Excerpt:
The IPO Math Forces the Issue
Both OpenAI and Anthropic are on IPO timelines for the second half of 2026. OpenAI completed the largest private funding round in history in April, $122 billion at an $852 billion post-money valuation. Anthropic has reportedly surpassed $30 billion in annualized revenue. Massive numbers, both of them. Also both attached to companies that are still burning cash at extraordinary rates.
Public markets will not tolerate the gap between subscription revenue and compute cost that has defined the past three years. The moment either company files, analysts will demand unit economics that show a path to margin. Usage-based billing is the fastest way to demonstrate that path.
None of this contradicts the repricing thesis. The pricing war is the last land grab before the gate closes. Both companies are spending aggressively now to lock in users whose switching costs will make them sticky when prices rise. OpenAI offers two months free. Anthropic offers 50% more capacity. Both expire in July. What comes after July is the real pricing.
So first they get devs hooked on AI.
Then they watch and wait as dev skills degrade. Like, the actual skills to get the work done without AI.
Then, once devs are unable to be productive without AI anymore, they turn down the screws, pulling profit from those who cannot do without anymore.
It’s a tactic of capitalism that is as old as time. But I think it’s been done too quickly this time, as there are still plenty of devs which have sat on the sidelines, waiting for things to shake out, and who haven’t had their skills erode away from AI usage because they just haven’t been relying on it or even using it.
EEE all over again.
Problem is there is tons of competition and anyone with a video card can run models on their PC for the cost of electricity.
Thats why theyre trying to buy up all the hardware-
No where as good or fast. And somebody has to make the model.
How fast did their skills decrease? We didn’t really even start hearing much about aAI until about a year ago. It was all theoretical up until then. They should be able to level up their skills again pretty quickly.
This is article slop, please don’t make me read things that were written by bots, have no authors, feature ai generated images, citing themselves as a source… People would happily Ai-generate anti-ai blog posts and sell them to us all day long if we let them.
This stuff is from press releases, primary sources are not hard to find.
My first thought is that companies that built dependence on AI workflows will just start hiring humans again.
But I think it’s more likely they’ll just switch to free, self-hosted AI models.
The whole point of cloud was that renting wholesale infrastructure is cheaper and less hassle than building your own.
There’s no way anything but a sliver of dependent companies are capable of hosting their current LLM usage.
Nah, LLMs are small and easy to run
But I think it’s more likely they’ll just switch to free, self-hosted AI models.
Implying anyone left there will even have the know-how to set one up
Wow, you just convinced me to learn how to do something with AI.
It’ll be a very useful skill to help companies migrate from third party to self hosted AI in the next decade
Still not “free” even if self-hosted and I would think perf and accuracy would be noticeably worse.
My company started using AI in coding process a lot more now but its only because the AI responses are useful 80% of the time. If its lower than that, not worth the time to use for most tasks.
I don’t really think it matters. Companies that have adopted AI are already fucked. It hurts more then helps.
But they’ll be willing to thow VC money at hiring people to setup AI systems, despite how much worse it leaves them
Honestly: those who based a business model on this are probably bad business people.
Yup. And vendor lock in? As soon as prices go up, people will realize they can use Zed, Continue.dev, OpenWebUI, or any other FOSS system and point them at Claude or OpenAI or whatever APIs, lock in won’t matter.
I feel for the employees whose lives will be tumulted by their moronic management who went “all in”
If you think this will stop at just harming the employees of those companies, youre wrong
Well everyone graduates with their CS degree and thinks that not only will they be making $150k - $200k in their first job, they will save up so much money and be so in demand that it won’t matter that they get laid off.
The idea of people investing billions based on “annualized revenue” tells me that rich people have too much fucking money.
They’ve literally never had this much more, proportionally. If this isn’t too much, there’s no such thing.
Money is 100% wasted on the rich.
There’s a mass move to Chinese models now, which are like 10 times cheaper already, and will probably be about 50 times cheaper when the new pricing kicks in for Claude and Codex.
Can’t wait for the American AI companies to go bust.
Chinese model labs are still doing token (monthly) bundles, too. Core problem is that bundles only work when most people don’t use up the limit. US models have high cash burn and low margins even at high token prices.
Not to mention China could subsidise them just to pressure the US companies even more.
I wonder if they are timing it for the price increase to let them stick to the promotions that cost them more before pulling the rug. Could even be a win to open source them if the US government tries to block it.
Enshittification in 5… 4… 3…
it started shit.
Microsoft has already raised prices with their Github Copilot service. In some cases the new price is 27x higher. And the actually useful models are only available in the super expensive plan (and then also incur high fees for actually using it).
How do you enshittify shit?
AI… now with CoPilot! 🙃🙂
It’s the turducken of shit!
The headline seems way off from the content. It talks about planned IPOs as the basis, and somehow a “time bomb went off” “in real time”. What kind of chronology is that. I’ve read better AI generated stories than this dogshit click bait.
https://pivot-to-ai.com/2026/05/18/github-copilot-ai-token-charges-to-go-up-10x-100x/
https://x.com/Sugar_develop/status/2054323790327492851
https://x.com/abebeos/status/2054445525030654461
https://x.com/wogikaze/status/2054365603994255639
https://www.reddit.com/r/GithubCopilot/comments/1tbfkui/ill_just_leave_this_here/
How does this work when “good enough” AI like Deepseek V4, GLM and such are so dirt cheap they’re basically free for businesses? And available from tons of providers, or even self hostable?
And that’s only going to get exponentially more dramatic. Bitnet models + nonlinear attention + ASICs alone is at least an order of magnitude cost drop, though there’s tons of lower hanging fruit to pick.
AI is a race to the bottom, not the top. Intelligence has basically topped out, efficiency has not. It’s never going to replace most workers, but just assist them or automate dumb tasks, increasingly cheaply.
And Tech Bros can’t make trillions off of that.
How does this work when “good enough” AI like Deepseek V4, GLM and such are so dirt cheap they’re basically free for businesses? And available from tons of providers, or even self hostable?
Typically what separates enterprise-grade products and services from alternatives is a contract with an SLA… but that generally means there’s some contractual requirements for the reliability and productivity of the product or service. I’m not sure that any of the overhyped chatbots are reliable enough to support such contractual obligations, or that there’s a useful way to measure their productivity.
contract with an SLA
Plenty of hosters provide that. Cerebras, for example, fabs their own ASICs (seperate from Nvidia), builds them into servers, hosts a number of open-weights models themselves in friendly jurisdictions, and offers SLAs for enterprise clients; it doesn’t get more “guaranteed” than that in AI Land, but there are tons of hosts to choose from.
https://www.cerebras.ai/build-with-us
The major sticking point is that the best open weights models are Chinese. This doesn’t actually matter from a security standpoint anymore than buying a Chinese tire does; they’re dumb weights anyone can finetune, host and run on whatever software/hardware stack one wants… But try explaining that technical distinction of “using Chinese AI” to executives responsible for entire corporations.
There are even attempts to “launder” Chinese models to make them palatable for western enterprise use. For example:
Plenty of hosters provide that. Cerebras, for example, fabs their own ASICs (seperate from Nvidia), builds them into servers, hosts a number of open-weights models themselves in friendly jurisdictions, and offers SLAs for enterprise clients; it doesn’t get more “guaranteed” than that in AI Land, but there are tons of hosts to choose from.
This makes sense for first-party hardware businesses like Cerebras that are renting or selling their platform to developer businesses (second party) for the purpose of creating AI-based software tools which they will then sell as services to other businesses (third party), and I can see that guarantees could be written in a contract for the first-to-second-party relationship.
What I don’t see is that any such guarantees can be effectively written or enforced in a second-to-third-party contract, where an AI SaaS company is selling their software service to companies that don’t do their own development, and expect that the service they have contracted will produce reliable results.
Actually, what Cerebra’s does is no different than any generic host. They provide API access to LLM weights, though most providers will do it with some standard open source serving software like VLLM or SGLang.
And they all use the same open weights LLMs. They arent the software developer.
Cerebras doesn’t train their own model. And I think this is fine for service guarantees as long as the weights do not change, hence will provide the exact same deterministic results at zero temperature (and generally perform the same when used as a service).
My experience is that a lot of “enterprise” LLM stuff is used in bulk, for results that can be “good enough” with a reasonable error rate. Like (for example) extracting info from literally millions of documents. Or as RAG/querying their own internal documentation.
The rich have already cashed out and secured their fortunes
It’s the suckers, losers, poors and those who think they know better who are all going to lose everything
They’re going to try to stick taxpayers with the debt in a “too big to fail” way, and the current administration will help them.
Another classic example of …
CAPITALIZE PROFITS!
SOCIALIZE LOSSES!
When the system … any system fails, the owner class like to congratulate themselves on how they made it work so good … but when it fails, they’re the ones always groveling to the masses asking them to pay for everything.
This reminds me of the “pet rock” craze in the early 70’s. Everyone as getting them. They were everywhere. Myself, and a few others, never got one and never did. The whole “pet rock” thing came and went and passed us by. I plan to do the same for “AI”.
Man, comparing AI to a pet rock is crazy delusional.
Or was it allegorical? Ah, you’re right, let’s call it delusion. Who am I kidding?
You missed on a lot with a pet rock, my daughter discovered them just a few years ago and was completely in heaven from the idea. Perhaps, the similar with AI?
They’re both essentially chunks of metal that people treat like living things. Except pet rocks don’t fuck the environment and siphon energy off the power grid.
Dammit! After hearing this maybe I should have tried harder to grab that brass ring. Life holds some regret for us all, I guess.
The real pricing will come when companies can’t go back to regular workers, as predicted here: https://feddit.org/post/30017257
Apparently that post is gone?
The point was that prices will rise when companies are dependent on AI to the point that the owners don’t have money to pay for their AI girlfriends.
The title was “they’re ransoming my Al girlfriend back to me at exorbitant prices”. But it’s not only AI girlfriends. Once AI runs the company, the regular AI can also demand ransom prices.
Didn’t Amazon sail on while losing money for years as a public company? Why would super-hyped businesses like these have issues?
Adding more context to what phcorcoran said, AWS was something that actually had an “end goal”. A new company offering cloud servers where you can host your stuff to be reached by the internet at large, which was already a proven necessity even back in 2003.
AI is still trying to sell itself as something useful. Not only that, the fixed monthly cost makes zero sense, because tokens have an actual monetary value - there is a cost in processing, cooling, network, etc, which can be attached to it. You’d need an army of low-usage users to pay for the power users[1] to have it make any sense. The alternative is actually charging per token, like pay 10 dollars and get 1k tokens or whatever.
you know those f2p games where the players who use their credit cards are above all others? The free players’ value is in being the punching bag of the paying players. Now imagine the reverse: you need to recruit 20 paying players, who are ok to take a beating, to keep one unprofitable whale in the game. It makes no business sense, but it’s exactly what’s going on at the moment with their monthly rates. It’s no wonder every other week, AI users are reaching their monthly limit faster and faster ↩︎
According to Ed Zitron, AWS cost about $52 billion between 2003 and 2017 (adjusted for inflation), and became profitable after that. OpenAI did 13 funding rounds at a total of $180 billion and the signs are that they are burning through that capital at an accelerated pace and nowhere near profitable. The scale isn’t quite the same
https://www.wheresyoured.at/premium-ai-isnt-too-big-to-fail/
a lot of that article is about too big to fail (title, even! though its mostly at the end). We do need to worry about the last gasps of the empire clinging to the bubble by grasping for such bailouts.
This assumes that their main product is the AI that they sell to the world and not the AGI that they’re developing to replace humans.
AGI ain’t happening, it’s a pipe dream. The best that’s going to happen is less power consumption through more efficient chipsets. These tech billionaires are delusional. For anything technical “ai” is useless if the answers don’t already exist in the training data.
Yes. In addition: AGI requires situated knowledge and human-level interaction which involve difficult research problems that are still a long way from being solved. The techbros have been finding out that just throwing data at these problems doesn’t work.
AGI “can’t” happen like man can’t fly or 256 KB is enough memory. Will or can it happen? No one is sure. Will it happen this way, through plugging more compute and power into LLMs, that’s a definite no. But it’s the money grab right now, and anyone who didn’t follow the leaders would have been left behind with nothing.
So you’re right for this discussion, but in the broader sense we can’t say that for sure. One thing I’m sure about is how LLMs and the profit scramble for it have ruined actual research into the real thing.
Here’s my prediction. So-called “AGI” will not happen at any point in the life of anybody alive today (including people who were born while I was typing this message). See, thing is, we can’t even define intelligence in any meaningful way that has general agreement among the academic stakeholders (philosophers, neural scientists, cognitive scientists, etc.) and there isn’t any plausible line of inquiry that will change this in the pipeline that I can find.
And you think a bunch of techbrodudes are going to successfully make an artificial version of something that can’t even be successfully defined?
Yeah. Not gonna happen.
I think the most “foolproof” method of creating AGI would be to copy the structure of the human brain. Yes, it’s extremely complex and is unlikely to be the most efficient solution for a given intelligence level, but we know it works.
That’s not going to happen in anybody currently alive’s lifetime either. The best we can manage right now is the Cedars-Sinai model of a neuron (note the singular!) that approximates the electrical behaviour of a single human neuron. (There are huge biochemical signalling networks in brains that are a large part of brain function as well.)
We are nowhere near close to emulating a single human neuron down to the molecular level, complete with ion channels and all the other raw complexity of the beasts. We’ve barely begun working out how the biochemistry interacts with and modifies that electrical activity. And this is presupposing that the Orch-OR hypothesis (the one that posits essentially quantum computing in the neuron via microtubules and assorted mechanisms held within those) is wrong. If that’s correct, we’re even farther behind on emulating a human neuron.
That’s A human neuron.
We have about 86 billion or so neurons in the brain.
Oh, and wait! Neurons may not even be the whole picture! It’s turning out that we’re finding some “thought” happening outside of the neurons in the brain.
Did I mention that this is all the static structures of the brain? As little as we know about those, we know even less about the dynamic interaction of things. Biochemical weighting. Plasticity vs. rigidity. Dendrite pruning and regrowth. We know none of this yet, except that they exist.
Oopsie!
Just 800 billion more dollars bro, I promise I’ll make AGI by next semester bro, trust me bro, we’re totally on the cusp of it bro
