Supply is booming. Demand is lumpy. Here is why neoclouds are adding a compute marketplace next to direct sales, what it adds, what it cannot, and how to avoid channel conflict.
Supply is booming. Demand is lumpy. That is the neocloud business model problem, and it is why the neocloud demand channel mix is changing, as a growing number of GPU clouds pair direct sales with a second demand channel (a compute marketplace that sells capacity to many smaller buyers), though the evidence is a pattern across examples, not a census. This piece sets out the evidence in public filings and independent research, then says what we think follows from it. Every market claim is sourced and dated, and our own opinions are marked "Our view." One disclosure up front: packet.ai runs a provider program that lets operators list capacity on a B2B marketplace, so we have a stake in this argument and you should weigh it accordingly.
Key takeaways
Start with supply. SemiAnalysis's ClusterMAX 2.0 (November 2025) tracked 209 GPU cloud providers, up from 169 in March 2025 and 124 a year before. It rated 84 of them and said those 84 cover 95% of the market by volume. By simple arithmetic that leaves the other 125 or so tracked providers sharing roughly 5%, which is what a long tail looks like.
AI compute demand at the top is large but concentrated. CoreWeave's FY2025 10-K reports $60.7 billion of remaining performance obligations at December 31, 2025, up from $15.1 billion a year earlier, and says Microsoft accounted for about 67% of its 2025 revenue. It also discloses commitments of up to about $11.9 billion from OpenAI through October 2030 and up to about $14.2 billion from Meta through December 2031. The same filing says the company expects backlog to grow "in steps, which may include periods of decline," and expects customer concentration to continue.
None of this means the gpu rental market is slack. SemiAnalysis's April 2026 analysis, titled "The Great GPU Shortage," reported its H100 one-year contract price index up almost 40% between October 2025 and March 2026, a measure of one GPU model on one contract length and not of every provider. Capacity can be tight for the largest buyers and thin for everyone else at the same time. A growing market is also not the same as a growing set of buyers that any one provider can reach: the totals describe aggregate demand, not the demand an individual operator can win.
Our view
The tightness is real at the top, where a few large buyers sign multi-year commitments with a handful of providers. Operators outside that group tend to see demand arrive in lumps: a customer takes a block of capacity, leaves, and the hours in between sit idle. That is an inference from the evidence above, not a measured statistic.
The chain of reasoning, in four steps (our view):
Look at how the largest deals are built. In August 2025, TeraWulf announced two 10-year hosting agreements with Fluidstack for more than 200 MW at Lake Mariner, about $3.7 billion of contracted revenue over the initial terms, with Google backstopping $1.8 billion of Fluidstack's lease obligations to support project debt. A few days later an expansion lifted contracted capacity to over 360 MW and about $6.7 billion of contracted revenue, with Google's total backstop rising to about $3.2 billion. A later TeraWulf filing described a 168 MW joint venture at its Abernathy campus, with a 25-year lease worth about $9.5 billion of contracted revenue and Google backing about $1.3 billion of Fluidstack's lease obligations. In each case the backstop covers specific lease obligations that support project debt, not Fluidstack's business in general.
Sacra describes the commercial side of the same company: its private cloud business runs on GPUs it owns and operates, sold on two to three year contracts with 25% to 50% paid up front. Fluidstack has also arranged GPU-backed debt capacity of up to $10 billion through Macquarie, and the two announced a GPU financing deal for European AI labs in April 2025.
Our view
Deals like these work because a hyperscaler's credit stands behind the tenant and because the buyer is large enough to justify a dedicated sales and legal effort. A small operator selling directly has neither. Each of those neocloud customers is a separate sales cycle, a separate contract and a separate support relationship, and one lost customer leaves a visible hole in utilization. Direct sales still matter, because price control, margin and customer relationships live there, but on its own it is a thin foundation for a small fleet. That is a structural disadvantage, not a universal rule: some small providers with specialist customer relationships or niche demand make direct sales work well.
A compute marketplace, sometimes called a gpu capacity marketplace, is a platform that aggregates GPU capacity from several providers, or from many hosts, and sells it to buyers through one account, usually self-serve. It is not the secondary market for used graphics cards that the broader search term "gpu marketplace" often returns. The names are used loosely, but they differ. In our usage, a marketplace lets providers or hosts list capacity that buyers choose from; an aggregator puts several providers behind one account or API and typically sets its own customer-facing price; a reseller or broker contracts for capacity and resells it on its own terms; and an exchange emphasizes standardized, comparable terms. A given platform may do several of these, so searches for a gpu cloud marketplace, gpu aggregator, gpu reseller, gpu broker or gpu exchange all land in the same territory.
The channel is already visible. A comparison report describes SF Compute as a marketplace that lets buyers resell unused capacity, creating spot and forward markets, and Compute Exchange describes itself as aggregating reserved capacity from neocloud partners for terms from one to 36 months. SemiAnalysis's ClusterMAX review noted that GPU.net advertised access to 42 providers and 121,000 GPUs. A comparison site notes that aggregators typically set the customer-facing price themselves, which may differ from the underlying cloud's direct rate.
Operators are on the supply side of these channels too. RunPod's documentation says its Secure Cloud runs in T3/T4 data centers operated by vetted partners, Vast.ai has a datacenter tier for providers with five or more GPU servers in a professionally managed facility, and Compute Exchange says its reserved capacity comes from neocloud partners. That is a pattern visible in several examples, not a census of the industry.
The mix matters too. Sacra estimates that at the end of 2024 about 62% of Fluidstack's revenue, roughly $112 million of about $180 million in annualized revenue, came from private cloud and about 38% from its marketplace. It estimates average contract values of over $100 million for private cloud against about $340,000 for the marketplace, and describes the marketplace as a lead source for the higher-margin private cloud. These are Sacra's estimates, not figures Fluidstack has published, and the company is private.
The table is our view of how the two channels differ, not measured data.
Our view
As a gpu cloud sales channel, a marketplace adds many small buyers, self-serve onboarding and short terms that fit idle windows. It does not typically add the long, predictable, creditworthy contracts that let an operator borrow against its hardware. What matters for financing is contract length, predictability and buyer credit, not the channel label, and marketplace demand tends to be shorter and less predictable. It is better thought of as a way to keep hardware busy between contracts, or to find first customers before any contract exists.
The obvious worry is that a marketplace undercuts your own customers. Because aggregators can set a customer-facing price different from your direct rate, the same GPU can appear at two prices. Add the platform constraints that exist on some marketplaces, such as Vast.ai's verification, which fails machines that run personal workloads, and the second channel needs deliberate design. The packet.ai guide to the five ways to turn idle GPUs into revenue covers those constraints.
Our view
Four rules we would follow. First, separate the inventory: decide which capacity is reserved for direct customers and which is sold through the marketplace, and never list the same GPUs in two places without a plan, because double-booking is how you lose both customers. Second, fence prices instead of fighting over them: give the marketplace shorter terms or interruptible capacity, and give direct customers commitment discounts. Third, let marketplace customers graduate: a buyer who grows past self-serve should have a clear path to a direct contract, a path Sacra describes at Fluidstack, where it calls the marketplace lead generation for private cloud. Fourth, measure each channel separately with allocated, busy and billable hours, as in GPU utilization explained, so you can see which channel earns more per GPU-hour.
Our view
An illustration of the rules, not a recommendation of specific numbers: an operator with 64 GPUs commits 48 of them to two direct customers under reserved contracts and lists the other 16 on a marketplace as interruptible capacity. When a direct customer expands, the operator takes GPUs from the marketplace pool, never the other way round, and no committed GPU ever appears in a marketplace listing.
Our view
If the pattern holds, we expect the hybrid to become the norm for operators that cannot land large anchor contracts: contracted anchor customers to finance the hardware, and a marketplace to fill the gaps, as Fluidstack's estimated mix suggests. That is a reasonable expectation, not a proven outcome for every provider. We expect reliability and ratings to decide who gets marketplace demand, since buyers already rely on third-party ratings (SemiAnalysis interviewed more than 140 end users for ClusterMAX 2.0). And we expect small operators to use a marketplace first and add direct contracts as they grow, not the reverse. The honest uncertainty is the cycle: if capacity stays tight, the marketplace matters less to the largest operators, and if it loosens, it matters more.
A simple decision framework (our view):
See the operator page for how the packet.ai provider program works, or apply to list capacity. For the vocabulary used here, the GPU provider glossary defines neocloud, take rate, reserved capacity and the other terms.
Retrieved October 12, 2026. Company descriptions of their own services are the companies' claims, and third-party summaries are noted as such.
Last reviewed: October 12, 2026. This is an opinion piece. Market facts are sourced and dated above, passages marked "Our view" are opinion, and nothing here is financial advice or a forecast. It contains no packet.ai fee, payout or earnings claims. For packet.ai's current provider requirements and terms, see the provider page.
Same models. Same API. Fraction of the cost. Start free — no credit card required.
Start Building →