Microsoft's GitHub grounds Copilot account sign-ups amid capacity crunch

Microsoft's GitHub has stopped accepting new Copilot individual subscriptions while the code hosting biz figures out how it can meet its service commitments without breaking the bank. The code locker has paused signups for GitHub Copilot Pro, Pro+, and Student plans, wrote Joe Binder, VP of product, in a blog post on Monday, …

[cyberdemon]

Get your short positions now

The AI bubble is ready to pop

(not that it'll help you, if your bank goes bust)

[Clausewitz4.1]

Re: Get your short positions now

” The AI bubble is ready to pop”

I am not sure, bcz I see real demand. It will all depend if people are willing to spend the new, much more expensive, costs associated with power comsumption.

[Doctor Syntax]

Re: Get your short positions now

It will all depend if people are willing able to afford to spend the new, much more expensive, costs associated with power consumption.

FTFY

That's the critical factor. If* the somewhat more realistic costs for users give them no return then they won't buy. At that point the whole AI supply chain ceases to be able to pretend that it could, itself, provide a return and that's the bubble popped.

* "If" is doing some pretty heavy lifting here by assuming the possibility that they might be able to afford it.

[Clausewitz4.1]

Re: Get your short positions now

” At that point the whole AI supply chain ceases to be able to pretend that it could, itself, provide a return and that's the bubble popped.”

I agree with your analysis. Lets see the behaviour in the next 2 months. By current datacenters project’s cancellations/delays, I believe you are right - 100%-200% estimated annual revenue growth seemed ilogical anyway.

[doublelayer]

Re: Get your short positions now

I think Clausewitz4.1 has a point relating to demand which may suggest the bubble has a bit more resilience in it. While Microsoft and a few others are starting to push the prices up, they haven't yet completely abandoned subsidizing things, just reduced it. At the same time, the people complaining haven't said they're going to cancel and try doing without, they're still mostly trying to make things work under different limits. That suggests that many of them are willing to pay somewhat more and will eventually do it when the providers give them no chance. It would take a while to figure out how much they have to pay to do what they were previously doing, likely at least a month or two, and by the time they decide whether they're going to do so, it's likely the providers will increase the prices again and start this cycle over again.

I think the bubble popping is going to happen when you have more people focusing on reducing their usage of AI rather than arguing over model choices or user caps. That's especially true because a lot of the load that the providers are making money from are from large companies which pay for token usage, so while noisy, the user subscriptions are relatively small, especially as providers keep screwing over those users by taking things away so it costs them less and less to meet their side of the constantly-changing arrangement. When you start seeing many companies that don't use per-user subscriptions complaining, that's a much stronger sign that they're growing unwilling or unable to keep putting up with the increases.

[Dan 55]

Re: Get your short positions now

If you can still activate Copilot Free on GitHub via dark UI patterns and have no way of turning it off except opening a support ticket then the bubble isn't going to pop yet.

[AtomicDog]

Re: Get your short positions now

The market correction that needs to happen is decentralisation - there are plenty of free models out there now that perform pretty well.

OpenAI/Microsoft/Claude start ramping up prices or diminishing capabilities to limit demand because they can't scale quickly enough.

They can't bring data centres online nearly fast enough to fix that problem, which means they'll not be able to utilise all the GPU capacity they've "committed" to buy.

That means GPU prices should hopefully drop to a sensible level and then the world and his dog that want to use AI but don't want to pay megabucks to the megacorps can buy their own hardware capable of running quality models, and then run them for just the cost of electricity.

It would get rid of the ridiculous need to build data centres and all the expensive infrastructure around them by spreading the load across the electricity grid.

Genuinely longer term I think they'd be better off developing and licensing their models and not serving them. Once things like the NVIDIA DGX Spark start becoming more mainstream and drop in price, AI at the edge will be the future. Not AI that the supplier decides you've used enough of today, thank you.

[Anonymous Coward]

Cry me a river.

[Pascal Monett]

Yeah, sure

"usage limits designed to maintain service reliability our monthly revenue"

There, FTFY.

[ecofeco]

Re: Yeah, sure

Fact.

[wpolarised]

One can't help but think the cracks are beginning to show in the fancypants AI future we were promised

[An_Old_Dog]

Two Questions

So, in addition to the large amount of resources it takes to run the AI itself, the agentic actions it creates also consumes many resources on the back-end.

Q1. Are these agentic actions doing a greater amount of useful, cost-justifiable work than directly-humanly-thought-out-and-created automations?

Q2. Is it less-expensive, in total (including human time costs) to automate via AI? (Just pushing the costs off onto someone else doesn't count here.)

[cschneid]

I am altering the deal. Pray I don't alter it any further.

Not sure what happens to all those companies that made so many of their programmers redundant in favor of drowning the remaining staff in work with Copilot as their floatation device. I _am_ sure I don't care.

[ecofeco]

Good time to repost this

Here's how M$hit has structured Copilot usage:

You don’t “have Copilot.” You have a stack:

Base license (required)

M365 Business Basic / Standard / Premium

M365 E3 / E5 (enterprise)

Then optional add-ons:

Microsoft 365 Copilot (main one)

Copilot for Business (same thing, SMB pricing variant)

Copilot Studio (build your own agents)

Copilot Credits (usage-based billing pool)

It's more of a structure problem and not a demand one.

[NewThought]

I see this problem resolving itself:

1. Users will realise that much of what they're using agents for can actually be done with standard code - so usage will moderate

2. Hardware is getting better suited to the task (e.g. Google's TPUs instead of graphics cards)

3. AI models are becoming more economical while still being just as good

[Dan 55]

3. AI models are becoming more economical while still being just as good

Are you sure? Claude has had a lobotomy sometime in the past month or two and Anthropic is raising costs. MS is raising Copilot costs.

[midnitet0ker]

Spoiler Alert: AI Is Cost-Prohibitive

I'm not the smartest man in the world but if AI companies are scaling back infrastructure while demand goes unmet, it suggests the fundamental business models are unsustainable.

This is probably because there's no practical way of making the numbers work: AI consumers will not/cannot afford to pay the actual cost of using their favorite models because they're actually ruinously expensive to operate.

I can't wait for the bubble to burst and Sam Altman asking if you want fries with your order.