Distributed training
Multi-node runs across NVLink and fabric, NCCL-tuned and checkpointed, with health-aware rescheduling when a node drops.
Bare Metal · Dedicated GPU infrastructure
Dedicated GPUs, nodes, and clusters.
Deployed with our base stack — OS, NVIDIA drivers, orchestration, and monitoring — and operated by Nestor. No spot pool, no shared capacity, no ticket queue. You run the workload; we run the infrastructure underneath it.
Dedicated single-tenant nodesNVLink & multi-node fabricSLURM or KubernetesRoot access, operated by Nestor
nestor / managed environment
env-7af · liveWorkload path
01
Workload
02
Managed env
03
GPU cluster
04
API · jobs · rollouts
Status
- Capacity reserved8× H200 · region us-east
- Drivers configuredCUDA 12.6 · NCCL 2.21
- Cluster onlineslurm · 8 nodes · healthy
- Monitoring activegpu · node · job
- Support channel openshared · 1 business day
operated by nestoruptime · 14d
Why Nestor
For workloads a console and a credit card can't solve.
Self-serve GPU clouds work until they don't: spot interruptions, shared pools, generic environments, and a support ticket queue when infrastructure blocks model work. Nestor is the other model — a dedicated environment scoped to your workload, operated by a team you can actually reach.
Dedicated hardware.
GPUs reserved for you for the full term — no spot pool, no shared capacity, no noisy neighbors.
Configured around your workload.
Drivers, runtime, orchestration, storage, and access patterns tuned to your model and framework — not a generic image.
A human operator.
Direct support channel for onboarding, environment changes, and incident triage.
What you get
The machine, not a slice of one.
Dedicated hardware, wired and tuned for GPUs that saturate — from a single node to a multi-node cluster, scoped to your workload and operated by Nestor.
- Single-tenant nodes
- Whole GPUs reserved for your term — root access, your kernel modules, your containers. No hypervisor slice, no shared tenant, no noisy neighbors.
- Multi-node fabric
- NVLink inside the node and a low-latency fabric across nodes, sized to the cluster — so distributed training scales instead of stalling on communication.
- High-throughput storage
- Local NVMe plus attached shared storage, provisioned to keep the GPUs fed — checkpoints, datasets, and caches without an I/O bottleneck.
- Managed orchestration
- SLURM or Kubernetes, configured and operated by Nestor — job scheduling, multi-node launches, and health-aware rescheduling.
- Base stack, preconfigured
- OS, NVIDIA drivers, CUDA, NCCL, containers, and monitoring installed and tuned before you log in. You get a cluster that's ready, not a bare rack.
Built for
The jobs a shared endpoint can't run.
Bare Metal is for the workloads you operate yourself — compute-heavy, long-running, and multi-node. You run the job; Nestor runs the infrastructure underneath it.
Fine-tuning & post-training
Full-parameter or adapter runs on dedicated capacity — your framework, your data pipeline, your schedule.
RL / continual training
Generation, reward, eval, and training loops on capacity planned for bursty, multi-stage post-training.
Large-scale batch
Embeddings, evaluations, and data processing at fleet scale — long jobs on GPUs that are yours for the term.
Self-managed inference
Run your own serving stack when you want the GPUs, not a managed endpoint. Want us to run the model instead? That's Inference Endpoints.
Managed scope
What Nestor manages.
Six operational layers between your workload and the hardware — scoped during onboarding and run by Nestor for the duration of the commit.
Layer 01
Capacity planning
Workload mapped to GPU type, cluster size, term, and region.
Layer 02
Provisioning
Dedicated environments, not a shared pool.
Layer 03
Orchestration
Managed SLURM or Kubernetes.
Layer 04
Runtime setup
CUDA, drivers, containers, frameworks, inference engines.
Layer 05
Networking & storage
Private access, attached storage, data movement.
Layer 06
Monitoring & support
GPU health, utilization, node status, and a direct channel.
Infrastructure
GPU capacity.
Capacity Nestor deploys through its partner network — scoped, contracted, and operated per deployment.
| GPU | Memory | Interconnect | Typical use |
|---|---|---|---|
NVIDIA B300 Blackwell Ultra. Built for trillion-parameter training. | 288 GB HBM3e | NVLink | Frontier training, large-scale post-training, high-memory inference. |
NVIDIA B200 Blackwell. Native FP8 / FP4 throughput. | 192 GB HBM3e | NVLink | Training, FP8 inference, large model serving. |
NVIDIA H200 SXM Hopper. 1.4× HBM vs H100, ideal for long-context workloads. | 141 GB HBM3e | NVLink | Long-context inference, fine-tuning, memory-heavy workloads. |
NVIDIA H100 SXM Hopper. NVLink, distributed training proven. | 80 GB HBM3 | NVLink | Distributed training, fine-tuning, high-throughput inference. |
NVIDIA A100 SXM Ampere. Proven workhorse for training and inference. | 80 GB HBM2e | NVLink | Fine-tuning, research, batch inference. |
NVIDIA RTX Pro 6000 Blackwell Blackwell. 96 GB GDDR7. Fits 70B FP8 on a single card. | 96 GB GDDR7 | PCIe | Cost-efficient inference, 70B-class serving, development. |
NVIDIA L40S Ada. 48 GB GDDR6. Steady inference and lightweight fine-tunes. | 48 GB GDDR6 | PCIe | Inference, embeddings, lightweight fine-tuning. |
Onboarding
From workload to running infrastructure.
A direct path from architecture review to a stable, supported environment — scoped with your team before any capacity is committed.
- Step 01Architecture reviewYou share workload, model, framework, and timing; we scope GPU config, orchestration, and term.
- Step 02Environment setup & acceptanceWe provision, configure, and hand off; your team validates before billing starts.
- Step 03Ongoing operationNestor stays on as the infrastructure operator: monitoring, changes, support, and future capacity.
Security & access
Private environments with controlled access.
Environments are scoped to one customer. Access patterns, network boundaries, and container provenance are agreed during onboarding and operated by Nestor.
- Environments scoped to a single customer
- SSH-key access; VPN or private connectivity where available
- Private containers and custom images supported
- Export control and acceptable-use restrictions apply
Note
Nestor is in private beta. We do not currently claim SOC 2, ISO, HIPAA, or other compliance certifications. Security and compliance roadmap is available on request.
Specific networking, storage, or access requirements are scoped with your team during onboarding rather than assumed by default.
Operated by
Who operates Nestor
Nestor is operated from New York by an infrastructure team with backgrounds building GPU cloud and edge infrastructure products at AWS and leading GPU cloud providers. We work directly with a vetted network of data center and capacity partners across the US.
Contact
Inference Endpoints
Want the model served instead of the machines delivered?
Bring a model or a container and we serve it from a private, OpenAI-compatible endpoint on dedicated GPUs — and operate everything underneath.
Onboarding selected teams
Running something a self-serve cloud can't hold?
Nestor is onboarding selected AI teams in private beta. Tell us what you're trying to run — we'll follow up within one business day.