OpenRouter for GPUs: the missing aggregation layer for compute

If you build with large language models, you probably know OpenRouter. It puts dozens of model labs behind one API, so you can switch models without rewriting your integration or opening a new account each time. People keep asking the obvious follow-up question: where is the OpenRouter for GPUs?

This post looks at why the aggregation layer worked for models, why it hasn't been built yet for the compute underneath them, and what an OpenRouter for GPU clouds would need to do.

Up front: we are Opengrid, and this is the layer we are building. Opengrid is pre-launch. The market view ships first and the routing API is in development.

What OpenRouter proved

OpenRouter sits between developers and model labs. You send a request to one API, and OpenRouter forwards it to the model you chose, with one bill at the end. According to Sacra, it charges about a 5% fee.

The market rewarded that layer quickly:

Date	Event	Value
June 2025	Series A	$547M valuation
May 2026	Series B, led by CapitalG	$1.3B valuation
August 2026	Reported acquisition by Stripe	$7B+

That is roughly half a billion dollars to a reported seven billion in about fourteen months. The lesson is not that every aggregator is worth billions. It is that when a market is fragmented and buyers need to switch often, the layer that removes the switching cost becomes very valuable.

GPU compute has the same problem, at a bigger scale

The model market was fragmented across dozens of labs. The GPU market is fragmented across hundreds of clouds. SemiAnalysis tracks 209 of them, up from 124. Each has its own console, contract, API and bill.

Prices are spread out and they move fast:

• An on-demand H100 lists from about $1.25 to $11.07 an hour depending on the provider.
• AWS cut its H100 on-demand price by 44% in one announcement in June 2025.
• One-year H100 contract prices rose about 40% between October 2025 and March 2026.

And the spend underneath is much larger. Gartner puts AI-optimized cloud infrastructure at $42B in 2026, up 96% year on year. That is the rentable GPU market an aggregation layer would sit on top of.

Switching providers today means a new account and contract, a new API and tooling, a new bill to reconcile, and re-testing who is reliable. So teams pick once and overpay for as long as they stay. That is the same switching cost OpenRouter removed for models. For more on it, see how to choose a GPU cloud provider, and why switching costs so much.

Why the tools that exist aren't it yet

Several companies cover part of this. None of them covers all of it:

• Aggregators like Shadeform give you one API across 20+ GPU clouds, but you still pick the instance yourself. NVIDIA DGX Cloud Lepton covers NVIDIA's cloud partners only.
• Orchestrators like SkyPilot and TensorPool run jobs across clouds you already hold accounts with. They don't run a marketplace or show you market prices.
• Marketplaces like RunPod, Vast.ai and SF Compute each sell their own pool of supply.
• Indices and futures from Silicon Data, Ornn and Compute Exchange price compute and sell large contracts, but don't route on-demand jobs.

The details are in our comparison of GPU cloud aggregators and marketplaces.

What an OpenRouter for GPUs needs to do

Models and GPUs differ in one important way. With OpenRouter, you usually choose the model, because the model is the product. With GPUs, the provider shouldn't matter. Nobody cares which cloud trained their model, as long as it was fast, reliable and fairly priced. So an OpenRouter for compute has to do more than forward requests. It has to make the choice.

That means:

1. Take a job description, not an instance ID. GPU type, count, region, price ceiling and deadline.
2. See the whole market live. Prices and availability from every connected provider, not a fixed list.
3. Score the options. Rank them on price, location, reliability and performance.
4. Place the job on the winning provider through its own API.
5. Bill once, whichever provider ran it.

Here is what that looks like as a request:

POST /v1/route · 8× H100 · US · < $2.50/hr · done within 12 hours

Illustrative request. The routing API is in development.

OpenRouter vs Opengrid

	OpenRouter	Opengrid
Layer	Models	GPU compute
Sits between	Developers and model labs	AI teams and GPU clouds
Who chooses the supplier	Usually you pick the model	Opengrid routes to the best provider
Fee	About 5% (Sacra)	Not yet announced
Status	Live	Pre-launch. Market view first, routing API in development

Where this goes

Routing is not the end point. Once real workloads flow through one layer, it sees real transaction prices instead of list prices. That becomes a live price for every GPU type and region, and a live price is something you can hedge. CME lists H100 and B200 rental futures from 5 October 2026, so that market is already starting. See CME GPU futures explained.

The model layer already has its aggregator. The compute layer doesn't yet. If you want one API key with every GPU cloud behind it, join the Opengrid waitlist.