Tell us the models and the users. We return a configuration sheet and a quote.
Spec a system →Model serving, access control and monitoring on hardware you own.
See the platform →
Thirty minutes with an engineer: your workflow, your data, and whether custom AI fits.
Book a scoping call →Custom hardware for on-prem LLM deployments, with the software and the engineers, delivered as one system you own.
Most vendors sell one of these and leave you to integrate the rest. We specify the hardware, ship the software on it and write the custom code on top, so one team answers for whether it works.
Vendor-neutral AI systems in four sizes, specified to the models you'll run and the people who'll use them.
| 01 | Desk, Rack, Cluster and Edge configurations |
| 02 | Sourced across vendors; no lock-in |
| 03 | Burn-in tested before it ships |
| 04 | Installed on site, supported after |
The cstmAI™ stack that ships on every system: serving, search, agents, evaluation, governance and connectors.
| 01 | Open-weight models served on your hardware |
| 02 | Cited answers from your documents |
| 03 | Agents with human approval steps |
| 04 | Audit logs, permissions and redaction |
Engineers who build the agents, applications and integrations on top, and hand everything over.
| 01 | Two-week discovery sprint |
| 02 | Fixed-scope builds against an evaluation set |
| 03 | Agents, apps, document AI, fine-tuning |
| 04 | You own the code, prompts, evals and weights |
Configurations are sized, not sold off a shelf. The ranges below are typical; the quote names every part and the reason for it.
Workstation-class, 1–2 GPUs, for a team or a pilot
| GPUs | 1–2 |
|---|---|
| GPU memory | 24–96 GB each |
| Users | ≈ 1–25 |
4–8 GPU rack server for production workloads
| GPUs | 4–8 |
|---|---|
| GPU memory | 48–192 GB each |
| Users | ≈ 25–500+ |
Multi-node, for large models and many users
| GPUs | 16 and up |
|---|---|
| GPU memory | 80–192 GB each |
| Users | Hundreds–thousands |
Compact or ruggedized, for offline sites
| Accelerators | Module or 1–2 GPUs |
|---|---|
| Memory | 16–64 GB |
| Users | One site |
Every cstmAI system ships with the same layers. Your applications sit on top; the platform and the metal sit underneath; governance runs through all of it.
Each phase ends with a review and a go/no-go. A first system is typically in production 8–12 weeks after discovery begins.
Pick the workflow, audit the data, size the hardware, write the evaluation set.
Software built and scored against the evaluation set, reviewed weekly with your team.
Hardware assembled and run under sustained load on our bench, with the software loaded.
Racked, cabled, networked and joined to your identity provider, on your site.
Your team runs it with our runbooks, or we run it for you. Models updated as they improve.
Custom AI earns its keep where the data is sensitive and the work is specific. These are the sectors we build for most.
Contract review, clause search and matter knowledge, with privileged files kept in-house
HIPAA · PHI · Minimum necessaryIntake, documentation and policy tools with PHI kept on your network
GLBA · Records retention · Model riskUnderwriting, KYC review and research, documented for model risk
Engineering IP · Quality records · OT networksEngineering search, quality records and plant-floor assistants
CUI · CMMC · Disconnected networksRecords, policy and analysis tools on networks you control, offline if needed
Client confidentiality · Independence · Knowledge reuseProposal, research and knowledge tools for consulting, accounting and engineering firms
NERC CIP · OT security · Field operationsField, maintenance and regulatory document tools for operators
All sectorsCustom AI is an AI system built around one organization's own work (its documents, procedures and systems) instead of a general-purpose tool used as it comes. At hardware.co that means the hardware it runs on, the software that serves and governs it, and the code our engineers write on top, all specified for you.
By quote, because scope drives cost. Every project starts with a two-week discovery sprint that ends in a written plan, a hardware configuration sheet and a quote for the build. You decide whether to continue at that point.
Discovery takes two weeks. A first system is typically in production 8–12 weeks after discovery begins, depending on hardware lead times and integration work. Multi-node clusters and facility work take longer, and we date the plan after the site survey.
Control over where data goes, predictable cost once usage is steady, and the ability to run with no internet connection. Cloud APIs are a good place to experiment. If your workload doesn't justify hardware yet, we'll say so in discovery.
No. We specify systems from multiple vendors and choose each component on workload fit, availability and support. The configuration sheet records why every part was chosen.
Open-weight models such as Llama, Qwen, Mistral, DeepSeek, Gemma and gpt-oss, plus embedding, speech and vision models. We test candidates on your own tasks and choose on score, speed and hardware cost.
You own the hardware outright, and the code, prompts, evaluation sets and any fine-tuned model weights we create are delivered to your own repositories and storage.
Our systems support HIPAA, SOC 2 and GDPR programs with on-premises deployment, access control, audit logging, retention settings and redaction. Compliance belongs to your organization's program, and we document how each control maps to it.
A buyer's tour of AMD, Intel, AWS, Google, Groq, Cerebras, SambaNova, Tenstorrent and Huawei silicon for running open-weight LLMs on your own hardware.
How to size on-prem GPU volume by seat count, concurrency and active use patterns, so your AI hardware matches real team load without overbuying.
The real failure modes in the first 90 days of on-premises AI, what each one costs to fix, and the scoping decisions that stop them happening.
Tell us 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.
