); flashed, configured, and validated before shipment.
  • Your assistant — the preinstalled personal AI assistant, configured per client before the appliance ships. It arrives with a name, and you can change it.
  • Agent runtime — the assistant runs through a Digital Twin Pro-tested configuration built on the Nemotron family and other leading open models: sandboxed isolation, guided onboarding, routed inference across local models, network policy, and lifecycle management.
  • Local model runtime — capable local open models for private everyday work, sized to each system's memory.
  • External provider path (optional) — bring-your-own-key connections to Claude, GPT, Gemini, and other providers when you choose. The key is an authentication credential, not an encryption key.
  • Pairing and dashboard — QR-based phone pairing; a browser dashboard for connections, settings, and backups.
  • Owner-approved memory — the assistant retains only preferences and project context you approve, stored on the appliance. You can review and remove retained items from the dashboard.
  • Services layer (optional) — one-time services: Life Upload, Expert Sessions (software refresh, fixes, migration), and integration setup. Pricing is on the services page.
  • Private large-job processing — for approved archive jobs too large for the appliance, private processing under a written scope returns finished results to your device.

Your assistant and its open models

Your assistant is preinstalled on every Digital Twin Pro: configured for you before shipment and supported through setup and optional one-time services. It thinks with the Nemotron family and other leading open models, run locally on the appliance and chosen per job.

Everything runs within a Digital Twin Pro-tested configuration: we pin and validate the specific model and runtime versions included with the shipped image, and upstream updates reach your appliance only after they pass validation — through Managed Private AI, never silently.

Capability and data boundaries

What runs where, what needs the internet, and what can leave the appliance.

Capability Runs where Requires internet What can leave
Local chat, drafting, summaries Appliance No No model-provider submission required
Local document retrieval Appliance No No model-provider submission required
Owner-approved memory and preferences Appliance No No model-provider submission required
Phone pairing & dashboard Appliance + your browser No (same local network) Stays on your network
Claude / GPT / Gemini (your own key) External provider Yes Selected prompt and attached context, under your account and the provider's terms
Messaging apps (Telegram and similar) Messaging provider Yes Messages routed through that provider
Life Upload Written project scope Yes, for transfer Only authorized archive material, erased from our systems after delivery
Remote Operations Approved secure channel Yes System data needed for the approved support scope
Expert Session (one-time service) Appliance Yes No content submission required beyond update delivery
Custom agent workflows on the appliance's open models Appliance Depends on workflow On request — advanced configuration (upstream alpha; we validate the shipped version)

Remote-support boundary

There is no default remote access. Owner-Authorized Remote Operations sessions are approved by you per session over a secured channel (Tailscale SSH, or a managed device connection where used), scoped in writing, and logged where supported. Only system data needed for the approved support scope is accessed, and no standing access remains after closeout. Details are on the security & privacy page.

Recovery, rollback, and updates

  • Factory image — every appliance includes a factory-reset image so you can return to the shipped, supported configuration.
  • Backups — you can copy backups to your own drive from the dashboard; a one-time service can set this up with you.
  • Update policy — your system continues to operate on its installed, supported configuration with the first three months of Managed Private AI included free. Covered fixes apply to the supported software version. Refreshing to the current supported image, with configuration preserved, is available as a one-time Expert Session.

Supported, not supported, and known limitations

  • Supported — the shipped image and its validated configuration; assistant workflows you enable; documented integrations; covered fixes for the supported software version.
  • Not supported — modifications outside the supported image (you own the hardware and may modify it, but modified configurations fall outside covered fixes); upstream alpha features not included in the shipped, pinned configuration; unsupervised legal, medical, financial, or tax decision-making.
  • The largest frontier models do not run locally on appliance-class hardware; the assistant routes those tasks to an external provider only when you connect one.
  • Local model quality and speed depend on the system's memory and the model class it runs. The lineup table above is the sizing guide.
  • AI output can be wrong or incomplete and needs owner review for consequential decisions.
  • Automations and integrations operate within configured, documented limits — we tell you where the ceiling is for your deployment.

Glossary

Every term used on this page, in plain English. This is the only page on the site where these words appear.

Module / compute module The single circuit board carrying the processor, graphics engine, and memory. It is the part that determines how much the machine can think about at once.
Carrier board The board the module plugs into, providing power, cooling, ports, and storage connections.
GPU Graphics processing unit — the part that does thousands of small calculations at once. AI models run on it.
CUDA cores / Tensor Cores The individual calculating units inside an NVIDIA GPU. Tensor Cores are specialised for the maths AI models use most.
Ampere / Blackwell Names of NVIDIA GPU generations. Blackwell is the newer of the two.
TOPS Trillions of operations per second — a headline measure of raw AI calculating speed. Higher is faster, but real-world speed also depends on memory.
Sparse vs dense Two ways of counting the same speed. Sparse figures assume the model skips predictable zeros; dense figures assume it does not. Dense is the more conservative number, so both are given here.
INT8 / FP4 / FP16 How precisely each number inside a model is stored. Lower precision runs faster and uses less memory, at some cost in accuracy.
PFLOP A quadrillion calculations per second. A theoretical peak, not a sustained rate.
LPDDR5 / LPDDR5x The type of memory used. The "LP" means low power, which is why these machines run quietly on very little electricity.
Unified memory One pool of memory shared by the processor and the graphics engine, instead of two separate pools. It removes a copying step, which matters for large models.
Memory bandwidth (GB/s) How fast data moves between memory and the processor. For AI work this is often the real speed limit, not TOPS.
NVMe / SSD Fast solid-state storage with no moving parts. This is where your documents and the assistant's memory live.
Self-encrypting drive A drive that scrambles its own contents in hardware, so the data is unreadable if the drive is removed.
Watts (W) Electricity used. For scale, a desk lamp is roughly in this range.
JetPack / DGX OS NVIDIA's operating-system bundles for these machines — the base layer our image is built on.
System image The complete, tested copy of all software installed on the machine. Yours ships with a known-good image and can be restored to it.
The assistant’s name Your assistant arrives with a name, and you can change it. Everywhere on this site it is simply "your assistant".
Open model families The open model families the assistant thinks with — NVIDIA's Nemotron family and other leading open models — run locally on the appliance inside a sandboxed runtime, which is what keeps one task from reaching into another.
Sandbox A walled-off space a program runs inside, so a mistake in one task cannot affect the rest of the system.
LLM / large language model The kind of AI model that reads and writes language. "Local model" means one that runs on your machine rather than someone else's.
Parameters (B) The number of adjustable values inside a model, in billions. Roughly, more parameters means more capable and more memory-hungry.
Inference The act of running a model to get an answer, as opposed to training one.
Document retrieval (RAG) Looking up the relevant passages from your own documents and handing them to the model before it answers, so the answer is grounded in your material.
Bring your own key (BYOK) Connecting an outside AI provider using your own account. The key is an authentication credential, not an encryption key.
Pairing The one-time step where your phone and the appliance recognise each other, done by scanning a code.
Tailscale SSH / secured channel An encrypted private connection used only for a support session you approve, and closed afterwards.

NVIDIA, Jetson, JetPack, DGX, and DGX Spark are trademarks of NVIDIA Corporation. Digital Twin Pro is an independent product company and is not affiliated with, endorsed by, or sponsored by NVIDIA unless otherwise stated.

Back to choosing the right one

The managed layer

All of the above describes the appliance you own. Managed Private AI is an optional operating layer on top of it: monitoring and health checks, scheduled model and software updates, configuration tuning, backup assistance, remote operations within an approved scope, and — at the Concierge level — a named specialist who knows your deployment. It changes who operates the platform, not who owns it: cancel and the whole stack below keeps running exactly as configured.