Credits for Runpod from OpenAI

[PersonalRec]

Did any of you guys receive any credits from OpenAI to run the cluster on Runpod?

I personally find the prices on Runpod very high in comparison to other cloud platforms. I have applied twice for credits from OpenAI (the smallest amount) and never received any response whatsoever, even though they promise to give an answer within 2 working days. I mean, renting 8x H100 is expensive and OpenAI claims that they have a money pool for this competition. The most frustrating thing is that I didn't even get a rejection for the credits request, just complete ignorance.

Has any of you received any credits so far?

[shalyhinpavel]

Same experience here. I applied twice and only received the minimum tier ($25), which covered environment setup and 2 out of the required 3 seeds. Waited ~1 week for approval each time.

The practical impact: I'm now running experiments on Google Colab (self-funded) instead of RunPod, simply because it's cheaper when credits aren't available. My submission PR #875 — a pure neural architecture (no caching, no n-gram postprocessing) — was trained entirely on paid personal compute.

The deeper issue is that GPU quota is a shared resource, and the current wave of cache-heavy submissions (which can require 8×H100 × multiple seeds × iterative tuning cycles) places a disproportionate load on that pool. This leaves researchers focused on neural architecture — arguably the core intent of the competition — waiting indefinitely or self-funding.

A few suggestions:

Priority queue or separate credit pool for pure neural submissions

Faster credit response (current SLA of "2 working days" is not being met)

Credit scaling based on submission type or parameter budget used
The competition format is genuinely interesting and I want it to succeed — but right now the resource allocation structure disproportionately disadvantages solo independent researchers without institutional GPU access.

[iverbovoy]

I received 2 credits from OpenAI — $25 and $500.

The first one came almost immediately after applying, within the first few days. The second one took about 2–3 days of waiting.

Judging by the fact that the leaderboard hasn't been updated for three days now, combined with the influx of PRs, I'd guess the number of credit requests is several times higher than expected.

I also ran RunPod on my own money. But the experience was interesting and well worth it. Don't lose heart, think they can't process everything fast enough.

[greqone]

I have applied a total of 3 times I think, got the very basic 25$ USD which went basically for setup, used my own 450 USD on Modal and then pushed around 350 USD of my own money into Runpod as well
The turnaround seems quite bad here

[ibarrajo]

same as @iverbovoy I received both credits, same timeline and also ran the runpods with my own money at the start, out of pocket I probably spent about 120-140.

I try to get a few different ideas in, do light weight testing, ensure the artifacts will result in the proper size by using just noise, and I have them queued up for testing in the 8xH100s so I can do multiple tests in sequence since provisioning the runpod takes a fair amount of time.

I'm considering waiting a couple of weeks now before my next set of submissions because I will probably be more efficient in the future and give more time for any shower thoughts / novel ideas to take place instead of repeating the same patterns as everyone else with a fraction of a gain.

[hegdeadithyak]

It's the same for me, I did not get the credits yet may be they are busy but it will be good if they allot the credits.

[asuramaya]

I have not touched runpod, all mlx. the h100s are cool but expensive prove the concept on apple silicon first.

[hegdeadithyak]

I have not touched runpod, all mlx. the h100s are cool but expensive prove the concept on apple silicon first.

Won't we do that if we have apple silicon?

[asuramaya]

I have not touched runpod, all mlx. the h100s are cool but expensive prove the concept on apple silicon first.

Won't we do that if we have apple silicon?

yes and no, everyone wants to be #1 on the leaderboard but pay-to-win is against the spirit of the competition, you can think around the money or spend it on inference. I submit a cheap mlx result and someone else pays for the h100s to verify or destroy it(haha)

[hegdeadithyak]

I get your point about avoiding a pay-to-win approach, and I agree that proving concepts efficiently matters. We’ve already validated our ideas on Colab atleast me and hopefully others as well for a very small subset of data, so the foundation is there.That said, if the goal is to compete seriously on the leaderboard, at some stage we’ll likely need stronger hardware like H100s—not as a shortcut, but to fairly benchmark and push the limits of what we’ve built.

[faisalmemon]

I suspect this competition has turned out to be more popular than anticipated. For what it's worth, my approach is to tune the model so it works on my DGX Spark so that I can still make architectural decisions and experiment. I am working on the expectation my fixes would apply to the H100 run.