How the long tail hides
The reason a sixty-provider list is plausible is that the providers most likely to publish standardised pricing in a scrape-friendly format are the same providers a comparison site can pick up most easily. The price-comparison sites are not wrong about the providers they list, they are simply optimising for a much narrower visibility surface than the market actually has.
The long tail is not made up of one kind of company. It is a mix. There are regional operators in markets the major neoclouds have no real footprint in. There are GPU marketplaces with capacity but no public hourly rate. There are resellers who package hyperscaler GPU regions under their own brand. There are sovereign-cloud providers whose pricing is by contract only. And there is a steadily growing set of new neoclouds that did not exist eighteen months ago, many of which are sub-scale today and meaningful tomorrow.
What we kept and what we cut
A larger directory is only useful if the entries are real. We index by operator domain rather than by listing, which means a candidate has to clear a qualify gate before its profile is created. The gate is mechanical and conservative: the operator's own site has to advertise GPU compute, in plain prose or in a structured offering page, before the profile lands in the directory. Aggregator listings, news mentions, and inferred presence do not qualify a provider on their own.
What we cut on the way to 1,009 is at least as informative as what we kept. We dropped dead domains, parked domains, marketing landing pages with no operating company behind them, broker pages with no underlying capacity, and a long list of candidates that mention GPUs only in a blog post or a careers ad. The numbers in this report are the survivors.
What the directory now covers
We track 19 accelerator models with at least three providers each, the threshold below which a model hub would be too thin to publish honestly. The largest model pools, unsurprisingly, are the workhorses that the rest of the market is currently structured around.
Geography is the harder dimension. Only 170 of 1,009 providers publish a serving-country list of their own, which means 23country landings clear the directory's honesty floor and the rest live behind the global model hubs only. We would rather show fewer country pages and have them be true than show many and have them be sloppy. This is the geography gap, and closing it is the next pass.
Why this matters for buyers
For a buyer the difference between a sixty-provider list and a thousand-provider list is not the providers at the top, it is the providers near the bottom of the shortlist. The well-known names will appear on both. The smaller direct operator in your specific country, the marketplace with unused capacity for the exact accelerator you wanted, the reseller running a more friendly procurement process around hyperscaler space, all of those live in the long tail that the short list does not show.
We do not believe the long tail is automatically better than the names everyone names. We do believe the long tail is automatically worth knowing about, because the right answer for a specific workload sometimes lives there and sometimes does not. Either way, you should be able to see what the market actually contains before you decide.
How to read the directory
The cleanest starting points are the model hubs (which list every provider that advertises a given accelerator), the operator-type hubs (which separate direct operators from marketplaces and resellers), and the country pages (where the geography signal is strong enough to publish honestly). Each provider profile shows the accelerator lineup, the claimed serving locations, pricing where the provider publishes any, and the operator-type label derived from independent signals.