Capacity planning first — access coordinated on request.
ArcaSys helps you define the footprint you need and coordinates delivery through the integration path. Specific GPU classes, pricing, and timelines are confirmed per engagement.
Capacity catalog
Verified configurations only.
Empty until rows are substantiated. No invented NVIDIA SKUs or performance numbers.
This catalog is reserved for configurations that ArcaSys can substantiate. Until a row is marked Verified, fields remain “On request” or “Not published.” We do not list unverified SKUs or performance figures.
Official
Vendor-published specification cited with product name and document version.
Verified
Configuration confirmed in a real engagement or lab note with date and environment.
Proposed
Planning suggestion only — not validated. Shown only when clearly labeled.
| GPU model | VRAM | Node topology | Interconnect | Max nodes | Suited for | Region | Lead time | Billing | Status |
|---|---|---|---|---|---|---|---|---|---|
| No verified capacity rows published yet. All GPU models, interconnects, regions, lead times, and billing modes remain On request until substantiated. | |||||||||
To request a configuration, send GPU class preferences, VRAM needs, node count, interconnect constraints, region preference, and target start date via Contact. We will respond with what can be confirmed — and what remains open.
Capacity shapes
Talk in shapes, confirm per request.
These are planning categories — not a published SKU or price list.
Single-node discussions
For prototyping, fine-tuning, and smaller jobs — we help size a single-node ask before requesting capacity.
Multi-node planning
For larger training or parallel pipelines — topology and node count are planned with your constraints in view.
Scheduled windows
When milestones matter, we discuss reserved windows as a planning item; confirmation depends on available supply.
Reference architectures
Suggested planning topologies.
Labeled as suggestions — not verified designs — until real deployment evidence exists.
Training — suggested planning topology
Suggested configuration only. Not a verified design unless marked with deployment evidence.
- GPU compute nodes (count TBD)
- Shared or parallel filesystem (TBD)
- Job scheduler / orchestration (TBD)
Interconnect, storage bandwidth, and checkpoint strategy are defined during scoping.
Inference — suggested planning topology
Suggested configuration only. Latency and concurrency targets require measurement before acceptance.
- Serving nodes (count TBD)
- API / gateway layer (TBD)
- Observability plane (TBD)
Engine choice (e.g. vLLM / TensorRT-LLM / Triton) is selected per model and only after compatibility review.
Render / media — suggested planning topology
Suggested configuration only. Throughput claims require a measured pipeline.
- GPU workers (count TBD)
- Asset storage (TBD)
- Queue / batch controller (TBD)
Codec, resolution, and batch size drive sizing; no public throughput numbers are published yet.
Higher-isolation — suggested planning topology
Suggested configuration only. Isolation controls must be named in the contract.
- Dedicated GPU nodes (TBD)
- Network / tenancy boundary (TBD)
- Access & audit controls (TBD)
Shared vs dedicated vs private is a contracting choice, not a default product feature.
Software stack
Software stack & compatibility
ArcaSys does not yet publish a verified CUDA / runtime / inference-engine matrix. When a stack is proposed in an engagement, we will cite vendor compatibility documents and record the versions and last review date. Until then, treat any stack discussion as planning input only.
Compatibility matrices will cite vendor documents with version and last review date when available.
How capacity discussions work
Three steps from inquiry to a delivery plan.
Clear sequence. Commercial and technical details filled in case by case.
01
Share the workload
Model family, expected scale, and when you need capacity.
02
Agree a plan
We outline options, open questions, and a practical next step.
03
Coordinate delivery
Once terms are clear, we help arrange access on request.