nestor
Talk to us
nestor
Talk to us

Infrastructure for the AI economy

Fastest way to train and serve models

Dedicated GPU clusters built to your spec in days, and managed inference that gets the most out of every token you serve.

Talk to us

Speed matters.

In the age of AI, where all things are converging, technology differentiation is getting narrower. For intelligence abundance, we owe it to ourselves not to wait months for GPU clusters to become available, or to spend countless iterations getting a model in front of an audience.

10 days

Brendan Eich built the first prototype of JavaScript in ten days in May 1995. It shipped in beta that September.

Source: Brendan Eich, a history of the language

3 weeks

Ken Thompson wrote the first version of Unix in three weeks.

Source: UNIX — A History and a Memoir

Nestor does what the hyperscalers can't

Nestor The hyperscalers
Getting capacity Moving fast, from a single conversation. Weeks spent in negotiation.
Contract structure Transparent, with one rate and no strings attached. Stacked, vague contracts that come with strings attached.
Support Real engineering and human support for your hardware. A two-day wait to talk to someone who doesn't control the hardware, then endless escalations.

Getting capacity

Nestor

Moving fast, from a single conversation.

The hyperscalers

Weeks spent in negotiation.

Contract structure

Nestor

Transparent, with one rate and no strings attached.

The hyperscalers

Stacked, vague contracts that come with strings attached.

Support

Nestor

Real engineering and human support for your hardware.

The hyperscalers

A two-day wait to talk to someone who doesn't control the hardware, then endless escalations.

Product 01 — GPU Clusters

Your cluster, built to your spec, in days

Blackwell-class compute on non-blocking InfiniBand, with a Weka parallel filesystem underneath so the fabric is never waiting on storage.

Reference architectureNVIDIA B300
Compute

NODE 01

8× B300

NODE 02

8× B300

NODE 03

8× B300

NODE N

scale to spec

Fabric
InfiniBand NDR · leaf / spine RoCE v2 available
Storage
Weka parallel filesystem checkpoint and dataset tier

Compute

NODE 01

8× B300

NODE 02

8× B300

NODE 03

8× B300

NODE N

scale to spec

Fabric

InfiniBand NDR · leaf / spine

RoCE v2 available

Storage

Weka parallel filesystem

checkpoint and dataset tier

Schedulers

Slurm and Kubernetes

Validation

NCCL benchmarks on handover

Access

Role-based, org-wide

Also available

B200, A100, RTX PRO 6000

Product 02 — Inference

Maximize your token performance

Push a container and get an endpoint. Routing, batching and LoRA caching are handled for you, with no servers to manage on your side.

Request routingContainerized · no servers to manage
Nestor inference request routing Chat completion requests and image generation requests both enter the Nestor inference engine. The engine routes chat completions to deepseek-v4-pro and deepseek-v4-flash, and image generations to the Flux ComfyUI endpoint. Chat completions Image generations deepseek-v4-pro deepseek-v4-flash flux · comfyui graph chat, agents, long context media endpoint 1.6T MoE · 1M context 284B MoE · low latency native ComfyUI graphs Nestor inference engine route · continuous batching LoRA + model cache FP8 decode on B300 autoscale to zero

Chat completions

chat, agents, long context

Image generations

media endpoint

Nestor inference engine

route · continuous batching

LoRA + model cache

FP8 decode on B300

autoscale to zero

deepseek-v4-pro

1.6T MoE · 1M context

chat

deepseek-v4-flash

284B MoE · low latency

chat

flux · comfyui graph

native ComfyUI graphs

media
Ship your first request
cURL 
curl https://api.serve.nestor.software/v1/endpoints/$ENDPOINT_ID/jobs \
  -H "Authorization: Bearer $NESTOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input":{"prompt":"Explain MoE routing"}}'
202 · request accepted Explore the API

Put your next training run on Nestor

Talk to us