What Is GPUaaS? The 2026 Definition, the Four Business Models, and Where the Price Gaps Are

◆ THE DEFINITION

What GPUaaS means in 2026, and what it stopped meaning

In 2023 GPUaaS mostly meant spinning up an H100 instance on a cloud console and paying by the hour. By 2026 the term covers the whole range of ways to get GPU compute without buying it: hourly on-demand, interruptible spot, reserved instances with a term discount, dedicated bare-metal clusters on six or twelve month commitments, rack-scale allocations sold a year ahead of delivery, and tokens sold per million without any GPU being visible at all.

What ties those together is that someone else owns the hardware, the facility, the power contract and the financing, and the buyer pays for use. What separates them is everything else: whether the capacity is guaranteed, whether it can be taken away, whether the GPUs share a NVLink domain, whether the buyer gets bare metal or a virtual machine, and who they are actually contracting with.

The useful way to sort that in 2026 is by business model rather than by product name, because the business model decides the price. There are four.

◆ THE FOUR GPUaaS BUSINESS MODELS

Model definitions as used in GPUaaS.com market analysis, October 2026. Some companies straddle categories; the model describes the dominant way they sell.

◆ MODEL BY MODEL

What each one is good at, and what it charges for

Hyperscalers sell GPUs the way they sell everything else: as instances inside a platform of managed services, with identity, storage, networking and billing already wired together. That integration is what the premium pays for, and it is real if the rest of the workload already lives there. The costs that surprise people are the ones around the GPU, which the hidden costs post covers: egress, storage, managed networking, and the gap between reserved commitments signed in 2024 and the on-demand price in 2026.

Neoclouds are GPU-first clouds that bought hardware at scale, usually on debt secured against customer contracts, and sell it with less platform around it and a lower price. They publish on-demand rates and negotiate committed clusters privately. The 2026 story for neoclouds is capacity: several have been sold out on current-generation parts for months, and the provider shakeout earlier this year removed some of the smaller ones.

Marketplaces and aggregators own nothing. They list capacity from many hosts, from individuals with a 4090 to data centers with H100 racks, and let buyers pick by price. Spot clears lowest here, near $2.50 for H100 through 2026, because the host is competing with every other host for the next hour. The price is excellent and the capacity can be withdrawn mid-job, which is the whole trade.

Matchmaking networks also own nothing, but they work the other way round. A buyer submits a spec, a term and a region; the network routes it to vetted providers who quote inside a published band; the contract that results runs between the provider, the network and the buyer. The capacity is dedicated and committed, the price sits at or near spot, and what the buyer gives up is the ability to spin it down next week. That is the model GPUaaS.com runs, and how it works covers the mechanics.

◆ THE PRICE GAP, SAME GPU, FOUR MODELS, OCTOBER 2026

Per GPU-hour. Hyperscaler and neocloud medians from published rate cards, late September 2026, as cited in the GPUaaS.com October benchmark. Spot from published marketplace listings and the Compute Desk B300 index, October 2026. Committed bands from the GPUaaS.com October 2026 benchmark. On-demand and committed are different products; the comparison is directional.

◆ WHERE THE GAPS COME FROM

Flexibility is priced, and so is who carries the risk

The 3x spread on H100 is not four companies disagreeing about what an H100 is worth. It is four different products. Hyperscaler on-demand is the right to start and stop any time, inside a platform, with the hyperscaler carrying all the utilisation risk. Neocloud on-demand is the same right with less platform and a provider carrying the risk on a thinner balance sheet. Spot is the right to use capacity until someone pays more for it, with the buyer carrying the risk. Committed is dedicated capacity for a fixed term, with the buyer carrying the utilisation risk and getting the lowest guaranteed price in exchange.

The two-price market post covers why spot and committed can sit at the same number and still be different products, and the wholesale pricing post covers why the committed band sits around 30% below what most enterprise buyers pay today.

One gap in the table deserves a note. On B300, the committed band through the network sits below both the neocloud on-demand record and the marketplace spot listing this month, which the B300 vs B200 post explains. That is unusual and it is the clearest case this year of the business model, not the GPU, setting the price.

◆ WHICH MODEL FOR WHICH BUYER

Match the model to how predictable the workload is

A team that does not know what it needs next month should be on-demand, and should pay the premium knowingly. A hyperscaler if the rest of the stack is there; a neocloud if it is not. A team running experiments that can be checkpointed and restarted should use spot and treat preemption as a feature of the price. A team with a production workload, a training run with a known end date, or an inference service with a known floor should be on committed capacity, because that is the only tier where the lowest price is also guaranteed.

Most organisations end up with more than one. The rent, reserve or buy framework sets out how to split a workload across tiers, and the billing models post covers how each one lands on a finance team's ledger.

For single GPUs and token-metered inference rather than clusters, packet.ai is the sister route: the same provider network, sold as tokens with 24/7 human support, which is a fifth way to buy GPU compute that does not need a GPU to be visible at all.

◆ FAQ

Frequently asked questions