Technology · HOLD.co portfolio
hardware.co — AI hardware, built to spec.
hardware.co builds custom systems for running large language models on-premises. One team specifies vendor-neutral GPU hardware, ships the cstmAI™ software stack on it, and writes the agents and applications on top — so an organization gets private AI on hardware it owns, with one contract and one point of contact.
A HOLD.co company
The metal under private AI.
hardware.co sits in the HOLD.co technology segment alongside DEV.co, LLM.co, ERP.io and CyberAttack.ai. LLM.co builds private language models and DEV.co builds the software around them; hardware.co supplies the machines they run on, for customers whose data cannot leave the building.
Organizations that want AI on their own terms usually assemble it from parts: a server from one vendor, a model from another, open-source serving software, and a consultant for the application — with nobody responsible for the whole. hardware.co closes that gap. Until October 2026 the business operated as cstm.ai.
What it does
Three parts, one system.
Most vendors sell one of these and leave the customer to integrate the rest. hardware.co delivers all three, so one team answers for whether it works.
Vendor-neutral GPU systems, sized to the workload.
Configurations are specified from the models a customer will run and the number of people who will use them, sourced across multiple manufacturers, burn-in tested before they ship, and installed on site.
Custom AI hardware →The cstmAI™ stack on every system.
Model serving, access control and monitoring, plus modules for cited search over internal documents, workflow agents with human approval steps, fine-tuning and evaluation, governance, and connectors to ERP and CRM systems.
Custom AI software →Engineers who build on top, then hand it over.
Custom agents, internal applications, document and vision AI, fine-tuning and systems integration — scoped in a two-week discovery sprint and built against an evaluation set. The customer keeps the code, prompts, evals and weights.
Custom AI development →The cstmAI™ lineup
Four sizes, every part chosen for the workload.
Systems are sized, not sold off a shelf, and every one is quoted from a written scope. Sourcing is vendor-neutral, with no lock-in to a single manufacturer.
- cstmAI™
Desk
Workstation-class, 1–2 GPUs, for a team or a pilot.
≈ 1–25 users
- cstmAI™
Rack
4–8 GPU rack server for production workloads.
≈ 25–500+ users
- cstmAI™
Cluster
Multi-node, for large models and many users.
Hundreds to thousands
- cstmAI™
Edge
Compact or ruggedized, for offline sites.
One site
Who it serves
Where the data is sensitive and the work is specific.
Regulated and security-conscious organizations that need AI on networks they control — offline, if required — and documentation their reviewers will accept.
Our thesis
Open-weight models are now good enough that the question is no longer whether a company can run AI itself, but who will stand behind the whole system when it does.
Spec a system on hardware.co.
Tell hardware.co the models you want to run, how many people will use them and where the hardware should live. An engineer replies with a first configuration and the questions that decide the quote.
Interested in the platform or the portfolio?
Whether you're planning an on-prem AI deployment or you're an owner, operator, or investor exploring a deal with HOLD.co, start a confidential conversation.