Nebius vs Replicate
Nebius is an AI cloud: you rent full GPU or CPU virtual machines, with root and disks that stay, and build on them yourself. Replicate runs AI models as containers behind a prediction API, on GPUs that scale up with requests and down to zero.
The companies and products
Nebius
Nebius is an AI cloud company, listed on Nasdaq and based in Amsterdam, that rents GPUs and the infrastructure around them.
You create a Compute VM from a preset (GPU or CPU only), add disks and an SSH key, and get a full Linux VM with root and an optional public IP. Around it Nebius sells Managed Kubernetes, Slurm clusters (Soperator), Serverless AI jobs and endpoints, object storage and networking. It is built for teams that train, fine-tune and serve AI models, not as a sandbox for agents.
Replicate
Replicate runs AI models with an API: thousands of community and official models, plus your own models packaged with Cog.
You call a public model with one API request, fine-tune one on your data, or package your own model with Cog, Replicate's open-source tool that builds a Docker image with an HTTP prediction server. Replicate runs it on CPU or NVIDIA GPU hardware as deployments that scale up with traffic and down to zero. It is built for teams that serve models, not for agents that need a machine of their own, so this page compares the places where the two overlap.
Which use cases each one fits
For each use case, Battleships prices every sandbox provider on the same workload and sorts them from the cheapest monthly bill up. “29th cheapest of 364” means 28 of the other 363 providers run that workload, with every feature it needs, for less. “With limits” means the provider runs it but misses something, so it is listed after every full match. Open a use case for its whole ranking and what each provider misses.
| Use case and its workload | Nebius: price rank | Replicate: price rank | bochka: price rank |
|---|---|---|---|
| Software factory 100 at once · 4 vCPU / 8 GB · 2,200 sessions of 8 h a month A Nebius VM gives a coding agent root, Docker and a disk you set up yourself; a Replicate model is a stateless container you reach only through its API. | 197th of 367, with limits $2,272 a month | 268th of 367, with limits $4,340 a month | 3rd cheapest of 367 $634 a month |
| Always-on AI employees 50 machines 24/7 · 2 vCPU / 4 GB An assistant that lives on one machine all month fits a small Nebius VM; on Replicate each prediction stops at 30 minutes and the files do not stay. | 210th of 366, with limits $2,261 a month | 335th of 366, with limits $11,300 a month | 5th cheapest of 366 $657 a month |
| Code interpreter bursts 100 at once · 1 vCPU / 2 GB · 200,000 sessions of 1 min a month Short code runs suit Replicate's per-second, scale-to-zero billing, while Nebius needs a VM per job and a quota raise for many at once. | 111th of 345, with limits $202 a month | 258th of 345, with limits $295 a month | 186th of 345, with limits $2,000 a month |
| Sandboxes for your users 300 at once · 2 vCPU / 4 GB · 30,000 sessions of 23 min a month Nebius can host your users' sandboxes after a quota raise above 12 VMs per region; Replicate gives them a model endpoint, and bochka gives each a full VM with hosting. | 81st of 361, with limits $716 a month | 317th of 361, with limits $3,705 a month | Cheapest of 361 $209 a month |
| RL rollouts 1,000 at once · 2 vCPU / 4 GB · 1,000,000 sessions of 3 min a month | 52nd of 360, with limits $3,049 a month | 174th of 360, with limits $17,900 a month | 34th of 360, with limits $2,000 a month |
| Agent evals 250 at once · 4 vCPU / 8 GB · 15,000 sessions of 20 min a month | 29th cheapest of 364 $610 a month | 67th cheapest of 364 $1,440 a month | 10th cheapest of 364 $200 a month |
| CI runners 40 at once · 8 vCPU / 16 GB · 6,000 sessions of 12 min a month A Nebius CPU VM works as a CI runner with root and Docker; Replicate has no shell and does not document Docker inside. | 27th cheapest of 377 $297 a month | 311th of 377, with limits $864 a month | 4th cheapest of 377 $86 a month |
| 24/7 dev boxes 1 machine 24/7 · 4 vCPU / 8 GB A dev box needs SSH and a disk that keeps your installs, which Nebius and bochka have and Replicate does not. | 68th cheapest of 368 $91 a month | 306th of 368, with limits $226 a month | 18th cheapest of 368 $26 a month |
| Computer-use agents 50 at once · 4 vCPU / 8 GB · 20,000 sessions of 15 min a month Neither Nebius nor Replicate has a desktop or browser product, while bochka streams a 60fps desktop with Chromium on every sandbox. | 187th of 369, with limits $610 a month | 263rd of 369, with limits $1,296 a month | 6th cheapest of 369 $200 a month |
| Browser-use agents 50 at once · 2 vCPU / 4 GB · 30,000 sessions of 10 min a month | 223rd of 367, with limits $305 a month | 322nd of 367, with limits $1,296 a month | 7th cheapest of 367 $200 a month |
| Scraping & bots (startup scale) 50 at once · 1 vCPU / 2 GB · 100,000 sessions of 3 min a month | 253rd of 346, with limits $302 a month | 272nd of 346, with limits $439 a month | 6th cheapest of 346 $90 a month |
| Mass scraping & bots (undetectable) 1,000 at once · 1 vCPU / 2 GB · 3,000,000 sessions of 2 min a month | 174th of 338, with limits $6,049 a month | 221st of 338, with limits $8,775 a month | 71st of 338, with limits $1,800 a month |
| Parallel red teaming 200 at once · 2 vCPU / 4 GB · 5,000 sessions of 30 min a month | 176th of 360, with limits $156 a month | 306th of 360, with limits $855 a month | 2nd cheapest of 360 $200 a month |
| GPU eval jobs 10 at once · 8 vCPU / 32 GB · 500 sessions of 30 min a month Both fit GPU eval jobs, and Nebius ranks 18th against Replicate at 22nd; bochka has no GPUs, so this one belongs to them. | 18th cheapest of 98 $1,127 a month | 22nd cheapest of 98 $1,338 a month | Can't run it |
| macOS agents 5 at once · 4 vCPU / 8 GB · 1,000 sessions of 20 min a month | Can't run it | Can't run it | Can't run it |
Key features
| Feature | Nebius | Replicate | bochka |
|---|---|---|---|
| Kind of product | AI cloud: GPU and CPU VMs, Kubernetes, Slurm clusters, Serverless AI and storage | Model inference API: public models, fine-tunes and your own Cog models as autoscaling deployments | Sandboxes for AI agents: full VMs with agents, a desktop and hosting ready |
| What you get | A full VM behind a hypervisor, with its own kernel and root | A model packaged as a Cog container; the docs do not name the isolation layer | A full VM with its own kernel on dedicated AMD Ryzen 9 9950X hosts |
| Shell access | SSH and root on your own VM | No SSH; you call the model through the HTTP API or client libraries | SSH, root and systemd, like a normal Ubuntu server |
| Docker and nested virtualization | Root on your VM and container VMs from Docker images; nested virtualization is not documented | The model is the container; Docker inside and /dev/kvm are not documented | Docker, systemd and /dev/kvm for Android emulators and VMs on every machine |
| GPUs | Nvidia L40S, RTX PRO 6000, H100, H200, B200 and B300, on-demand or preemptible | T4, L40S, A100 80GB, H100 and H200, up to 8 per instance; H200 and most multi-GPU need a committed spend contract | None, CPU only |
| Ready-made models | None in Compute; you bring the model, or run it as a Serverless AI job or endpoint | Thousands of public and official models for images, video, speech and text, plus fine-tuning | None; your agent calls whatever model API you choose |
| Scale to zero and wake | No; you start and stop VMs yourself | Yes; deployments scale to zero and start again on the next prediction | No; a stopped sandbox is started by API |
| Files after a stop | Disks stay, and keep billing on their full provisioned size while the VM is stopped | No persistent disk; instances start fresh from the image, and API inputs and outputs are deleted after an hour by default | The whole disk is saved every minute and on stop, and a stopped sandbox keeps it for free |
| Snapshots and fork | Manual disk snapshots; to fork, make a disk from the snapshot, then a VM from the disk | No snapshots or fork; you add replicas of the same image | Snapshots every minute and on stop; fork a running sandbox in one call |
| Run time | Regular VMs run until you stop them; preemptible ones can be stopped at any time | Instances can stay on, but each prediction times out after 30 minutes unless support raises it | No session cap on paid plans; you set the time to live |
| Biggest machine | CPU VMs up to 256 vCPU / 1,024 GB | CPU tiers of 1 or 4 vCPU; GPU instances up to 8 GPUs | 16 vCPU / 32 GB on the $100 plan and up |
| Reaching the machine | Optional public IPv4 per VM, SSH and Nebius Tunnels; you set up TLS yourself | One prediction API per model behind API token auth; no own IP or ports | HTTPS URLs for up to 50 ports, token-protected; raw TCP; dedicated IPv4 or IPv6 |
| Agents, desktop and keys | No agents or desktop built in; SecretStash hands secrets to the VM itself | No agents or desktop; secrets are sent to the model as inputs and redacted from logs | Seven coding agents with a prompt, steer and events API, a 60fps desktop, and a credentials proxy that keeps keys outside |
| Regions | Finland, France, Spain, United Kingdom, Israel and the US | Infrastructure subprocessors listed in the United States; no region choice documented | EU only: Germany, Finland, France |
Every feature Battleships tracks (43)
| Feature | Nebius | Replicate | bochka |
|---|---|---|---|
| Full VM (own kernel) | Yes | No | Yes |
| Docker inside | Yes | No | Yes |
| Nested virtualization | Not published | No | Yes |
| Root | Yes | Not published | Yes |
| Systemd | Not published | Not published | Yes |
| Desktop GUI | No | No | Yes |
| Browser | No | No | Yes |
| Computer-use API | No | No | Yes |
| SSH | Yes | No | Yes |
| Persistent disk | Yes | No | Yes |
| Pause/resume | Yes | No | Yes |
| Automatic snapshots | No | No | Yes |
| Fork/clone | Yes | No | Yes |
| Live fork (no pause) | Yes | No | Yes |
| Memory fork | No | No | No |
| Custom image (Docker or snapshot) | Yes | Yes | Yes |
| ≥24 h sessions | Yes | Yes | Yes |
| Idle auto-stop | No | Yes | No |
| Wake on request | No | Yes | No |
| HTTPS preview URLs | Yes | Yes | Yes |
| Raw TCP inbound | Not published | Not published | Yes |
| Public IPv4 | Yes | No | No |
| Egress allowlist | Yes | No | No |
| Secret proxy | No | No | Yes |
| Secret proxy for any API | No | No | Yes |
| Preinstalled agents | No | No | Yes |
| Agent harness API | No | No | Yes |
| MCP server | No | Yes | No |
| GPU | Yes | Yes | No |
| Windows | No | No | No |
| Arm64 | No | No | No |
| Self-hosting | No | Yes | No |
| BYOC | No | No | No |
| Open source | No | Yes | No |
| Scoped API keys | Not published | Not published | Yes |
| Spend limits | Not published | Yes | Yes |
| Usage per sandbox (API) | Yes | No | Yes |
| Audit logs | Not published | Yes | No |
| SSO | Yes | No | No |
| SOC 2 | Yes | Not published | No |
| HIPAA | Yes | Not published | No |
| EU data residency | Yes | No | Yes |
| Zero data retention | Not published | Not published | Yes |
Performance
There is no comparable measurement for either: neither Nebius nor Replicate was part of the public StarSling benchmark run, so we cannot say which is faster at CPU work. bochka did 26.69 Node.js build-and-test runs per second at 4 vCPU / 8 GB in that run, on dedicated AMD Ryzen 9 9950X hosts. Nebius CPU VMs run on AMD EPYC Genoa or Intel Ice Lake, and Replicate's hardware is chosen for GPU inference.
Node.js build-and-test loop at 4 vCPU / 8 GB, higher is better. Measured with the open hpc-sandbox-benchmarks suite. Nebius and Replicate have no published run in this suite yet, so there is no like-for-like number to compare.
Pricing in five scenarios
Each bill is priced from the providers' public price lists by the open Battleships engine, with the cheapest plan that runs the scenario.
| Scenario | Nebius, a month | Replicate, a month | bochka, a month |
|---|---|---|---|
| One coding agent, workdays 4 vCPU / 8 GB, 50 GB disk, 8 hours a day, 22 days a month, files kept between days | $25 · Pay as you go | $57 · Usage · with limits | $20 · $20/mo |
| One agent that never stops 4 vCPU / 8 GB, 50 GB disk, running 24/7 all month | $91 · Pay as you go | $237 · Usage · with limits | $26 · $20/mo |
| 50 agents in parallel 50 machines of 4 vCPU / 8 GB at once, 8-hour sessions, 1,100 sessions a month, Docker inside | $1,127 · Pay as you go | $2,851 · Usage · with limits | $317 · $20/mo × 3 seats |
| Sandboxes for your users 2 vCPU / 4 GB, 30,000 sessions of 10 minutes a month, up to 300 at once | $305 · Pay as you go | $1,620 · Usage | $200 · $100/mo × 2 seats |
| 1,000 short jobs at once 2 vCPU / 4 GB, 1,000 at once, 3 minutes each, 200,000 jobs a month | $610 · Pay as you go | $3,580 · Usage | $2,000 · $2000/mo · with limits |
Nebius: Nebius bills running VMs per second for vCPU and RAM with no plan fee, and bills disks by provisioned size even while a VM is stopped. From 1 October 2026 AMD Genoa costs $0.015 per vCPU-hour plus $0.0045 per GiB-hour, and an H100 is $4.50 per GPU-hour on demand.
Replicate: Replicate bills per second by hardware tier, from prepaid credit that expires after a year. Public models bill only while they run; private models and deployments bill setup, idle and active time.
Replicate: Plans start at $20 a month, and the whole fee is usage credit spent per second on running sandboxes. A stopped sandbox costs nothing and keeps its disk.
Limits in those scenarios
Nebius starts each region at 12 VMs and, in most regions, 200 vCPU across CPU-only VMs, so 50 agents in parallel, 300 user sandboxes and 1,000 short jobs all need a quota raise first; its smallest CPU VM is 2 vCPU / 8 GB, so the table prices it with twice the memory. Replicate does not fit the workday agent, the always-on agent or the 50-agent fleet, because it has no persistent disk, and the fleet also needs Docker inside, which is not documented. Each Replicate prediction stops after 30 minutes unless support raises it, and it publishes no disk or egress price. bochka queues the 1,000-job burst, because its $2,000 plan allows fewer starts per day than that needs.
| Scenario | Nebius | Replicate | bochka |
|---|---|---|---|
| One coding agent, workdays | None that bite |
| None that bite |
| One agent that never stops | None that bite |
| None that bite |
| 50 agents in parallel | None that bite |
| None that bite |
| Sandboxes for your users | None that bite |
| None that bite |
| 1,000 short jobs at once | None that bite |
|
|
Which one should you use?
Use Nebius when
you need GPUs or a cluster to train or run models, a full VM with root and a lasting disk, a very large machine, or a region in the UK, Israel or the US.
- The newest GPUs and real clusters. For training and big GPU jobs, Nebius rents L40S up to B300 and runs Slurm and Kubernetes clusters around them.
- A whole VM that is yours. When you need root, SSH, Docker images and a disk that stays, a Nebius VM is a normal Linux server you control.
- Big machines, more regions, network rules. For one huge job or a region outside the US, Nebius VMs reach 256 vCPU and support egress rules by IP range.
Use Replicate when
you want to call or serve an AI model on GPUs, from its catalog or packaged with Cog, as an API that scales to zero when nobody calls it.
- Thousands of models, one API call away. Teams that want an image, video, speech or language model without hosting it can start with one request.
- Scale to zero, billed per second. Models that wait for traffic scale down to zero and wake on the next prediction, and public models bill only while running.
- Cog runs anywhere. Cog, the open-source tool that packages models for Replicate, builds standard Docker images you can also run on your own servers.
Where bochka fits
Nebius and Replicate are both built for GPU work: renting GPU machines, or serving models. If what you are really running is agents that need a machine of their own, bochka gives each one a full VM with coding agents, a desktop, Docker and /dev/kvm ready. Keep the GPU part on Nebius or Replicate and let your agents call it from bochka.
Agents and a desktop come ready
On Nebius you install the agents, a desktop, a VNC server and a browser on each VM and keep them working. Replicate has no shell, desktop or agent harness, and each prediction stops at 30 minutes.
The disk is saved every minute, and you can fork it
Nebius snapshots only when you take one, and a stopped VM still pays for its disks. On Replicate an agent's files are gone when the instance scales down, so each run sets itself up again.
Real keys stay outside the machine
bochka's credentials proxy works with model providers and any HTTPS API. Nebius SecretStash hands the secret to the VM, and Replicate passes secrets in as inputs, where a prompt-injected agent could read them.
Use bochka when you run agents that need a machine of their own: a full VM with coding agents, a desktop, Docker, /dev/kvm, HTTPS ports and a credentials proxy, saved every minute and free while stopped. Compare it one to one: bochka vs Nebius and bochka vs Replicate.
Questions people ask
Is Nebius cheaper than Replicate?
In every scenario we priced, yes. One agent that never stops is $91 against $237 a month, 50 agents in parallel are $1,127 against $2,851, and 1,000 short jobs at once are $610 against $3,580. Replicate also does not fit the three agent scenarios, because it has no persistent disk.
Which is faster, Nebius or Replicate?
We cannot say: neither was part of the public StarSling CPU benchmark run. bochka did 26.69 runs per second on that test at 4 vCPU / 8 GB.
Can Nebius or Replicate run Docker?
Nebius gives you root on your own VM and can start container VMs from Docker images. On Replicate, Cog builds your model into a Docker image, but running Docker inside that container is not documented. A bochka sandbox runs Docker, systemd and /dev/kvm like a normal Ubuntu server.
Do my files stay after a stop?
On Nebius, yes: disks stay, but they keep billing on their full size while the VM is stopped. Replicate has no persistent disk, and API inputs and outputs are deleted after an hour by default. bochka saves the whole disk every minute and a stopped sandbox keeps it for free.
Which has more GPUs?
Nebius rents L40S, RTX PRO 6000, H100, H200, B200 and B300. Replicate offers T4, L40S, A100 80GB, H100 and H200, and H200 and most multi-GPU setups need a committed spend contract. bochka has no GPUs, so your agents in bochka can call either one.
Is bochka SOC 2 compliant?
SOC 2 is in progress. We're open about where we stand: live status at status.bochka.xyz and our data retention policy at docs.bochka.xyz/data-retention.
Try bochka
Install the CLI, then run bochka new.
Product preview