Repository avatar
AI Tools
v0.6.0
active

gpu-price-index

io.github.gridhopai/gpu-price-index

Live GPU rental prices on Vast.ai, RunPod, Nebius, AWS, Azure and Google Cloud, with grid carbon.

Documentation

GridHop GPU Price & Carbon Index

A free, read-only MCP server that tells an AI agent what a GPU costs to rent right now, across several clouds and marketplaces, and how clean the electricity is where that GPU sits.

No account, no API key, nothing to install. It cannot rent anything; it only answers questions.

Want your agent to actually run the job, not just price it? That is GridHop, coming next. Join the waitlist.

https://gridhop.dev/v1/mcp

Add it to your agent

Claude Code

claude mcp add --transport http gridhop https://gridhop.dev/v1/mcp

Cursor (~/.cursor/mcp.json)

{ "mcpServers": { "gridhop": { "url": "https://gridhop.dev/v1/mcp" } } }

Any other MCP client: streamable HTTP, URL above, no authentication.

What you can ask

  • "What's the cheapest RTX 4090 I can rent right now?"
  • "Find an 8×H100 machine in a datacenter, under $25 an hour."
  • "Which A100 offer runs on the cleanest grid?"
  • "Compare T4 prices on AWS, Azure and Google Cloud in Europe."
  • "What does an 8×H100 machine cost on demand on AWS and on Azure, and how much cheaper is spot?"
  • "Which GPUs with at least 48 GB can I rent, and what do they cost?"
  • "Has the H100 price gone up or down this week?"
  • "Which European grid is cleanest right now, and when is power cheapest in Texas?"

The tools

ToolWhat it answers
find_offersThe cheapest or cleanest offers for a GPU class right now.
list_gpusWhich GPU classes are on offer: VRAM, providers, machine sizes, cheapest on-demand and spot price.
price_historyHow the cheapest price of a class has moved, per day or per hour.
grid_carbonHow clean and how expensive the electricity is on each power grid, now and at its best hour.

find_offers

ArgumentMeaning
gpu_class (required)GPU model. Loose forms work: 4090, RTX 4090, H100, A100-80GB.
countGPUs on one machine. Default 1.
vram_minMinimum VRAM per GPU, GB.
region_filterTwo-letter country codes, e.g. ["US", "DE"].
tierdatacenter drops independent marketplace hosts.
max_price_hrCeiling in USD per hour for the whole machine.
providersAny of vastai, runpod, nebius, nebius_spot, aws, aws_spot, azure, azure_spot, gcp_spot. aws, azure and nebius are on-demand prices; the _spot ones are spot prices.
include_interruptibleOmit for on-demand prices plus AWS, Azure, Google Cloud and Nebius spot. true adds marketplace minimum bids. false drops all spot, which removes Google Cloud entirely.
min_reliabilityDrop hosts below this provider reliability score. Default 0.9.
max_carbon_g_kwhOnly hosts on a grid at or below this carbon intensity.
sort_byprice (default) or carbon.
limitUp to 50 offers. Default 10.

Each offer comes back with provider, GPU class and count, price per hour, country and region, host tier, reliability score, whether it can be interrupted, and where known: grid carbon intensity (g CO₂/kWh), how old that reading is, the grid zone it refers to, and a rough estimate of grams of CO₂ per hour for the GPUs at full load.

One offer from a real answer to "cheapest RTX 4090 in the US" (15 November 2026):

{
  "provider": "vastai",
  "gpu_class": "RTX4090",
  "gpu_count": 1,
  "vram_gb": 24,
  "price_usd_hr": 0.575,
  "price_per_gpu_hr": 0.575,
  "region": "Texas",
  "country": "US",
  "tier": "community",
  "reliability": 0.995,
  "interruptible": false,
  "carbon_g_kwh": 410,
  "carbon_source": "eia",
  "carbon_lag_h": 48,
  "est_gco2_per_hr": 184,
  "grid_zone": "US-TEX-ERCO",
  "carbon_uncertain": false
}
FieldMeaning
price_usd_hr, price_per_gpu_hrList price for the whole machine, and per GPU.
tierdatacenter, or community for an independent host on a marketplace.
reliabilityThe provider's own score, 0 to 1. Empty where the provider publishes none.
interruptibletrue for spot and bid prices: the machine can be taken back.
carbon_g_kwhGrid carbon intensity, g CO₂eq per kWh. Empty where no open data covers the grid.
carbon_sourceentsoe_gen (Europe), neso (Great Britain) or eia (US).
carbon_lag_hAge of the carbon reading in hours. 0 to 3 is live; US readings are 24 or 48 (same hour of day, one or two days earlier).
est_gco2_per_hrRough grams of CO₂ per hour: carbon intensity × nameplate GPU power × GPU count.
grid_zoneThe power grid the reading refers to. Pass it to grid_carbon for more.
carbon_uncertaintrue when part of that grid's generation has no fuel type and the mix is partly assumed.

The answer wraps the offers with observed_at (snapshot time), total_matching, warnings (things the agent should tell you, such as stale carbon readings), and sources.

list_gpus

One optional argument, vram_min (GB). Returns one row per GPU class, cheapest first.

price_history

ArgumentMeaning
gpu_class (required)As for find_offers. H100 returns one series per variant.
granularitydaily (default) or hourly (last 72 hours).
daysHow far back, 1 to 90. Default 14.

Each point has the cheapest on-demand price seen, the typical one (median of that day's hourly lows), the cheapest AWS, Azure, Google Cloud or Nebius spot price, all per GPU, and how many hourly snapshots the day has. A summary gives the low, the high and the change over the period. Prices follow find_offers' defaults: no marketplace minimum bids, no hosts under 0.9 reliability. History starts 1 November 2026 and is hourly from 10 November.

grid_carbon

ArgumentMeaning
region_filterTwo-letter country codes.
grid_zoneOne grid by id, e.g. SE-SE3, US-TEX-ERCO. The ids match grid_zone on offers.
sort_bycarbon (default, cleanest first) or power_price.
limitUp to 60 grids. Default 20.

Per grid: carbon intensity now and how old the reading is, the low and high of the last 24 hours and when, the hour of day that is usually cleanest, the day-ahead wholesale power price now and the cheapest hour in the next 24. Wholesale prices cover Europe (ENTSO-E), Texas (ERCOT) and California (CAISO); a GPU host may pay a fixed tariff, so a cheap hour on the grid does not by itself make a GPU cheaper.

Where the data comes from

Refreshed every hour. Every result carries observed_at, the time of the snapshot.

  • Prices: Vast.ai and RunPod public listings, Nebius spot prices from Nebius's public spot-price API and its on-demand list prices via the SkyPilot catalog (checked against Nebius's pricing page), the AWS EC2 on-demand and spot price feeds, the Azure Retail Prices API (pay-as-you-go and Spot), and Google Cloud's Billing Catalog (Spot; vCPU + memory + GPU, added up per machine type).
  • Carbon intensity: computed from each grid's generation mix with IPCC lifecycle emission factors. Generation data: ENTSO-E Transparency Platform (Europe), NESO Carbon Intensity API (Great Britain, CC BY 4.0), US Energy Information Administration (US).

What to keep in mind

  • Prices are list prices at snapshot time, before any fee, and they move. A marketplace machine can be rented by someone else a minute later.
  • AWS, Azure and Nebius have both on-demand and spot prices; Google Cloud has spot only. Its on-demand prices are not in the index yet, so every Google Cloud row is interruptible and include_interruptible: false returns none of them.
  • Nebius spot prices move fast: with demand, as often as every 15 minutes, several-fold within a day, and they can exceed the on-demand price. The index reads them once an hour.
  • Cloud rows are list prices, not machines confirmed free. On AWS and Azure the large GPU types usually need a quota increase first, and Azure Spot capacity can be unavailable. Some Azure sizes are left out until their GPU count is confirmed (NCads A10 v4, RTX PRO 6000, GB200), as are sizes with a fraction of a GPU.
  • Google Cloud rows are what a machine would cost, not proof it is available. Spot GPU capacity is often unavailable; the tool says so when it shows those rows.
  • Marketplace bid prices are minimum bids, a floor rather than the price of keeping the machine, so they are left out unless you ask for them.
  • Carbon values cover about half of the offers: Europe, Great Britain and most US states. Elsewhere the field is empty rather than guessed.
  • US carbon readings are one or two days old (same hour of day), because the source publishes late. carbon_lag_h says how old each reading is.
  • Some readings are estimates. Where a grid operator reports part of its generation without a fuel type, the offer is marked carbon_uncertain.
  • Carbon intensity counts what is generated in a grid zone, not imports. The CO₂-per-hour figure uses nameplate GPU power only; it is an estimate, not a measurement.
  • Outside the US, Vast.ai gives only a host's country, so countries with several grid zones use one default zone. grid_zone shows which.

What is logged

To see whether the server works and what people look for, each request is logged with: the name and version your agent reports for itself, the country Cloudflare saw the request come from, the tool called, its arguments, and the outcome. No IP address, account, or anything else that identifies a person is stored.

Limits

60 requests a minute per client. Beyond that the server answers 429 with a Retry-After header. A normal agent session uses a handful.

Status

Early and free. The server may change; breaking changes will get a new path (/v2/mcp).

What comes next: GridHop

This index is the first public piece of GridHop, a broker for AI agents. Your agent hands over a GPU job with a budget and a deadline; GridHop picks the machine and the hour across marketplaces, runs it, and returns the results. One prepaid balance, spend caps enforced on our side, scheduling by price and by grid carbon.

It is not open yet. Join the waitlist to hear when it is, and to tell us what you would run.

Questions, wrong prices, a GPU class you could not find: write to support@gridhop.dev or open an issue.

No installation packages available.