The same H100 costs $2.69 or $12.29 an hour depending where you rent it | Hacker News Reader

The same H100 costs $2.69 or $12.29 an hour depending where you rent it

github.com

166 pointsdatareceipts84 comments

You are completely mad if you are running any kind of valuable workload on Runpod community or Vast. Hosts are absolutely peeking inside the VM's (you can see it openly discussed in their Discords).

Could you please share screenshots?

Sam is welcome to peek at my generative furry porn all he wants ( ͡° ͜ʖ ͡°)

( ͠° ͟ʖ ͠°)

There's got to be a ton of valuable workloads that don't care much about secrecy.

Who cares if someones spying on the process of fine tuning a model to help us pick strawberries? Even if they could somehow find a competitor to sell it to - the competitor wouldn't be interested because the risk of violating trade secret law is surely greater than whatever they can glean from our training run...

> violating trade secret law

There is law for patents. Trade secrets exist outside the realm of enforcement.

Trade secrets are protected by civil (https://www.law.cornell.edu/uscode/text/18/1836) and criminal law (https://www.law.cornell.edu/uscode/text/18/1832). There's a history of enforcement actions under both.

Completely incorrect:

https://en.wikipedia.org/wiki/Trade_secret#United_States

That's a common misunderstanding. If you learn of a trade secret through improper means, you can be liable federally under the DTSA or similar state statutes like CUTSA in California.

These were the statutes cited in the Anthony Levandowski lawsuit 9yrs back, where Uber settled for a 0.34% equity stake in their company.

If you give away trade secrets by intentionally giving them to another party without an explicit agreement of secrecy no statute in the world will protect it.

You have to take “reasonable” measures to protect trade secrets. Uploading them to a service without any confidentiality agreement or worse, where the provider officially has the right to use that data, defeats the definition of a reasonable measure.

> Who cares if someones spying on the process of fine tuning a model to help us pick strawberries?

Privacy should be a default. It doesn’t matter what you’re doing.

It’s also unlikely they know what’s in the VMs until after they look at it.

Sure, but if you have the option of training your strawberry picking model for 1/5th of the price knowing that you will be spied upon, that's probably still a good deal.

Just like you can find all of my 3D models online because not being able to make them private is the price for the free license of Onshape, and I'm happy to let you look at those models rather than pay whatever a commercial license would cost.

Exactly. I've recently been training a number of models, and the value to any potential competitor would be approximately zero.

It's so integrated with the specific inputs of my product, and the outputs that it needs, that the model just has no value to anyone else.

If a competitor needed such a model, they need to train their own that matches the form of their inputs and outputs, which is just going to be different.

As long as your training data itself isn't sensitive. If that's e.g. photos of products and you're training a quality assurance photo recognition software, that's probably fine, if it's unredacted support tickets, not fine.

You can process your data through some learned feature extractor offline first, just to be sure.

I used to work at a company who used to do this, but I'd argue that it almost never does what you want re: concealing the input space. There are so many ways to recover the input space from the latents, and often they are constructed expressly to contain exactly enough information to do precisely that.

But how much effort would that require? ... is what I was going to say but in this age of "claude reverse engineer this doohickey" maybe the better question is how much compute and how many tokens it would require.

> a competitor to sell it to - the competitor wouldn't be interested because the risk of violating trade secret law is surely greater than whatever they can glean from our training run

Unless the competitor is in a country like China.

>Unless the competitor is in a country like China.

But how likely is that?

IANAL, but it seems to me that to have something as a trade secret usually requires showing that the secret keeper put in significant effort to keep the thing a secret, usually there being some contract in place that defines the domain of knowledge that is considered a secret if the knowledge (i.e. models) need to be shared.

Uploading the models to an unscrupulous hosting provider where there was only some boilerplate terms of services agreed to.. would not seem to be that.

Being careless in how you handle your corporate data will be evidence against your trade secret claims. You don't just get to claim anything as a trade secret, you need to actively go about trying to keep it a secret. Going with vendors who you know or had reason to believe were going to mishandle your "trade secrets" likely makes them no longer trade secrets.

Coca-Cola can't go dropping leaflets around with the recipe for Coca-Cola Classic and then claim the recipe is a trade secret.

IANAL. Talk to your own lawyer before you mishandle your data.

A lot of Vast.ai GPUs are also firmware-limited, so it's a big gamble when you run your workloads. You also can't get a cluster.

It's fine for comfyui workloads or simple LLM inference, but you lose both privacy and stability.

GPUs set to lower power states and CPUs with frequency scaling turned down. That's the biggest hit and miss for what I'm doing. Honestly don't care if people are looking at my workloads as it's just processing public data. But do care if it's someone's laptop that's actively using 1/2 the vram.

each machine's power limit and performance are displayed and filterable on vast. not sure about other marketplaces but vast is pretty transparent about the measured specs of their machines

It might have been fixed, but I had a problem a few months ago where I would get 2x5090s, but they were limited at 80% usual clock. Nothing was different on the rental card.

> (you can see it openly discussed in their Discords).

Can you explain? I haven’t used Runpod and don’t have time to search through a Discord to figure out what someone is talking about.

I haven't seen any Discord discussions, but the nature of the platform guarantees there will be leakage third parties. RunPod consists of random people renting out their own GPUs for use. They aren't owned by RunPod.

In any system where large numbers of people are renting out their homelab equipment, It's to be expected that there is no consistent data privacy policy. I can't even imagine trying to enforce some sort of system that would guarantee privacy of workloads.

I hold the inverse risk assessment- I think acting on a peek risk is much higher with larger operations such as OpenAi or eg Amazon -as they've allegedly done the same on their product marketplace and deciding on generic products, in addition to open source-alike service offerings.

Larger entities have the resources, the inclination and the protective legal teams to act much differently than a resource tight n-tier platform operator

It's really quite the opposite.

Large firms have a much harder time hiding that kind of behavior, and they have much more to lose.

Large firms require processes and controls around everything. If Amazon is peeking into corporate customers' data, then there have to be mechanisms designed to allow that, a whole bunch of people know, it's part of engineering designs, and a whistleblower is going to leak it to the press.

And then Amazon is going to lose all of its business customers once its privacy guarantees were shown to be intentionally circumvented.

It is simply not economically rational for a company like Amazon or OpenAI to look at the data of its corporate customers.

Regular consumers are a different story, because you need to respond to abuse, illegal activity, etc. So they don't have the same privacy guarantees.

But corporate customers, that stuff is watertight. No cloud provider is going to risk losing their entire business because they violated their privacy guarantees.

> If Amazon is peeking into corporate customers' data, then there have to be mechanisms designed to allow that, a whole bunch of people know, it's part of engineering designs, and a whistleblower is going to leak it to the press

An awful lot of this sort of thing in the corporate world is based on pinky-promises and legal threats, more than on robust technical controls.

Every production system I have ever seen has at least one set of dev-accessible keys that has access to the production DB. Mature systems at least had audit logging around who retrieved those keys, less mature systems often didn't have any kind of internal access logs.

Is "Amazon the corporate entity" snooping on your data? Absolutely not. Is someone with Amazon internal access snooping on your data? You'll never know either way.

That's exactly how amazon decides what becomes an amazon basics product.

AWS has also strip mined open-source software, elasticsearch, redis, mongo, browser rendering tools, etc.

Given that even non-Amazon vendors can use Amazon.com search features to discover bestselling products from third-party sellers in various categories, I doubt there's anything contractually prohibiting Amazon from using the same information to drive their own product mix.

Same goes for selling prices. Published retail prices are not trade secrets.

> If Amazon is peeking into corporate customers' data, then there have to be mechanisms designed to allow that, a whole bunch of people know

According to this comment, they did have a mechanism and a process for that: https://news.ycombinator.com/item?id=23929959

Yet it seems to have had negligible impact on their ability to obtain and retain customers.

What are places in your view to rent GPUs that are trustworthy they will not peek in your opinion?

UpCloud's [0] spot instances sometimes match the pricing of Vast.ai instances. I mostly use Vast.ai if my dual 3090s can't handle it anymore for non private things and move to UpCloud spot instances if it needs more.

[0]: https://upcloud.com/

I think it's best to operate under the assumption that anything you're sending to a third party is going to get spied on at some point.

You are ignoring the real usecases here. Not everything is large scale, secretly, b2b enterprise applications. Instead, many consumer applications are fully fine with being out in the open.

As an example, midjourney use to just publish ever generated image to their website.

If you are just doing inference things are public/open from the start, you dont have to worry much about the stuff you are talking about here.

> As an example, midjourney use to just publish ever generated image to their website.

Which is kind of genius, if you think about it. Users can't get away with generating smut, CSAM, or copyrighted material, because it goes straight to the public record.

Why would anyone in the not un-nice parts of the world aged past the teens feel like others seeing the smut they've generated being something to get away with? The people not into discussing their sexual lives with you are highly unlikely to be comfortable bringing up even the existence of yours anyways, and the people you have shared interests with now just have a new springboard for curiosity and conversation. Sure, you have the Swedish municipality director that a couple of weeks ago chose to resign after it came to light that he didn't quit his onlyfans work after all. But even that was more about lack of openness, breach of trust in a trusted position, than the nature of the activities in and of themselves.

I don't know man. I agree that people probably shouldn't care, but the US figures sort of highly in discussions about the internet, and it's a place with pretty regressive attitudes towards that kind of thing.

There's a lot of use cases that are not "commercially sensitive" that would require data center level privacy. E.g. a good percentage of academic research and experimentation, long running "boilerplate processing" for monitoring logs for e.g. a bird watching system or large scale web crawling where the individual website itself is mostly public. There are also use cases where the data itself is useless without other data to correlate. E.g. many log analytics use cases, if you are missing the main database data itself then the logs have zero value when intercepted. There are also use cases where you are doing general industrial level hotdog/not hotdog classification where unit economics is the most important and there's no real value to privacy e.g. on a factory floor (though sabotage is a bigger risk here).

Could confidential computing help here?

What is it good for if hosts spy on their customers?

Public datasets, and public data analysis.

I've used them a bit to prototype stuff where the data didn't matter.

Enthusiast use, probably.

Honestly I think every one is peaking at everything, I don't want to be slanderous but I think even the more trusted cloud providers peak, though they do so more quietly and carefully.

I work at a big cloud provider and this is absolutely not happening on the regular. More and more safeguards have been put in place over the years to prevent that.

Of course it doesn't happen on the regular.

But it definitely does happen.

The 1.8x difference between the cost in AWS and Azure is still relevant though.

I only use their serverless stuff, but you can choose which security and compliance you want for their serverless stuff.

One thing that I'm not seeing mentioned very frequently is that it's really hard to actually get the GPUs that hyperscalers advertise depending on your region, use of cloud credits etc.

On Azure if you have credits (even hundreds of thousands of USD of credits through programmes like MSFT for startups) you just can't get any GPU quota in any region, even if you beg and plead with support reps. Even if you have cash it's often difficult to get any of the cheaper cards. You can forget about spot instances etc because the regions are essentially all at 100% capacity.

I've heard on other hyperscalers that it's also challenging.

This is super chilling for anything other than GenAI inference/training because if you have some more traditional workload that requires GPUs (traditional ML, other kinds of modelling) you're SOL.

And not just hyperscalers... even the neoclouds who profess to have fantastic availability and GPUS NOW just tell you to jog on. Though of course they do that after they've signed you up and promised preemptible instances :)

There are a couple of new indices out there.

[1] "Silicon Index". Built against pricing observations, including listed prices and transaction data.

[2] "Ornn Compute Price Index (OCPI)". Claiming to use only executed transactions.

Your mileage may vary, but I find these new indices to be quite compelling.

[1] https://www.silicondata.com/products/silicon-index

[2] https://data.ornn.com/

And yup, its not just how much these resources cost, but how much people are willing to pay for them.

(edit: formatting)

Ornn is complete BS. Tried dealing with them and they were ridiculous.

what'd they do?

> 4.57× between cheapest and dearest, for the same chip

> What the 4.57× is not

> - InfiniBand

> - A datacentre SLA

> - Enterprise compliance, committed capacity...

Worth nothing that neither of those costs 2x the price of running GPU itself

Yeah but they have the infra and hence the leverage.

I have 0 idea why a 8 entry MD table with a side of AI blather is being upvoted to top of HN.

Oh god I dealt with this recently while training my own 1B model.

Runpod - mostly okay, except sometimes instances get dropped on smaller ones. Prices float around 3.20 mostly for H100s. Cheaper ones exist but are PCIe. Best DX out of the bunch honestly (well, AX, as my agent was the one dealing with it). Didn't end up doing the 1B run on it, but did 100M+ runs and some minor experimental ones. Great for agentic exploration, "RSI", minor training loops and research.

Vast.ai - a bit pricier for the same H100s and it didn't have any clusters. So I could at best split my workload across 2-3 different H100s and spend time syncing. Not worth it.

Akash - since I'm trying to stay in a reasonable budget, I decided to try Akash, which was the worst experience out of all. Cards would be "available" but would never respond to your request. After trying to contact support via Discord, I got dozens of fake support accounts adding me to scam discords and asking for my wallet key. Support answered after a while, but without a real answer.

Mithril.ai - Ended up using this one, while it seems quite startupy, the actual offer is nice, it was easy to get going and they offer quite cheap H100s, offering singles and clusters.

Tbh while Skypilot is great, the marketplace fragmentation is leaving room open for a openrouter-like compute provider.

This is why ClusterMax (from Semianalysis) was created. Not all hosts are the same, even with the same GPUs.

https://newsletter.semianalysis.com/p/clustermax-30-the-indu...

Oh nice, did not know about this.

Looking at the report, I didn't even know one could use CoreWeave without an enterprise account, seems it could be cheaper per hour than the same cards elsewhere. Might have saved me a few hundred bucks, thanks!

One interesting thing going is device fraud on the 3rd party market places - i.e changing the clock speed, dividing devices up and renting same one at same time and hoping rentees won’t notice, everything you can think of someone is trying.

lots of interesting dynamics in the preemptible (spot) markets as well. If you’re smart i’d say there is a >10x cost advantage available over hyperscalar OD pricing with basically no downside. All easy with agents now as well, so it will likely soon be competed away (you already see this with spot rates increasing significantly in last 3 months and converging to just below the od prices)

This is premise for https://skypilot.ai/, right? Cloud provider arbitrage.

Partially. SkyPilot now targets managing reserved clusters mostly for larger-scale users. Smaller-scale users tend to use elastic.

There’s much more data here, with more providers and splits by the different H100 types: https://cloud-gpus.com/nvidia-h100/

I tried the TensorDock, which now offers H100 "from $2.25/hr", two years ago. The SSD IO was terrible, maybe oversold. My training workers was stuck at loading the dataset and cannot fully utilize the GPU.

The spread is indeed very big.

Lowest $1.90/hr Average $7.08/hr Median $7.91/hr

https://gpus.skypilot.co/?gpu=H100&count=8

Yes. And the prices a lot of times don’t even make sense

I’ve run into many hosted GPUs with processes stuck, with outdated drivers/libraries, or poor Internet connections

I've only used Modal.com -- They give you $30 month of usage free. Otherwise, $3.95/hr for an H100 SXM5.

Speaking of Runpod.io GPU availability, it is definitely increased in the past 6 months, getting 80Gb+ vram on demand takes few mins (not 5-6 hours like it used to be) at 1-3 usd an hour rate ... which made me thought a week ago - finally GPU bubble peak arrived

Another thought - i want to audit the actual usage of GPU like actual efficient inference time not billed one without relying on "trust me bro" billing. Especially on AWS. Had no trust issue with Runpod.io

Any effective inference audit ideas are appreciated

lol watch the $12 people just subcontract out to the $2 and hope no one notices lolll

I guess the author hasn't discovered vast.ai which is a minefield.

Vast.ai seems like a scam to me. I funded my account with a little bit of money ($25) and tried repeatedly to reserve an instance, but it would just hang: no response, just waiting to provision it.

Gave up after 4 attempts and wrote the $25 off as a cost of the learning experience.

These prices don't seem real. For me, runpod is 8.99 an hour for b300 right now.

I wish there was some kind of GPU marketplace so one could just arb this opportunity.

Nice that they clearly explained why it is different

But a ps5.48xlarge has 8x H100s. Are the other options here comparable?

runpod is incredible, im a big fan, but good luck with availability on any specific GPU above 20gb of vram when you actually want to use it.

last time I was trying to run some workloads there the pricing of an h100 was largely irrelevant.

Hmm, if it works. Was interested so thought I would go check it out. Their site seems busted atm.

Same feeling. It’s great when GPUs are available. Also helps if you can use non data center cards in your workflow.

if they're all just an API away is anyone buying for 2.69 somewhere and then selling at 12.28 elsewhere arbitraging things?

I mean after the initial capex you’re just paying for electricity.

Seems once the government controls it the prices go haywire

Reply on news.ycombinator.com