Compute Comparison
About

Real-time GPU & LLM price comparison

Compute Comparison aggregates live GPU cloud pricing and LLM inference costs so AI engineers, researchers, and teams can find the cheapest compute for their workload — without manually checking dozens of provider dashboards.

94+
GPU cloud providers tracked
509+
GPU instance types
33+
LLM inference providers
274+
LLM models priced
15 min
Price refresh interval
87
GPU rental landing pages

Why we built this

GPU compute pricing is fragmented across dozens of providers, each with different instance naming conventions, billing models, and regional availability. Finding the cheapest H100 for a training run used to mean opening 15 browser tabs.

We built Compute Comparison to solve that — a single place where you can compare on-demand and spot rates across every major GPU cloud, updated every 15 minutes.

The same problem exists for LLM inference APIs: token prices change frequently, new providers launch constantly, and the cheapest option for GPT-4-class capability shifts month to month. Our LLM price tracker covers 274+ models across 33+ providers so you can make cost-informed decisions.

Who uses Compute Comparison

AI engineers & researchers
Finding the cheapest GPU for fine-tuning, training runs, and inference serving.
ML platform teams
Benchmarking cloud spend and negotiating contracts with data to back it up.
Startups & indie developers
Stretching compute budgets further by routing workloads to the lowest-cost provider.
Finance & procurement
Tracking GPU price trends and forecasting AI infrastructure costs.

How the data works

01
Provider API polling
We query provider APIs and public pricing endpoints every 15 minutes. Where APIs are unavailable, we scrape structured pricing pages and normalise the data into a unified schema.
02
Normalisation & deduplication
Raw listings are normalised to a common GPU model taxonomy — mapping provider-specific names like "p4d.24xlarge" to "A100 SXM4 40GB". Duplicate listings and stale entries are filtered automatically.
03
Pricing served via API
Normalised prices are served through our public REST API at /api/prices (GPU) and /api/llm (LLM). The same API powers the live tables on every page. Rate limits apply — see /api-docs for details.
04
Freshness monitoring
Each provider listing carries a last-seen timestamp. Listings not refreshed within 24 hours are flagged as potentially stale. Our internal price audit dashboard monitors data quality across all providers.

Methodology

How we define "live", what we include, and where the data comes from.

What "live" means

Provider pricing endpoints are polled on a 15-minute cycle. Each price record carries a fetched_at timestamp visible in the API response and shown in the "Updated" label on live tables.

"Live" refers to the most recent successful fetch from that provider's API or pricing page — not a real-time websocket stream. If a provider's endpoint is temporarily unreachable, the last known price is shown with a staleness indicator after 60 minutes.

LLM token prices are refreshed on the same 15-minute cycle. Model availability (whether a model is currently accepting requests) is not tracked — only the advertised price per million tokens.

Data sources

GPU prices are sourced from provider REST APIs where available (Lambda Labs, CoreWeave, RunPod, Vast.ai, and others), and from structured pricing pages for providers without public APIs.

LLM inference prices are sourced from provider pricing pages and, where available, from OpenAI-compatible /models endpoints that return pricing metadata.

All prices are denominated in USD. Providers that publish in other currencies are converted at the daily exchange rate from a public FX feed.

GPU model normalisation

Providers use inconsistent naming: AWS calls an H100 instance p5.48xlarge; Lambda calls it gpu_1x_h100_sxm5. We maintain a hand-curated mapping table that resolves these to a canonical GPU model name (e.g. NVIDIA H100 SXM5 80GB).

The mapping is updated manually when providers launch new instance types. If you spot a misclassification, let us know.

Multi-GPU instances (e.g. 8× H100) are normalised to a per-GPU hourly rate so they're directly comparable to single-GPU listings.

What's not included
  • Egress & storage costsData transfer and object storage fees vary widely and are not included in any price shown.
  • Reserved / committed-use discountsPrices shown are on-demand or spot rates. 1-year and 3-year reserved discounts are not tracked.
  • Negotiated enterprise ratesLarge-volume contracts are private. Actual enterprise pricing may be 20–50% below list.
  • Networking & interconnect feesInfiniBand, VPC peering, and cross-region transfer costs are excluded.
  • Software licensingSome providers bundle CUDA, drivers, or OS licensing into the hourly rate; others do not. We do not normalise for this.
Update cadence by data type
Data typeRefresh intervalStaleness thresholdSource type
GPU on-demand prices15 minutes60 minutesProvider API / pricing page
GPU spot prices15 minutes60 minutesProvider API
LLM token prices (input)15 minutes60 minutesProvider pricing page / API
LLM token prices (output)15 minutes60 minutesProvider pricing page / API
GPU price historyDaily snapshot at 00:00 UTC48 hoursAggregated from intraday polls
LLM price historyDaily snapshot at 00:00 UTC48 hoursAggregated from intraday polls
GPU specs (VRAM, TFLOPS, etc.)Manual — updated on new GPU launchN/AVendor datasheets
Data accuracy notice

Prices shown are estimates based on publicly available provider data and may not reflect negotiated enterprise rates, reserved instance discounts, or egress/storage costs. Always verify pricing directly with the provider before committing to a workload. Compute Comparison is not affiliated with any GPU cloud or LLM provider listed on this site.