Changelog

Compare pages carry only the two machines being compared

Every comparison page used to ship the full panel dataset baked into its HTML - 473 pages at 1.85MB each. Pages now prerender with just the rows for the two GPUs or providers being compared and load the full feed in the browser as before. Same pages, same numbers, roughly 90% less weight per page.

Location pages carry a slim world-floor seed

Every geography page used to embed the full panel alongside its own rows - 125 pages at ~2MB each. The world-floor comparison now ships with just the five fields it reads per row. Same tables, same floors, roughly 75% less weight per page.

Refreshes commit only when the panel actually changes

The half-hourly refresh used to commit on every run because fetch timestamps and the exchange rate always differ, which forced a full site rebuild every 30 minutes. Runs now compare the listings themselves and leave the repo untouched when nothing moved; prices and the freshness stamp publish with the next real change.

Pages cache at the edge for the refresh window

Prerendered pages now carry a 30-minute edge cache aligned with the refresh cadence, so repeat views and crawler visits are served from the nearest edge instead of re-reading the origin every time. Content still updates with each real data change.

Verda (formerly DataCrunch) billing quantum verified: 10-minute increments

docs.verda.com confirms Pay As You Go billing in pre-paid 10-minute increments with the unused portion refunded - the BILLING line now shows on the provider's source receipt. Also checked this round: Akamai (explicit hourly model, rounded up to the nearest hour - matches the panel's default, no line needed) and Hyperbolic (no published billing increment - skipped, not guessed). Coverage is now 12 providers.

Billing-quantum coverage widened to 11 providers

Four more first-party verifications joined the BILLING line on provider source receipts: RunPod per-second (docs.runpod.io/pods/pricing: pods billed by the second for compute and storage), Lambda per-minute (docs.lambda.ai/public-cloud/billing: hourly usage in one-minute increments), Salad per-second (docs.salad.com: usage tracked per second) and Thunder Compute per-minute (thundercompute.com/docs/billing: compute billed per minute while instances run). Coverage now: per-second at RunPod, Salad, gpu.ai, VoltageGPU, AWS and Google Cloud; per-minute at Lambda, Thunder Compute, Machine0, SpotGPUs and Hinode.

Provider pages now show the verified billing quantum

The source receipt on every provider page gains a BILLING line where the billing unit is verified first-party: per-second at gpu.ai, VoltageGPU, AWS (partial instance-hours, Linux) and Google Cloud (1-minute minimum, then per-second), per-minute at Machine0, SpotGPUs and Hinode. Providers that publish no billing quantum show nothing - assume hourly. The analysis behind it: https://www.compute.cafe/blog/the-rounding-tax-on-gpu-hours

New analysis: the INR 20 L4 - India's GPU market is cheap exactly where it matters

Nine India-region providers are on the panel (AIC Cloud, NeevCloud, inhosted, E2E Networks, Cyfuture, Shakti Cloud, hostnot, JarvisLabs, Neysa). The piece splits the market: inference-class cards where India IS the global floor (Mumbai L4 INR 20/hr, second-cheapest of 128 providers; L40S INR 54, third-cheapest) versus H100s at INR 180-356 against a $0.89 global floor but below the $4.50 rentable median - then the three buyer questions (user latency, INR-vs-USD accounting with GST input credit, catalog vs live rows).

New analysis: the rounding tax - a $2.00 H100 that costs more than a $3.24 one

The billing quantum is the hidden second number in every GPU price. With verified billing terms across the panel (per-second at gpu.ai, VoltageGPU, AWS and Google Cloud with a 1-minute minimum; per-minute at SpotGPUs, Machine0 and Hinode), the piece prices the same 20-minute job and the same 8-session interactive day across five machines: a hypothetical hourly-billed $2.00 H100 ends up at $16 for 96 minutes of real work, more than AWS's per-second $11.01 at triple the sticker. Checklist: match quantum to job shape, compute effective cost with the ceil, read minimums, treat silence as hourly.

New analysis: the same H100 costs $1.19, $6.88, or $11.06 tonight

With AWS and Google Cloud joining the panel, the market's full spread for one chip sits on one screen: IO.NET's $1.19 DePIN floor (sold out at verification), AWS's single-GPU p5.4xlarge at $6.88, and Google's 8-GPU-only A3 block at $11.06 per GPU-hour. The piece works through what the cloud premium buys (data adjacency, interconnect fabric, compliance), where it collapses (AWS undercuts VoltageGPU's in-stock H100 at $6.95; Google's L4 at $0.71 and RTX PRO 6000 at $4.50 beat specialist floors), and the H200 head-to-head (AWS $7.91 vs GCP $10.60 vs the $3.29 specialist floor).

Novita AI joins the panel, numbered 129

Novita's GPU pods now come from its own no-auth products feed - the API behind the pricing page's GPUs tab: on-demand 1-GPU hourly with live stock (RTX 4090 $0.67/hr, RTX 5090 $0.56, H100 SXM 80GB $3.39, L40S 48GB $0.55; all four read zero available at verification and show sold out). Spot prices and exact 8x linear multiples stay excluded per panel policy.

3 new providers, numbered 126 through 128: IO.NET, QuickPod and Hinode

IO.NET joins through its own no-auth market-snapshot API - the feed behind its IO Explorer - with per-GPU-hour rates across five deployment tiers and live device stock (H100 SXM $1.19/hr, B300 $4.48, L40S $0.79, RTX 4090 $0.25; 9 rows, most fleets hired out at verification). QuickPod adds its full 42-model live rate card at the current floor of each range (H200 $2/hr, H100 80GB $2.50, RTX 5090 $0.50, RTX 4090 $0.10, V100 32GB $0.14). Hinode adds per-minute GPU desktop machines at hourly from-floors (L4 24GB from $1.75/hr, L40S 48GB from $3.30, RTX PRO 6000 Blackwell 96GB from $8.05); its fractional workspace product stays excluded per panel rule.

AWS joins the panel, numbered 125

AWS on-demand GPU prices now come from AWS's own no-auth price feed - the JSON that powers the EC2 on-demand pricing page, refreshed daily: 18 whole-instance hourly rows for US East (N. Virginia), Linux (1x H100 80GB p5.4xlarge $6.88/hr, 8x H200 p5en.48xlarge $63.30, 8x A100 80GB p4de.24xlarge $27.45, 8x A100 40GB p4d.24xlarge $21.96, 1x V100 p3.2xlarge $3.06, 8x V100 32GB p3dn.24xlarge $31.21, T4 $0.53, L4 $0.80, L40S $1.86, A10G $1.01, T4G $0.42, Radeon Pro V520 $0.38, M60 $1.14, 8x Gaudi HL-205 dl1 $13.11). GPU specs verified against AWS's own docs. Excluded per panel policy: exact linear multiples, duplicate 1-GPU configs, fractional-GPU sizes, and Trainium/Inferentia (not GPUs).

Google Cloud joins the panel, numbered 124

GCP's accelerator-optimized pricing page (cloud.google.com, JS-rendered, no no-auth API) was verified in a real browser: 7 on-demand hourly whole-machine rows for Iowa (us-central1) - 8x H200 $84.81/hr, 8x H100 SXM $88.49 (A3 High; A3 Mega $93.40), A100 80GB $5.07, A100 40GB $3.67, 16x A100 40GB $55.74, L4 $0.71, RTX PRO 6000 Blackwell $4.50. Excluded per panel policy: commitment (CUD), spot and DWS columns, A4 8x B200 (no on-demand price in us-central1), fractional-GPU G4 rows, exact linear multiples, and per-GPU attachment add-on prices (T4/P4/V100/P100) that exclude the VM.

Latitude.sh prices corrected to real on-demand hourly, now per region with stock status

Every Latitude.sh row read 2x low: the connector collected the marketing cards' 'starting at' figures, which are monthly-commit rates expressed per hour. Verified against the page's own Hourly toggle: on-demand is H100 80GB $3.37/hr (1 GPU, US), RTX PRO 6000 $48/hr (8 GPUs, US/Netherlands; $52 Japan; $66 Australia) and B300 $128/hr (8 GPUs, US). Rows now come from the plan data embedded in the pricing page, carry their region, and show sold-out status - all three GPU plans were out of stock at verification.

SpotGPUs expanded to its full 47-model on-demand catalogue

The catalogue table is JS-rendered, so the on-demand column was verified in a real browser: per-GPU-hour rates, never reclaimed, billed per minute (B200 SXM $2.49, H200 SXM $1.09, H100 SXM $0.89, MI300X $1.45, A100 PCIe 80GB $0.19, RTX 4090 $0.09, down to RTX 3070 $0.05). Spot (reclaimable) and reserved tiers stay excluded per panel policy.

2 new providers, numbered 122 and 123: Cyfuture and VoltageGPU

Cyfuture's JS-rendered instance tables yield 9 on-demand INR rows (H100 SXM INR329/hr, B300 INR570, MI300X INR274, A100 80GB INR198, RTX PRO 6000 INR256, Gaudi 2 INR101; reserved columns excluded, converted at the observed rate). VoltageGPU joins with its confidential-VM tier on its own Intel TDX hardware: RTX PRO 6000 Blackwell 96GB $3.80/hr and H100 80GB $6.95 in stock, H200 141GB $8.08 sold out at verification; its standard tier is a distributed provider network and stays excluded per panel policy.

3 new providers, numbered 119 through 121: SpotGPUs, gpu.ai and Zettabyte

SpotGPUs publishes an on-demand H100 SXM rate ($0.89/GPU-hr, never reclaimed, per-minute billing); its spot and reserved tiers are excluded per policy. gpu.ai adds a live marketplace rate table (24 models at their lowest tier: H100 PCIe $1.98, H100 SXM $3.49, H200 SXM $3.59, B200 $8.13, RTX PRO 6000 $1.99, A100 80GB $0.81, RTX 4090 $0.41). Zettabyte adds an on-demand H100 from-floor at $2.51/GPU-hr.

1 new provider, numbered 118: AtmosCompute

Consumer RTX cards at hourly from-floors on the homepage (RTX 4090 from $0.14/hr, RTX 3090 from $0.20/hr, both 24GB); rows carry the (from) label. A100 80GB and H100 SXM are quote-gated ('On request') and excluded per panel policy.

2 new providers, numbered 116 and 117: 1Legion and Theta EdgeCloud

1Legion publishes hourly bare-metal rates on its homepage cards (RTX A4000 $0.39/hr, RTX A6000 $0.72). Theta EdgeCloud publishes per-GPU-hour on-demand rates (RTX 3090 $0.14, A100 80GB $1.99, RTX 4090 $0.26, RTX 5090 $0.49, H100 80GB and H200 $2.29, V100 16GB $0.99); its page states 'Prices are on-demand rates'.

2 new providers, numbered 114 and 115: Machine0 and Deep Infra

Machine0 publishes a per-minute-billed VM rate card (RTX 4000 Ada $0.836/hr, L40S $1.727, RTX 6000 Ada $1.727, MI300X $2.849, H100 $4.851, H200 $4.917; 8-GPU SKUs are exact multiples, so 1-GPU rows carry the same information). Deep Infra's GPU Deploy table adds per-GPU-hour on-demand rates (A100 $0.89, H100 $2.20, H200 $2.69, B200 $3.69, B300 $4.89); its LLM token pricing is out of panel scope.

Akamai rows now come from its official live API, with multi-GPU configs

The provider moved from a browser-verified pricing-page snapshot (3 rows) to Akamai's official no-auth Linode API: on-demand hourly USD read directly (RTX 6000 x1/x2/x3/x4 at $1.50/$3/$4.50/$6, RTX 4000 Ada x1 at $0.52, x2 at $1.05, x4 at $2.96 - cheapest config per GPU count). RTX PRO 6000 Blackwell stays a gated row: it is request-access only and absent from the API catalog.

New analysis: the $1.68 H100 that actually costs $3.37

Why the cheapest number on a GPU pricing page is usually a monthly commitment in disguise - the Latitude.sh case, the three-rate ladder every provider runs, and the 365-hour break-even that decides whether the discount saves or costs you money.

New analysis: 113 providers publish prices, 19 will show you a machine

Of 2,192 tracked listings only 711 come from a feed that can confirm stock. The rentable single-GPU H100 market tonight: 86 machines across 31 providers, median $4.50/hr, floor $1.30 - and why a sold-out $1.68 is not cheaper than an available $4.50.

Thunder Compute rows now come from its live public API, with stock status

The provider moved from a pricing-page scrape to its documented open API: hourly rates read directly (A6000 $0.35, A100 80GB $1.09, H100 PCIe $3.20, L40 $0.79), each row now carries a live available-or-sold-out signal, and L40 joined the panel for this provider. Multi-GPU instance totals are exact linear multiples, so single-GPU rows carry the same information.

GPUBrazil rows now come from their live public API, with per-region stock

The provider moved from a pricing-page scrape to its documented open API: prices read in dollars directly, each row carries its region and a live stock count, and coverage widened from 10 to 58 GPU models. Spot and fractional-slice listings stay excluded by policy.

Yotta Labs and Thunder Compute outbound links now route through referral programs

Clicks to these two providers carry our referral codes (Yotta Labs: tiered credit rewards per activated referral; Thunder Compute: 3% lifetime credit share). Ranking is unaffected; the methodology page disclosure covers every wired program.

Comparison pages render full content on first load

Every provider-vs-provider and GPU-vs-GPU comparison page now ships its head-to-head scoreboard and rate table in the initial page instead of a loading screen.

RunPod referral terms corrected

The disclosed cash affiliate tier now reads 10%, matching RunPod's current program page; the record previously said 12%. A full audit of all 113 providers' public affiliate and referral programs was completed alongside - no other corrections needed.

Origin tracking added to the panel feed

The dataset now carries a stable origin marker, so copies that show up elsewhere can be traced back to the source.

Location pages render full content on first load, with every model's local floor vs the global floor

Each geography page now ships its stats, per-model comparison (cheapest local rate, who holds the world floor, and whether the region beats it) and FAQ in the initial page instead of a loading screen. The comparison now covers every tracked model in the region, not just headline chips, and the structured FAQ data is computed by the same code as the visible answers.

Changelog page scrolls again

The page sat outside the app's scroll container, so anything past the first screen was unreachable. It now uses the same scrolling shell as every other page.

Provider market-position numbers now match everywhere

The visible position block and the structured FAQ data on provider pages are computed by the same code, so the counts agree row for row.

Provider pages render full content on first load

Every provider page now ships its stats, model table and price position in the initial page instead of a loading screen. New on each page: where the provider's prices sit against the panel - models where it beats the typical price and models where it is the cheapest option.

Model pages go deep: trends, provider notes, buy guides

Every GPU model page now carries a model-level price trend from the daily snapshots, per-provider notes, and a buy guide with node-vs-single economics and monthly cost - all computed from the live panel and server-rendered. New FAQs (trend direction, monthly cost, node vs single) ship in the page and the structured data together.

The full dataset stays public

An overnight experiment with gating the raw panel was rolled back within the hour after it broke pages. The complete row-level feed - every listing, every receipt - remains published at its usual endpoints.

Counts in titles update weekly

The provider and listing counts baked into page titles and meta now refresh on a weekly snapshot instead of every data build; prices and feeds stay hourly.

Real logos for 48 providers

Official site icons replace placeholder initials, and every future provider gets its real mark automatically.

India and China price pages for every tracked GPU

Per-model landing pages in rupees and yuan, local providers first, each row linked to its published source.

Live price and current month in every title

Model and location titles carry the current floor price and stamp; the homepage leads with the live provider count.

Docs and API pages tell the truth

The published endpoints, license (CC-BY) and refresh cadence now match what the site actually serves.

This changelog

A readable record of what ships, newest first. The price panel itself refreshes hourly.

14 new providers, numbered 100 through 113

A China-focused sweep: MatPool, Baidu BCC, Kingsoft Cloud, Gitee AI, CTYun, Featurize, AutoDL, OpenBayes, PPIO, Paratera, Lanyun, AI Galaxy, Suanjia Cloud and 4090Xianka - including the first RTX 5090 listing in China.

Full market refresh now runs every 2 hours

Doubled from every 6. Live-API providers still update hourly.

Answers readable without JavaScript

FAQ and pricing schema now render on the server, so crawlers and AI agents see the same facts visitors do.

Phantom providers removed

Providers that merged but never ran no longer appear with empty panels.

Taiga Cloud pricing corrected

An 8-GPU bundle price was stored as the per-GPU rate; it read 8x low. Fixed at the source.

Dead pages now say so

Unknown GPU, provider and location URLs return real 404s instead of empty pages.