Shadeform Alternatives in 2026: Marketplace vs Matchmaking for Committed GPU Clusters

◆ WHAT SHADEFORM IS

One API over many clouds, billed once

Shadeform sits on top of a few dozen GPU clouds, from hyperscaler regions to small neoclouds, and presents their on-demand inventory through one console and one API. A developer can see which cloud has an H100 available right now at what price, launch it, and pay Shadeform rather than opening an account with the underlying provider. Templates, SSH, and basic orchestration are built in. For a team running experiments across whatever is cheapest this afternoon, that is a real reduction in friction.

The pricing model is the underlying cloud's on-demand rate, passed through, sometimes with a margin and sometimes with a platform fee. The capacity is whatever those clouds have unsold at that moment. The commitment is hourly. Those three facts are what make it good at what it does, and they are also the boundary of what it does.

Shadeform does list reserved and multi-node options from some providers, and it would be unfair to say it is on-demand only. But the product is built around instance availability across clouds, and that is how most of its customers use it.

◆ FIVE WAYS TO SOURCE GPU COMPUTE, COMPARED

Descriptions from each product's public documentation, October 2026. Pricing from published rate cards and the GPUaaS.com October 2026 benchmark; on-demand and committed are different products and the comparison is directional.

◆ WHERE AN AGGREGATOR STOPS

An API can find eight GPUs now. It cannot promise 256 on one fabric for a year.

The aggregator model works by exposing unsold inventory. That inventory is, by construction, the capacity nobody has committed to, scattered across clouds and regions, in the quantities each cloud happens to have free. For one to sixteen GPUs on a given afternoon that is plenty. For a 256-GPU training cluster that has to be contiguous, InfiniBand-connected and held for twelve months, it is the wrong search space entirely. That capacity is not unsold; it is allocated by a provider to a buyer on a contract, and it does not appear in an on-demand API because nobody is offering it by the hour.

The price follows the same logic. An aggregator passes through the on-demand rate, which for H100 on the clouds it fronts runs around $2.50 to $5.00 per GPU-hour in 2026 depending on the cloud. The committed band through the GPUaaS.com network is $2.50 to $3.00 for H100 SXM, and it is paid on every hour of the term by a buyer who has given up the right to spin down next week. The two-price market covers why those are different products at similar numbers.

The third thing an aggregator cannot do is the human part. A committed cluster request involves interconnect topology, power and density at the facility, SLA terms, replacement policy, residency documentation and a three-party contract. None of that is an API call, which is why the four business models post puts aggregators and matchmaking networks in different categories rather than treating one as a better version of the other.

◆ THE OTHER THREE

SF Compute, Vast.ai and going direct

SF Compute runs an exchange for short-term blocks of cluster time, typically H100 and H200, where buyers bid for contiguous capacity over days or weeks and the market clears the price. It is the closest thing to an aggregator for training runs rather than instances, and for a two-week run on 64 H100s it is often the cheapest option available. It is less suited to twelve-month terms, to the newest Blackwell parts, and to residency-constrained requests, because the exchange is optimised for liquidity rather than for a specific facility.

Vast.ai is the floor of the market: a peer marketplace where anyone from an individual with a 4090 to a data center with H100 racks lists capacity, and spot clears near $2.50 for H100 and often well below. For interruptible single-GPU work it is hard to beat on price. What it does not offer is an SLA, a guaranteed fabric, or enterprise contract terms, and a later post in this series covers what enterprise buyers lose on a marketplace in detail.

Going direct to a neocloud such as CoreWeave, Lambda or Nebius means picking a provider and negotiating with them. That suits a buyer who already knows who they want and values the relationship. It limits the request to that provider's fleet and regions, and it means getting one quote rather than several, which the market intelligence post argues is the most expensive habit in GPU procurement. The top 10 providers post covers the direct options.

◆ WHEN MATCHMAKING FITS

A spec in, several quotes out, one contract, dedicated capacity

GPUaaS.com is the other end of the spectrum from an aggregator API. A buyer submits a spec: GPU, count, term, region if residency requires it, interconnect if training requires it. The request is routed to 20+ vetted partners who quote inside the published band within a few hours. The buyer picks, and the contract runs between the provider, GPUaaS.com and the buyer. The capacity is dedicated, the term is fixed, there are no buyer fees, and the minimum is an 8-GPU node. Capacity can be sourced in any region except China and Taiwan, and presold against a future delivery date.

That is slower than an API by hours and faster than a direct negotiation by weeks, and it produces something an aggregator cannot: several quotes on the same committed spec, side by side, from providers who have already been vetted on interconnect, facility, support and residency. The how it works page has the mechanics and the October benchmark has the bands.

Most organisations end up with both. An aggregator or marketplace for burst and experiments, a committed cluster for the production floor. For single GPUs and token-metered inference, packet.ai covers the part of the market an aggregator serves, from the same provider network.

◆ FAQ

Frequently asked questions