Compose your cluster.
Train your model.
Serve it.

They are infrastructure providers — they rent you GPUs from their own datacenters and regions. ResonTech is the layer above: a decentralized datacenter you compose yourself. Plug in capacity from any of those providers, plus your on-prem racks, your reservations, or partner clusters, and treat them as one pool. Submit a single job; the platform places workers across the mix, coordinates training, and serves the resulting model — with minimal cross-cluster overhead. We aren't competing with their infra; we sit on top of it.

Not a rewrite, but a thin adaptation — typically 50–100 lines. You define a model class, a dataloader, and the training step the way you normally would in PyTorch / TensorFlow / JAX, then expose them through a small Executor + Persistor scaffold so the platform can shard data, dispatch workers, and aggregate model deltas across clusters. We're shipping a Claude plugin that ports an existing training script to that scaffold automatically — paste your repo, get back a submit-ready job in minutes.

Public Pool bills per-minute on allocated GPU time, the same model as RunPod or Lambda — you pay for the minutes a worker is reserved to your job, with no monthly minimums and no pre-purchased credits. Managed Cluster is a flat monthly reservation for dedicated capacity. Private Fabric is per-cluster licensing on your own hardware. Inference endpoints bill per active replica-minute and can scale replicas down between traffic. Egress, storage, and NAT are itemized line-by-line — no hidden tail in the invoice.

The fabric detects the failure, evicts the bad node from the worker pool, and resumes from the last checkpoint automatically. You receive a notification but no manual intervention is required. For long multi-cluster runs this can save dozens of GPU-hours you would otherwise rerun from scratch.

You can host the bucket yourself (any S3-compatible store). Workers fetch only their assigned shard through short-lived presigned URLs; the control plane never proxies raw bytes. Managed Cluster lets you pin storage per region. Private Fabric runs the entire control plane inside your perimeter, with air-gapped mode for sensitive environments. Data-sovereign mode keeps data anchored across organizations for federated training and federated inference.

Install the Python SDK with one pip install, point it at your model or training script, and submit. First job typically runs within minutes on the public pool. No infrastructure provisioning, no cloud-account setup, no support ticket to request quota.

H100 SXM5, A100 80GB, and A40 nodes depending on availability and priority tier. Managed Cluster reserves specific GPU types — H100 NVLink, A100 PCIe, L40S — for your team. Private Fabric runs on whatever you bring (B100/B200, H200, A100, L40S, RTX-class, mixed pools all supported).

Across clusters the platform runs federated-style coordination — each cluster trains a local round (multiple SGD steps), then exchanges only model deltas with the central aggregator over gRPC + TLS. That replaces per-step all-reduce, which would die on WAN latency. The aggregator combines updates using a chosen algorithm — FedAvg by default, with FedOpt, FedProx, Scaffold, and DiLoCo selectable per job. Deltas are compressed and the sync interval auto-adjusts to the observed link bandwidth, so it tolerates anything from 10 Gbps cloud interconnect down to public internet. Final model quality typically lands within a few percent of centralized training; the wall-clock cost is the round-trip latency between sites.