Reserved GPUs. Contract terms. Salad orchestration.
Salad Dedicated is reserved infrastructure: bare metal servers and virtual machines you hold on contract, with no preemption, deployed across Salad's distributed footprint. The first GPU is the NVIDIA RTX PRO 6000 Blackwell, and thousands of them are coming online.
Dedicated is sold by quote, not self-serve. Allocations are reserved in order of signed agreement.
For workloads that have outgrown interruptible compute.
Community Cloud is the lowest-cost GPU compute in the market. Some workloads still need more than it can give: more VRAM, a fixed footprint, a data-center boundary, or the certainty that an instance will be there at 3 a.m. That is what Dedicated is for.
Predictable capacity
Your GPUs are yours for the term. Plan launches, SLAs, and budgets against a fixed footprint instead of a marketplace.
96 GB of VRAM
Serve 70B-class models in production quantizations on a single card, run longer contexts, and fine-tune without sharding gymnastics.
Distributed by design
Reserved cards are placed across Salad's distributed footprint, not concentrated in one building, so an outage at any single site never takes your whole allocation with it.
Bare metal or VM
Take a whole server for maximum control, or per-GPU virtual machines when you want to start smaller and grow.
Salad orchestration, optional
Run Dedicated nodes as a private GPU class in your SaladCloud organization and schedule container groups onto them with the same API, or manage the machines yourself.
Burst to the network
Hold your baseline on Dedicated and overflow to Community Cloud for spikes. One scheduler, two pools, the lowest blended cost available.
NVIDIA RTX PRO 6000 Blackwell.
The workstation-class Blackwell part with data-center memory: 96 GB of GDDR7 with ECC, fifth-generation Tensor Cores with FP4, and PCIe Gen 5. It runs the models that matter today on one card, at a cost per GPU-hour that H100-class clouds cannot approach.
| Architecture | NVIDIA Blackwell |
|---|---|
| Memory | 96 GB GDDR7 with ECC |
| CUDA cores | 24,064 |
| Tensor Cores | 5th generation, FP4 / FP8 / FP16 |
| Memory bandwidth | ~1.6 TB/s |
| Interface | PCIe Gen 5 x16 |
| Form factors | Single-GPU and multi-GPU servers; per-GPU VMs |
Specifications per NVIDIA published data. Additional GPU classes will be added as capacity comes online.
Four workloads that belong on Dedicated.
Sustained inference
Models serving traffic around the clock, where a fixed fleet is cheaper than paying for elasticity you never use.
Large models and long contexts
70B-class LLMs, video generation, and long-context RAG that need 96 GB on one device rather than sharding across consumer cards.
Fine-tuning and training
Multi-day jobs that cannot tolerate interruption and benefit from ECC memory and consistent hardware.
Single-tenant requirements
Workloads that must run on hardware no one else touches, with a contract that says so.
From conversation to capacity.
Scope
Tell us the workload, VRAM, GPU count, and term. An engineer confirms fit and recommends bare metal, VMs, or a hybrid with Community Cloud.
Reserve
Sign a contract for your allocation. Terms range from monthly to multi-year; longer terms carry better rates.
Provision
Servers or VMs are provisioned with your OS image or as Salad-managed nodes in your SaladCloud organization.
Run
Deploy with SSH, your own tooling, or Salad Container Engine. Monitoring, support, and hardware replacement are included.
Pick the pool, or use both.
| Community Cloud | Salad Dedicated | |
|---|---|---|
| Hardware | Consumer RTX 3060–5090, 8–32 GB | RTX PRO 6000 Blackwell, 96 GB (more classes coming) |
| Placement | Residential and small-business nodes in 190+ countries | Reserved, single-tenant hardware across Salad's distributed footprint |
| Interruption | Interruptible; priority tiers and automatic reallocation | None. Reserved for the term |
| Access | Managed containers via SCE | Bare metal, VMs, or SCE-managed |
| Pricing | Per second, from $0.015/GPU-hr, no commitment | Contract, by quote; monthly to multi-year terms |
| Best for | Stateless inference, batch, simulation at the lowest cost | Sustained serving, large models, training, single-tenant needs |
What buyers ask.
When can I get capacity?
RTX PRO 6000 Blackwell capacity is being brought online now and allocated in order of signed agreement. Talk to sales to reserve a position and get a delivery window for your GPU count.
What does it cost?
Dedicated is priced by quote based on GPU count, term, and form factor. Longer terms and larger allocations carry lower per-GPU-hour rates. We publish Community Cloud prices openly and will show you the blended cost of a Dedicated baseline plus Community Cloud burst.
Can I use my own images and tooling?
Yes. Bare metal and VMs come with your choice of OS image and full root access. If you prefer, we manage the nodes as a private GPU class in your SaladCloud organization so you deploy containers with the same API you already use.
Will other GPUs be offered?
Yes. RTX PRO 6000 Blackwell is first because it offers the best VRAM per dollar for inference today. Additional classes will be listed here as they come online; tell sales what you need.
What about support and hardware failures?
Monitoring, a named engineering contact, and hardware replacement are included in every Dedicated agreement. Service levels are specified in the contract.
Reserve your allocation.
Tell us the GPUs, the workload, and the term. We'll come back with a quote and a delivery window.