GPU cloud aggregators and marketplaces compared (2026)

Buying GPU compute in 2026 means choosing between a lot of tools that sound alike. Some are GPU cloud aggregators and some are marketplaces. Others are orchestrators, price indices, futures venues or decentralized networks. They solve different parts of the problem, and most people searching for "the best GPU marketplace" end up comparing products that were never meant to do the same job.

This guide sorts the main GPU compute tools into six groups. For each group it covers what the tools do and where they stop, so you can tell which one fits your workload.

Up front: we are Opengrid, and we are building in this space. Opengrid is pre-launch. The market view ships first and the routing API is in development. We have tried to describe every company here fairly, using public information as of September 2026.

Why there are so many kinds of tool

SemiAnalysis tracks 209 GPU clouds, up from 124. Each one has its own console, contract, API and bill. Prices for the same chip vary widely: an on-demand H100 lists anywhere from about $1.25 to $11.07 an hour depending on the provider. Prices also move fast. AWS cut its H100 on-demand price by 44% in a single announcement in June 2025, and one-year H100 contract prices then rose about 40% between October 2025 and March 2026.

Every tool below handles part of that mess. None of them yet handles all of it.

1. Price comparison sites

Examples: getdeploying, ComputePrices, gpus.io

These sites list GPU prices across providers so you can compare them in one place. They are a good starting point for research, and they are free.

Where they stop: you can look, but you cannot buy. Once you have found a price, you still go to that provider, open an account and launch the instance yourself.

2. Orchestrators

Examples: SkyPilot (raised a $20M seed in July 2026), TensorPool (Y Combinator W25), dstack

Orchestrators sit between your code and the clouds you already use. TensorPool, for example, is a CLI that runs a training job on the cheapest GPU it finds across clouds. These tools are popular with ML engineers because they cut down the scripting needed to run jobs in more than one place.

Where they stop: they run on cloud accounts you already hold. They are not marketplaces, and they don't tell you what a fair price is across the whole market. See TensorPool vs Opengrid for a closer look.

3. Aggregators

Examples: Shadeform (Y Combinator S23), NVIDIA DGX Cloud Lepton (launched May 2025)

Aggregators put many GPU clouds behind one account and one API. Shadeform connects 20+ GPU clouds through one API and console. NVIDIA DGX Cloud Lepton is a marketplace across NVIDIA's cloud partners, where the partners set the prices.

Where they stop: you still pick the instance yourself, and nothing routes the job by your constraints. Lepton only covers NVIDIA partners, so it isn't neutral across the whole market. Also, aggregators reach about 20 of the 209 GPU clouds. More detail in Shadeform vs Opengrid and the DGX Cloud Lepton comparison.

4. Marketplaces

Examples: Vast.ai, RunPod (valued at $1B), SF Compute (valued at $300M), Mithril (formerly Foundry, valued at $350M)

Marketplaces sell GPU time from their own pool of supply, often at the low end of the price range. RunPod is a developer GPU cloud. SF Compute sells tradable blocks of GPU time. Mithril runs spot GPU auctions. Vast.ai is a marketplace where hosts list GPUs for rent. Developer clouds and marketplaces cluster at roughly $1.25 to $2.89 an hour for an on-demand H100.

Where they stop: each one sells its own pool of supply. None of them works across providers, so a better price or more capacity at another cloud is invisible from inside the marketplace.

5. Indices and futures

Examples: Ornn ($33M seed, June 2026), Silicon Data ($30.5M Series A), Compute Exchange, Liquid Compute

This group treats compute as a financial commodity. Silicon Data publishes GPU rental price indices, and CME's H100 and B200 rental futures, listed from 5 October 2026, are built on them. Ornn runs the OCPI price index, ICE futures and a marketplace for dedicated GPU capacity. Compute Exchange runs auctions and forward contracts for blocks of GPU time.

Where they stop: they offer indices, futures and big dedicated contracts, but no on-demand API routing. They tell you what compute costs, or let you lock in a price for a large block. They don't place a job for you today. See Silicon Data vs Opengrid, Ornn vs Opengrid, Compute Exchange vs Opengrid and our explainer on CME GPU futures.

6. Decentralized networks

Examples: io.net, Akash, Aethir

Decentralized networks pool GPUs from many independent operators and coordinate them with a token. They can reach supply that the big clouds don't list.

Where they stop: each one is a token-based supply silo with its own tooling. Moving between them, or between one of them and a traditional cloud, means starting over.

The full comparison

Category	Examples	What it does	Where it stops
Price comparison sites	getdeploying, ComputePrices, gpus.io	Lists GPU prices across providers	You can look. You cannot buy.
Orchestrators	SkyPilot, TensorPool, dstack	Runs jobs across clouds you already use	Needs your own cloud accounts. No marketplace, no price discovery.
Aggregators	Shadeform, NVIDIA DGX Cloud Lepton	One API and account across many clouds	You still pick the instance. Lepton covers NVIDIA partners only.
Marketplaces	Vast.ai, RunPod, SF Compute, Mithril	Sells GPU time from its own supply	Each sells its own pool. None is cross-provider.
Indices and futures	Ornn, Silicon Data, Compute Exchange, Liquid Compute	Prices, futures and large contracts	No on-demand API routing.
Decentralized networks	io.net, Akash, Aethir	Token-coordinated GPU supply	Token-based silos, each with its own tooling.
Opengrid (pre-launch)	Opengrid	Market view, then routing by price, region, reliability and deadline	Not live yet. Market view ships first, routing API in development.

Where Opengrid fits

The table splits into two halves. Some tools sell access to compute but give you no way to judge the price or the provider. Others publish a price but won't place your job. Nobody yet tells a buyer where a job should run and then runs it.

That gap is what Opengrid is being built for. You describe the job once (GPU type, count, region, price ceiling, deadline) and Opengrid routes it to whichever provider wins on price, availability, region and reliability. You get one invoice no matter who ran it. The same live market view that informs routing also gives you a fair-price benchmark, and real routed jobs build up reliability data over time.

How to choose today

• You only need to know prices: start with a price comparison site.
• You already have accounts at several clouds: an orchestrator like SkyPilot or dstack fits.
• You want one account across many clouds and are happy to pick instances yourself: use an aggregator like Shadeform.
• You want the lowest list price and don't mind one supply pool: a marketplace like Vast.ai or RunPod.
• You need to lock in a price for a large block or hedge costs: look at the index and futures group.
• You want the decision made for you across the whole market: join the Opengrid waitlist.

For pricing context, read why the same H100 costs so much more at some providers and how GPU rental prices moved in 2025 and 2026.