Insider: Red Hat is capping devs' bot budgets

EXCLUSIVE The Register has seen information that IBM’s Linux subsidiary is moderating its previously gung-ho AI stance by imposing limits on token spend. The Reg FOSS desk received a tip-off this week from a well placed company insider that a new edict has been sent down from Red Hat management to its legions of hard-working …

[Paul Herber]

"And no sharing your allowance!"

Share and Enjoy!

[Anonymous Coward]

This reminds of funny old country story. Dmitri he had only 3 potatoes for wife and 3 kids, I have 3 potatoes just me. I give Dmitri one potato. Dmitri and family next day execute by firing squad for steal potato. Is funny story because potato not my potato, potato I steal from Sergei when he not look. Sergei have many potatoes he no counted them. He not eat all potato.

Red Hat like home country for me. Is glorious mother company with red flag. I go drive taxi for Red Hat.

[FIA]

Seems par for the course I'm afraid.

Our place pushed AI like crazy... just long enough that we all forgot how to code... now we have token limits.... so we just tell our customers how good AI is instead... ;)

[Cris E]

We spent so much learning to AI code that we're going outsource our AI work to India to afford it. Or something. Leadership is in a bit of a muddle, but I can see my retirement from here so Dang the torpedos, AIm for that reef!

[Anonymous Coward]

Human are cheaper than AI … esp. the offshored and outsourced ones in Bangalore.

They’ll be getting rid of the automated looms and knitting machines and bringing the Luddite’s back next.

[IGotOut]

And don't forget...

...there WILL be more price rises to come.

[BoHu]

Re: And don't forget...

Yeah, especially as the "ROI calculator" (TFA link) tells us (in dark red letters) that anything above the $64/mo median (eg. $300/mo) already drags ROI below average and should prompt one to "Renegotiate pricing or confirm this cost is correct before scaling further" ...

I'd hazard that at $300/mo there's too much futzing around with the "self-futzing system" to prevent a right ROI capsize... (per the Glyph Lefkowitz TFA link)! ;)

[Paul Herber]

Re: And don't forget...

I'm currently paying $0.73/month for about 200GB of storage to a US cloud based provider (clue = Glacier) and I'm trying to find a way to reduce that to as close to zero as I can. So the chances of me paying anything for AI is ...

[Michael Strorm]

Re: And don't forget...

Bingo. Even with the price rises and restrictions that have caused a lot of fuss in recent months- and which seems to have been forced by skyrocketing use, particularly in the demanding AI/vibe-coding area- it's been estimated that most AI services are still being massively subsidised.

And that the real price they'd require to break even and make even a moderate profit as things stand will, in some cases, be several times that.

Of course, that's before they even start making back the money they invested in the first place... if they ever will.

Because I've said before that they can get people using this if they shove it in faces for free, or with massive subsidies, but it's open to question how many people will be willing to pay "full whack" for it.

Even if the strategy is to get companies dependent upon, and locked into, such services before saying "fuck you, pay me", that assumes there are enough companies out there *able* to do so. They might not care about customers going under if they can't, but they'll still need enough of them to survive to get paid what they want/need.

And the issue is that it's not clear that the market is- or ever was- big or rich enough to support that.

[David Hicklin]

Re: And don't forget...

> Of course, that's before they even start making back the money they invested in the first place... if they ever will.

is the plan not to float the AI company on the stock market just before it implodes so that the investors get their money back and all the mugs who buy the shares are left with nothing ??

[Anonymous Coward]

Run it locally

Download ollama and Gemma and run it locally.

And if you say “it’s not good enough”… …nor is the expensive version. So at least dont waste money!

[Liam Proven]

Re: Run it locally

> run it locally

https://bsky.app/profile/phineas.bsky.social/post/3m4qd6xyz5s2l

«

Using Generative AI?

You’re prompting with Hitler!

GenAI is a Fascist Project!

Try using your brain instead

Don’t surrender your creativity to the tech billionaire’s control

»

[Dan 55]

Re: Run it locally

Post is restricted to people who are logged in, so link for those without a Bluesky login.

Similar.

[that one in the corner]

Re: Run it locally

Fascinated by the weird biasses cropping up in those comments, such as people agreeing that

>> I love that it even has the piss yellow background as most AI images. Brilliant.

As most AI images? Which AI images would those be? Looks at the reams of AI junk that creeps into the corners of my web browsing, including, oh dear, a certain online news website: nope, not seeing any particular love of a yellow background.

Do you think they are referring to AI images that they have generated for themselves? Or are oblivious to the way that ads are targetted & personalised, so The Machine reckons they like a yellow image, and everyone else gets it in shades of blue?

[Dan 55]

Re: Run it locally

If you give an image to an AI and tell it to reproduce it, it will do (more or less - it could turn into nightmare fuel) but in a yellower hue.

Apparently many images are slightly yellow and have to be corrected in Photoshop.

[Anonymous Coward]

Re: Run it locally

That sounds like 99% of Photoshop work is safe then.

[FIA]

Re: Run it locally

Don’t surrender your creativity to the tech billionaire’s control

Erm... isn't having a box you can transcribe an idea into, which is then writ large in image form, empowering your creativity, not surrendering it?

You're probably surrendering your creative output, but as a tool to allow you to realise ideas without having to have a certain skill then surely it allows more people to realise their creativity?

Like how sampling let a lot of people that couldn't play an instrument write music?

(Mind you, maybe I'm being too naive, I remember this thing called 'photograpy' we used to have before Photoshop became a thing... go ask your parents...)

[Anonymous Coward]

We all (most of us) know that AI/ML is going to be very expensive but I do think they have started the price rises way too soon. Obviously they can't keep throwing money at it but it's already got to the point where companies cut back like this and some could even start to question whether outsourcing work to AI/ML is actually cost effective versus employing people. You could argue that it's great for finding bugs or security holes which it can be but once you've done that what's the ongoing use case for it? Why would anyone ask it to do something rather than show me how I can do something myself where possible? That's what I think anyway.

Maybe we'll get to the point in a decades time where people are begging for tokens on the street? "Spare any tokens guv'nor, I got some work to do to feed me wife and kids". That's a nice dystopian thought.

[Anonymous Coward]

You could argue that it's great for finding bugs or security holes which it can be but once you've done that what's the ongoing use case for it?

Finding more bugs and holes?

I think it's likely, especially in countries with an efficient way of improving their electrical grids and native populations not manipulated by foreign adversaries, that hardware co-design will cause the price of AI to significantly decrease.

Running Gemma 4 locally will result in a more predictable expenditure; however, it may not be cheaper nor result in the same productivity gains. My question is when, if ever, will the quality benefits of agentically written software outperform well-trained human teams of horses.

[Michael Strorm]

> "Obviously they can't keep throwing money at it but it's already got to the point where companies cut back like this and some could even start to question whether outsourcing work to AI/ML is actually cost effective versus employing people."

Yeah, but that was always the problem with the (supposed) strategy of subsidising the massive costs of gen AI in order to get people and companies hooked on and locked into using it.

Namely that you have to be not just willing but *able* to keep on doing so until you've achieved that aim. The problem being that they're not there yet, but are already clearly running into problems in that area.

That's a particular issue for the "pure AI" companies (those unlike- e.g.- Microsoft which has massive, profitable divisions elsewhere) which have always operated deeply in the red and been dependent upon a continued stream of investor cash contingent upon the assumption of future profits. But investors won't knowingly throw good money after bad, and there are going to be limits.

So it's likely that many companies- which were already shoving huge amounts into AI- know this, but are finally ramping up prices now not because they want to, but because they've been forced to.

(Ironically, this has probably been exacerbated by their success, particularly in the area of AI/vibe coding which seems to be one of the most demanding uses.)

[Fruit and Nutcase]

Cheech and Chong?

The contrast between the bot evangelism that we reported upon back in March is striking. It’s almost as if the optimism of “Chris & Ashesh”

Looks like whatever they were on in March has worn off and now they are stone cold sober.

Chris & Ashesh[Cheech and Chong]

[rgjnk]

If they think that's expensive...

...wait until they see what developers cost.

As long as it's productive and used effectively those spends are pretty tiny compared to just keeping a bum warming a seat.

I'm hardly a big advocate for throwing money at certain AI companies (all mine is run in house) but I do know what an hour of engineer time costs and if I can stretch some of that via some AI tokens that's a fair wedge of expense I'm balancing against.

[Dave K]

It's the typical approach. Sell it cheap/unlimited to encourage adoption, then once it's embedded into many companies workflows, you start to jack up the price and bring in the limitations. And as usual, too many companies were awed by the cheap shiny-shiny and didn't think to ask themselves what would happen when the price started going up and the limits came in.

Still, with any luck it'll put an end to companies gauging employee performance simply on account of how many tokens they consume. Maybe it'll even prompt the reverse - those that meet deliverables without consuming a shit-ton of tokens just might be the more skillful and productive developers...

[GoneFission]

Enshittification with a "captive audience" used to require a good product that people actually want to adopt & use; the initial low-barrier-to-entry is just an enticement. Then you could redistribute all efforts towards squeezing the juice out of your user base.

Then they found out you can simply force people to use your not-good product anyway and accomplish the same thing, as long as you promise the suits magical RoI numbers.

We (the collective humanist "we") could have done so many cool things with those resources instead, yet here we are.

[Tron]

A more flexible solution.

Cap your costs by offering devs a salary+tokens package. The more tokens they use, the less they are paid. The fewer tokens they use, the more they are paid.

I think GAFA are aware that they need to get people more hooked than they are, and they have endless amounts of cash, so it is odd to see any cutting back. But this is a leak, not a press release. They presumably don't want us proles to know that AI is not really viable at any level above abject addiction.

[druck]

Re: A more flexible solution.

Excellent idea.

[Liam Proven]

Re: A more flexible solution.

> offering devs a salary+tokens package

Seems to me that this punishes the workers, not the bosses. Is it just me?

[Anonymous Coward]

Reduced bodgets might be a good thing.

[Petalium]

Conclusion, the productivity gain is max 300/ month and doesn’t scale with increased use. If it was more profit with more use, there wouldn’t be a limit (unless the managers are completely financially illiterate)