Aedilic

tl;dr gpudeploy.com is Airbnb for GPUs. It is a marketplace and easy-to-integrate software for renting out the over $1bn in cheap, high-quality compute that is sitting around unused (see e.g. gpulist.ai). We are two engineers from Cambridge and ETH Zürich with experience in Robotics and Deep Learning.

—

Hi, we’re Nico and Lukas, the cofounders of Aedilic (doing business as GPUDeploy).

We met in high school when we were both 14 and bonded over our interest in coding. Our first project was a cookie-clicker game that we wrote during a sleepover two weeks after meeting for the first time.

Nico has since received a Masters in Econometrics from the University of Cambridge with a research focus on the mathematical theory of Machine Learning algorithms (Random Forests, DBSCAN). Lukas holds a Masters from ETH Zürich, where he trained robot dogs at the Robotics Systems Lab, using the latest Machine Learning models. During his Masters, he wrote the official ETH Zürich RSL open-source Reinforcement Learning “algorithms” library. His Master's thesis was published in the highest-ranked Robotics conference (ICRA) this year.

There is a lot of cheap, high-quality, unused compute sitting around

gpulist.ai alone lists over $1bn in unrealized contract value. These are the latest Nvidia models in cutting-edge data centers, with Infiniband interconnect and perfect internet. They are cheap because nobody books them. Meanwhile, research labs and AI companies are looking for cheap, high-quality compute but will not go through unknown, unclear websites and the associated sales process. We allow them to book pre-configured GPU instances on-demand at wholesale prices that they would normally need to negotiate.

We’ve built software that makes idle GPUs in data centers accessible and super simple to book and use via our website. Data-centers and AI companies with idle compute have the option to list instances on the spot market as soon as they are available, it takes 5 minutes to set up. Unlike other solutions, our software enables fractional use of GPU boxes, so 8 GPUs can be rented out to 8 different people.

Other options are either expensive, booked out, or a pain to use

We know how inconvenient current solutions are because we’ve dealt with them, both at university and for the previous iteration of this startup (robotics-related).

For example, Lukas, during his time at ETH, was given a blank check by his lab to book GPUs for his research. He spent the entire night with his supervisor going through Google Cloud to scrape together a few GPU instances; Google will only tell you that an H100 is unavailable in Region X after you configured and launched your instance. Plus, they were expensive, with an A100 SXM 40GB at $3.60/gpu/hour (compare this to $1.28 on our site).

Solution: GPUDeploy

GPUDeploy is a software solution and marketplace for renting GPUs. We have made our site simple, to the point, and opinionated - there are no hurdles between you and booking a GPU. We offer very competitive pricing and ensure that the compute instance you book is reliable and meets our benchmarks.

We both have a spot offering and an Airbnb-like experience for long-term bookings. We match your spot request with available compute resources that meet our standards. You can then simply connect to the machine to run your workloads.

We launched with a Show HN four weeks ago, which immediately got #1 and stayed on the front page for 2 days.

Try it out!

Our ask:

• If you need GPUs, try it out.
• If you have compute, try renting it out on our site. It’s really simple (2 commands) and flexible, and we help with custom onboarding. For example, you can make a profit of ~$3000k a year with an RTX4090. Currently, the majority of our computing is booked out. Anyone is able to add compute, but we will sort out unreliable providers immediately.
• We collaborate with data centers. If you have a cloud with idle GPUs, reach out.

Message nico@gpudeploy.com or contact@gpudeploy.com for any questions.