AI's economics don't make sense(wheresyoured.at) Related: AI's biggest critic has lost the plot - https://news.ycombinator.com/item?id=47934353 |
AI's economics don't make sense(wheresyoured.at) Related: AI's biggest critic has lost the plot - https://news.ycombinator.com/item?id=47934353 |
I soured on him when he could not calculate cumulative revenue on an exponential curve, ignored everyone who showed him how to calculate it, and then kept writing that Anthropic’s revenue numbers are fake based on his inability to do math.
It’s too bad because any heavily hyped industry needs good critics (think Ida Tarbell to Rockefeller) but they should be honest critics, and he’s not, which really undermines not only his but others’ criticism of the industry.
Because, comparing vs GPUs
~16k–17k tokens/second per user
<1ms latency
10x power efficiency
20x cheaper production
Model to Si ~ 60 to 90 days
We have every reason to believe SW_to_Si will facilitate improving economics
[0]: https://www.wheresyoured.at/why-are-we-still-doing-this/
Economics Don't Make Sense.
I mean, seriously... our current late-stage capitalist economy is the chaotic sloshing of excess capital or inverted debt in a shallow tub within which clumsy giants are stamping like toddlers, and a parasitic kleptocratic oligarch class balances its efforts biting the toddler ankles in hope of more stamping judged advantageous, and, bagging what water they can.
A $20 subscription 2 years ago is not providing the same level of intelligence you're getting today.
Every major lab knows open source models are 6 months behind (See Google's "We have no moat") and none of them plan to make money on inference. Companies are subsidizing users to create moats that persist when models are essentially free for most everyday use.
That subscription was then and is now likely still subsidized.
The truth is that the AI companies are gambling that inference cost will continue following a hyper version of Moore's Law, e.g. Google TurboQuant.
The countervailing thesis is that frontier models are consuming more and more compute.
The deepest truth: you often don't need a frontier model to get commercially acceptable results from AI. Thus, bring on the true pricing! and I'll just switch models to something financially sustainable.
It seems he realizes he was wrong about that and has pivoted slowly to, "well, maybe they work sometimes, but the cost isn't justified." Which is a reasonable question! I just find his style of never admitting when he is wrong off putting and the way he presents things as absolute fact, when he's guessing like the rest of us. He was right about a lot, wrong about a lot, it's okay to admit that, I don't think his fan base would care.
Which is to say, it's easy to scapegoat this guy, but I think his approach is not any different from other "opinion piece" bloggers that we all tend to reshare.
internet rewards provocative takes - plural.
this mate has a single take and writes more about this one thing than jrr tolkien did in all his works combined
We need better critics of the industry.
Give the man a 2000 word budget and he could probably write a better article and cover the same information
There's plenty, but they don't have enough material to post once a week. And if you don't post once a week, you don't end up on HN once a month. As simple as that. Looking at the blogs that show up on HN regularly, the usual hit rate is 10-25%.
I often wonder if there are people promoting people like Zitron because they want the poor quality criticisms to be prominent enough to be the ones that they face most often. It must be a lot easier than having to address valid criticisms.
[0] - https://www.reddit.com/r/BetterOffline/comments/1p5zv33/why_...
1) Nearer-term investment returns on AI businesses and data center build-outs.
2) Claims that LLMs are now (or soon will) rapidly displace most/all senior positions in certain high-skill professions (eg software engineering, music/film making, etc), leading to less overall jobs for those kinds of workers and mass unemployment.
3) The "Foom" overnight takeoff hypothesis that AI will soon be able to iteratively sustain substantial self-improvement directly yielding profound new fundamental capabilities across infinite generations with no human involvement.
I've never thought that AI isn't already quite useful for some things today, or that no investors will ever make money on AI, or that AI won't displace some workers in some types of jobs, or that using AI isn't already helping accelerate the development of AI. Just that there's been a lot of hype, exaggeration and over-estimation about how much impact, how soon and how broad. There will be a few instances of rapid, large impacts but the majority of it will be slower, more gradual and less disruptive than extreme predictions - and many of the most over-the-top predictions may not ever happen. Not because they can't happen but probably for more mundane economic, logistic and human-factors reasons along the lines of why we're no closer today to the 1950s visions of a flying car in every driveway.
But they're still quite useful tools and accelerators or force-multipliers.
And you're still going to need humans in the loop.
And I'm very worried that the capex buildout will implode once we hit diminishing returns and good-enough models can be run on substantially smaller footprints.
It all isn't going away, though, and it will still continue to improve.
https://www.theargumentmag.com/p/ais-biggest-critic-has-lost...
We get hype bubbles. They’re (nearly?) always bigger than the thing they’re about, in a given time and place.
It’s reasonable to think the AI hype train is one of those, to some degree or another. It’s also reasonable to see great utility in llms, now and in the future.
Let's be fair here, the endgame is not "a few hundred bucks a month." Not for how much money has been invested. How much extra you have to spend to make developers how much more productive, and will companies go along with it is the trillion dollar question.
Maybe Ed is right even if he's wrong on some things?
This reminds me of a Bitfinexed blog situation. That guy researched and proved Tether token scam for years and he was right. But he didn't account for a tiny nuance - Tethers are useful for financial crime and are propped by that public regardless of the financial viability or rejection by every decent financial institution. Turns out you can have a hundred billion of unbacked tokens, if they are "alternatively backed" instead. I suspect LLM monsters may turn out the same way (or not).
Serious question - are there any LLM bubble critics with more sane and to the point style of writing and not just posting unsubstantiated hype for views like most on YT?
I have been pleasantly surprised at its utility knocking out grunt work. It's not super smart, but it's great at things like writing a python script to edit characteristics of a jsonl file or sorting structured data. I didn't ever expect it to be useful beyond extremely limited output and it's actually kinda good when you know how to narrowly target the tasks. The constraints of code make it a more suitable category than all the other stuff.
It's still a bs hype machine with Elon saying it might save all of humanity in court today. That's pretty unlikely.
How many tokens can you realistically burn through in one chat session? Opus and many other frontier models do maybe 60tok/s, less 250k/hr out. In you can use more, but in most cases cache is 5-10:1 cheaper than new input. Say you average 500ktok in, 90% cache, per request. That amounts to 100-150ktok in new input-equivalent costs, which in most cases is ~20-30ktok in output-equivalent costs. Do a request every minute, that's a total of about 1.5-2Mtok/hr. At API prices that's $50/hr for Opus, but really it probably only costs Anthropic $10/hr to serve that.
That said, even if a developer is burning $50/hr, many, many employees at large companies cost more than $100k/yr to employ all costs considered, so making them say 20-30% more productive can easily make that worth it for most. If the labs shave their margins ultimately to more like 20-30%, you'd have ~$15/hr in costs to use the services, and nearly every white collar job is way over 30k/yr to employ. If your salary is 80k, you probably cost the company 200k all in, so making you 15% more productive offsets the $15/hr cost.
So first party providers are not in a horrifying position or anything from a subsidization standpoint. The people in bad shape are Cursor and Perplexity, who don't have frontier models and are dependent on the open source community, which is typicly 6-12 months behind the frontier. They have to pay full freight API costs at 80% margin for the big boys to serve their harnesses, which is indeed untenable, and they'll have to either force users to use open source models and/or in house models they can serve at-cost or they will have to charge vastly more.
Gemini, Claude, and ChatGPT first-party services like Antigravity, Codex, and Claude Code are not in serious trouble though.
I just don't think that LLM business models can survive the allure of advertising dollars, any more than Search could, or TV, or Radio, or Movies. Ignoring the talk of copilot putting ads into pull requests, there is just no way that publicly hosted LLMs will not end up inserting ads into the output.
This looks like what I remember. https://freakonomics.com/podcast/is-google-getting-worse/
More seriously for software engineering it’ll just cost a lot.
> On an economic basis, a monthly subscription only makes sense with relatively static costs.
Running a data center is a fixed expense. Whether or not people use that data center to it's capacity doesn't change how much the operator pays (electricity use factors into this, since a GPU running at 100% will use more watts than an idle one, but it doesn't move the needle much on other fixed and variable costs of a data center).
> They also assumed, I imagine, that the cost of tokens would come down over time, versus what actually happened — while prices for some models might have come down, newer “reasoning” models burn way more tokens, which means the cost of inference has, somehow, gotten higher over time.
This is backwards. When the cost of something goes down, people use it more. This is basic supply and demand. Inference has gotten cheaper already, and will continue to do so.
Companies subsidizing costs for growth happens all the time. Yes, switching to usage-based pricing instead of subscriptions sucks for customers, but enterprises will continue to pay.
I wonder what the rough costs of a data center look like over the lifetime of one GPU generation?
10% building
60% GPU
30% power
I haven't gone looking for that information, but I haven't run across it either.
I pray this happens soon, but I feel I've been hearing some version of it for a while.
This tech has uses. It has quite a lot of them in fact. However there is no usage of ChatGPT or Claude that makes OpenAI or Anthropic worth anything fucking close to what they're valued at right now, and both firms are scrambling to figure out how to get down from the top of the AI house of cards without detonating in the process.
Meanwhile DeepSeek is coming out with more capable models that run on far less onerous hardware and with far less compute requirements that does basically exactly what the vast majority of users actually want it to do.
This is going to be a financial bloodbath. Not for anyone actually responsible for it, of course, they'll be fine. It'll be everyone else getting soaked which is the only reason I give two shits.
It's interesting to compare it to electricity. Basically Anthropic was selling a flat fee electricity subscription, and when someone started connecting expensive washing machines (OpenClaw) to their subscriptions, instead of changing the pricing model, they banned washing machines...
I wonder if we will get to "electricity" style pricing for AI. What makes electricity predictable is relatively constant average usage over time + price is manageable. I'm just not buying electrical house heating and manage my electricity spending within some bounds.
With AI the problem is that we are only now getting to useful AI, and for now it's still too expensive to be useful, so they subsidize until they can stabilize at "cheap enough and smart enough" level. But it feels like that's still 2 years away while they are stopping to subsidize now. Will be interesting.
No? It was flat, but with ambiguously stated limits (eg. 5x, 10x 20x). They were discriminating on how the "electricity" was used, but that's not that much different than how power companies have different rates for residential users vs industrial users.
Is this an actual issue aside from people letting their autonomous agents run overnight?
EZ might have incautiously and incorrectly called the peak several times, but his newsletter is nearly always stacked with citations and insights that, at least to my cursory but frequent inspection, pan out.
His argument(s) have evolved over time, but what of it? That just shows he's not the dogmatist the author wants him to be. Discourse evolves, get over it.
2026 Zitron has a good sense of the scale at which AI is requiring enormous financial complexity and volume to realize, and his basic point is that it isn't sustainable in the medium term.
He is self-evidently correct.
I disagree. It really reads as conclusion is fixed argument change as they are disproven.
I'm sorry but telling me that this is what AI can do is a sad state of affairs. Like this is google level stuff.
Also, I didn't read this whole thing, but I have yet to see Zitron respond to the strongest AI financials claim, which is that the models themselves are profitable on a life-cycle basis, even if the companies are not profitable on an annual basis due to capital expenditure. Dario made this claim exactly, and it more or less blows all of Zitron's financials arguments up.
They went from GPT 2 a text only, goldfish-esque memory at a 8th grade reading level to what we have today, GPT 5, multimodality + a token window encompassing a enclyopedia and a Doctorate/Masters level of mastery in major subjects.
The economics are probably betting on this exponential growth to continue, which if it fails, the cash would burn.
I am not sure if you would call claude code "an auto loop", but you don't need to be running something crazy like gas town to spend a lot of tokens with Claude.
If something is cheaper than alternatives, spending patterns change. People subsidize corn or power and so consumers alter behavior to take advantage of those prices.
Doubtless some people will reduce usage as a result. But Ed seems to find the idea that a 10 man developer team might spend 80K a year on tokens ridiculous. I don't understand this. Has he seen how much developers are paid? If you get a 20% productivity boost from coding agents, then that's two developers for 80K - effectively very good value.
Where things could go wrong is in comparison to cheaper models. If it's 5K a year for Qwen, and it's 2/3 as good will you pay 75K extra for Opus? Perhaps not.
Actually I think it'll be rare for a manager to be choosing between either a junior developer or a coding assistant, since each are going to benefit the team in very different ways and it'll often be obvious which you need.
What I mean is that at the price levels in the article the coding agent still had a realistic chance of positive ROI. People will pay for things with positive ROI.
> Don't attribute to malice what can be attributed to incompetence.
We're currently used to SAAS billing models that are either all-you-can-eat subscriptions, or metered around some easy-to-understand metric like # of users, or otherwise number of gigabytes consumed.
The SAAS economics work that way because the compute consumed is typically too cheap to meter. Some customer uses a little more than average, some customer uses a little less than average; it's not worth the time to even it out to the penny.
AI is so darn CPU (GPU? AIPU?) intense that will only be profitable, and affordable, if it can be metered like electricity and billed with a small margin.
In SAAS, we're not used to metering billing computations this way.
The internet seems to be saying that 70%+ of Anthropic revenue is per-token metered API, which would largely invalidate the article, but I can't find a solid source.
Customer: “I don’t want to pay more than $100/mo for my website” Developer: “What are your goals?” Customer: “1M daily visits, 1,000 monthly signups.”
And we've spent the past 25 years offering serverless compute, auto-scaling, pay-as-you-go for AWS and Internet infrastructure. And the economics are still a hard sell.
Companies don’t sell their goods/services at cost. A model’s being priced at, say, $30/M for output tokens doesn’t say anything about what it costs the company to provision the 1M tokens via the model.
And no, you cannot extrapolate from any company margins that someone may have overheard in an SF coffee shop onto individual product line margins or their trajectory either. This information is usually unknowable even in most SEC filings for public companies.
It’d be great if people who wrote these articles used, say, AI to look up some basics on how a business operates. It’s really easy to do, believe me ;)
And the funny thing is the estimate pure CAPEX Spend of AI companies needs them to earn about $20B to $40B a month to cover cost of capital alone of their trillion dollars of investments.
That's exactly what is going to happen. India/China prices will be $100-200/month, US prices will be $5000/month. Keep in mind that most of these costs will be covered by the employer. It'll put downward pressure on dev pay, of course.
This all becomes extremely visible when trying to do agentic coding with local language models - you quickly realize that controlling context length and model size is just as important as avoiding wasted effort. The real scam is not AI Q&A ala ChatGPT, that's actually quite viable - though marginally less so as conversations grow longer. It's agentic coding with SOTA models and huge contexts.
You can look at: https://sebastianraschka.com/llm-architecture-gallery/ and see how much things have changed.
I've used single digit billions in a couple days, FWIW.
That's with 86 interactions spread intermittently over a couple of hours so if I did a full working day like that I'd be looking at maybe 40 to 50 million.
Truth of the matter in most companies large enough is if you make your devs 30% more productive, then that'd mean 30% more code going through "change management" hell for months. You're not even paying to stand still, you're just pushing even more down a bottleneck. The price most people are willing to pay to make things worse is close to zero.
Do you perchance have a source for this? Is the profitability assessment comprehensive, including hardware amortization? I've found it hard to track down actual hard numbers for the cost of inference.
This seems to be the lynchpin of your argument.
It makes me wonder if I have been living under a rock, because I have never heard of frontier labs making money. AFAIK all AI firms are simply burning money to acquire customers at this stage. Is this wrong?
You're confusing the profit from the marginal token and overall profit (basically gross margin and operating margin). The comment you're replying to is calculating that AI labs are probably making a substantial profit per paid token. It's just that so far that profit has not been able to overcome the ongoing R&D and capex costs.
I do not understand how the companies can end up in positive, unless something fundamental changes
do you think per token prices will go up or down in the long term? will the price per task trend down or up?
what about the price of human labor?
Prices going up or down depends on what labs decide and what users demand. Strong models being profitable at lower prices than what frontier labs offer is a fact.
Nobody including the connected article is making the argument that this cannot be profitable ever. People are saying "there is no way this admittedly quite interesting tool is going to be able to make back all of this money" and I think they are completely right to say that.
You can absolutely make money with this stuff, just not at this scale. The buildout for this shit has been certifiably crazy and a number of the involved firms are overleveraged for tens and even hundreds of billions of dollars.
How in the sweet fuck are you paying that off, plus giving investors dividends, selling this at $15/hour/user??? That math does not math. A quick google says there are between 1.5 and 4.4 million developers in the US alone, let's say it's 5 million, to be generous, and each of them is subbed to this for 8 hours per day, continuously. That's 600 million per year in revenue. If you took ALL that revenue, and put it towards paying down this debt, not leaving any for employee salaries, upkeep, ongoing development, it would take DECADES to pay down what OpenAI already owes.
And yes I'm sticking directly to code, because that's the only thing I've seen it be really good at. Are we really proposing that every knowledge worker on earth and every manager of such workers is going to have an autonomous agent running all the time!? To do what, make sure they don't have to read or write email? Which even just that example is bringing in a fucking mess of legal, compliance, and security violations because LLMs are not intelligent and are not capable of being properly secured.
Like I'm sorry, I cannot take this industry seriously when even the most basic back-of-napkin math is saying, nay, screaming from the rooftops that they are FUCKED.
That math is not mathing. $15/hour/user, with 5M devs, 8hrs and 240 working days per year that is 144B in revenue.
Of course people don't work every day, but even with European-level holidays that number is off by a factor of 240 or so.
According to your math, that's $600 million per day
He does in this [0] article.
It doesn't even feel like particularly strange accounting to me. Aren't there plenty of companies that spend a lot in one year and realize the gains in the next year? If I build a house this year and sell it next year, the house was still profitable, even if next year I'm building 3 more houses to sell in the year after.
The TL;DR is that Dario likes to talk about imaginary/hypothetical companies a lot in interviews, and those companies' financials don't have a direct basis in reality.
But it's not that unusual, right? If I build a house this year and sell it next year, the house might still be profitable even if next year I'm building 3 more houses, so the company as a whole is still in the red on an annual basis.
I mean, I'm not a financial expert but that doesn't seem all that unusual to me.
Until they file an S1 to go public and show the world the books, take everything they say with a grain of salt. The amount of financial engineering going on in this space is astounding, and I'll believe it when I see an objective 3rd party release an audit confirming this claim.
The problem is that Zitron has charm, an authoritative voice and a very aggressive online presence. That's a difficult combination to compete against.
The landscape has changed a lot, but his content has remained mostly the same, maybe much more aggressive and less curious (in the beginning he would entertain other viewpoints more often), but since the tech itself has changed around him, so his repetive shtick falls a lot more flat than it used to because he is completely unwilling to entertain any other position other than the one that established his blog/show.
iHeart is so antiAI they added "Guaranteed Human" in the middle of every podcast they stream.
Does apple run additional ads on podcasts?
I think the real problem with that approach is you wouldn't be able to label the sponsored part, which I guess is a legal requirement in some places.
You can drive this car, but the last mile of this trip will use as much gas as the first 20 miles.
I think it’s in anthropics interest to keep this fact hidden from CEOs who push for ai adoption.
It's sad because such extreme, over-broad views presented as absolutes save AI zealots the trouble of creating straw men of skeptical positions. It's easier to just lump all AI skeptics together with Zitron and Marcus. I guess it's time to call myself something else, like maybe "AI Realist." My skepticism around AI has always been more specifically targeted to questioning more extreme claims about the degree of impact and how soon it will be meaningfully felt across broader society. I've also tried to be clear my concerns are centered on LLMs and not AI or machine learning in general.
My position regarding the long-term (5-10 yrs) has always acknowledged that LLM-based solutions will continue to improve substantially, find more real-world, meaningful use cases and that the currently unsustainable cost-to-value will eventually normalize to a sustainable equilibrium enabling profitable businesses (after some major financial pain); but, that LLMs as a technology still have some fundamental limits on what they can do which aren't separable from how they innately work. Practically, this means I doubt that LLMs, as one type of AI, can ever fully replace an experienced, highly-effective human's ability to self-develop fundamental new knowledge from novel contexts then reduce that learning to high-value abilities in applied practice and then iteratively build on that loop to discover entire new areas of knowledge which weren't even visible without the prior layer of new knowledge - and then do that over and over. I've never thought that goal is categorically impossible for AI, just that it will require a new and different approach beyond LLMs. While that new approach may incorporate LLMs as an essential component, just evolving, refining and expanding LLMs alone won't get us there. I'm encouraged that recently several top AI research luminaries have been saying similar things.
I used to be very bearish, my posting history indicates so - but not quite to the degree zitron did. I was just bearish on all the "developers will be gone by 2026" hype kind of thing, I adopted LLM use as soon as it was available and it has been in some part of my workflow since then, and I have watched it improve, so I took a similar view as you. The breaking point though, is sometime in the last 6 months I believe that something fundamentally changed and they became very good to use in almost every part of my workflow now, which I could not do before.
I dont know whether the skeptic psychosis is a thing innate to certain types of people, or if AI itself kind of encourages it, or a combination of both. Same with the other side of it, the people unrealistic about what it can do, or think it's become conscious, that kind of thing. I can see as I use it more myself, I am kind of becoming more optimistic than I probably should about what it can do, and then I stop myself, usually because something breaks.
But I've seen people, coworkers especially, go through that curve of "oh my god I'm gonna use it for everything" and then hit that "break," where it does something awful, and then jump right on over to the extreme like "this can never be useful." I wish people had more pragmatic views about it.
AI companies aren't selling coding tools. Claude Code is not a coding tool! It's a tool that does coding, which is subtly different. The total addressable market for a coding tool is all developers, which is maybe 25-30M people worldwide, the total addressable market for people who need code written is potentially around a few billion or so, maybe more.
Is the new data that models are more useful for coding than they once were?
Imagine being in 2003 and saying compute costs won’t go down. That’s Ed lol.
EDIT: Some quick research on this so you guys have actual numbers: https://gist.github.com/dwaltrip/a037be938d2b5ecc8b8b238736e....
There's multiple separate angles that all contribute to token-costs going down: chip improvements, engineering improvements for running inference in general, AI architecture and training advances that give similar intelligence in a smaller model, improvements in the quality of the training data, data center design / economies of scale, networking and rack-level improvements that are multiplicative with chip advancements, and so on...
If you analyze the situation for 5 minutes, it's blindingly obvious that price-per-token will continue to improve. And there's a very similar case for intelligence-per-token as well.
And don't get me wrong -- I have many concerns about how this is all unfolding and how it will impact society. But let's get our basic facts straight.
But if it helps, no, the data being discussed is surrounding the economics of running inference and R&D, nothing to do with the utility of models for coding.
That still feels incredibly optimistic given how split the community at large seems to be about how good this tech is, and it assumes all those developers also all work for firms large enough to pay for all of that.
However we are still very much in back of napkin math. We haven't even gone into what it costs to provide these services, how much it's going to cost yet for all these datacenters to be built, how much electricity and water they're going to rip through, their own employees and basic overhead, and all the rest. So IMO, we've now elevated it from "hopeless" to "this could work if a whole lot of other things line up really well."
I dunno I think about the language some people are using about AI investment and it is reminiscent of the many years where people were saying Amazon was a bad buy because they never turned a profit. Admittedly AI companies are investing more than the money they've already brought in, but I would be very hesitant to predict that it's all froth given the usefulness I've gleaned from the tools.
Don't get me wrong, I'm not unconcerned, but I think there are good reasons to suspect that at least some of the AI companies are making sound investments.
Sure, it's back of napkin math, and I also think that several of the companies we see today won't survive and/or will only survive due to consolidation, but I also think the spend is going to be immense.
With respect to the datacentres, I expect we'll see inference costs crash over the coming years - we're only seeing the beginning of what dedicated ASICs will do to inference, and what work to make models more efficient will do to the need for the very largest models, and while that might drive down the spend on individual subscriptions, I think it will drive up the total spend dramatically as cheaper models become capable enough to put them "everywhere".
But, yeah, ultimately we're guessing. I'm happy to put my guesses on the record, though, and look forward to look back and see how wrong I got it in a couple of years.
And the cost of not-quite-paid tokens.
What seems to actually be happening for white collar workers is that the price they can charge for their labor is dropping, but the price of their expenses (housing, food, gas) continues to rise.
Over a few centuries better tools and technology made it so that <5% of the population in rich countries are farmers. They use tools like million dollar harvesters.
Second, it seems a hard to believe that hundreds of billions of dollars would be spent and untold numbers of data centers would be built just to gain a measly couple hundred dollars per seat.
The second prong of the argument is basically that, when you invest in Anthropic, you can't just invest in one model and then collect the profits from that model. You're investing in a whole company in the hopes that they can be profitable overall; at some point they'll need to stop spending so much money on training and give it back to the investors instead. Zitron argues that this isn't going to happen because training is actually something that companies need to do to retain customers at all. An analogy here might be the fact that Microsoft has to spend a certain amount of "R&D" budget fixing security vulnerabilities in Windows Server just to retain their current customer base; if attackers found out about a serious security hole but Microsoft didn't fix it, everyone would need to stop using Windows Server. LLM companies do the same kind of thing to fix "jailbreaks" and other unexpected model behaviors.
The third prong of the argument is that, in general, there's a long history of companies using creative accounting to try and make themselves look profitable and then collapsing because they're not actually profitable. For example, WeWork's "community-adjusted EBIDTA" figured claimed the company was profitable using very similar arguments to Dario, and then the company went bankrupt. If you're already cooking the numbers, you have almost arbitrary flexibility to report whatever "margins" you want by excluding some of your costs from the calculation.
Construction companies capitalize and depreciate over many years so they can answer "yes" they are profitable even when they are very cashflow negative. This is exactly Dario's point: model training costs are treated as expenses but in practice are much closer to construction costs. Model training effectively produces an asset, the model weights, which will generate revenue for many years into the future.
> Zitron argues that this isn't going to happen because training is actually something that companies need to do to retain customers at all.
This is exactly why Dario's point about each training run being profitable is so important. It suggest that this is not true. Customers are happy to use old models long enough to fully pay off their costs.
> there's a long history of companies using creative accounting
Zitron seems to know very little about accounting evidenced by him using terms like "gross margin" wrong in this article. He's pattern matching against his limited exposure to company financials to find superficial similarities between the AI labs and famous frauds. Find me a company that doesn't report non-GAAP measures. Google search claims 96% of SP 500 companies do it. Are they all frauds too? Sometimes non-GAAP adjustments are eye roll inducing but they are tolerated because they can be genuinely useful to get a fuller picture of the business.
I guess I don't blame you or anybody for having a deal of cynicism, but these arguments just don't seem very concrete. Like, if Dario was lying or not, he probably wouldn't share the actual numbers, and he probably would propose a "model lifecycle" accounting. And if the business model had potential or not, there probably would still be vast investment in the next model. Zitron has had nothing but cynicism towards AI from the start, and it's his whole shtick, and so these arguments don't seem very credible coming from him, even though they seem reasonable coming from you.
Taking a look at the sources in your link, the MIRI's "Observations About LLM Inference Pricing"report [0] seems like one of the least biased ones (forgive me if I don't believe everything a16z has to say about the economics of AI).
Some choice quotes from the report:
"Imagine you went to the gas station and the price was $4.00, and you look across the street at another gas station and the price is $40.00 — that’s basically the situation we currently see with LLM inference."
"Overall, LLMs do not appear to be priced like other commodities."
"It's possible that some providers are slightly modifying a model that they are serving for inference, for example by quantizing some of the computation"
"Unfortunately, it is difficult to make strong conclusions about the underlying costs of LLM inference because prices range substantially across providers. The data used in this analysis is narrow, so I recommend against coming to strong conclusions solely on its basis."
Another source you linked, Don't Panic Labs[1], seems to agree with Zitron:
"It is a little unclear as to why the price per token is dropping, and I am still a little worried that the price per token will, at some point, go up."
"According to another graph at Epoch AI, the cost to train a model doubles every eight months. This tends to align with the common wisdom that we are getting a really good deal right now, while everyone is fighting to build market share."
If your inference costs come down due to quantization, that doesn't count, since you're cutting costs by offering a worse service, and there's only so much you can do that before your customers walk away. If your inference costs come down due to subsidization, that doesn't count either, since that obviously won't last forever. If your inference costs come down but your training costs double every eight months, that poses a significant problem for your business. If your argument to that is "training costs won't continue to increase at this rate forever". Well, inference costs won't continue to come down at this rate forever, either.
From what I can tell, there still isn't enough data to draw a strong conclusion either way.
[0]: https://techgov.intelligence.org/blog/observations-about-llm...
[1]: https://dontpaniclabs.com/blog/post/2025/12/02/the-price-per...
It's obviously a fuzzy scale.
In a free market like that it's not guaranteed that everything ends in monopoly. Actually mostly it won't. Monopolies that do occur are due to high costs of entry and are usually temporary.
Basically free market advocates fail to see more that a single step in the complex web of dependencies, which tries to prevent neo-feudal monopolization of everything by unchecked, unelected and being above most laws and taxes, robber barons.
I dislike unnecessary bureaucracy and excessive government control as much as anyone, I was born in the authoritarian USSR after all and I do study history. But I fear neo-feudalism even more. I certainly have zero self-delusions about being in a "ruling class" in that potential free market dystopia.
And the effect by which regulation actually strengthens incumbents and reduces competition is well known.
A common problem in these discussions is conflation of different goals. You talk about companies selling "poisonous shit". That's not a competition related goal so has nothing to do with anything I've been saying. It's an environmental goal. Governments often pass environmental law fully accepting that it will reduce competition and might strengthen or even create new incumbents - and they don't care! In fact most environmental law is like that because it's exactly as you say, other countries like China don't pass such laws and out-compete local firms as a consequence.
But that's not a failure of the free market. It's a failure of environmental law. Or, sometimes not even a failure, just a known tradeoff.
As a general rule it's hard to find markets that are controlled by monopolies over the long run without government regulation being to blame. Temporary monopolies can arise naturally and there's nothing wrong with that, but over time they usually fall by the wayside unless a law is preventing that from happening.