Skip to content
GPU Finder

Always improving

Changelog

New features, improvements, and fixes for GPU Finder.

Subscribe again after cancellation

  • Billing now shows when a fully settled subscription ended. Choose Paid or Unlimited to start a new monthly subscription on the same account and API keys.

Updated Nebius spot pricing guidance

  • Nebius spot pricing guidance now reflects the November 2 removal of its price ceiling: spot rates can exceed on-demand rates.
  • Removed the outdated Nebius floor-to-ceiling bars from the spot GPU price comparison.

New GPU rental comparisons

  • Compare B300 with MI355X and RTX 5090 with RTX PRO 6000 using exact-count rental quotes, stock checks and 24-hour compute budgets.
  • The MI355X rental guide now explains missing single-GPU quotes, memory requirements and whole-node budgets, with a link to the B300 comparison.

Prices first on GPU pages and the homepage

  • GPU pages now open with the cheapest in-stock offer and its Visit button; the notes on how stock is checked sit behind a link below it.
  • The homepage price table now appears directly under the search, ahead of the price alert and the chart.
  • The AMD MI355X page has a new rental checklist covering node size, ROCm support, budget and stock.

Clearer AWS GPU rental budgets

  • Compare whole-instance hourly and 24-hour AWS compute budgets with instance sizes, regions and price observation times. Catalog prices clearly distinguish unverified capacity.
  • Fractional accelerator catalog entries no longer appear as zero-GPU rentals or set whole-GPU price floors.

Compare H100 and H200 rental budgets

  • The H100 vs H200 comparison separates one-GPU and eight-GPU on-demand stock quotes, shows 24-hour compute estimates, and explains memory fit and completed-job cost.

In-stock offers first for everyone

  • Price tables now open with in-stock offers for every visitor; out-of-stock and unknown-stock offers stay one tick away in the offer visibility filters.
  • 4x and 8x server pages now show the cheapest in-stock node and its Visit button directly under the page title.

Choose between L40S and RTX 5090 rentals

  • The L40S vs RTX 5090 comparison now explains 48 GB versus 32 GB memory fit, shows single-GPU on-demand stock quotes and 24-hour budgets, and links to rental listings.

Clearer MCP conversation continuity

  • The standard MCP endpoint now includes its conversation handle and reuse guidance in the main structured and text results, making it easier for clients to keep related calls and retries together.

MI355X rental guidance and clearer stock summaries

  • The MI355X rental guide covers GPU count, ROCm compatibility, memory, rental budgets and stock checks.
  • Stock summaries now identify their on-demand price scope so spot-only listings are not mistaken for absent capacity.

B300 rental guidance and whole-node budgets

  • The B300 rental page now explains single-GPU versus eight-GPU pricing, memory and connectivity checks, Blackwell software compatibility and total session costs.
  • Eight-GPU B300 pages show 8-hour, 24-hour and 7-day compute budgets from the lowest on-demand whole-node rate with fresh reported stock. Estimates are withheld when no qualifying offer exists.

Clearer search loading feedback

  • Searches no longer briefly show a failure message while results are still loading.

Hinode links updated

  • Hinode website and documentation links now use hinode.sh.

Hinode pricing collection restored

  • Hinode pricing now accepts its smaller published catalog and includes the current workspace charge in each machine's hourly price.

Better Safari compatibility

  • Adjusted page metadata to avoid a Safari error on spot GPU prices and other pages with multiple structured-data entries.

Less personal data in MCP responses

  • MCP replies now omit internal account, receipt and request identifiers, and raw collector errors while preserving usage limits and retry guidance.
  • The dedicated /mcp/openai endpoint omits conversation tracking; the existing /mcp endpoint retains it. Both share account usage limits.
  • The privacy policy explains the data shared with connected assistants and MCP operational analytics.

More private browser error reports

  • Browser error reports now remove URL query strings and fragments from request details, navigation history and source filenames while keeping useful code locations.

Better diagnostics for browser failures

  • Rare browser stack-overflow reports now include limited source and browser-state details to help identify the cause without collecting page content or form values.

Clearer responses to invalid requests

  • Malformed MCP handshakes now return an invalid-parameters response without consuming data units.
  • Invalid page-navigation requests are rejected before page rendering, allowing the browser to recover with a full page load.

More resilient Meshive stock checks

  • Meshive price and availability checks now retry temporary empty responses while preserving the last known stock if the feed remains unavailable.

Recover from page errors

  • The page error screen now offers a full reload to recover from browser translation conflicts.

Runpod pricing-source referral links

  • Runpod pricing citations on its provider page and the egress-fees guide now preserve our referral attribution while still opening the pricing source.

More reliable translated searches

  • Fixed a homepage crash when searching or changing filters after translating the page in your browser.

API launch waitlist retired

  • The API launch waitlist is closed. Old confirmation links now point to the API access guide, where you can get started directly.
  • Existing email unsubscribe links continue to work.

Improved billing service monitoring

  • Improved detection of repeated billing failures and external service timeouts to help us respond to service problems sooner.

A consistent GPU Finder email design

  • Sign-in messages, subscription confirmations, GPU alerts and API usage notices now share GPU Finder's grid mark and green styling, with layouts for mobile and dark mode.

Reliable private performance measurement

  • Fixed performance measurement startup while retaining numeric-only data collection and blocking external scripts.

Email delivery insights

  • Email delivery receipts help us detect failed, delayed and bounced messages. Analytics exclude message content and sign-in links.
  • Known email campaigns can be measured through public-site visits and provider clicks without identifying newsletter or alert recipients.

Better detection of interrupted stock updates

  • Empty PrimeIntellect feed responses now register as collection failures while preserving the last recorded stock information.

Better page performance measurement

  • Added public-page timing and numeric performance measurements to help identify slow browsing experiences, while preserving privacy opt-outs and excluding account pages.

Private page-depth measurement

  • Added an anonymous numeric measure of how far public pages are viewed, respecting privacy opt-outs and excluding account pages.

Compare provider model coverage, regions and billing

  • The providers page now answers which clouds list NVIDIA flagship GPUs, have the widest tracked NVIDIA range, list European regions, and bill per second or minute, with links to compare listings and estimate costs.

Better search failure diagnostics

  • Added privacy-conscious diagnostics to help investigate failed GPU searches and offer lookups.

Account analytics transparency

  • Updated our privacy policy to explain how verified sign-in emails identify developer accounts in PostHog. Public-site browsing remains anonymous.

More reliable analytics checks

  • Improved production analytics verification and kept public-site privacy checks separate from API account analytics.

More accurate browser traffic measurement

  • Fixed analytics incorrectly classifying ordinary browser visits as automation and clarified user-agent collection in the privacy policy.

More private public-site analytics

  • Public-site analytics explicitly disable geographic enrichment.
  • Anonymous events help us understand unsuccessful searches, region choices, subscription form visibility and recognized AI referrals without collecting raw referral URLs.

Analytics quality checks

  • Added automated checks to protect analytics privacy, QA traffic exclusion and search conversion measurements.

API analytics transparency

  • Our privacy page now explains how API account activity and Stripe billing records are used to measure activation, repeat usage, and revenue, and how to request exclusion from account analytics.

More reliable API analytics

  • Added checks for API activity and billing analytics, including missing deliveries, privacy protections, and dashboard accuracy.

Application error monitoring

  • Added error and performance monitoring to help diagnose website failures, with account and request data collection disabled.
  • Updated the privacy policy to explain application monitoring.

Anonymous subscription confirmation counts

  • We now measure completed news and price-alert confirmations without collecting email addresses or linking confirmations to earlier browsing.
  • Repeated confirmation clicks count once, and marked QA signups remain separate from public analytics.

GPU comparisons show where each price is

  • GPU-vs-GPU pages now name the provider behind each cheapest on-demand price, show the price per GB of GPU memory, and list providers that rent both GPUs with each one's price side by side.

Prepare an offer visibility experiment

  • Added support for comparing fresh-stock defaults with a broader offer catalog. The experiment is inactive until launched.
  • Public-site measurement uses a random anonymous browser cookie lasting up to 90 days to keep experiment assignment consistent. Do Not Track and Global Privacy Control remain respected.

Privacy controls for session replay

  • Public-page session replay helps us investigate usability problems, with input values and subscription email addresses masked. Developer pages remain excluded, and browser privacy preferences are respected.

Public-site analytics privacy controls

  • Added PostHog analytics for public-site searches and interactions, with no session replay, account tracking, or persistent browser storage.
  • Updated the privacy policy to explain the new analytics and browser privacy controls.

More reliable search analytics

  • Search analytics now distinguish successful results, empty results, and request failures while keeping canceled requests out of failure counts.
  • Explicit QA visits can be excluded from public analytics without storing cookies or visitor location.

Better service failure detection

  • Improved detection of API failures and missed billing and collection health checks, with sensitive request and payment details excluded from monitoring.

Meshive added

  • Meshive GPU Cloud is now listed: RTX 5090 pods from 1× to 6× and the RTX PRO 5000 Blackwell, run in Meshive's own facility in Seoul, South Korea, billed per minute with spot prices.
  • Free GPU counts come from Meshive's public pricing feed every hour and are shown next to each row.

More accurate GPU models and provider availability

  • Akash, Runpod and Lium now keep SUPER GPUs separate from their base models. Lium nodes with zero free GPUs are no longer treated as rentable.
  • Runpod availability is checked separately for Secure and Community Cloud. Lambda, AceCloud, Vultr, DigitalOcean and Shadeform stock is matched to the exact offer and region.
  • Complete inventory checks now withdraw disappeared offers, including Lium, Fluence, Theta community, Akash and EmpirioLabs listings. Lium also reconciles stock when hourly prices refresh.

Exact Vast GPU variants and rental links

  • Vast listings now keep RTX 2080/4070/4080 SUPER, RTX 4070 Ti SUPER, RTX 4090D, laptop RTX 3060 and A800 separate from their base models.
  • Rental links select the exact GPU model and count across the audited Vast inventory, including larger RTX 4090 and RTX 5090 configurations.
  • Every complete Vast price refresh now updates availability too, clearing configurations that disappeared from the latest inventory.

Vast availability matches its console filters

  • Vast prices and stock now exclude hosts that do not meet the console defaults: 7 days remaining rental time, 90% reliability and 16 GB disk.
  • RTX 2060 SUPER and RTX 2070 SUPER are listed separately from the base models. Unavailable listings no longer claim that the machine was rented.

Open matching GPU offers on Vast

  • Listings for 17 more GPUs now open Vast with the selected model and GPU count. Marketplace browse links show the model at any count.
  • The A100 and A100 80GB pages also filter by memory per GPU.

Vast listings rented out right now

  • GPU pages now note how many Vast.ai listings for that GPU were rented out at the last check, with their last listed price and a link to browse live Vast offers. These listings stay out of the default table because they cannot be rented right now.

A clearer API and MCP guide

  • Find setup steps and reference sections with a persistent desktop guide, clearer connection details, and more readable layouts on mobile.
  • See every MCP client option without scrolling the tab bar, with copy controls beside each configuration filename.

GPU Finder is in the Claude directory

  • Claude users on the web, desktop and mobile can now add GPU Finder from the Claude connectors directory and sign in with their GPU Finder account, with no API key to paste.
  • The developer guide links to the listing from the agent setup steps.

Google sign-in for developer accounts

  • Continue with Google using the same verified email to keep your developer account, API keys, usage and plan.
  • Google accounts with third-party email addresses confirm their mailbox before connecting. Email sign-in remains available.

App names on MCP consent screens

  • MCP connection requests now display the registered application name before you allow access.

A more consistent public site

  • Refreshed the GPU catalog, provider directory, comparisons, picks, blog, About, changelog, research index, and cost calculator with clearer page introductions and consistent spacing.
  • Made catalog cards easier to scan, added About section shortcuts, and improved blog and release layouts on desktop and mobile.

Availability Index: October 2026 edition

  • New: the GPU Cloud Availability Index October 2026 edition (/research/gpu-cloud-availability-index/2026-10) covers 1 October through 1 November UTC: 181 provider, GPU and node-size records, 91 of them ranked, with CSV and JSON downloads
  • Changed: Akash, Koyeb, OCI and Verda are ranked for the first time; Digital Ocean falls below the coverage floor and Vast.ai stays unranked

Automatic budget H100 pick

  • The Budget guide selects the lowest-priced eligible single H100 on demand and shows its region, stock check, availability score and selection rules.

Live cost picks for inference, fine-tuning and local LLMs

  • The Inference, Fine-tuning and Consumer guides now show the lowest in-stock on-demand rate for their chosen GPU, with the winning provider, stock check and selection rules. The fine-tuning pick requires per-second billing.

Marketplace stock and rental selection

  • PrimeIntellect's two-H200 selection now requires fresh in-stock evidence; other listings retain the general provider link.
  • Vast prices and stock now cover verified, rentable hosts, matching the console's verified-host filter.
  • PrimeIntellect and Vast prices now refresh hourly alongside marketplace stock checks.

Find a GPU rental from your MCP client

  • Use the find_gpu_rental prompt to compare the cheapest listed node with the cheapest node recently observed in stock. The pricing and stock guide explains hourly rates and how to interpret missing observations.

Live European 8× H100 training pick

  • The Training guide shows the lowest on-demand hourly price for an in-stock eight-H100 server in Europe, along with its provider, location, last stock check and selection rules.

Keep your RTX 4090 selection when visiting Vast

  • Vast rental actions for one or two RTX 4090 GPUs now open with the GPU model and exact GPU count selected.

Clearer stock now and availability history

  • Availability labels now distinguish stock now from the percentage of tracked time reported fully available, with visible coverage hours.
  • Tables and stock summaries explain why limited stock can be rentable beside a 0% history. History coverage uses listing-hours instead of implying continuous calendar days.

Consistent GPU rental summaries

  • Single-GPU recommendations now use single-GPU price and stock history, with recorded catalog prices clearly separated from current stock. Charts label their own node size.
  • About’s data-source breakdown now uses the same visible provider total as its headline and the homepage.

Hinode GPU pricing added

  • Compare on-demand L4, L40S and RTX PRO 6000 machines in US East and Frankfurt. Each hourly price includes one running workspace. Storage and outbound traffic are billed separately.

Keep your GPU selection when visiting Lium

  • Lium rental links now open with the listing’s GPU and GPU count selected, plus its country when the location can be mapped reliably.

Keep your H200 selection when visiting PrimeIntellect

  • The verified DataCrunch two-GPU H200 141GB listing now opens PrimeIntellect's matching offers with the GPU variant and count selected. Other listings keep their existing destination.

Provider-specific GPU pricing pages

  • Compare Runpod B200, Lium RTX 5090 and Nebius H100 pricing by GPU count and region, with reported stock and clearly explained stock history.
  • Pages enter the sitemap only when at least two distinct priced configurations are tracked.

Recurring GPU availability checks with REST or MCP

  • Market snapshots can now be scoped to a provider and exact provider-reported region before listings are combined.
  • Download a watcher that compares stock and whole-node budget eligibility across runs, stays quiet when unchanged and preserves its baseline on failed checks. The API guide includes REST/MCP setup, a scheduled-client prompt and call-budget examples.

Keep your RTX 5090 selection when visiting Vast

  • Vast rental actions for one or two RTX 5090 GPUs now open with the GPU model and exact GPU count selected.

Password sign-in for provisioned accounts

  • Provisioned developer accounts can sign in with an existing password and continue through the standard MCP connection consent flow.

Clearer MCP tool names

  • All eight MCP tools now have human-readable titles to make GPU pricing, availability, history and account usage easier to identify in AI clients.

Compare rental configurations and whole-node job budgets

  • RTX 4090 and H100 rental pages now explain host, allocation and billing checks before choosing a provider.
  • Eight-GPU H100 and H200 pages estimate 8-hour, 24-hour and 7-day compute budgets from the lowest on-demand whole-node rate with fresh reported stock.

Clearer verified stock answers and provider catalogs

  • GPU and node answers distinguish fresh self-service stock from catalog coverage, show the selected stock-check time, and explain observed reliability coverage.
  • Provider pricing pages include clearly labelled unknown-stock catalog prices by default. GPU pages remain accessible when catalog listings have no confirmed stock.

EmpirioLabs AI added

  • EmpirioLabs AI GPU Cloud is now listed: 27 GPUs from RTX A4000 to B300 in 1× to 8× sizes, billed per second, with spot prices where they undercut on-demand.
  • Stock comes from EmpirioLabs' public catalog every hour and is shown next to each row.

Hot Aisle shows sold out instead of going stale

  • When every Hot Aisle MI300X is rented, its rows now read out of stock with a fresh check time, instead of keeping their last in-stock reading

Lium GPU rows open every variant on Lium

  • Lium rows for H100, H200, B300 and RTX PRO 6000 now open Lium's page for the whole GPU family, so you see every variant we priced (for example H200 and H200 NVL), not just one
  • Transparency: Lium GPU-row links now carry our Lium referral code and are marked rel="sponsored". Rankings, prices, availability, and sort order are unaffected

Focus comparisons on available self-service offers

  • Comparisons hide out-of-stock, unknown or stale stock, and sales-only offers by default. Limited stock stays visible.
  • New offer visibility filters let you include those listings again. Default from prices use the same eligible offers.

Clearer listed prices and stock claims

  • GPU page previews now use the same confirmed single-GPU offer as the page answer, or explicitly label a listed price when stock is unconfirmed.
  • GPU price ranges and structured offers distinguish confirmed stock from catalogue listings.
  • Provider descriptions now explain when stock is unknown, including Google Cloud, instead of promising hourly availability for every provider.

Add GPU Finder to Cursor or VS Code in one click

  • When you create or rotate an API key in the developer portal, Add to Cursor and Add to VS Code buttons install the MCP server with that key already filled in
  • A Copy Claude Code command button gives you the claude mcp add command with your key in place

Azure prices now come from Azure's own price list

  • Azure GPU prices are read from Azure's public Retail Prices API instead of the SkyPilot catalog, which had not changed since September 2025 and covered only 8 US regions. Azure now shows on-demand and Spot prices in about 50 regions, and H200, MI300X, GB200 and RTX PRO 6000 VMs appear for the first time
  • The ND H100 v5 VM is listed with its real 8 GPUs instead of 12, so its per-GPU price is no longer understated

One tabbed setup block for every MCP client

  • The API page now shows MCP setup for Claude Code, Cursor, VS Code, Claude Desktop, Codex CLI, Gemini CLI and Windsurf/Cline in one tabbed code block with a copy button; your last choice is remembered
  • Setup now starts with getting a free key, so the snippets you paste already have what they need
  • Clearer layout: larger type, numbered steps, example prompts you can try, and protocol details moved to the end
  • New Pricing section on the API page: Free, Paid and Unlimited side by side, with the detailed usage rules one click away

REST examples in cURL, JavaScript and Python

  • The API page's REST quickstart now has cURL, JavaScript and Python tabs, a sample response and a short guide to reading it
  • Errors are now a table of status, code and what to do, plus fixes for common MCP connection problems

Run the MCP server locally with npx gpufinder

  • New open-source npm package: npx -y gpufinder mcp starts the GPU Finder MCP server over stdio for Claude Desktop, Cline, Windsurf and other clients that run local servers. It forwards to the hosted endpoint with your API key, so tools and usage are the same
  • The API page's Claude Desktop setup now uses the package instead of mcp-remote

Price alerts one click from the answer

  • Every GPU page now says under its opening answer that stock is re-checked every hour, with a link that opens the price-alert form for that GPU
  • Provider pages with a stock signal show how many configurations were in stock at the last check and link to the cross-provider price table

Verda prices now come from Verda's own API

  • Verda GPU prices are read from Verda's own API instead of the SkyPilot catalog, which only listed the instance types that happened to be in stock when it was generated. Types such as 4× H200, 2× RTX PRO 6000 and 8× L40S now appear, in every Finnish location where they can be deployed
  • A Verda instance that sells out keeps its price and is shown as unavailable, instead of disappearing until the next catalog update

Digital Ocean prices now come from DigitalOcean's own API

  • Digital Ocean GPU Droplet prices are read from DigitalOcean's own API instead of the SkyPilot catalog, which listed only H100 and at outdated prices. H200, L40S, RTX 4000 Ada, RTX 6000 Ada, MI300X and MI325X now appear with on-demand prices, and B300, MI350X and MI355X with spot prices

Fixed: Digital Ocean GPU stock

  • Fixed: Digital Ocean's GPU stock is read from its full size list again (it had grown past one page), and a GPU Droplet that DigitalOcean no longer lists is shown as out of stock instead of keeping its last known status

Fixed: "most reliably in stock" named providers that were out of stock

  • Fixed: the "most reliably in stock" line on GPU pages judged each provider by its single best listing, so a reseller with dozens of listings could qualify on one well-stocked listing while the rest were out (Shadeform for H100). Providers are now scored on the average of all their listings for that GPU, and only those with confirmed stock right now are named

Fixed: Nebius RTX PRO 6000 listed as RTX 6000

  • Fixed: Nebius's 96 GB RTX PRO 6000 Blackwell Server Edition was filed under the older 24 GB RTX 6000, so it was missing from RTX PRO 6000 pages, spot prices and price history. It now appears there, history included

Fixed: reliability scores counted stock we had stopped checking

  • Fixed: when a provider's stock check stopped covering a listing, its last known status was still counted every day, so the 30-day reliability score kept rising on stock nobody had checked (Digital Ocean H100 read 100% on a check from August 25). A status now counts only while it is less than 36 hours old

New: /spot-gpu-prices — spot, preemptible and interruptible GPU prices from every provider

  • New: /spot-gpu-prices compares the cheapest spot price per GPU-hour from every provider that publishes one, for each GPU at least two providers price on spot, cheapest first: the provider's own term, how the price is set, discount vs on-demand, what happens when the capacity is reclaimed, and Nebius's published floor-to-ceiling range. Only listings read in the last 36 hours count. CSV and JSON downloads included
  • Linked from the footer, /cheapest-cloud-gpu, the spot section of every GPU page and the Runpod vs Vast comparison

Oracle Cloud prices now come from Oracle's own price list

  • Oracle Cloud (OCI) GPU prices are read from Oracle's public price list instead of the SkyPilot catalog, which had not changed since September 2025. Existing shapes keep their prices, and B200, B300, GB200, GB300, MI300X, MI355X and RTX PRO 6000 bare-metal shapes now appear with on-demand prices

OVHcloud prices now come from OVHcloud's own API

  • OVHcloud GPU instance prices are read from OVHcloud's own price list instead of the SkyPilot catalog, which had not been updated in a year and listed only A100 and V100. H100, H200, L40S and L4 instances now appear, and V100 prices reflect OVHcloud's current rates

Scaleway prices now come from Scaleway's own API

  • Scaleway GPU Instance prices are read from Scaleway's own API instead of the SkyPilot catalog, which listed only H100 PCIe, P100 and RTX 3070 in Paris, at 2025 prices quoted in euros as if they were dollars. L4, L40S, H100 SXM and B300 now appear too, in Paris and Warsaw, with euro list prices converted to US dollars at the daily ECB rate
  • Scaleway stock is now tracked per region, so a shortage in Warsaw no longer hides behind capacity in Paris

Spot price history on GPU price history pages

  • Added: each GPU's price history page (e.g. /gpu/h100/price-history) now plots the weekly spot / preemptible / interruptible floor next to the on-demand floor, with a table of the last 8 weeks, when at least two providers publish spot prices for that GPU
  • Added: CSV and JSON downloads of every price history page's data
  • Changed: the spot section on GPU pages now compares the cheapest spot price per GPU from every provider that publishes one (previously only Runpod and Vast), cheapest first, with each provider's own term and, where the provider documents it, what happens to the instance when it is reclaimed

Spot prices: 7-day range per provider, and links from every comparison

  • New: each row on /spot-gpu-prices shows the lowest and highest of that provider's cheapest spot price over the last 7 days, so you can see whose spot price moves and by how much. Same rule for every provider; hourly-read prices show intraday moves. Also in the CSV and JSON downloads
  • Every provider comparison where either side sells spot now lists both sides' cheapest spot price and links to /spot-gpu-prices; provider spot sections link there too

Spot prices show each provider's own term, how the price is set, and when it was read

  • Improved: every spot price in the tables now carries the provider's own name for it (Spot, Preemptible, Interruptible or Bid) and a grey badge for how the price is set: fixed, dynamic, bid or marketplace
  • Added: each spot price says how often GPU Finder reads it and when it last did, e.g. "updated hourly · as of 10:21 UTC"; the line turns amber when a reading is overdue
  • Added: a preemptible pricing section on /providers/nebius covering Nebius's dynamic spot pricing (live since October 8, 2026), with an on-page disclosure that GPU Finder's founder works at Nebius

Wider pages on large screens

  • Pages now use up to 1280px of width instead of 1024px, so price tables and comparisons show more at a glance on laptop and desktop screens. Articles keep their shorter, easier-to-read line length

Withdrawn marketplace offers no longer show up

  • Fixed: Vast, Lium and PrimeIntellect list only offers you can rent right now, so an offer missing from their feed for 36 hours is now removed from every table, floor price, chart and API answer instead of lingering for 14 days at a price nobody offers. Other providers keep the 14-day window, so a SKU that briefly drops out of a catalogue does not flicker

Fluence GPU marketplace added

  • New: Fluence is listed as a marketplace with on-demand VM and bare-metal prices from H100, H200, B200 and B300 down to L4 and RTX A4000, in 1, 2, 4 and 8 GPU configurations across the US, Canada, Europe, India, Japan and Singapore
  • New: every Fluence row names the upstream data center operator it is resold from (Spheron, Massed Compute) in its instance type. Offers resold from DataCrunch are left out because Verda (formerly DataCrunch) is already listed directly
  • New: Fluence prices are quoted in USDC and shown as USD 1:1; launching an instance prepays the first 2 hours. Availability is refreshed hourly from Fluence's live offer list

Live prices in search results, and an embeddable price badge

  • New: GPU, GPU-vs-GPU, provider comparison and cheapest-GPU pages now show the current cheapest on-demand price per GPU-hour in their search-result title and description, refreshed hourly. The cheapest-GPU page quotes only prices reported in stock
  • New: a live-price badge for any GPU or provider at https://gpufinder.dev/badge. It is a small image that shows the cheapest on-demand price per GPU-hour and links to the matching GPU Finder page; copy the HTML or Markdown snippet to add it to a README, docs page or pricing page

Provider stock history counts GPU models, not configurations

  • Improved: the 7-day stock squares on /providers now show the share of a provider's GPU models with at least one configuration in stock each day. Before, every GPU count and region counted separately, so a provider with many sold-out sizes of an in-stock GPU showed red

SaladCloud GPU pricing and live stock added

  • New: SaladCloud is listed with one row per consumer RTX GPU class it prices, from RTX 3080 to RTX 5090, including RTX 5080, RTX 5070 Ti, RTX 4070 Ti Super and RTX 3090 Ti. Prices are read daily from Salad's published price list and include vCPU and RAM
  • New: SaladCloud's High priority (never preempted by other workloads) is shown as the on-demand price and its Lowest priority (interruptible batch) as the interruptible price. Medium and Low priority sit between the two and are explained on the SaladCloud provider page
  • New: SaladCloud availability is checked hourly and shows the number of free High-priority nodes for each GPU class. Placement is automatic across Salad's network, so every row is listed as Global

Weekly price history on GPU pages

  • Changed: the Price History chart on GPU pages now covers the last 6 months and plots one point per week, so price moves within a month are visible. Weekly prices are tracked from August 10, 2026; earlier months show the month-end price and are labelled as such in the tooltip
  • The full multi-month record is still on each GPU's price history page, e.g. /gpu/h100/price-history

The MCP server works with Cursor, VS Code, Claude Code and Claude Desktop

  • New: MCP clients that open with the 2025-era handshake, which is most of them today, can now connect to https://gpufinder.dev/mcp with an API key and call every tool. Before this, only clients on the newest protocol version could
  • Same price on either protocol: listing tools and connecting are free, and each data call spends one unit of your monthly allowance
  • New: copy-paste setup for Cursor, VS Code, Claude Code and Claude Desktop on the API page

Akash Network GPU pricing and live stock added

  • New: Akash Network is listed with one row per GPU model on its decentralized marketplace, including H100, A100, RTX PRO 6000, RTX 5090, RTX 4090, RTX 3090, T4 and V100. Prices are read daily from Akash's public price feed
  • New: Akash availability is checked hourly and shows the number of free GPUs of each model across every provider on the network
  • Note: the Akash price is the capacity-weighted average of what its providers recently bid for a single GPU, the same figure Akash publishes. Your own deployment is priced by the bids it receives and can be lower or higher. CPU, memory and storage are sized per deployment and billed separately

Koyeb GPU pricing and capacity levels added

  • New: Koyeb is listed with its self-serve NVIDIA GPU instances: RTX 4000 SFF Ada, RTX A6000, L40S, A100, RTX PRO 6000, H100 and H200, in 1x, 2x and 4x configurations where Koyeb sells them. Prices are read daily from Koyeb's public catalog and are billed per second
  • New: Koyeb availability is checked hourly. Koyeb publishes a capacity level (low, medium or high) for each instance type and region rather than a stock count; low is shown as limited, medium and high as available
  • Note: Koyeb instances that need access approval, such as its 8x nodes, are not listed

PrimeIntellect prices and stock come straight from the PrimeIntellect API

  • Changed: PrimeIntellect prices are now read from PrimeIntellect's own API instead of the SkyPilot catalog. The catalog's PrimeIntellect file had not been updated since September 2025, so it listed configurations PrimeIntellect no longer sells, at old prices. The listing now shows what PrimeIntellect can rent out today
  • Fixed: PrimeIntellect rows that its live feed no longer listed kept their last stock status and showed a stale availability badge. They are now shown as out of stock, and drop off the site after 14 days
  • Changed: PrimeIntellect availability is reported per upstream cloud, configuration and country instead of one status per GPU model and count

Nebius prices come straight from the Nebius API, including floating spot prices

  • Changed: Nebius on-demand and preemptible prices are now read from Nebius's own API instead of the SkyPilot catalog
  • New: Nebius moved preemptible GPUs to floating spot pricing on 8 October, set per GPU type and region. The Nebius spot column shows the spot price at our latest daily price collection, so the same GPU can show different spot prices in different regions

Availability Index methodology 1.1: retired listings no longer count as unavailable

  • Changed: from methodology 1.1 a configuration counts for a month only if the provider offered it then, meaning it was priced in the catalogue or reported in stock at least once. Listings a provider has retired are excluded instead of being scored as 0% available
  • New: September 2026 is published under both versions. The edition first published under methodology 1.0 is unchanged (/research/gpu-cloud-availability-index/2026-09); the methodology 1.1 restatement (/research/gpu-cloud-availability-index/2026-09-v1.1) uses the same history and freeze time and reads H100 27%, H200 48%, A100 51% and B200 15% of covered hours, with its own CSV and JSON
  • Changed: editions from October 2026 onward use methodology 1.1 only; each edition page states which version it was computed under

Availability Index: September 2026 edition and findings

  • New: the GPU Cloud Availability Index September 2026 edition (/research/gpu-cloud-availability-index/2026-09) covers 1 September through 1 October UTC, frozen on 1 October: 120 provider, GPU and node-size records, 92 of them ranked, with CSV and JSON downloads
  • New: a findings post, H100 Availability, September 2026 (/blog/gpu-cloud-availability-september-2026), with the first month-over-month comparison: H100 was reported available 25% of hours, down from 34% in August
  • Changed: Lium and Packet are ranked for the first time; Cudo, delisted on 19 September, is not in this edition

Google Cloud pricing guide, and comparison pages that link to each provider's full catalog

  • New: the Google Cloud (GCP) provider page has a pricing guide with the cheapest B200, H200, H100, A100, L4 and T4 rows, plus answers built from live data on H100 and A100 per-GPU cost, Spot VM prices, egress and TPUs
  • Improved: every provider comparison links straight to both providers' full GPU pricing pages, and its search snippet now shows the models both providers list, which side is cheaper on how many, and a live H100 (or lead model) price pair
  • Clarified: AWS vs Lambda compares AWS EC2 with Lambda (formerly Lambda Labs), the GPU cloud, not the AWS Lambda serverless service

Lium B300 and A100 80GB listings

  • New: Lium B300 rows, including whole 8x B300 hosts, and Lium A100 80GB rows — both labels Lium uses for B300 (SXM6 AC and SXM6 PC) and its A100-SXM4-80GB label were previously skipped
  • Improved: Lium's B300, A100 80GB and H200 rows link to the Lium explore filter with the most hosts for that GPU

Cudo removed: it no longer offers on-demand GPUs

  • Removed: Cudo Compute has moved to large-scale AI infrastructure and no longer rents GPUs on demand, so its provider page, prices and stock tracking are gone; the Cudo vs Runpod comparison now redirects to /compare

X daily post: floors now come from the full price history, and a quiet market no longer skips the post

  • Fixed: the @GPUFinder_dev post considered only offers repriced in the last eight days, so stable-priced offers were invisible to it — H200 read Vast $2.61 while PrimeIntellect $2.00 was in stock, B200 $5.60 instead of Lium $5.50, and MI300X (Runpod $2.39) was missing from every post since launch. The post now uses each offer's latest recorded price, exactly like the market band on the homepage, so the two agree
  • Fixed: the post's freshness check measured the age of the last price change instead of the last collector run that verified the offer, so a day without repricing (11 September) skipped the post. It now checks when the offer was last seen; the 36-hour limit is unchanged

llms.txt now follows the llmstxt.org format

  • Changed: /llms.txt has a single H1 and a blockquote summary, and every page, API endpoint, dataset and provider entry is an absolute Markdown link (- [name](https://gpufinder.dev/…): notes) instead of a bare path — site audits had flagged the file as containing no links, and an agent reading it out of context can now resolve every URL without knowing the host

Node-size and region pages: their own overview, Product markup, and links from the hub, home and footer

  • New: every /gpu/<model>/8x and /4x page opens with a short overview computed from its own rows — the node-vs-per-GPU-hour math with the median listing and the provider spread, how many providers actually have that size in stock versus the listed floor, and spot and egress economics for that node size (what 1 TB of checkpoints out costs in hours of the cheapest node)
  • New: every /gpu/<model>/us, /europe and /asia page explains the regional floor against the global one per GPU-hour, which providers have endpoints in the region (and which global pools are counted), the node sizes on offer there, and stock and egress for those listings
  • New: node and region pages carry Product + AggregateOffer structured data (whole-node USD/hr for node pages, per-GPU-hour for region pages), matching the model pages
  • Improved: the /gpu hub lists every submitted node-size and region page, the home page and the site footer link the most-searched ones (8x/4x H100, 8x H200, 8x B200, H100 by region), node pages link their model's region pages and vice versa — until now each of these pages had a single inbound link
  • Changed: node-size pages are submitted in the sitemap only from 3 priced providers (they still render from 2); OG card images and hashed build assets now send X-Robots-Tag: noindex so they stop appearing in Search Console's crawled-not-indexed report

Availability Index August 2026 revised: Theta EdgeCloud artefact removed, Vultr B200 re-keyed to 8x; new post on the findings

  • Fixed: the August 2026 Availability Index edition is regenerated from corrected history at its original 1 September freeze time. Theta EdgeCloud had ranked second for H100 and H200 at 100% because its static cloud-VM catalogue was recorded as available until the 23 August fetcher fix; those 322 records (1-23 August, all Theta cloud VMs) are now unknown, Theta no longer ranks, and every other record is unchanged. The page and the JSON carry a revision note
  • Fixed: Vultr's vbm-256c-3072gb-8-b200-gpu bare-metal plan is an 8x B200 node, not 9x — the count was derived from 1,536 GB of GPU memory at 180 GB per card; B200 is now 192 GB per card and the existing row was re-keyed so its history follows
  • New: /blog/gpu-cloud-availability-august-2026 — what a month of hourly stock checks says: H100 reported available 34% of covered hours across 12 ranked clouds (25% outside one marketplace feed), H200 and A100 about 52%, B200 5%; 8x nodes far scarcer than single cards; a third of provider × GPU × size combinations never available. With the caveats that make the numbers safe to quote
  • Improved: the Availability Index generator accepts --generated-at and --revision-note so a published edition can be regenerated from corrected history without moving its freeze time

Cudo stock badges: Out instead of Stale, A100 rows gain stock data

  • Fixed: Cudo node sizes that no single host can currently fill (for example 8× H100 when the largest host has 4 free GPUs) now show a fresh "Out" badge; previously the hourly check skipped them, so their badge froze on the last snapshot and read "Stale · 64d ago" while the price stayed current
  • Fixed: Cudo's 80 GB A100 listings now receive stock data — the stock feed labelled the card differently from the pricing catalogue, so all 13 rows showed "No data" despite free capacity

GPU Finder is on X: a daily cheapest-in-stock price post from @GPUFinder_dev

  • New: follow @GPUFinder_dev on X for one post every morning (07:15 UTC) with the cheapest in-stock price per GPU-hour for H100, H200, B200, A100 80GB, MI300X and RTX 4090, plus the biggest 24-hour move among them — the same numbers as the market band on the homepage, generated from the same data and never edited by hand
  • New: the footer, the Organization metadata and llms.txt link the account; Twitter cards now name it as the site handle

Search titles: brand-first homepage, cloud-intent H100/A100, server-price node pages; Lambda pricing guide

  • Changed: the homepage title now leads with the brand — "GPU Finder — Cloud GPU Pricing Comparison (2026)" — after the 24 August retitle pushed the name to the end of an 82-character title and the page slid from position 2 to 12 on its own name in Google within a week
  • Changed: the H100 and A100 pages are titled for cloud pricing per hour ("H100 Cloud Pricing per Hour (2026) — Rent H100 GPUs") with matching H1 and description; the old rental template ranked 55-77 on retail card-price queries and never for the cloud phrases the pages answer
  • Changed: node-size pages say "8x H100 Server Price (2026) — from $12.80/hr" to match how people search for multi-GPU node prices; the RTX 5090 page is titled "Rent RTX 5090 in the Cloud (2026) — Prices & Live Stock"
  • New: a Lambda GPU pricing guide on /providers/lambda — cheapest H100 and A100 per GPU-hour from live data, per-minute billing, no egress fees, and how Lambda compares with AWS — alongside the AWS guide
  • New: /providers answers the questions people ask about GPU cloud providers (who we compare, which is cheapest, who shows live stock) with visible FAQs and structured data
  • Changed: /picks/budget is now "Budget GPU Cloud Picks for AI Workloads" and /compare points readers to the cheapest cloud GPU table, so three pages stop competing with /cheapest-cloud-gpu for the same search
  • Fixed: price-history pages now require at least two providers with priced history as well as six months of data — the single-provider pages that entered the sitemap overnight on 4 September (K80, M60, M4000, P4000, TPU v3, Inferentia, a Virtex FPGA, RTX PRO 6000 96GB) are removed again; 29 price-history pages remain
  • Improved: the egress fees post is refreshed for 30 providers and covers the other hidden GPU cloud costs — billing increments, idle storage, and re-runs from unreliable capacity

Availability Index: complete August 2026 edition

  • Improved: the GPU Cloud Availability Index August 2026 edition (/research/gpu-cloud-availability-index/2026-08) now covers the complete calendar month, 1 August through 1 September UTC, generated after the period closed; it replaces the month-to-date freeze of 24 August that was published at launch, and the CSV and JSON downloads carry the same 124 provider × GPU × node-size records
  • Improved: the Availability Index hub and edition pages describe complete and month-to-date editions accurately and state when an edition supersedes an earlier freeze at the same URL

Data-correctness fixes: marketplace rows, Lambda regions, price guards

  • Fixed: Vast.ai listings are one row per GPU, GPU count and location — the cheapest SXM/PCIe/NVL variant no longer opens a new row and leaves the old one showing a stale price for two weeks; Vast stock is now reported per location instead of one global count on every row
  • Fixed: Theta EdgeCloud community rows keep one stable row per GPU and count instead of moving to a new row whenever the cheapest host changes country
  • Fixed: Lambda regions that drop out of the capacity list now show "unavailable" within the hour instead of staying "available" until the listing ages out; Theta cloud VMs no longer claim 100% availability from a static catalogue; Hyperstack, Hot Aisle, Cudo and Theta community stock use the same thresholds (>5 available, 1-5 limited)
  • Fixed: a $0 or broken on-demand price can no longer become a floor, a chart bar, the homepage's first row or a "near its 6-month low" claim; spot prices equal to or above on-demand are no longer stored or shown as spot
  • Fixed: GPU-vs-GPU pages compare per-GPU-hour floors on both sides (an 8×-only model is no longer measured against the other side's single-GPU price) and say so in the table; copy now states prices refresh daily and availability hourly
  • Improved: /api/v1/status reports, per provider and per collector (prices, availability), the last successful run, the current failure streak and the last error
  • Improved: one shared pricing pipeline behind the homepage, GPU pages, provider pages and the API — region merging, spot labelling and pool-level stock de-duplication now behave identically everywhere; /api/v1/prices rows come back sorted by on-demand price
  • Improved: monthly price history is materialized every morning (previous + current month, the price in effect at month end) so the trend chart and /api/v1/history always reach the current month — including GPUs that had no 2025–26 backfill

Design refresh: one price table, answer cards, mobile cards, dark mode

  • New: every pricing table on the site (homepage, GPU pages, provider pages, comparisons) is the same component — prices in monospace numerals, sortable headers with a visible sort state, a sticky header and first column, a per-row Visit button, and marketplace listings collapsed behind "+N more listings" so one provider no longer fills the page
  • New: on phones the tables become cards with the price first, and the homepage filter collapses to one line ("H100 × 8 · All regions · Edit") so the prices are in the first screen
  • New: an answer card — cheapest in-stock price, provider, region, stock badge and a Visit button — opens the homepage, every GPU page and every comparison, and the same card is drawn in the link-preview image
  • Improved: one availability vocabulary everywhere: In stock / Limited / Out / Stale · Nh ago / No data, with the 30-day reliability share on a second line; heatmap cells carry a ✓ / ~ / ✕ glyph so colour is never the only signal
  • Improved: /providers shows each provider's logo, category, one-line description, billing granularity, live-stock status, egress, cheapest H100 per GPU-hour, a 7-day availability sparkline and top GPUs, and can be sorted by cheapest H100, most GPU models or name
  • New: dark mode follows your system setting; a new 3×3 availability-grid logo and favicon; JetBrains Mono for every number; the primary colour is now a deep green shared with the "in stock" signal instead of generic blue
  • Improved: picks deep-link to the matching live price table; the footer gains a Product column; the dropdown menus work with arrow keys; every page passes Lighthouse accessibility at 100 (focus rings, 24 px targets, reduced-motion support)

H100 hub: cheapest H100 by GPU count, in-stock list, upgrade guidance

  • New: /gpu/h100 opens with "Cheapest H100 right now by GPU count" — the lowest on-demand 1×/2×/4×/8× node at each size, normalised to per-GPU-hour, with stock status and the cheapest spot per GPU-hour — plus the providers reporting H100 stock on the last hourly check and a short "when to upgrade to the H200" section linking the live H100 vs H200 comparison
  • Improved: footer GPU links and the buttons at the bottom of every GPU-vs-GPU page now read "H100 cloud pricing" style anchors, so each GPU hub is the one page that term points to
  • Improved: the link-preview image for the homepage, /gpu, /providers, /compare and the picks pages is a neutral card with the cheapest H100, H200, B200, A100, MI300X and RTX 4090 prices instead of the H100 chart

New: /cheapest-cloud-gpu — the cheapest price per GPU-hour for every GPU

  • New: /cheapest-cloud-gpu lists every GPU model we track with its cheapest on-demand price per GPU-hour and the cheapest listing reported in stock on a check in the last 6 hours (provider, node size, stock badge, 30-day reliability, provider count, spot floor), ordered by the price you can actually rent at; flagship cards for H100, H200, B200, A100, MI300X, L40S, RTX 4090 and RTX 5090 sit on top. Linked from the footer, /gpu and /picks/budget
  • New: /tools/gpu-cost-calculator estimates a job's real bill — node price × hours, egress after the provider's free tier, and the rounding overhead of per-hour vs per-minute vs per-second billing — prefilled with the cheapest in-stock H100 and linked from every provider page's egress section
  • New: region pages /gpu/<model>/europe, /us and /asia for H100, H200, A100, A100 80GB and B200 — every listing in that region with the cheapest in-stock option, the regional price range against the global floor and 30-day reliability; the homepage region filter now recognises many more provider region spellings (marketplace host countries, Shadeform city names, Hyperstack, PrimeIntellect country codes) and global pools count for every region
  • New: nine more GPU-vs-GPU pages with real shared catalogues — B200 vs H100, A100 vs H200, H100 vs L40S, B200 vs B300, GH200 vs H200, L40S vs RTX 5090, A100 vs A100 80GB, RTX 4090 vs RTX A6000, L4 vs L40S — and four provider comparisons (Google Cloud vs Nebius, AWS vs Azure, Hyperstack vs Lambda, Runpod vs Shadeform), each with at least three GPU models priced on both sides
  • New: price-history pages /gpu/<model>/price-history for every GPU with at least six months of data — the cheapest on-demand price each month (chart and per-provider table), 6- and 12-month low/high, month-on-month change, lowest and highest ever, and a visible FAQ; linked from the GPU page's history section
  • New: node-size pages /gpu/<model>/8x and /gpu/<model>/4x for every GPU that at least two providers sell at that size — cheapest in-stock node, per-GPU-hour and per-month cost, the full node-level price table, the 7-day availability heatmap for that size, and a size switcher; linked from each GPU page under "<GPU> by node size"
  • New: the homepage gains a "GPU cloud pricing per hour" table under the main results — the cheapest in-stock price per GPU-hour for H100, H200, B200, A100, MI300X, L40S, L4, A10G, RTX 4090, RTX 5090 and T4 with the provider, node size and 30-day reliability

Shorter page titles, generated llms.txt, count hygiene

  • Improved: page titles lead with the price ("H100 Cloud Pricing from $0.41/hr (2026)"), stay under 60 characters and carry the brand once; RTX 5090 / 4090 read "Cloud Rental", B200 / B300 "GPU Rental & Cloud Pricing", and TPU, Trainium, Inferentia and Gaudi pages are titled as accelerators on their platform ("TPU v5p Cloud Pricing (Google Cloud)") instead of "Cloud GPU Pricing"; the blog index is "GPU Cloud Pricing Blog — Guides & Comparisons"
  • Improved: /llms.txt is generated from the provider registry (counts and the /go link list can no longer drift), links the sitemap and RSS feed, and documents the public /api/v1 endpoints and their rate limit
  • Fixed: the homepage description no longer claims "200+ GPU models"; the about page derives its provider counts from the registry

Structured data pass, visible Q&A, thin-page gating

  • Improved: every page's structured data is now complete and representative — GPU pages carry a Product with image, URL and brand; provider pages an Organization with its offer range; blog posts an Article with image, publisher logo and a real "Updated" date; the homepage a WebSite search action; /about an Organization with logo, founder and profiles — and every FAQ block on the site is visible text, not markup-only
  • New: the market snapshot on GPU pages answers its three questions in plain sight ("Is now a good time to rent…", "What is the cheapest…", "Which providers most reliably have it in stock")
  • Improved: provider comparisons with fewer than three GPU models priced on both sides are no longer offered to search engines and sit behind a disclosure on /compare, which now lists comparisons by catalogue overlap; provider pages open with a short factual blurb (billing, stock source, egress, referral) and breadcrumbs appear on picks, blog and about pages

Faster pages, answer-first homepage, GPU pages reordered

  • Faster: GPU and provider pages now send only the first 20 pricing rows and one availability view in the page payload and load the rest when you press "Show all" or switch the GPU-count view — /gpu/h100 is about half the HTML it was; the homepage likewise ships its first 20 rows
  • Faster: the price charts load only when you scroll to them, and the GPU page opens on the provider comparison tab so no history request fires on load; the comparison bar chart no longer clips the tallest bar, the history chart defaults to the six cheapest providers plus "Others" with a labelled $/hr axis, and chart animation respects reduced-motion settings
  • Faster: provider comparison pages render their editorial and FAQ on the server (FAQ answers are now visible in the page, matching the structured data) instead of shipping the whole page as JavaScript
  • Improved: the homepage leads with the answer — the cheapest in-stock 8× H100 right now, provider count and last update — with one alert sign-up under the summary bar; the explanatory copy moved below the table
  • Improved: GPU pages are reordered — prices first, then availability, history, spot and specs — with a jump nav; the market snapshot states the unit next to every price ($/GPU-hr vs per-node) and no longer shows "mid-range" in red
  • Fixed: the 7-day availability heatmap uses the same time-weighted math as the reliability score; Google Cloud tpu-v6e placeholder catalog rows ($700/hr flat across every topology) are excluded; provider pages and /api/v1/providers quote the smallest-node price range like the provider index
  • New: RSS feed for the blog at /feed.xml; GPU-vs-GPU pages carry share images; the mobile menu closes on Escape and shows a close icon

Accuracy, sitemap and mobile fixes from the August site review

  • Fixed: Vultr fractional GPU slices (2–4 GB of a shared card) are no longer listed as full 1× GPUs — the A16 "from $0.059/hr" and A40 "from $0.075/hr" floors were slice prices; existing slice rows age out within 14 days
  • Fixed: GPU model pages with no current offers now return 404 instead of "0 providers, from $0.00/hr", and the sitemap only lists models with live prices, with real last-modified dates from the data instead of the deploy date
  • Fixed: GPU URLs are plain lowercase slugs (e.g. the Virtex UltraScale+ page no longer 404s); legacy forms redirect
  • Fixed: the model-page price comparison chart only compares instances at the labelled GPU count (no more 1× rows inside an "8x" chart) and takes on-demand and spot from the same instance
  • Fixed: the /providers index refreshes hourly again (it had been frozen for over a day) and quotes each provider's smallest-node price range — the Google Cloud "$27,399/hr" TPU-pod figure is gone
  • Improved: availability percentages and the 7-day heatmap are now time-weighted (a one-hour stock-out reads ~96%, not ~50%), and "cheapest in stock" is gated on current stock status
  • Improved: on phones the price columns come first and vCPU/RAM columns hide on narrow screens, so On-Demand and Spot are visible without scrolling; sortable headers announce their sort state to screen readers; muted text meets AA contrast
  • Improved: homepage title and heading say what the site is — cloud GPU pricing comparison — and every provider count on the site comes from one constant
  • Security: price-alert sign-ups only accept known GPU models and sane counts/prices, reject cross-site posts, and send at most one confirmation email per address every 10 minutes; weekly database backups no longer include subscriber data

Lium referral link

  • Transparency: Lium links (provider page, /go/lium, API providerUrl, alert emails) now carry our Lium referral code — this supersedes the Aug 20 note that Lium links had none. Rankings, prices, availability, and sort order are unaffected; outbound links stay rel="sponsored" and the About page discloses that some provider links are referral links

Digital Ocean comparison pages

  • New: four Digital Ocean head-to-head pages are live — AWS vs Digital Ocean, Digital Ocean vs Lambda, Digital Ocean vs Runpod, and Digital Ocean vs Vast
  • Improved: sparse-provider comparisons now explain limited catalogs near the pricing table, so missing Digital Ocean rows show as unavailable data instead of implying a cheaper winner
  • Transparency: Digital Ocean links may use Awin/referral tracking, but rankings, prices, availability, and editorial verdicts remain price- and data-driven

EcoHash region now shown

  • Fixed: EcoHash RTX PRO 6000 listings now display their region (US) from EcoHash's availability feed instead of "N/A" — on the provider page, the RTX PRO 6000 page, and the public API
  • Unchanged: price and availability history for these listings carried over; prices ($1.89 per GPU-hour, billed per second) were already flowing live from the feed

Lium marketplace pricing and live stock

  • New: Lium (lium.io), a GPU marketplace where independent hosts compete on price, is now the 31st provider — RTX 3090, RTX 4090, RTX 5090, RTX 6000 Ada, RTX PRO 6000, L40S and H200 listings with per-second billing, sourced from Lium's public nodes feed
  • New: Lium rows carry live stock (available / limited / unavailable) and a host count per GPU, configuration, and country, refreshed hourly — hosts that allow partial rentals also appear at the 1/2/4-GPU steps they sell
  • Transparency: Lium listings linked to lium.io with no referral code at launch (a referral code was added Aug 22 — see above); rankings, prices, and sort order are unaffected by how a provider is added

Packet Dedicated pricing added

  • New: Packet (packet.ai) is now listed as a pricing-only provider with Dedicated full-card GPU rates for L40S, RTX 4090, A100 80GB, and B200, sourced from Packet's public pricing page
  • Transparency: Packet Dynamic prices are shared multi-tenant slices, so they are not mixed into GPU Finder's on-demand full-card tables until a dedicated shared/fractional pricing surface exists
  • Unchanged: Packet availability is not collected yet; its dashboard GPU pages are a future candidate source for stock and provisioning signals

Clearer affiliate disclosures

  • Transparency: pages that contain affiliate-network links now state "We earn commissions when you shop through the links below" above the links — on GPU model pages, comparison pages, and provider pages
  • Transparency: price-drop alert emails carry the same disclosure before the provider link
  • Unchanged: rankings, prices, and sort order are never affected by commissions; referral links remain marked with rel="sponsored" and disclosed on the About page

Faster, steadier page loads

  • Improved: the homepage comparison table now arrives as ready-made static HTML instead of loading in behind a placeholder, removing the biggest layout jump on the site
  • Improved: price chart placeholders now reserve the exact space of the loaded chart, so the page no longer shifts when charts appear
  • Improved: every GPU model page and provider page is now pre-rendered, cutting server response times on first visits

Direct provider links on every pricing row

  • New: every row on GPU model pages and comparison pages now has a Visit button that takes you straight to the provider, so you can act on a price without hunting for the provider's site
  • New: price-drop alert emails include a direct link to the provider offering the new lowest price
  • New: Runpod's own signup bonus ($5 credit when you add $10 via our link) is shown next to its Visit button
  • Transparency: some provider links are referral links and are labeled as such — rankings, prices, and sort order are never affected by commissions

EcoHash live availability and corrected specs

  • New: EcoHash RTX PRO 6000 listings now show live stock status (available / limited / unavailable), refreshed hourly from EcoHash's public availability API
  • Fixed: the EcoHash 8-GPU configuration now lists its correct 128 vCPUs and 512 GB RAM (previously estimated at 64 vCPUs) — per-configuration specs and pricing now come directly from EcoHash's authoritative feed

Stable provider links for AI assistants and the API

  • New: every provider now has a stable outbound link at gpufinder.dev/go/{provider} (e.g. /go/runpod) that redirects to the provider's site — safe to share anywhere a provider URL is needed
  • New: the /api/gpu-instances API returns a providerUrl field on every result row, so API consumers and AI assistants can link users straight to the provider
  • New: llms.txt now lists the provider links so AI assistants answering GPU pricing questions can pass them along
  • Transparency: a few of these links carry our disclosed referral code — rankings, prices, and sort order are never affected by commissions

Provider data reliability release

  • Fixed: availability collection now keeps updating healthy providers when one provider fetcher fails or times out, so one flaky API no longer blocks the rest of the stock refresh
  • Improved: availability history and collection heartbeats are recorded more consistently, giving reliability scores cleaner coverage after provider interruptions
  • Fixed: Hot Aisle and Hyperstack pricing ingestion now handles their current API response shapes, keeping those provider listings refreshed instead of silently dropping changed rows
  • Improved: provider dependency maintenance keeps the ingestion pipeline current without changing the public comparison workflow

EcoHash RTX PRO 6000 pricing added

  • New: EcoHash now appears on RTX PRO 6000 pricing with 1/2/4/8-GPU configurations normalized from its public pricing endpoint
  • New: EcoHash's RTX PRO 6000 listing links to the provider product page and starts at $1.98/GPU-hour, with per-second billing noted by EcoHash on its site
  • Unchanged: EcoHash has pricing data only right now, so no availability fetcher or stock signal was added

Reliability depth, spelled out on every GPU and provider page

  • New: a '30-day reliability depth' section on each /gpu/[model] and /providers/[name] page shows how many rows have a 30-day reliability score, what share of visible rows are covered, and the deepest history window on the page — turning the availability data into a readable trust signal instead of a bare badge
  • New: each page now calls out its strongest reliability signals (the providers or GPUs held available for the most covered time) and states the limits plainly — scores stay hidden until a row has enough history, and some provider APIs report coarse capacity levels rather than exact stock counts

Reliability scores now span a 30-day window

  • Improved: the reliability score on every /gpu/[model] and /providers/[name] row now reflects the last 30 days of availability history instead of 7 — a steadier signal that smooths over one-off stockouts and better separates consistently-available providers from flaky ones
  • Improved: GPU-vs comparison pages and the 'most reliably in stock' callouts now read from the same 30-day window, so the whole site tells one story
  • Unchanged: price-drop and reliability-change alert emails still fire on a tighter 7-day signal, so a slow multi-week drift doesn't get lost in a month-long average

Spot & interruptible pricing for Runpod and Vast.ai

  • New: Runpod and Vast.ai now show explicit spot / interruptible prices pulled straight from their APIs — Runpod's Secure and Community Cloud spot rates, and Vast.ai's marketplace bid floors — each labelled by its real model (spot, interruptible, or bid)
  • New: a spot discount badge on GPU pages shows how much you'd save versus on-demand, with a reminder that spot capacity can be reclaimed — handy for checkpointed batch and fine-tuning jobs
  • Improved: spot prices appear only when a provider exposes a distinct spot/interruptible figure — values equal to on-demand or without a real source are hidden instead of shown as a fake discount (Lambda stays on-demand-only until a verified spot source exists)

Hyperstack H200, B200 and RTX PRO 6000 now listed

  • New: Hyperstack pricing now comes straight from its own API instead of the shared catalog, so its latest GPUs are listed — H200 (141GB SXM5), B200, and RTX PRO 6000, alongside the existing H100, A100, L40 and RTX A-series
  • Improved: Hyperstack on-demand and spot prices now update with the rest of the site instead of waiting on third-party catalog refreshes

News digest signup and stronger availability coverage

  • New: GPU Finder now has a sitewide weekly news digest signup in the footer, plus an inline banner after the GPU results table so readers can subscribe while comparing live prices
  • Improved: Runpod availability now covers every tracked GPU/count combination, including no-stock cases that previously stayed stale, so reliability scores and in-stock signals have cleaner coverage
  • Improved: AceCloud availability is now pulled from its public regional API, while unavailable Fluidstack and Hyperbolic endpoints are skipped explicitly instead of pretending stale data is current

More accurate price comparisons

  • Fixed: provider comparison pages now compare per-GPU prices — previously a provider selling a GPU only in 8x instances was compared at the full 8x instance price against a competitor's single-GPU price, overstating differences and occasionally flipping the 'cheaper' verdict. Tables now show $/GPU/hr
  • Fixed: delisted instances no longer appear in pricing tables — SKUs a provider removed from its catalog used to keep showing at their last-known price indefinitely, and could set a GPU's displayed floor price. Listings now drop out after 14 days unseen
  • Fixed: price history charts and market snapshots no longer carry delisted SKUs' prices forward into new months
  • Fixed: price-alert emails now use the same duration-weighted reliability metric the site displays (the alert evaluator was still using a record-count average that understated stable providers)
  • Fixed: the RTX A6000 page now shows its full specs and pricing-context write-up (a key mismatch meant it never rendered)
  • Improved: Scaleway availability now reflects real capacity from their scarcity API — previously every listed GPU showed as available regardless of stock
  • Fixed: availability history no longer records conflicting statuses when a provider lists multiple variants of the same GPU (e.g. H100 SXM/PCIe) — reliability scores and heatmaps get cleaner from here on
  • Improved: the homepage now serves from cache with a 5-minute refresh instead of querying the database on every visit — noticeably faster first paint
  • Fixed: GPU model pages no longer show a 'spot' price that's identical to on-demand (same suppression the homepage and provider pages already had)

GPU vs GPU head-to-head comparison pages

  • New: side-by-side GPU comparison pages at /gpu/{a}-vs-{b} — 20 curated matchups (H100 vs H200, A100 vs H100, RTX 4090 vs 5090, and more) with live on-demand and spot price floors, GPU memory, provider counts, and 7-day availability reliability
  • Each comparison opens with a live 'which is cheaper right now' verdict and a data-driven 'when to pick which' breakdown — all numbers update hourly from the same feed as the rest of the site
  • Every GPU page now links to its head-to-head matchups, so you can jump straight from a single GPU to how it stacks up against the alternatives

Billing granularity on picks + two new provider comparisons

  • Fixed: price-history charts (and the per-GPU market snapshot) now extend to the current month — recent months are read from the live price feed instead of stopping at the last bulk-imported snapshot
  • Every pick on /picks now shows the provider's billing increment (per-second / per-minute) — per-second avoids rounding a 14-minute job up to a full hour, a ~4x difference the listing $/hr hides. Verified from provider docs across Nebius, Runpod, Vast, Lambda, Hot Aisle, Hyperstack, and Prime Intellect
  • New comparison page: AWS vs Google Cloud — hyperscaler head-to-head across A100 / H100 / H200 / B200, with the TPU and committed-discount trade-offs spelled out
  • New comparison page: Cudo vs Runpod — budget/mid-tier matchup across H100 and L40S, serverless vs committed-use

Consumer GPU picks + biggest-price-spread comparisons

  • New 'Consumer / Local LLM' picks page — the best cloud GPUs for running local LLMs on consumer cards (RTX 5090 / 4090 / 3090), with live per-hour pricing and 7-day reliability
  • Every /gpu/[model] page now links to the use-case picks it's recommended for ('Best for these workloads')
  • The Compare index now shows the GPUs with the biggest price spread across providers right now — the same card can cost 4-10x more on one cloud than another
  • New blog post: 'Cheapest Cloud RTX 4090 & 5090 for Local LLMs (2026)' — live consumer-GPU pricing table, VRAM fit guide, and rent-vs-buy math
  • Provider comparison pages now link to the use-case picks each provider is recommended for ('Best for your workload')

Market snapshot: 'is now a good time?' on every GPU page

  • Every /gpu/[model] page now opens with a market snapshot built from data we already collect — no guesswork
  • Cheapest configuration confirmed in stock right now, with the provider and live $/hr
  • Price-trend read from 30+ months of history: whether the floor price is near its 6-month low or high, and the change vs last month
  • 7-day reliability narrative — which providers most reliably hold the GPU in stock, and which are frequently waitlisted
  • Plain-language answers are also emitted as FAQ structured data so search and AI assistants can cite them directly

AceCloud added as 28th provider

  • AceCloud (India-based GPU cloud) live with 43 SKUs across H100, H200, A100, L40S, L4, RTX PRO 6000, RTX 6000 Ada, RTX A6000, RTX 8000, RTX 4090, A30, A2 — Mumbai region
  • Pricing pulled from AceCloud's public REST API, refreshed daily with the rest of the catalog
  • AceCloud reached out via email to be listed — first inbound provider integration
  • Noida and US regions to follow once region slugs are confirmed

Email alerts: price drops + reliability changes

  • Sign up on any /gpu/[model] page or homepage to get an email when your watched GPU drops below your threshold price OR when its 7-day reliability score changes meaningfully
  • Double opt-in (confirmation email) — alerts only activate after you click the link
  • Daily evaluator at 08:00 UTC; max one email per change per subscription per 24 hours
  • One-click unsubscribe in every email
  • Powered by Resend; alerts@gpufinder.dev

Per-region availability + consistent pricing tables

  • Per-region availability for Nebius — H100 in eu-north1, H200 in eu-west1, etc. now show their actual regional capacity instead of a fleet-wide aggregate. Stock is summed across data-center fabrics within a region, not across regions
  • Same per-region precision rolled out for Hyperstack, Lambda, Cudo, and Shadeform
  • When a displayed row spans multiple regions, the row now reflects the best status across those regions and the sum of their stock — previously it froze whichever region was first in the result set
  • Show-all pagination on /providers/[name] and /gpu/[model] — top 20 rows by default, then a 'Show all N instances' link. The AWS provider page (924 instances) no longer renders in one giant DOM dump
  • Unified column order across Home, GPU model, and Provider pricing tables: Provider/GPU, Instance, GPUs, vCPUs, RAM, On-Demand, Spot, Availability, [Egress], Region. Availability has its own column on the home table now (previously stacked under the price)

Availability heatmap + reliability scores + price history on every GPU page

  • 7-day availability heatmap on /gpu/[model] — provider rows × day columns, colored by % of availability checks that reported 'available' that day. Hover any cell for percentage + sample count
  • Reliability score on every row of /gpu/[model] and /providers/[name] — % of the last 7 days an instance was reported available, derived from 100k+ availability snapshots
  • Score color-tiered: green ≥95%, neutral ≥80%, yellow ≥50%, red below. Hidden when fewer than 50 data points (low-sample guard)
  • Per-GPU detail pages now include the full price history & comparison chart
  • Chart defaults to the Price History tab — landing visitors see months of price trends immediately
  • Chart auto-selects the GPU count that has the richest history (1x, 4x, or 8x)
  • Spot vs on-demand bar comparison shipped alongside on the same chart

About page, H100 vs H200 blog & provider links

  • About page — learn how we collect data, our methodology, and who built GPU Finder
  • New blog post: H100 vs H200 comparison with live pricing table from 27 providers
  • Share buttons on blog posts — copy link, share to X or LinkedIn
  • Visit provider buttons on all provider detail pages — click through to sign up
  • Expanded availability coverage — more GPU configurations tracked for Runpod, Cudo, and Scaleway

Theta EdgeCloud & UX improvements

  • Theta EdgeCloud added as 27th provider — cloud GPUs (H100, H200, A100, L40S) and community GPUs (RTX 3060-5090) with real-time availability
  • Inter font with tabular numerals for better price column alignment
  • Sticky first column on GPU model and provider page tables for easier mobile scrolling
  • Keyboard-accessible sortable table headers
  • Availability status now shows 'Unknown' when data is stale

Provider pages, comparisons & availability expansion

  • Provider detail pages — browse all 26 providers with GPU offerings, pricing, availability, and egress costs
  • Provider comparison pages — compare any two providers side by side with live pricing and egress
  • Navigation update — Providers and Compare in header, mobile hamburger menu for all pages
  • Real-time GPU availability expanded to 13 providers — added PrimeIntellect, Shadeform, improved Runpod, Vast.ai, and Hyperstack data
  • Availability now covers multi-GPU configs (1x, 2x, 4x, 8x) — not just single-GPU
  • Request a Provider form — tell us which providers to add next
  • Custom 404 page with GPU shortcuts for easy recovery
  • Dynamic OG images for comparison pages showing head-to-head pricing

GPU model pages, egress costs & Hot Aisle

  • 20 dedicated GPU model pages — browse pricing for H100, A100, B200, MI300X, and more
  • Egress cost comparison across all 26 providers — see who charges for data transfer and who doesn't
  • Hot Aisle added as 25th provider with live API integration
  • Vultr added as 26th provider with pricing and real-time availability
  • Real-time GPU availability from 10 provider APIs — see what's in stock right now
  • Availability checks every 30 minutes
  • Popular GPUs section on homepage for quick navigation
  • Improved mobile experience — responsive tables, better form layout

Launch

  • GPU Finder is live at gpufinder.dev
  • Compare GPU cloud prices across 24 providers

Price alerts & API access

  • Sign up for price drop alerts via email
  • API early-access interest page — tell us how you'd use a GPU pricing API

Price history charts

  • Price history line chart — track how GPU prices change over 14 months
  • Popular GPU shortcut buttons for one-click comparisons
  • Dynamic social preview images with provider price charts

Charts & visual improvements

  • Price comparison bar chart — instantly see which provider is cheapest
  • Faster page loads with server-side rendering

24+ providers

  • Expanded from a handful of providers to 24+ including AWS, GCP, Azure, Lambda, Runpod, Nebius, and more
  • Accurate GCP pricing — GPU and base VM costs combined correctly

Initial release

  • Compare on-demand and spot GPU prices across cloud providers
  • Filter by GPU model, count, and minimum VRAM
  • Sortable results table with provider, instance type, and pricing