$ hardware·capabilities
Right-sized hardware,
not oversized appliances.
We size the system to your workload, not to the vendor's commission. The same GB10 silicon ships at $2,999 to $5,079 across vendors. The difference is markup. We disclose ours.
spec_sheet·3 tiers·markup disclosed
> tiers
Sized to the work.
Three configurations. One disclosed markup. The price scales with the workload, not the sales target.
- 01
Build
Developer workstation / prototyping
- workload
- Code-gen, internal RAG prototypes
- price
- $9K–$18K all-in
- hw markup
- 10–15% disclosed
- svc margin
- 30–50%
- 02
Run
Production single-node inference
- workload
- Steady load, 7B–70B open-weight models
- price
- $90K–$180K all-in
- hw markup
- 10–15% disclosed
- svc margin
- 30–50%
- 03
Floor
Multi-GPU, high-volume or edge-hardened
- workload
- Warehouse / floor deployment
- price
- $250K–$450K all-in
- hw markup
- 10–15% disclosed
- svc margin
- 30–50%
$ fit_check
We will tell you when not to buy.
Owned hardware is the wrong answer more often than vendors admit. Here is the honest gate we run before quoting anything.
[ DO NOT ]When you do not need owned hardware
- Bursty or coding-agent workloads
- Frontier multimodal needs
- Facility not ready for power / cooling / network
- IT bandwidth insufficient to operate it
- Sustained utilization below 60%
[ DO ]When you do
- High-volume, steady-state inference
- Regulated or contractually-bound data
- 7B–70B open-weight models fit the task
- 60%+ sustained GPU utilization
- 60%
- minimum sustained GPU utilization
> bundles
Pre-tuned, not blank-box.
Every tier ships with at least one agent already tuned to the work, so the box earns its keep on day one, not month three.
- A1LLMOps tuning
Coding agent
- Repo-aware code-gen
- Review + refactor
- Test scaffolding
- A2Workflow tuning
Operations agent
- Carrier onboarding
- Freight-bill audit
- Exception management
- A3Internal RAG
Knowledge agent
- Document retrieval
- Grounded answers
- Source citations
> architecture
The destination is hybrid.
Frontier models in the cloud for reasoning and burst; owned open-weight models on local hardware for the steady, sensitive baseline load.
$ next
Find out if the floor is even the right place for it.
We run the workload-fit assessment before we quote a box. If owned hardware is wrong for you, that is the answer you will get.
