AMP PBC founder Anjney Midha says cloud GPU rental prices have more than doubled since January 2026

It's so over...

FWIW there was a period 2-3 years ago where people were paying $5-7 per H100 hours for certain reservations. The crunch now is real, but it’s been worse.

@tmophoto umm. i'm not so sure. it might have more to do with everyday people gobbling up coding and other tokens like cookie monsters

@AnjneyMidha I think a large part of this is that ai crypto mining thing. There is a chance this is just a temporary blip

@azzabazazz @AnjneyMidha the change way too fast for that

@AnjneyMidha @tmophoto We need models on a chip, stat.

@tmophoto @AnjneyMidha And they'll continue to as well at the hyperscaler level. But if researchers get priced out of doing anything with a hyperscaler that doesn't deliver Oil Industry-grade profitability we'll need them to have a mechanism to continue to advance the bleeding edge

@AnjneyMidha My strategy plan.

⤵️

@AnjneyMidha We have availability! Especially with 5 whole boxes of 8xMI300x. Minimum one month. Same price as H100’s, which we beat with 2.4x more vram.

http://Admin.hotaisle.app

@AnjneyMidha What would you recommend today for someone with these two use cases? a) single H100 for a few hours a day. b) 8 or 16 H100s for a week, for a small training run.

runpod used to be reliable, but frequently out of capacity now.

@AnjneyMidha sf compute has h100 rentals averaging about $1.96 per gpu hour right now. frontier skus got expensive fast but a clean 2x on mainstream h100 since january is hard to prove from public marketplaces

As I recall there was never a shortage of actual toilet paper, it was just that a huge percentage of it was in commercial channels unused and no one wanted to bridge to consumer distribution when they thought it would just be a few weeks of disruption. I don't think there's a ton of unused but ready capacity hiding somewhere to unwind this, the rampocalypse or whatever is next, though I'd be happy to be wrong about that.

@AnjneyMidha AI gold rush. GPUs will be scarce until we solve energy. @FuseEnergyTech can't wait for fusion.

@AnjneyMidha @ANRFIndia Hope you take note

Our vendors are not agreeing to give us the GPUs at the agreed quote

@azzabazazz @AnjneyMidha buy a few sparks or gpus, its not that hard to bring it all under your control. you can do anything you can do on an h100 on a couple of sparks, just takes a bit more time

@AnjneyMidha Sticking with the analogy: if the industry uses a whole roll per trip, “shortage” becomes inevitable

@AnjneyMidha This is exactly when infrastructure assumptions break. Always-on GPUs made sense when capacity was cheap. While industry optimized for faster inference, not efficient inference. Now the bill is arriving. And this likely gets worse before it gets better.

@AnjneyMidha Expensive compute makes waste visible: sloppy retrieval, oversized contexts, idle tool calls, and unbatched experiments stop being implementation details and start becoming product bugs.

@AnjneyMidha I'm not part of a lab, so it's basically impossible to run experiments on my own anymore

@AnjneyMidha GPU rental prices are up 2x+” is not true for every GPU or every rental market. It is much more defensible for scarce frontier Blackwell / GB-series cloud capacity and some quoted dedicated-capacity deals than for the broad H100/H200 market