How AIHostCompare ranks and verifies providers

Prices, GPU types, VRAM, regions, billing, storage, deployment and management fields are captured with source URLs and timestamps. This is what the dataset holds today:

Every value comes from the highest source on this list that states it:

1. the provider’s official API;
2. the official pricing page;
3. the official documentation;
4. the official region or status page;
5. a reliable secondary source, only when no official source states the value, and labelled as secondary.

Each value points to the source it came from, with the date and time the source was read. A value that no source states is stored empty and shown as Not verified.

Verification workflow

1. Research. Provider records are compiled by AI research agents that read the official pages live and must cite a source for every value they record.
2. Adversarial re-check. A second pass tries to disprove each value against its cited source: units, per-GPU or per-node pricing, pricing type, region scope, list, introductory or renewal price, currency and tax. Values that fail are corrected or set to Not verified.
3. Validation. Every record must pass a strict schema. A value that cites a missing source, an unknown GPU or region, a duplicate id or an invalid date fails the build, so it cannot be published.
4. Confidence. Each record carries a confidence level (high, medium or low) and notes that explain anything less than certain, such as a GPU variant the provider does not name exactly.
5. Automatic price checks. Where a provider publishes a machine-readable price list (today Scaleway and OVHcloud), an automatic check, scheduled daily, compares our prices with it. It does not change prices on the site: every difference is reviewed before a price is updated. The price status page shows which prices this covers and when they were last checked.

Units, billing periods and plan labels are normalized so a workload can be compared fairly.

• USD per GPU-hour. Every GPU price is stored as US dollars per single GPU per hour.
• Per node to per GPU. When a provider sells whole multi-GPU instances, the instance price is divided by the number of GPUs and the node size is kept in the record. 56 of the 117 GPU offers in the dataset are multi-GPU nodes.
• Currency. Prices listed in euros are converted to US dollars at the European Central Bank’s euro reference rate for the verification date, or the latest rate published before it. The original euro price and the rate stay in the record’s notes. Prices exclude VAT unless stated.
• GiB to GB. Some providers price storage or data transfer per GiB (2³⁰ bytes) rather than per GB (10⁹ bytes). We store the price per decimal GB: 1 GiB = 1.073741824 GB, so a per-GiB price is divided by 1.073741824, and the provider’s published per-GiB figure stays in the record’s notes. RAM and disk sizes keep the number the provider publishes.
• Introductory and renewal prices. When a plan is cheaper for a first term and renews at a higher price, both prices are recorded. The introductory price is always labelled “Introductory” and shown next to its renewal price. It is never the headline “from” price and never feeds a ranking, a lowest-price figure or a cost estimate; those use the renewal or standard price.
• Pricing types stay separate. Every GPU price is labelled with its type: on-demand, secure cloud, community cloud, serverless, spot, reserved or marketplace. A figure labelled on-demand uses on-demand list prices only: the provider’s standard price without a commitment, secure cloud included. Community cloud prices (GPUs on third-party hosts) and serverless prices never count as on-demand. The pay-as-you-go “from” prices on hubs and provider pages are the cheapest on-demand, secure cloud, community cloud or serverless price, so they can be lower than an on-demand figure, and they always show the price type next to the price. A marketplace’s lowest listing is never presented as a list price.
• Billing units stay visible. Each GPU price records its billing increment (per second, per minute, hourly or monthly). Monthly plan prices are the provider’s own monthly figures where it publishes them. Where we derive one from an hourly price for a server that runs around the clock, we use 730 hours a month (8,760 hours a year ÷ 12); the GPU cost calculator otherwise uses the hours you enter.
• Dated prices. A price applies from the day it takes effect. A price change a provider announces applies from 00:00 UTC on its date, because providers rarely state a time zone. Before that day every page, tool and the hosting advisor use the checked price; from it, the announced price. When a fixed price turns into a dynamic one (Nebius’s preemptible H100, H200 and B200 from 8 October 2026; their fixed rates applied through 7 October), there is no price to compare from that day: the offer is left out of lowest prices, ranges, budgets, rankings and the hosting advisor’s suggestions for the days it applies. A cost estimate for a month that includes those days shows no monthly total: only the known costs are listed and labelled as such, compute for the days before the change at the fixed rate plus storage and egress for the whole month. Cost estimates are made for one billing month, day by day.
• Shared allowances. Where free usage is shared across an account, such as Modal’s egress allowance per workspace and billing cycle, an estimate charges only what your workload adds on top of the rest of the workspace’s usage: [max(0, U + W − A) − max(0, U − A)] × the rate, with U the other usage, W your workload (all its instances together) and A the plan’s allowance, counted once. When the plan or the other usage is unknown, the estimate shows a range instead of one figure; an unknown cost is never counted as free.
• Billing rules in estimates. The GPU cost calculator bills each day’s run in the provider’s own increment: a started hour counts in full where billing is hourly and a started minute where it is per minute. Where hourly billing stops at a published monthly price (Hetzner’s dedicated GPU servers), a server costs at most that price a month. One-off setup fees are shown with the offer and never added to a monthly estimate.

Important data blocks show when they were checked and where the claim came from. Freshness is computed when a page is rendered: data pages are rebuilt on the first visit after they are an hour old, so a check date older than 7 days gets its “Stale” mark without anyone editing the page. A table is only as fresh as its oldest row.

Not verified means not verified. A missing value is never shown as zero, estimated or copied from a list article. Offers without a verified price are left out of every lowest, median and highest price and are sorted last.

Filters and the sort you choose decide the order. Sponsored placement never silently determines it.

• Category, GPU and framework pages sort providers by lowest verified price by default, comparing like with like: GPU-hour prices with GPU-hour prices and monthly plans with monthly plans. VPS and agent-hosting pages lead with monthly plan prices, the other pages with GPU-hour prices. Pay-as-you-go GPU prices (on-demand, secure cloud, community cloud and serverless) and spot, reserved or marketplace GPU prices are ranked separately, and hub pages show them as labelled groups. Providers without a verified price follow, then providers whose prices we may not republish. You can re-sort by last verification date or by name.
• Framework pages list hosts only. A provider appears on a page such as OpenClaw or vLLM hosting when the framework itself runs on its infrastructure with an official template, documentation or a community guide. Installing it yourself on a generic server does not count. Providers that only serve the models a framework calls, or connect to it as an integration, are labelled as model provider or integration and are not listed as hosts. 11 of the 100 framework entries in the dataset are model-provider or integration entries.
• The provider directory is alphabetical. Alternatives on a provider page share at least one hosting category with it, other than AI hosting itself; the most relevant come first (shared categories, then shared GPU families, then shared frameworks), and the picks are spread so that every provider is an alternative on at least three and at most six other provider pages. Ties go to the provider listed least often, then to the name. Affiliate status is never an input.
• Filters for GPU, memory, region, billing and features only narrow the list; they never re-weight it.
• Side-by-side comparisons show the same fields for every provider, with Not verified where a value is missing. Head-to-head pages exist only when both providers have verified data.
• Workload fit labels on provider pages always come with a written rationale.

How the tools calculate

The tools compute estimates from your inputs and the verified dataset, with the same price rules as above, and each tool page publishes its formula and assumptions:

• GPU Cost Calculator: how it calculates
• VRAM / Model Fit: how it calculates
• Agent Hosting Selector: how it calculates

• We don’t score or rate providers. AIHostCompare publishes no provider scores: lists are sorted by price, verification date or name. A score would appear only after its formula, inputs and weights are published on this page, and every score would map to underlying data.
• Providers can’t pay for placement. Paid placement does not exist today; if it ever does, it is labelled “Sponsored” and kept out of rankings, sorted tables and recommendations.
• We don’t estimate prices that aren’t published. A value no source states reads Not verified.

Affiliate links and money

• Active affiliate programs today: 1.
• An affiliate link carries rel="sponsored nofollow noopener"; the line under every page's heading discloses it and links to the disclosure.
• Outbound “Visit” buttons (such as “Visit DigitalOcean”) are the main action on every provider page, for every provider, whether or not the link earns a commission. In the hosting advisor, the first suggestion’s “Visit” button is the main action only when that suggestion is a confirmed match with a complete price. That rule reads the match status and the price, never affiliate status.
• A provider without an affiliate program is listed and ranked exactly like one with a program.
• Affiliate status lives in a separate configuration that no ranking, sorting, filtering, recommendation or score reads.

Details are on the affiliate disclosure page.

• Prices move between checks. Providers change prices, availability and terms without notice, so a value can be out of date even when it was correct when it was checked. Every price shows its check date; confirm it on the provider’s page before you buy.
• Most prices are re-checked in research passes, not automatically. 10 of 296 published price fields are checked automatically against a provider’s own price list; the rest are re-checked in research passes. The price status page lists which.
• Estimates are not bills. Prices exclude VAT unless stated. Real bills also depend on taxes, currency, billing increments, storage, network traffic and how you actually use a service, so they can differ from an estimate.

Providers whose terms restrict a comparison

We list a provider only when its terms allow a comparison service like this one to use the facts we publish. Some providers’ website terms restrict commercial use of the site or its data, ban reading their pages with automated tools, or require a licence before their prices or availability appear in a comparison. We do not list those providers: none of their prices, plans, regions or features appears on any page, in any tool, data file or structured data, and the daily price check does not read them. A provider is listed again once it gives written permission or a legal review confirms the use, after the same verification as every other record.

A corrections log records material changes and avoids silent retroactive edits.

1. Report an error with the contact form (topic “Data correction”) or by email, with the page, the field, the correct value and a source.
2. We check the cited source and your source. A confirmed error is fixed and the value re-dated.
3. Material corrections (a wrong price, availability, region, feature or compliance claim, or a calculation error) are logged below with the date, what changed, why, and the pages that showed it. Routine price updates are not corrections: the new price replaces the old one and shows its new verification date.

Corrections log

• 27 Sep 2026

  Correction

  Hetzner cloud plans

  Hetzner’s plans now start at $13.67 a month (Cloud CPX12) instead of $6.83 (Cloud CX23). Hetzner showed every Cost-Optimized cloud type (CX and CAX) as not available on 26 Sep 2026, so the CX23 could not be ordered. Its limitations now say so.

  Pages: Hetzner, AI VPS comparison, AI agent hosting comparison, OpenClaw hosting comparison, Agent Hosting Selector

• 27 Sep 2026

  Correction

  On-demand figures

  The lowest on-demand prices and ranges on the homepage, and the on-demand columns of the GPU price snapshot and the VRAM tier table, now use on-demand list prices only. They also counted community cloud prices (GPUs on third-party hosts) and, on the homepage, serverless prices, which are pricing types of their own.

  Pages: Homepage, How to choose a GPU cloud provider, How much VRAM do you need to run an LLM?

• 27 Sep 2026

  Correction

  Region count in Lambda’s summary

  Lambda’s summary now names 16 regions: 14 On-Demand instance regions and 2 US regions that appear only as filesystem endpoints. The count did not match the region list on the same page: it counted only the 14 instance regions.

  Page: Lambda

• 27 Sep 2026

  Correction

  RTX PRO 6000 editions

  Hetzner’s GEX131 servers now show the RTX PRO 6000 Blackwell Max-Q Workstation Edition and Nebius’s RTX PRO 6000 plan the Server Edition, each with that edition’s specifications. Modal does not say which edition it rents, so its offer names no edition. They were shown as the Workstation Edition, whose specifications differ from the Max-Q and Server editions.

  Pages: Hetzner, Nebius, Modal

• 27 Sep 2026

  Correction

  V100 form factor

  Lambda’s and OVHcloud’s V100 offers, and OVHcloud’s V100 plan, now read V100 16GB. They were shown as V100 SXM2, a form factor neither provider states on its current pages.

  Pages: Lambda, OVHcloud

• 27 Sep 2026

  Correction

  Uptime SLA scope

  Uptime SLA percentages now show what they cover directly under the figure, on provider pages and in comparisons. Hetzner’s covers its Cloud servers, which have no GPUs, Scaleway’s covers GPU and dedicated CPU Instances only, and DigitalOcean’s 99% covers GPU Droplets only. Comparisons showed only the start of each SLA note and provider pages showed the percentage on its own, so Scaleway’s 99.5% appeared next to development instances that have no SLA.

  Pages: Hetzner, Scaleway, DigitalOcean

AIHostCompare is an independent publication and a registered trade name of Loadiqo, a Dutch sole proprietorship; its owner, address and registration numbers are in the legal notice. Provider records are compiled by AI research agents and re-checked before they are published, as the verification workflow above describes. Guides carry the byline “AIHostCompare”; no data page carries a personal byline, and we do not attach invented names, personas or teams. The editorial standards say how facts, provider claims, estimates and opinion are kept apart.