Hacker News new | past | comments | ask | show | jobs | submit login
Ask HN: Lots of AWS credits expiring soon, ideas?
72 points by anonymouslenz 10 months ago | hide | past | favorite | 97 comments
Hey HN,

Long story short I have a lot (~$10K) of AWS credits expiring in a few months. Any ideas on the best way to make money / do some good with those? I can't directly sell them. Thanks!

Find a PhD student that needs GPUs for their research but isn’t in a first world institution. You could reach out to a few top professors and ask for a referral to someone promising.

Ie be a science benefactor

As a PhD candidate myself, I have to say this is an amazing idea. Meet with your local university's CompSci department and they will have many good uses for them.

I don't need them, I have actually accrued a large collection of credits myself for the program. But another university may not be as fortunate, so this can be a huge help and be the difference in many research projects being completed or not. Overall, research benefits everybody, so this is a better use of the credits than using it to middle-finger Bezos that many comments are suggesting.

Especially because credits are more beneficial and needed than ever before right now. I can tell you from being in academia that they are pushing almost all research projects to find AI/ML angles on everything we do. Thats the type of research in demand right now by the major journals and publications so all the universities are pushing for this twist on all current research projects to improve the likelihood of being published. And as you know, doing anything with AI/ML can 10x - 100x the cost. So credits are more valued right now than ever before.

Good idea, but make sure they don't accidentally blow past $10k, leaving you with a monster bill. People can be a little reckless with "free money".

This. AWS Budgets to get notifications when they reach certain spend thresholds will be a must. Also you should probably setup an AWS organization and have them operate in a sub-account if you go this route.

wow this is some real advice. I fully agree!

it's damn near impossible to get approved for enough GPU to do any serious research with random free credits it won't help much :/

absolutely, unfortunately you are right.

The suggestion of donating these concerns me.

IIUC these credits aren’t transferable so you’d have to set them up with access to an AWS sub account attached to your credit card. Last time I used AWS personally (it’s been a while) there was no way to limit spending in a subaccount.

Giving someone access to your account to burn through exactly $10k in credits where you are on the hook sounds… financially dangerous.

Am I missing something? Is there a way to make sure they only spend the $10k in credits and not a penny more?

A prepaid credit card, or virtual card (like those provided by Wise) would do the trick.

Unless AWS doesn't allow those.

It’s payment on account after use. Just because the card bounces doesn’t mean you’re not still on the hook for it

Every time this is mentioned the usual counterargument is that not paying your bill does not free you of responsibility for it. You can still be taken to collectors or end up in court.

they dont. thats why this situation sucks

Only $10k? Time to run a massively over engineered and over priced micro service based web server for like one month.

CQRS time baby.

one of the few times when CQRS is used properly

Should be enough to run something like https://vecka.nu for two months even! (vecka = week, nu = now)


Set up a server to help render up to date OpenStreetMap map tiles: https://wiki.openstreetmap.org/wiki/Servers/Tile_Rendering

Usually needs dedicated servers, but you should be able to contribute if you use a beefy VM

Hi! I'm the OSM Foundation board chairperson. We have a tile server running on AWS, generously sponsored by Amazon. It's quite a big machine, and the budget mentioned would only pay for a few months of runtime. It's currently one of the two tile servers we have in North America, and the fastest of the two by far.




Hetzner also has great servers for a budget, maybe they would be up to sponsor one or two servers for your project. Not a real cost for them, but great for your project. Just an idea.

I've been working on a project evaluating the Python and Javascripr programing performance of open-licensed large language models [1] but don't currently have the resources to run the larger 65B ones.

Any chance you'd be able to help run some evaluations?

[1] https://github.com/the-crypt-keeper/can-ai-code

Maybe run one of the large distributed computing projects like Folding@Home? https://foldingathome.org/

I thought Deep Mind had solved all of the folding we would ever need.

Did some digging and the answer is "apparently not": https://foldingathome.org/2020/12/08/protein-folding-and-rel...

That article is from 2020, but it seems there has been a lot of progress since then – https://www.deepmind.com/research/highlighted-research/alpha.... I think this might now be a solved problem, or at least it's not clear that Folding@Home is going to be the next breakthrough.

Well, if Folding@Home's project still exists, I'm going with Deep Mind is not working with them or there's still space to help.

Or Folding@Home is barely active enough to decide to stop, or they have a sunk cost fallacy... there are other options.

While Folding@Home has some mindshare among techies of a certain age, I suspect it's no longer practical given the wide availability of much higher performance cloud computing, and I don't think it has ever been cost effective. I suspect it needs reevaluating in the modern era. The last time I was looking at the project at all seriously was in ~2006 I think.

> Folding@home is one of the world's fastest computing systems. With heightened interest in the project as a result of the COVID-19 pandemic,[8] the system achieved a speed of approximately 1.22 exaflops by late March 2020 and reached 2.43 exaflops by April 12, 2020,[9] making it the world's first exaflop computing system. This level of performance from its large-scale computing network has allowed researchers to run computationally costly atomic-level simulations of protein folding thousands of times longer than formerly achieved. Since its launch on October 1, 2000, Folding@home was involved in the production of 226 scientific research papers.[10] Results from the project's simulations agree well with experiments.[11][12][13]

Source, https://en.m.wikipedia.org/wiki/Folding@home

Find a large prime number https://www.eff.org/awards/coop More ideas in an 8 year old thread https://news.ycombinator.com/item?id=9546609

I wish there were more mirrors to wikipedia dumps but not sure if few months make a difference https://meta.wikimedia.org/wiki/Mirroring_Wikimedia_project_...

Or calculate pinto a lot of digits. For $10k you might hit the record.

Not close:


108 days on a 2021 supercomputer is a lot more compute than $10k.

Beaten a year later in 2022 by GCP once again


Also for a lot more than $10k.

If I had a thousand gallons of unusable gasoline that was going to expire in a few months, I would just be proud that I found a way to not expend it

My metaphor falls apart if those 10,000 dollars of AWS credits are all coming from green energy, though

You are implying that AWS turns off the machine if this person doesn't spend his $10k. I don't know exactly if that is correct, but it doesn't sound like what actually happens.

If you run it through AWS Montreal the energy is 100% hydro, so if they did care about that, it's possible to avoid.

> You are implying that AWS turns off the machine if this person doesn't spend his $10k. I don't know exactly if that is correct, but it doesn't sound like what actually happens.

That is more or less correct.

Machines that aren't running anything are powered off to save money. so if you spend $10k in credits running a bunch of useless compute, that more or less converts into say $3k of wasted electricity (if we say 30% of AWS running costs are electricity, i just made that up, but it will be some non negligible percentage)

So if you have 100 servers running EC2 at 99% compacity, they turn off all the other machines? That doesn't sound correct. I imagine they have a good buffer and then maybe there is a subset that are powered down. I don't think you can power up a machine, launch the kernal VZ layer, connect to AWS and broadcast yourself, and what ever else it does in the 20 seconds it takes for a EC2 instance to be allocated to you, so there has to be a good buffer for people launching/tearing down EC2s constantly.

So basically there may be a few servers difference. So we are talking like a few kws, which is basically like 40 cents power wise using 4 cents a kwh.

Sometimes it does take significantly longer than 20secs.

I don't think they necessarily turn OFF said machines, but most computers for the past half decade will do lower power modes with a lot of efficiency. Most of the mini computers I've seen based on laptop skus (similar for servers, just many more cores) will run around 8-12W in idle state, but 10x or more that under load. I would asume many servers will be at least similar, with a higher top end.

I would agree unless the gasoline could potentially be used for scientific research that would allow less gasoline to be produced in the future. Which seems plausible for AWS credits.

What do you do with expired gasoline? I'm sure it does not become much less of an environmental issue after expiry? It seems to me that it would just become a lot harder to dispose of safely.

old gasoline works just fine as long as it is mixed with new stuff...

This “solution” defeats the very intend of GP proposal to not use it :-)

How about diesel?

Diesel isn't so bad about age, but you still need to watch out for water and rust building up in a tank of stored fuel over time.

You can just put it in an older vehicle that doesn’t care

nah - old fuel is especially bad for old vehicles. They really don't like having the fuel separate into waxy components in carburettors and clog everything.

Just mix it 50/50 with fresh fuel and you're good to use it anywhere.

Put an email in your bio so folks who need to get something done can contact you. Even if you cannot sell the credits directly, who says you cannot run something for someone else? Say, as a consulting contract.

If that’s too complicated, find some open source AI project that needs compute. RWKV. OpenAssistant, or something like that.

Like previously mentioned. Definitely give it to a PhD student *NOT* in the first world institution.

Maybe this is a bad use, but finding one of those really high-end AI voice suites and running off a bunch of audiobooks for books that never got narrated editions would be cool. Every time I get access to a better one I make a slightly improved version of "Sentenced to Prism."

This is really cool, and with the right distribution can really help people who can’t see, to access the material!

The sad state of affairs is that licensing complexities mean there are actually a lot of audiobooks that only exist to be accessed by blind people and aren't distributed for commercial sale.

Couple Ideas:

- build some sub accounts and run a hackathon for a local school/ kids non-profit - reach out to some non-profits and see if they have any processing jobs that could be helpful (you'd be surprised!) - train a weird CNN and open source the model (like one that can find wheres waldo, or something silly - do some benchmarking (really hard to find good benchmark data) - do some ML research - Run a job for a small startup that might struggle with compute costs - the PhD idea previously mentioned is fine, just don't send it out of North America, and don't do tier 1 universities... if your endowment is larger than the GDP of first world countries you can afford to give your PhD students some compute

best bet probably https://foldingathome.org/ unless there something specific you're interested in. Ai being all the rage good budget to play around with some gpus

You might have enough to run up a few electron apps for a day or two. Maybe.

post your public key to github, the rest takes care of itself.

_public_ key?

You can join an AWS partner, give them the credits in the exchange get virtual credits for a later consumption.

Donate it for learners. There's plenty of people doing certification. I'm sure a lot of them would like to do "hands on". If I were you I would go to places like examtopics and advertise. 100 of dollars would be enough, and it's easy to add subaccounts to AWS

It's easy to add subaccounts. It's hard (to impossible) to limit them to an exact budget.

Filter the CommonCrawl data for webmanifest file URLs and publish a concise URL index as a simple text file (on say GitHub).

We can use this to bootstrap a Progressive Web Apps (PWA) index webpage (also as a PWA) and give the platform app stores some competition!

CommonCrawl has an AWS Athena SQL query sample here: https://commoncrawl.org/2018/03/index-to-warc-files-and-urls...

We can filter by

  content_mime_type             STRING,
  content_mime_detected         STRING,
fields for the MIME type 'application/manifest+json'

Try train your own LLM.

If it's not against the terms, maybe you can sell it at a big discount?

I have the opposite problem, we have just finished a massive amount of AWS credits from our VC and unfortunately cannot stack these. Any advice on how to get more credits?

Head over to the elueuther.ai discord and discuss with some of the folks there. Tons of small experiments with LLMs can use the $10k in compute

Create a service which matches users with expiring credits to people who need compute at a discount. Spottier instances?

Tor relay or exit node until you find someone who could use the computing power better, like the PhD student as mentioned elsewhere (though I'd say any good idea that would get published can be worthy, don't need to sign with a university to do research as a hobby)

I work for a nonprofit that spends quite a bit on AWS. You could donate them and get a tax credit.

AWS promotional credits are non-transferrable.

The only way for the OP to "donate" those credits is to create an AWS Organisation and allow the non-profit create a sub-account to use the $10k credits.

But the OP has to be vigilant to ensure that they do not go over the $10k otherwise they will be on the hook for any additional charges that are racked up.

In Canada, gift cards aren’t allowed to expire any more. Is this equivalent at all?

it's probably a "promotional credit" that OP didn't pay for, aws doesn't really do gift cards

Try to come up with the fastest way to burn through it and then make a YouTube video discussing how you did it. Preferably something both very clever and very stupid at the same time.

Host a kernel.org or Debian or gentoo or some other mirror for awhile. Talk to the maintainers first so they realize it'll be a "short term" mirror which they may not want to hassle with.

Contribute to open source LLM training. Any projects?

just wondering: could you buy instances with the full-upfront payment?

you'd not only get to spend the money, you'd also get a significant discount.

Easy, create your own saas, do a promotional services, put a time limit, there’re a lot of FOSS projects

Capture auth packets of as many neighbours' wifi networks as you can and crack them on a GPU :)

I'm a bit confused why nobody suggested cryptocurrency mining? not feasible?

It's not permitted under any of the major cloud's ToS.

The big cloud providers attempt to detect and stop crypto mining. I'm not sure if there are ways around it, but if so, they probably involve mining incredibly inefficiently.

Personally I would look into standing up one of the popular AI models like stable diffusion and running a paid service on it.

I think this was a lot of people's first thought as well, but maybe we can brainstorm some other uses that provide a more direct benefit to persons in need, dare I say to society.

Find some worthy open source projects and run mirrors/torrents for them?

we had same issue, at that time it was we will not hire few developers or pay the bill to big corp, we switch in 1-2 weeks to hetzner on dev and stg, and saved good amount of money.

Just let ‘em expire.

Finetune an LLM for your own application.

make a REST API interface to the old fortune linux command. You could probably run it forever.

Download Common Crawl archive.

Train a new llm.. or btc mine

You probably mean Monero since it's mining Algo is CPU friendly

Spawn a bunch of a100

Mine a Bitcoin

buy reserved ec2 or rds capacity

Unfortunately credits don't usually apply to that. Learnt the hard way .

Train a model then burn it to the ground.

Fine tune a llm to make your mom jokes

comedy mine dodge coin option

Depending on the altruism level of yourself, you could do something like BOINC or just mine coins?

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact