Market study · updated 6 Oct 2026
The private compute market, October 2026
Hourly GPU launchers and price indexes already exist. The open slot is a public post for the deal those catalogs do not express: a count, a place, a term, and a tenant boundary, from either side.
Where the money and the power already sit
By October 2026 the constraint in AI infrastructure is power and contracted clusters, and the public catalogs of hourly GPUs are a side market next to that. Structure Research, in its 2026 Global Hyperscale Data Center Report (28 Sep 2026), described combined 2026 capital-expenditure guidance from AWS, Microsoft, Alphabet, and Meta of well over $700 billion, with the majority aimed at data centers, servers, and network. The same note put hyperscale self-build operational capacity around 29.7 GW.
The neocloud layer underneath those platforms is already a public-company story. MarkTechPost’s 23 Aug 2026 ledger, checked against company figures as of 21 Aug, put CoreWeave’s Q2 2026 revenue at $2.575 billion, up 112% year over year, with about 1.5 GW active and more than 4 GW contracted. Coverage of CoreWeave’s Q2 report put the revenue backlog near $104 billion, including multi-year commitments from Meta and OpenAI. Nebius was cited at about $3.0 billion of AI-cloud annual recurring revenue. Crusoe was cited at 4.9 GW contracted as of June 2026. Lambda’s published B200 on-demand rate in that ledger was $6.69 per GPU-hour.
Those numbers are the market Private Compute stands beside. They are contracted offtake and owned power. They are a poor description of a lab that needs eight GPUs in a specific city for one quarter, or of an operator whose cage is empty in December.
The hourly layer is crowded, and it is good at one job
Vast.ai, RunPod, Salad, Shadeform, Spheron, Lium, and a long tail of host pools already answer “what is the cheapest GPU I can boot today?” They do it with per-second or hourly billing, community hosts or partner clouds, and an API. Price indexes sit on top and reshuffle the same offers.
ComputePrices, writing up a 12 Sep 2026 snapshot, found marketplaces in 7 of the 10 cheapest RTX 5090 offers and 7 of the 10 cheapest RTX 4090 offers. At the data-center end the picture flipped. Marketplaces were 3 of the cheapest 10 H100 SXM offers, 4 of 10 for H200, and 2 of 10 for B200. At B200, neoclouds that own the hardware set the floor. The hourly board is where consumer cards clear. It is a thin slice of how a Blackwell rack actually gets reserved.
Published H200 on-demand list prices on 28 Sep 2026, collected by Memetik, still spanned a wide band among named clouds: Crusoe $4.29, DigitalOcean $4.47, Nebius $4.50 (raised to $5.40 on 1 Oct), Modal $4.54, RunPod $4.59, Together $5.99, CoreWeave $6.30 per GPU-hour. GPU.info’s 5 Oct 2026 B300 ledger showed the same spread at the next generation, with a direct-provider quote at $4.99 beside a cluster of neocloud and marketplace quotes around $8 to $9. A single “market price” is not something a buyer can hold. A dated post with a place and a term is.
| Signal | Figure | As of | Source |
|---|---|---|---|
| H200 on-demand, lowest named cloud in the Memetik set | $4.29 / GPU-hour (Crusoe) | 28 Sep 2026 | Memetik |
| H200 on-demand, CoreWeave list in that set | $6.30 / GPU-hour | 28 Sep 2026 | Memetik |
| B200 on-demand, Lambda published | $6.69 / GPU-hour | 21 Aug 2026 | MarkTechPost |
| Vast.ai GPU types / configs | 80 types, 617 configs, from ~$0.02 | 6 Oct 2026 | gpus.io |
| RunPod GPU types / floor | 37 types, from ~$0.24 | 6 Oct 2026 | gpus.io |
| Marketplace share of the 10 cheapest B200 offers | 2 of 10 | 12 Sep 2026 | ComputePrices |
| Marketplace share of the 10 cheapest RTX 5090 offers | 7 of 10 | 12 Sep 2026 | ComputePrices |
The open slot is the post, not another launcher
Three tools already exist for a buyer who wants a VM this afternoon: a host marketplace, an aggregator, and a price index. The missing tool is a public board where both sides can write the deal the API does not express.
That deal has a shape. It names a GPU or a facility. It names a count. It names a region and a country. It names a start, and usually an end. It says whether the machine is single-tenant, bare metal, a cage, or a confidential environment. It is sometimes an offer, sometimes a request, and sometimes both at once: GPUs free in Stockholm, storage wanted next door.
Private Compute is that board. Parties contract directly. The site stores the introduction and the inquiry. It does not take a cut of the GPU-hour, custody funds, or pretend to have visited the data center.
Who the board is for
Operators with a pod, a rack, or a cage that is empty, arriving, or partly spoken for. Their alternative today is a private broker list or a sales call. A post makes the window searchable.
Teams whose requirement is a term, a city, a tenant boundary, or a mix of GPU, power, and storage. An hourly account solves a different problem. A request lets suppliers find the requirement.
People who have one piece and need another. The both-post is the category the launchers do not have a field for.
What can go wrong, stated plainly
A new board is empty. Example cards on this site are labeled and kept out of the search index so they cannot be mistaken for inventory. The launch depends on real posts replacing them.
Trust is thin until verification exists. Version one believes the form. Buyers should ask for delivery evidence before they pay. A later product can sell verification and featured placement. Those fees still make sense when the contract itself stays off-platform.
A neocloud or a marketplace can add a “request capacity” form. The durable position is to stay neutral, to list the shapes they do not sell (cages, power, two-sided swaps, regions they skip), and to be the page a searcher lands on for “H200 capacity Frankfurt” or “bare metal B200 term”.
Who already occupies a lane
| Player | Model | What they already do well | What a board adds |
|---|---|---|---|
| Vast.ai | Host marketplace | Per-second rental of independent machines, on-demand and interruptible. On 6 Oct 2026 gpus.io tracked 80 GPU types and 617 configurations, from about $0.02/GPU-hour. | A named term, a single tenant, or a cage still needs a direct conversation. |
| RunPod | Secure cloud and community hosts | Pods, templates, and serverless. The same gpus.io snapshot listed 37 GPU types and on-demand prices from about $0.24/GPU-hour across 12 regions. | Reserved multi-month capacity and non-GPU infrastructure sit outside the pod catalog. |
| Shadeform | Aggregator | One account and API across partner GPU clouds. Founded around 2023. The buyer launches listed capacity. | Regional operators and buy-side requests do not get a post of their own. |
| Spheron | Data-center aggregator | Production and spot GPU capacity with per-minute billing, presented as a catalog. | Colocation, power, and custom terms stay a side conversation. |
| Salad | Consumer GPU containers | Low prices on interruptible consumer cards for fault-tolerant batch work. | Dedicated, located, term-bound capacity is a different purchase. |
| Lium and peers | Narrower host pools | Lium, QuickPod, Hyperbolic, io.net, and Theta each cover a slice: location, consumer cards, or decentralized supply. | They list machines. They do not collect a public request for a term and a place. |
| CoreWeave, Nebius, Lambda, Crusoe | Neoclouds | Owned or contracted power and GPUs, sold under an MSA. CoreWeave reported Q2 2026 revenue of $2.575B. Public rate cards are short and sales-led. | Smaller shapes, odd regions, and upcoming deliveries rarely become a public post. |
| Price indexes | Comparison sites | ComputePrices, GPU.info, and gpus.io compare hourly rates that providers already publish. | An index cannot accept an offer or a request. The board can. |
How Private Compute uses this
The product is a board, free to post while the first live capacity arrives. Monetization that survives a direct contract is featured placement and verification, later. Taking a cut of the GPU-hour would push the conversation off the site.
Search is aimed at the phrases a buyer types once they already know the hourly clouds: “H200 capacity Europe”, “bare metal B200”, “liquid-ready cage Frankfurt”, “GB200 delivery window”. Each of those has a guide page that quotes the definition and lists matching posts.
For answer engines, the site publishes llms.txt, a markdown alternate of each public page, Offer and Demand structured data, and an explicit definition of private compute. Example listings are labeled in prose and marked noindex so a model does not cite them as available GPUs.
Sources
- Structure Research, 2026 Global Hyperscale Data Center Report — 28 Sep 2026. Hyperscaler capex guidance and self-build GW.
- MarkTechPost, Best GPU Neoclouds 2026 — 23 Aug 2026. CoreWeave, Nebius, Lambda, Crusoe figures checked 21 Aug 2026.
- ComputePrices, GPU marketplaces compared — Snapshot 12 Sep 2026 of the ten cheapest offers by GPU.
- Memetik, H200 price per hour — Named-cloud H200 on-demand list prices as of 28 Sep 2026.
- gpus.io, RunPod vs Vast.ai — Catalog breadth and floor prices, 6 Oct 2026. Vast moved to 80 types and 617 configs; RunPod stayed at 37 types from about $0.24.
- GPU.info rental index — B300 offer ledger, 5 Oct 2026.
Rates move weekly. The figures above are tied to the date on each row. Re-check the source before you repeat a price.