AI inference for your team or practice
Self-hosted LLM hosting on a dedicated GPU rig: the best open-weight coding and reasoning models, served as an OpenAI-compatible API. We run the infrastructure. You call the endpoint. No per-token bills. No data sent to a public API. Flat monthly fee.
You are paying for AI twice, and handing your data to someone else to do it.
The per-token bill is unpredictable. One runaway agent loop and a team that seemed frugal turns into a four thousand dollar invoice. Meanwhile your client data, your patient data, and your proprietary code are flowing into a public API you do not control, and your lawyer or compliance officer is the one who eventually has to explain why. The honest answer to "should we self host" is usually "yes, but we do not want to hire an AI infrastructure person to babysit a GPU server."
Private, flat-priced AI is the version of AI your finance team and your legal team will both sign off on.
A dedicated AI inference rig. We run it. You call it.
A 128GB AI rig running open-weight models like Qwen3.6, DeepSeek V4 Flash, and GLM 5.2, served as an OpenAI-compatible inference endpoint. Single-tenant, hosted on dedicated hardware we operate, with your data isolated from public APIs and never logged. You get an API key and a URL. Point Cursor, Cline, Continue, LangChain, or the OpenAI SDK at it and you are running. No per-token fees, no rate limits, no content refusals on legitimate professional text, and no surprise invoices.
Pick the rig that fits who you serve
Both are 128GB, but built for opposite jobs. The Practice rig is for one professional. The Workgroup rig is for a whole team. Same flat-fee, no-per-token model either way.
- NVIDIA DGX Spark, 128GB unified memory
- 1 to 3 users, one professional or a tiny team
- Runs 70B to 120B models in one flat memory pool, no offload
- HIPAA-ready, BAA available, privileged-data safe
- Your fine-tunes, including domain or uncensored weights
- OpenAI-compatible API, we run the box
- 4x RTX 5090 (128GB VRAM) + 512GB DDR4 offload
- 30 to 40 concurrent interactive sessions on the fast model
- Qwen3.6 35B A3B for the whole team, DeepSeek V4 Flash + GLM 5.2 for hard tasks
- Dedicated, no noisy neighbors, no shared cloud pool
- Your fine-tunes, unlimited tokens, no per-token overages
- OpenAI-compatible API, we run the stack
No per-token charges. No overages. No egress fees. No rate limits. Month-to-month after 60 days.
Open-weight models that fit, served OpenAI-compatible
The rig runs the best open-weight coding and reasoning models that fit in 128GB. For most professional and development work, these now match or beat the frontier on the tasks that matter, at a fixed cost. Switch models in one line; bring your own weights.
Built for teams that cannot afford "the cloud AI did a weird thing"
Why this, not the usual options
Why not just use OpenAI?
Per-token volatility, data leaving your control, rate limits, content refusals, and model deprecations that force rewrites. Use them for the 15% of calls that need frontier reasoning. Use us for the 85% that need to be private, cheap, and predictable.
What if the hardware goes down?
We target 99.5% uptime with auto-restart, and route your traffic to a standby OpenAI key if the rig is ever unavailable, so your team stays online the whole time.
Can I trust a small operator?
The founder is on your account directly, no ticket queue. Signed contract, DPA and BAA, 30-day money-back. Trust comes from the contract, not headcount, and you can reach a human who knows your setup.
Common Questions
Can it run Claude, GPT, or Gemini?
No. Those are closed models. The rig runs open-weight models that fit in 128GB - Qwen3.6, DeepSeek V4 Flash, GLM 5.2, Llama, and your own fine-tunes. For the small share of work that genuinely needs a frontier closed model, keep a low-limit OpenAI or Anthropic key and we can route to it as a fallback.
Is it actually private?
Yes. Single-tenant hardware, no shared GPU pool, no prompt or completion logging on our side, your API keys, your access logs, and your data is wiped at contract end. Your data does not go to a public API.
Is it HIPAA-ready?
The rig is single-tenant, no-logging, data-isolated hardware hosted in Kingston, Ontario - built for regulated workloads. We say infrastructure, not a certified product, so we do not claim "HIPAA certified," and we do not promise a BAA out of the box. Whether it fits your HIPAA or regulated-workflow program is your compliance call. We operate the private hardware and work with your reviewer on what you need.
How many people can use it at once?
It depends on the models you run and how demanding the work is. A small fast model serves a whole team at once, while large reasoning models handle a few hard queries at a time with a queue. Tell us your team size and workload and we size the rig and model mix to fit.
What if the hardware fails?
We target 99.5% uptime with automated healthchecks and auto-restart, and we keep a low-limit OpenAI or Anthropic key on file so your traffic keeps flowing if the rig is ever down. Your team stays online while we sort it out.
Am I locked in?
No. 60-day initial commitment, then month-to-month with 30 days notice. No per-token overages ever. Bring your own editor (Cursor, Cline, Continue) or use the API directly.
How fast can we get started?
A first endpoint is usually live within a few days of a signed agreement. We load the models you want, hand you an API key, and you point your existing tools at the URL.
Reserve a rig
Tell us a bit about your team or practice and what you want to run. We respond within one business day with the right tier and next steps. No per-token surprises, ever.