Open data · updated daily from primary sources
What does an hour of GPU actually cost?
A daily index of on-demand GPU rental prices across hyperscalers, neoclouds and marketplaces. Every number traces back to a public price list, and the full history is versioned in git.
Market index
Median of provider medians across every segment, in USD per GPU-hour. Each provider gets one vote, however many regions it lists.
Where you rent matters more than what you rent
Segment median per GPU. Hyperscaler list prices run well above neoclouds and marketplaces for the same silicon.
View as table
Price-performance
Market index per dense BF16 PFLOP-hour. Lower is better.
Peak vendor specs, no sparsity. Delivered throughput depends on the workload, interconnect and software stack.
Hyperscaler premium
How much more the hyperscaler segment median costs than the neocloud median.
Explore a GPU
History, provider-level prices and regional spread for one GPU model.
Price history
View as table
Providers
Median and cheapest listed on-demand price per provider, with the SKU and region behind the cheapest.
By region
Median of region-specific list prices. Neocloud and marketplace prices are mostly global.
Cluster cost calculator
Price a block of capacity at today's list prices. Useful as a first-pass budget check before you ask providers for committed-use quotes.
At the market index
On-demand list prices, compute only. Excludes storage, networking and egress, and the committed-use discounts most large buyers negotiate. Marketplace capacity at this scale is rarely available as a single block.
Build vs rent
Should you buy servers or rent on demand? Owning is mostly fixed cost, so the answer turns on utilization. Compare the full monthly cost of owning 8-GPU servers with renting the GPU-hours you actually use at today's index.
Cost per useful GPU-hour
Owning gets cheaper the more you use it. Renting doesn't.
Cumulative cost
Monthly cost breakdown
Owning: straight-line depreciation with no residual value, interest on the average capital balance, power drawn at idle plus a utilization-proportional share, and colocation and operations as fixed monthly costs. Renting: on-demand list price for the GPU-hours used. Excludes networking fabric, storage, staffing beyond the operations line, and committed-use rental discounts.
Pipeline health
A GitHub Actions job collects every source at 06:17 UTC, validates the data, commits it to the repository and redeploys this page. If a source breaks, the job opens an issue.
Latest run
AI brief
No Claude usage recorded for today's brief.
Recent runs
One square per daily run. Hover for details.
Get the data
Methodology
What's measured
On-demand list prices for data-center GPUs, normalized to USD per GPU-hour. An 8-GPU VM at $98.32/hr counts as $12.29 per GPU-hour. Spot prices are tracked as a separate series. Reserved and committed-use pricing isn't public, so it's excluded.
How it's aggregated
Two-stage median: first each provider's median across its SKUs and regions, then the median across providers. Azure's hundreds of regional rows get the same weight as Lambda's single price list. Marketplaces are summarized to the median of all rentable offers.
Ask an agent
The index ships with an MCP server, so Claude or any MCP client can query prices and price out clusters:
claude mcp add gpu-index -- uvx --from \
"gpu-index[mcp] @ git+https://github.com/ewyluda/gpu-price-index" \
gpu-index-mcpThen ask: "What would 512 H100s for 90 days cost at neoclouds vs hyperscalers?"