Curious, how do services like this stay compliant with export controls on these devices? I assume a credit card is all I need to access these things?

They do not export devices, they export services.

Precisely. BIS does not consider cloud compute services to be exports.

There are some rules around when you knowingly know that a customer is military or WMD, and you obviously can’t supply services to someone on the entity list; but KYC-through-credit card kinda works at small scale.

Not the cheapest. Many providers have spot H100s for $8-12, and H200s for slightly less than the above prices

2.038 < 8

Per hour per 8x node (so $1-1.25 per hour per gpu)

Do you mean $8-12 per day? Otherwise I'm confused how it's cheaper.

I assume these are prices for 8x nodes.

Per hour per 8x node (so $1-1.25 per hour per gpu)

vast.ai has H100 at 1.93/h, H200 at 3.8/h but no 2500$ minimum. It's also possible to get a DPA with them, see where the machine is, this feels a bit comparatively dodgy

I like thundercompute for my personal stuff, especially a A6000 for 0.35/h when i want to do some training for example.

$1.09/GPU/hr for a H100!!

Their 6000 is never available.. been waiting for over 1 month now... it's always out of stock...

Its available right now for me, but yes i have had a few days where i have used the L40 @0.79 if i require the VRAM as the 6000 is unavailable.

If interested in compute, please let me know, got capacity at market.

Seems fraudulent. Who’s the operator/owner?

If it sounds too good to be true, it probably is.

So its ok for Anthropic and OpenAI to run inference at a loss, but criminal when others do it? :-)

For such a service it's legitimate to ask where those GPUs are working. Where's the data center?

It's more about where your money is going, since it's a prepaid offering.

This company would be more trustworthy if it listed an HQ, about page, history founder info, or anything...

great timing with codex, claude, cursor, and grok being down

long live opensource