Over the last month I have seen companies scrambling to measure deliverables against cost. Most of the back room talk is to the affect of giving devs a small allowance ($500 a month) and then making them prove their own productivity increases (again, based on deliverables, not LoC) before they either take it away or give them more.

Obviously this won’t be on an individual basis but some kind of unit.

Either way, with how much I see these companies cutting back I have no idea how the big AI companies are going to be profitable.

For awhile it was every 2-3 years you'd start a hardware refresh. As companies moved into more and more training, this timeframe started to shrink. It went from 36 months to 24 months. From 24 months to around 16-18 months. Last I checked last year, it was at 12 months. I think things may have slowed because of component availability, but otherwise whole data centers would be 6-12 months into full operations before they would start a refresh cycle.

Not to mention the massive increase in power density demand and cooling demand per rack that entails.

So no, "AI costs" have not gone down, in fact they are more expensive on training AND inference than ever.

This is why many are concerned about the heroin drip of api costs into orgs. For the companies that are public, look into their financials. It's gonna hit companies and high volume users like a ton of bricks.

I work at a Fortune 200 company. At first, it was the Wild West. Need an LLM? You got it. Need to or want to build an army of agents? Done and done. We literally had everything at the tips of fingers for about 3 months. Teams were building their own internal tools, the team I work on canceled contracts with several software vendors because teams were building the same tools for what they thought was nothing.

Then they signed contracts with Anthropic and Google because I would assume they saw the token usage was through the roof. One month later? They completely cut off access to everybody for both Claude and Gemini. If you wanted access? Suddenly it was several forms, along with several approvals and a rock solid business case why you needed it. And before you got to the forms? You were added to a waiting list that was thousands of people long.

The entire company is now in damage control after trying to get the genie back in the bottle. I'm guessing someone saw how much we would be paying for the tokens we'd been using and decided to shut the party down so to speak.

The math doesn’t math.

I think you forgot what super-intelligence means…

Btw, some Chinese corporates have already seen this and increased their price. Zhipu AI & Tencent for example. Alibaba, Baidu, and Tencent also announced multiple price increases for their AI services.

- if AI costs go down you can ask how the companies will make profit and then suggest the bubble popping

- if AI costs go up you can ask how people will afford it and then suggest the bubble popping

- if companies actually do make profit then you can say the companies are getting too big and powerful so it’s a bad thing for consumers

Essentially you have left zero to a small narrow path where you are happy with the outcomes.

Likewise, the quality of what I can get from a local model like Qwen 3.6 on an RTX 5090 is light years ahead of what I could get a year ago on the same hardware.

Once moat is achieved, you don't have to compete on price. Of course it'll be academic because the AI will probably destroy all of us.

And, even with the price increases, Z.ai and Tencent are still much cheaper than Anthropic or OpenAI models. I think there's an efficiency focus among the Chinese models that is absent at OpenAI and Anthropic, and in the end I suspect efficiency will be the winning feature. Google seems to understand that. Gemini 3.5 Flash is pretty competitive with the big guys, and it's small enough for Google to run it profitably (I assume) for a price that's much less than the frontier models. Gemma 4 models are showing off a bunch of efficiency techniques (MTP, QAT, the 12B encoder-less vision model that soundly outperforms much larger vision models, DiffusionGemma), and I assume they have several more techniques that aren't published.

If apparently the only way you can make money with your product this early is to dilute and adulterate it behind the scenes, it strongly suggests you want the customer to continue to believe they are getting value that you can't afford to supply.

More prosaically: if either of these firms could prove that they were even really close to profitable on inference, they would have bloomin' said so while they were trying to raise more money.

I would assume when price hikes happen either 1) less non technical people would vibecode as it doesnt impact the work that much 2) people use the cheaper chinese models 3)we're jamming ai into everything because were exploring. We will just niche down into use cases that provide high roi

Only reason deepseek is so cheap is because well I don't know, but actual pricing should be around their initial price which was 4x, at that price you have a healthy 25-50% margin based on occupancy, given the deepseek v4 is a very sparse moe model.

GLM 5.2 for example doesn't have more than 30-50% margins that's assuming old pricing for GPUs, current inflated GPU pricing well I am certain the margins must be lower. Ofc you can host for cheaper with quantization, and if you have very consistent capacity/utilization, which is not the norm with AI workloads.

Overall for large models like GPT 5.5 or Opus there must be healthier margins of around 50-70% assuming GPU pricing didn't increase for these companies. Even if it did 30-40% margin should be possible, even in worst case assuming all GPU they had saw a jump in pricing.

For smaller models it's hard to say, I would guess 20% but these models might be much smaller than I suspect, then it might be double that.

Note the issue is less intelligent tokens don't linearly scale down in memory usage, which is the biggest pain point of serving models. Context sizes have fucked us all.

Also anyone claiming OAI makes less margins on APIs or stuff might be wrong given they are on much lower context size, 1M context definitely is a lot more expensive to serve especially with smaller models like sonnet.

  > Neither Anthropic nor OpenAI allow Business nor Enterprise customers access to the high value $200/mo plan. 

Please tell more :). Do you pay per token from bedrock / openrouter / somewhere else? How many tokens you use over the month, and how many for each task? Which harnesses?

The big push for regulation and export controls is only going to ensure OpenAI & Anthropic are more like the automakers. Only in business because of protectionism, left to screw over US consumers meanwhile the rest of the world gets to enjoy cheap EVs

How do you know that the other models you are referring to aren't subsidized?

This is the crisis point for vibe-coders. A developer can go back to writing code by hand, as horrible as that might sound. Someone who hasn't learned to code but builds with AI can't go back. They either pay or they stop. That will be an painful choice whichever way you fall.

If true then why are neither Anthropic or OpenAI dropping their API pricing to gain market share when both are clearly doing all sorts of political and PR maneuvering to compete in a cutthroat market?

Since they aren't dropping the API usage prices (and are in fact raising them in a lot of subtle ways) then one of these options almost has to be true: they are still subsidizing inference, training costs are so ridiculously high that they need to make huge profits off inference or collapse in on themselves, or they are price fixing.

Here's a concrete example. Does some random AI company make operating profit on inference? I.e. if you only kept marginal costs, would you make a profit?

Well, depends what you account as your costs. If you're using hand-me-down hardware from previous generation's training, how much do you charge yourself internally for it? Maybe you show less, so investors take solace in profitable inference, even if you're losing money overall. How exactly are you accounting for electricity costs between training and inference? Is your army of SREs mostly servicing training new models (R&D expenditure) or inference (operating cost)?

This even has a name, and is called the "big bath" approach. If investors expect one part of your business to be a fiscal black hole, just shove all your costs there. They are accepting of it, and you make the rest of the business look better.

I'm not accusing AI companies of cooking the books, rather I'm trying to highlight you could see all the cash flows and still not know how much money is made or lost where.

Eventually the frontier labs will try to cut out the middle man once these models prove themselves and start doing partnerships with big firms in the domains, so they can take a % of the profits in perpetuity rather than just taking a one time payment. For example, after Anthropic Galen, they'll do a partnership with Pfizer to generate Ozempic-Superjacked and take 20% royalties on global sales.

Having growth up in the 90s, it is weird seeing companies share their technology secrets publicly.

If you zoom out to the year 2100, it becomes a little pimple on the economy that is ready to pop, but in the here and now it can cause a lot of damage to real people's wages and finances over the next 3 years.

But next year we could be in the middle of a massive $600B/yr capital-spending bubble deflating hard with unemployment accelerating towards 10% (or higher).

The internet never failed, but the telcom/dotcom collapse still happened in 2001.

I didn't get the sense this was LLM-written, but typo-signalling is... I donno a bit weird. Firefox is underlining some of the words as I write. I'm leaving "donno" unchanged even though it's flagging it as a misspelling but I suppose I'd still opt to fix something like "maiinstream" even at the risk of potentially seeming more LLM-ish!

Cheap, but gave them a massive user base they can claim is using AI

We have a pretty good idea of how much it costs to serve these models. You can pencil out the economics and guess at the model sizes and we know pretty decently how expensive the hardware is.

This like claiming it's meaningless to guess the margins of a restaurant without going into their books and seeing the exact recipets and recipes.

They ain't doing dark arts in the back. You can guess at what goes into the food based on similar recipies and how much that costs based on what you pay at the grocery store.

The funniest comment here. Have you seen the prices of the technical shit for the past two years? Dang, GPUs are not getting any cheaper, but more expensive with each year.

As a localLLM evangelist, I am hopeful this will bring more attention to the joys of rolling your own sovereign AI.

Microsoft adding Deepseek support already as I recall?

That is - for any definition of "they are behind X months" then eventually they get to the point Claude was in January when the world freaked out, but at 1/10th the cost. A lot of firms are going to mandate that is good enough for their developers.

Do these knowledge jobs have a significant corpus of not only knowledge but discussion and problem solving, all conveniently labelled for the AI to train on? Probably not. Coding has stack overflow, what does, say, advertising use?

If you decided to boycott every company that replaced staff with automation, you would be forced to exit the economy. Every company does this to some degree and the customers who vote with their wallet do not seem to care about a reduction in force.

[1]: https://arstechnica.com/ai/2026/06/gm-installs-robots-at-fla...

Pay for OpenAI Pro directly, but I’m the only guy that uses Codex. $100 a month. My nontechnical partner likes to talk to ChatGPT 5.5 Pro for image related tasks (think generating interior decorating pics).

The nontechnical staff use a Gemini account on a Google family AI Pro sub. I use Antigravity when working on Android or Google Cloud API codebases.

Everyone gets OpenCode Go. The cost is trivial. $10 a month per person.

Pay for MiMo directly. We use it during Chinese off peak hours though. Total spend so far $25 in last month.

We run a few Qwen models locally and pretty much have them pegged all day. RTX 5090 on a PC and a Mac Studio.

There’s also Grok which is used for Imagine for artistic / graphic design related work. I also use the subscription for a vision model in my oh-my-pi harness.

We’re having discussions about how to pull in GLM-5.2 cost effectively. We compete with third world development shops so we can’t really pass on inference costs, but we can benefit from getting jobs done for customers faster. But ⅔ of our work is either internal or open source projects we can’t bill for.

The market for open weight model hosting gives you an idea of the profitable price floor, it's pretty clear there's markup baked into OAI/Anthropic's APIs.

They are? In the before times of 2025, Opus 4.1 was $75 per million tokens. Opus 4.8 is $25, and Fable is/was $50.

And it does, nowadays, give you a bit of a veneer of mere curiosity when you're being accused of massive theft.

This is a delusional take. Sorry, but anyone claiming this hasn't used Fable and compared it to the current best open source models. I see a lot of hype posting about GLM5.2. I see absolutely ZERO people using it in production compared to GPT 5.5 or Opus 4.8.

All depends on who is holding the bag, and how big the bag is.

I do hope that a day will come where you can buy the nvidia spark thingy for 5k that can run the equivalent of Opus 4.6 or 4.5 locally and that would be a massive thing.

There isn't one AI intelligence S curve, there are thousands of them, and they're mostly invisible in the major benchmarks, but for someone trying to do work in that specific area of capability, the progress is transformative.

Edit: to the commenter below . It was widely reported that these companies were unprofitable 1 from last year. I am asking question to this specefic comment because they made a very specific claim about part of plan thats profitable . something only an insider would know.

1. https://www.wsj.com/tech/ai/openai-anthropic-profitability-e...

Considering how much they spend on sales, marketing and R&D that doesn't sound that absurd

The people have a right to make and use whatever models they want, protected by the constitution. At a minimum, the models are described in research papers that are unquestionably protected speech. Skilled devs turn those into programs, also protected speech.

You are way too deep in the HN bubble.

It does remind me of the time a chef told me when he puts lemon juice over a dish, he would intentionally not remove any seeds that went on it because it was a signal of quality. I wonder if future slop chefs will intentionally place seeds on dishes that came from a box...

I'm actually curious if this works, haven't tried but I assume it would.

And of course the C-suite will have unlimited access to Mythos tier models, which they'll use to summarize reports, while passing down mandates to rank and file to increase usage of less expensive models.

Advertising has centuries of print ads, 100 years of radio advertising, 70 years of TV commercials, etc. And modern AI does not necessarily need labeling.

The same is not true for the software industry execs.

I worked at Verizon during their layoffs last year. Biggest layoffs in the USA.

As someone who’s been laid off before, I knew that it generally boosts the stock price.

I bought VZ because of that. It’s up 15% since the layoffs.

Microsoft, an AI stock, is down 30% in the same timeframe.

I believe this hasn't been confirmed yet but I think it speaks to a bigger problem for the AI companies which is, if you give capable developers a good reasoning LLM, they can make it work like it was a really expensive model.

I believe we are 100% at the stage of good enough for the vast majority of tech companines. Fable and others will be more valuable for non-traditional tech companies.

I read somewhere that the chinese AI companies are sharing knowledge and it would not surprise me if the government is applying pressure by saying work together or else. If they work together, they can truly commoditize LLMs and with China ramping up hardware support for AI, I see the future being inference speed and hardware being the moat.

Maybe I should be aiming for something targeting 48gb of memory?

OpenRouter is the best guide to real costs.

That’s usually a sign that sales are not “just fine”.

Due to the fact that we’ve already done this before (Enron, Global Crossing) -

I’m willing to bet that there are contracts in place ALREADY, that define what happens in the event of a default.

In particular, I’ll bet that the buildings, the GPUs, the patents, etc…

All of these have probably been accounted for.

I worked at a data center that closed during the WorldCom era, and when they put the padlocks on the door, there were still websites “hosted” from the building.

I don’t know if they killed the power or what. I’d cleared out my desk long before they locked it all up. I wouldn’t be surprised to learn that these websites couldn’t get their own servers, since ownership was tied up in the courts.

In the Bay Area during that time, there were row upon row of empty office buildings.

A wealth transfer from the working class to a handful of billionaires bigger than any the world has ever seen (and the world has seen a lot of wealth transfer from the working class to billionaires).

If AI was around in the early 2000s Countrywide.ai would have been a thing.

I don't see how.

So depending on how literally we interpret Darios comment, OpenAI & Anthropic need to get to Apple+Google+Meta revenue numbers in like single digit years?

The banks aren't has exposed this time, as in 2008, most of it is tied up in private credit, its more akin to the fiber buildout in the 90s.

I know because I see how people went over the 4o model. I can see opus behaving clearly differently enough that I pick it for certain tasks.

How?

* Moores Law is almost over. The 5090 improves over the 4090 mostly because of quant improvements.

* even if the hardware improves, there’s a huge incentive to slow roll the next generation. Nobody wants to end up like Sun Microsystems. Sun’s used hardware was faster than its new hardware, once you considered price. Sun ended up competing with its own used equipment.

The most obvious place for improvement is RAM, network and storage.

If someone can bring more RAM onto the market, that will unstick things.

Maybe you're somehow legally allowed to distribute and download the weights, but most of us can't run GLM 5.2 at home.

And much more informative than the speculation and guessing in the article.

Myself and several other devs were laughing about the whole thing. The company was so amped about what AI could do they never even bothered collecting any analytics that would affirm or deny any of this had a positive impact. Even some of my team members were talking about the placebo effect AI has had on a lot of C-Suite folks.

https://carteakey.dev/blog/local-inference/local-llm-optimiz...

https://botmonster.com/ai/self-hosted-ai-agent-frameworks-20...

Personally I find myself swapping models depending if I am engaged in “trad-development” vs building agentic probes or apps involving imagery. Tailscale the LLM to your deployments and ta-da!

Which makes sense to me. Selling a chatbot interface/model access to the general public was never going to be a viable long term play. You still need developers to wrap the models into specialized tools. Queue the Jobs quote "It's a feature, not a product."

There is significant room to make more specialized neural network accelerators with new compute-in-memory architectures.

If the brain can run 86 billion neurons on 30W it must be possible.

  > but they do have the power to constrain commerce

Otherwise I don't see the comparison.

If I'm intelligent enough to use a tool, but I don't have the tool, that doesn't mean anyone who does have the tool is automatically more intelligent than me.

Likewise, comparing my performance without the tool against someone's performance with the tool wouldn't be benchmarking their performance, only benchmarking them with the tool's performance. The fairer comparison would be against me also with the tool.

This is the video I watched that explained the shenanigans (from the guests' perspective, not illegal, obfuscated)

https://www.youtube.com/watch?v=YrJzjC4kKCY

Certainly, the best models have gotten better since then, but I wouldn't consider DeepSeek V4 Pro or GLM 5.2 to be a big enough downgrade to be worse than coding by hand. I'm willing to spend a premium for the best model for coding because it wastes less of my time with dumb stuff, so I've got a Claude subscription. But, there is a limit to how much of a premium I'll pay. 10x over Chinese models? OK, fine. Opus saves me enough time to make it worth a couple hundred bucks a month. But, 100x, or more? Nah. I'll go a little slower, review the PRs a little more carefully.

And, open weights models do keep improving. DeepSeek V4 Pro is a notable improvement over earlier DeepSeek models, and the first DeepSeek model to cross the "better to work with it than without it" threshold into Opus 4.5 (or better) territory. GLM 5.2 is somewhere in the ballpark of Opus 4.6 (though without vision, a notable limitation for anything that requires a UI).

I can manage this budget with the chinese models in AWS BedRock. However, in my experience, they aren't as good as claude today.

And.. now I feel the need to look again. Darn, there goes my afternoon

I built my career on Solaris and it got rugpulled by Linux.

That wasn’t because of software, it was because of hardware. Linux’s cost advantage existed because Sun hardware had huge margins, because their software was basically free.

AI will probably be a repeat of this. Whoever can come up with the hardware solution that minimizes the cost per token will win.

I believe the 5090 still holds this crown, but someone certainly knows better than I do.