Compute that follows the clean grid.
GridHop is a GPU broker for AI agents. Your agent hands over a job, a budget and a deadline. GridHop picks the machine and the hour across six clouds and 48 power grids — where compute is cheapest and the electricity is cleanest — runs it, and hands back the results.
Not open yet. Free today: add the price & carbon index to your agent in one line.
Same GPU. Different hour. Different bill.
GPU prices move hour by hour across a dozen markets, and the power under each machine ranges from hydro-clean to coal-heavy. Today you pick one provider, pay its rate and babysit the job. Your agent can do better — if something is watching the whole grid.
Waiting pays
In our index data, letting a 10-hour H100 job start any time in the next 72 hours cut its median cost by two-thirds versus starting now.
Cheap capacity vanishes
Of the hours we checked, the cheapest H100 offer was gone an hour later 97% of the time. You can't catch it by hand. An always-on broker can.
Grams of CO₂ per kWh
Switzerland vs the Netherlands, same morning. The machine you choose decides how dirty your training run is — not just what it costs.
Flight search shows the cheapest fare. GridHop books it — and rebooks.
Think of a travel agent on autopilot: "Be in Berlin by Monday, under $20." Except the traveller is a training job and the airlines are GPU clouds.
Submit a job
Container, data, GPU class, rough duration. Plus the two numbers that matter: a budget cap and a deadline.
Scan the grid
Prices and availability across every provider, right now and forecast ahead — joined to grid price and carbon by hour and zone.
Run & hop
Launch in the cheapest, cleanest window. Checkpoint as it goes; if the host disappears or a better one opens up, move it.
Get a receipt
Results back to your agent, plus a receipt: what it cost, energy used, and the carbon footprint — ready for your ESG report.
Built for agents. Governed by you.
Your agent talks to GridHop over MCP. You set the rules once; GridHop enforces them on its side — never trusting the agent to behave.
- One prepaid balance. No account with every provider, one bill at the end of the month.
- Spend caps & approvals. Per job, per day, auto-approve thresholds — enforced server-side.
- Carbon as a constraint. Optionally cap gCO₂/kWh and the scheduler won't run you on a dirty grid.
# "fine-tune this, done by Monday 9am, max $20" run_job({ image: "ghcr.io/acme/finetune:1.4", gpu_class: "H100", est_hours: 10, deadline: "Mon 09:00", budget_cap: 20.00, carbon_cap: 100 // gCO₂/kWh }) → plan: Nordic host, start Sun 02:00, hydro-heavy grid → checkpoint every 30 min · 2 fallbacks queued
Illustrative example of the planned run_job flow. Job running opens with the first users; the price index is live today.
We've been watching the grid every hour.
GridHop starts from an open, hourly index of GPU offers joined to grid price and carbon intensity. Here is what flexibility is worth, and how far apart the grids are.
Median saving from a 72-hour start window
10-hour job · cheapest window within 72 h vs cheapest offer at submit time · all providers
Show as table
| GPU class | 72 h median | P25–P75 | 24 h median |
|---|
Simulated on 145 hourly snapshots (Nov 9–15, 2026). Early data with wide spreads: with only a 24-hour window the median saving is 0–12% for most classes, and 43% for A6000. Numbers will change as history grows.
Grid carbon intensity, same hour
gCO₂eq/kWh · cleanest and dirtiest zones · Nov 15, 08:00 UTC
Computed from each grid's generation mix × IPCC lifecycle factors (ENTSO-E, NESO, US EIA). US values lag about a day.
A broker, not another marketplace.
We don't own GPUs or recruit hosts. Supply already exists — GridHop is the brain on top that picks time and place, and guarantees the price. What we are building, in roadmap order below:
Time-shifting, automatically
Power is cheapest at night and when wind and sun are strong. GridHop forecasts prices and availability per provider, region and GPU for the next 72 hours and schedules into the dip.
Fixed-price jobs
Quote first, run second. "This job: $14, done by Monday." If a host drops, GridHop restarts from the last checkpoint on the next machine — your price doesn't move.
Carbon receipts
Measured GPU power × hourly grid intensity, per job. Monthly exports for Scope 3 reporting.
Datacenter-only by default
Private data never lands on community hosts unless you allow it. EU-only or US-only residency filters.
Teams & audit
Agents are first-class principals with their own caps. Every decision logged: offers considered, chosen, and why.
Every major source of GPUs
Marketplaces and hyperscalers in one view, on-demand and spot, with each host's reliability score.
Works where your agent works
One hosted MCP endpoint. No local binary, nothing running on your machine.
Give your agent eyes on the grid — in one line.
The GridHop price & carbon index is a free, read-only MCP server. No account, no API key. It can't rent anything; it answers questions.
claude mcp add --transport http gridhop https://gridhop.dev/v1/mcpOpen source: github.com/gridhopai/gpu-price-index-mcp
From index to autopilot.
Price & carbon index
Hourly GPU prices across six providers joined to 48 grids. Free MCP tools for any agent.
Run jobs from your agent
find → estimate → run → results. Prepaid balance, spend caps and approvals.
Fixed price & migration
Quotes with a deadline, forecast-driven scheduling, checkpoint-and-hop on host loss.
Teams & carbon reporting
Org policies, audit log, SSO, monthly Scope 3 exports, agent wallets.
Questions, answered.
Is GridHop a GPU cloud?
No. GridHop doesn't own hardware. It buys compute from the marketplaces and clouds that already exist — Vast.ai, RunPod, Nebius, AWS, Azure, Google Cloud — on your behalf. You talk to one API and pay one bill.
How does it save money?
By choosing when and where. GPU prices and availability swing hour to hour, and cheap capacity is snapped up fast. If your job can wait — even a few hours — GridHop schedules it into the cheapest window before your deadline.
How is carbon measured?
For each grid zone we compute hourly intensity from the generation mix (ENTSO-E in Europe, NESO in Great Britain, EIA in the US) with IPCC lifecycle emission factors. Per job, energy is GPU power draw sampled every minute plus host overhead; emissions are energy × that hour's intensity.
Can my agent overspend?
No. Caps per job and per day, and an auto-approve threshold, are set by a human and enforced by GridHop's control plane. Jobs above the threshold wait for your approval. You can revoke an agent's access at any time.
What about my data?
Data moves only between the job's machine and object storage; the control plane never runs your workload. Jobs with private inputs default to datacenter-tier hosts, never community hosts.
When does it open?
Soon, for a small first group. Join the waitlist and we'll write once, when it opens. Meanwhile the price & carbon index is free to use.
Let your next training run hop to the clean grid.
First users get job running before anyone else. Tell us what you'd run and we'll build for it.