I get the structural comparison they are trying to make.
But mortgages are not a frontier AI lab.
They try to draw a comparison to the valuation of the real estate and the valuation of the hyper scalers in the markets.
I would argue that the demand and valuation of a house is less elastic than AI. While a house’s value may continue to appreciate in the market there is an upper bound for the price of a house set by people’s income. We don’t know yet what the value of AI is. The underlying product, the model keeps improving and therefore increases its value. A house is still fundamentally a house a year later and doesn’t intrinsically appreciate in value.
From gpt-3 to gpt-5.5 there’s been a massive change in the underlying value of the product and company in a way that simply doesn’t happen with a house. That’s where the analogy breaks down.
Funny how you're echoing exactly what the author said in the article
> This is not precisely 2008. GPUs are not houses; take-or-pay contracts are not mortgage-backed securities; OpenAI is not a subprime borrower in Stockton, and artificial intelligence may well be the most consequential technology of the century, which is more than anyone could ever say for a McMansion in the Inland Empire.
> The bear case in this piece is not that artificial intelligence will fail, or that the demand is fake, or that the technology disappoints. It is narrower: that the financing structure can break before the demand arrives, because the obligations are fixed and front-loaded in commencement while the revenue is variable and back-loaded in adoption - and a fixed obligation meeting a lagging revenue stream is a solvency problem regardless of how transformative the underlying technology turns out to be.
> The industry will spend the next eighteen months debating whether artificial intelligence is a bubble, which is the wrong question, asked at the wrong layer. The technology is real; so were the houses. The question is narrower: what happens when instruments underwritten at the teaser meet their reset schedule, and who is holding the paper when the obligations cannot be met as written
Ie, if you spent $10M buying a house, it doesn't matter if it will be worth $100M in the future. If you're unable to make your mortgage payments in the interim, you're going to lose everything
"The underlying product, the model keeps improving and therefore increases its value."
That's not exactly true. Yes, the fundamental capabilities of the models do seem to be growing dramatically, but the economic value of any particular model may be steady, or even falling, because of commoditization, or other issues external to the model itself.
Without a moat, improvement in model capability does not necessarily translate into economic value--and the labs need economic value to pay their obligations.
The economic value may be real but the profits may not be.
One thing that's clear is that there is no leading vendor in this space and there may never be one. To some extent premium models can charge a premium price but it's going to be a competitive market and the likes of Anthropic and OpenAI will not be able to sustain monopoly pricing.
100%. I have sometimes wondered if China is playing the long game with the open weight models trying to tank the margins of the US labs so these profits don’t materialize and the US economy (currently predicated on the net that they will) suffers. I think this is a coherent strategy beyond just “don’t let the US control AI as a strategic asset.”
The AI debate has often been flattened into "Do you believe in the long term viability of the tech?" when there is also another question that needs to be asked in "Do you believe in the long term viability of these companies' business models?" It's a lot easier to believe the former than the latter.
It's entirely possible for this tech to be humanity altering in the long term while we are also in a huge bubble that could pop at any moment. In that way the housing analogy is apt. The utility of the houses themselves didn't change, the problem was purely with financial markets until eventually those markets made it everyone's problem.
Notably a lot of those companies went down, a lot of investors lost money, but the internet and e-commerce proved to be just as big as people expected back then, maybe even bigger.
That is much less of a systemic issue though. It could kill OpenAI or Anthropic, but it wouldn't render these large GPU data centers obsolete since whatever replaces them, be it open models or cheaper closed models, would probably still require a lot of GPU compute.
Less bad, yes. But still bad. There will definitely be demand for the compute, the question is whether it will be enough to keep the value of it at the levels you’d expect if the tenants were in a monopoly/duopoly world. Competition puts downward pressure on margins and that’d translate into pressure on data center leases.
"probably still requires a lot of GPU compute" is something that would be very desirable to undercut.
There's a galaxy of potential AI markets that are local-only for compliance/privacy/deployment environment reasons, and an equally large and overlapping set of use cases where you won't be able to bolt a rack of thousand-watt GPUs on the side of the device.
What if the next "DeepSeek shock" is something we can run for non-toy use cases on our box of old Android phones? The GPU data centres could be expensive albatrosses very quickly.
There's probably a case that right now, the bigger-is-better paradigm protects incumbents-- a smaller model will always have a FOMO factor unless it can be proven competitive, so that leaves the playing field to those who can afford to train and deploy a new Fable or Sol every few months.
The dynamics are interesting. If all businesses get productivity increases from AI, the margins they could have claimed are competed away. The model companies also have their margins competed away because of open models. The only companies that have a moat are the ones with capital as a barrier to entry and even then there is cut throat competition.
We might end up with massive consumer surplus from AI because no business will be able to raise prices due to competition. This is why it's so important that we don't allow for regulatory capture in this space.
> The underlying product, the model keeps improving and therefore increases its value.
I think this is true, but a customer's willingness to spend is based on _perceieved_ value, not actual value. For many companies, the _perceived_ value of AI has been trending down as internal projects fail and cost skyrocket, even as models on paper improve.
The value is going up, but the pricing is going down, right? At least at the token level. So the usage would have to go up dramatically to compensate for that.
> there is an upper bound for the price of a house set by people’s income
Isn't that essentially true here too? The money to pay these expected future AI prices is coming from someone's income. Sure, the pie will be growing at the same time, but enough?
Individuals and families buy houses(ideally). Entities like corporations buy work, often intelligent work. If they can crank up outputs and profits via more intelligent work done by models, I think the ceiling is global demand for the entity's product. Which is still a bound, and ultimately set by individual consumers.
The question is whether the consumers will have the income to spend. So I kind of agree in the end I guess.
> We don’t know yet what the value of AI is. The underlying product, the model keeps improving and therefore increases its value.
i dont know. i think we already know how far these things go. Buying tons of data on mercor to slightly improve one domain has not really even displaced ppl in that domain. i really cant tell the difference between opus 4.8 and 5
very few domains in the world are closed like math.
I mostly agree but at the same time the change in underlying value from opus4.6 to Fable has not been as dramatic, and I'm not really sure if a "better Fable" is something most people even need. To the point where a faster and more cost effective model is preferable
Many would argue that opus5 is a regression in value despite what benchmarks say.
1. The massive amounts of GPUs being purchased have far shorter valuable lifespans than a house.
2. A model's value seems to be depreciating at an unbelievable rate. The most expensive top SOTA models (GPT-5, Opus 4.1) a year ago are far less capable than GPT-5.6 Luna. Compared to when those models were new, Luna costs 85% less than GPT-5 and 98% less than Opus 4.1. That's good for us consumers, but if a lab stumbles for 6-12 months, a lot of their value goes away. Especially with open models only months behind the SOTA closed models.
Even if it is 5 or 7 years, houses last decades. Even cars last longer than GPUs. GPUs are a highly depreciating asset. They'll last longer than 3 years, but might not be very cost effective (tokens per Wh) compared to the latest AI accelerators that are available.
I wonder if it will be that the U.S. (and allied countries) will be using the latest 1-2 generations of AI accelerators, and if the 3-5+ year old stuff will be sold to Chinese datacenters since that will still be the most powerful tech we'll be allowed to export to them?
You pointed out correctly that house prices are bounded by incomes, then immediately made the same mistake with AI - someone has to want to pay for this stuff. It also has to compete against free models, which are not only almost as good, but they pull ahead sometimes.
Saying “the product increased in value” is only true if someone buys it!
I'm still paying OpenAI $20/month for a product that has gotten massively more valuable. That's the problem. They aren't getting any more money from me for a product which is much more valuable.
GPUs depreciate at a far, far faster rate than houses do. And then they need to be replaced. Who's paying for that? OpenAI and Anthropic don't even have positive free cash flow.
I think its well corrected by the increase in competition that meets or exceeds the quality. The house may have gotten nicer but now you are selling a single room
I don't agree with the luddite-esque AI views like using ChatGPT makes you dumber and it thinks for you, but- this is a real case of why people are angry and why they should be concerned about deployment of AI.
The article says it's basically to handle high call volumes maybe during some kind of widespread incident. But in those cases why would AI even be needed- if calls are flooding in for something and people need to be put on hold why do you need AI to ask them if they are calling about incident x? Just automate a voice for that and ask them that while they're on hold / when the system first picks up.
This smells like a bandaid over an underfunded system, and a way to sneak in further cost-cutting in the future where the real calls will be actually answered by AI.
My $0.000002: You dont need a study to know delegating task X can make you worse at task X, nor do you need one to know its vacuously true and avoidable.
The world is more complex, though: maybe people delegate thinking in some areas to be able to have more resources to think about other areas that are more relevant to them?
This is not necessarily true if you spend more than 2 seconds thinking about it.
If I'm raking leaves all day I eventually reach diminishing returns. My arms might even start to get tired and actually make me lose skill. Now I got a leaf raking robot which is not perfect but does 50% of the job and gives me free time to recover my hands and research new leaf raking techniques, maybe even invest into a leaf blower. Did I not become better at raking leaves?
People are not calculators - all these simplistic takes are frankly insulting to human capability.
You don’t need to go to irrelevant analogies such as removing leaves. Humans develop expertise by overcoming their struggles. When you delegate thinking and engaging with source material to a software you’re skipping the actual thing that makes us improve intellectually speaking.
Have you never had to learn a new language or a music instrument? AI can just translate a text into your native language, or tell you how the sentence should be built, but you have to actually actively engage with the low level process of trying to do yourself all the little boring things, a lot, for a long time, to develop the intuition and expertise
You're missing my point entirely. To counter you music example - grinding the same chord over and over does not make you a better musician more than watching youtube videos of other people playing. Skill growth is not as simple as it is in video games - you need variety and rest to actually grow your skill rather than mindless task repetition. This is where AI tools, when used correctly, can result to great growth.
You need both, the task repetition and the resting. Even better if the task repetition isn’t mindless, I don’t know why you added that qualifier.
And actually, grinding the exact same chord over and over can improve your music. You won’t become a grandiose composer, and even unlikely to become a descent musician from that alone, but you will definitely evolve from a messy play to something more comfortable over time. And likely get ideas and motivation for what to do next.
I really don’t think AI tools are good to use as a non expert, they nudge you away from meaningful engagement
> To counter you music example - grinding the same chord over and over does not make you a better musician more than watching youtube videos of other people playing.
It does if you are still learning fingerings. Repetition absolutely helps build skill.
Would you argue an athlete standing on a mark and practicing the same shot over and over is not improving their skill?
Only in a sense which is obviously not at all what's being discussed. If you hire a personal chef, have you gotten better at cooking? What about a nanny, have you gotten better at raising your children?
Of course automating some cognitive processes can free your time for other, more fulfilling aspects. But there's certainly no guarantee that that's the case.
> People are not calculators - all these simplistic takes are frankly insulting to human capability.
Funny example. Which do you think is more common, people who are innumerate because they believe using a calculator is sufficient, or people who are highly numerate and believe learning arithmetic was a waste of time?
> If you hire a personal chef, have you gotten better at cooking
Likely yes, as you learn cooking now only by repeatedly cutting carrots the same way you always did but expanding your horizons to new cooking techniques and socializing in the niche will yield much bigger growth.
This Bruce Lee level misconception of "practice the same kick for a thousand times" is extremely outdated perspective on how people actually progress effectively.
So how best to develop a kick with good form? YouTube is a better way? I really cannot relate to what you are trying to express. I practiced music by drills, it is not doing the same thing repeatedly, you get a little better each time.
Same way you learn anything well - with balance and diversity. You kick, go for a walk, rest, watch youtube, get kicked by other people, talk about kicking on a kicking fans forum.
The point is that you can learn faster and be better with AI tools as you can choose when to kick and when you have space for all of those other activities that will still train your kick.
Not only that, but this clearly seems like superior human condition where we get to choose our actions rather than be pigeon-held into kicking 8 hours a day. I'm a better programmer now that I can choose not to refactor python for 3 days even though after that I'd surely learn all library APIs by heart. Instead, I can spend 1 day working on initial refactor somewhat learning the APIs and have LLMs take over the rest while I spend other 2 days on other new tasks.
I think the confusion here is people assuming that if option of not doing thing exists then people will stop doing it and regress which is a real danger but it's one that can be solved for rather easily. I'd be willing to wager that in 5–20 years we'll have more domain experts than fewer.
> Likely yes, as you learn cooking now only by repeatedly cutting carrots the same way you always did but expanding your horizons to new cooking techniques and socializing in the niche will yield much bigger growth.
You are not engaging in good faith. Obviously the point of the argument is that if you hire a personal chef, you are outsourcing the actual cooking and investing little to no time yourself on improving at cooking. Interpreting it as having more freedom to cook, and your contractor providing a service being your friend/teacher, is absurd.
And even in your contrived hypothetical, there's two basic issues. First, mastery can be highly specialized. Second, if you only ever did the trivial and repetitive to begin with, you never intended to improve.
> This Bruce Lee level misconception of "practice the same kick for a thousand times" is extremely outdated perspective on how people actually progress effectively.
This is another straw man. Nobody is promoting this meme that the only thing you need for self-improvement is mindless repetition. That doesn't even make sense, considering the actual topic is thinking for yourself instead of only using AI. The point is you can't get better at something without actually practicing, and "difficult" or "tedious" doesn't equate to "unimportant."
For the people who believe this, I think it deeply reflects how they think about all knowledge and the kinds of things they would use this technology for- "finding the answers to things without having to think about them"- that's only one way to use it, and the least compelling way, in my opinion.
Struggling with things is important for understanding and thinking about things, and AI gives you a way to bypass that struggle. People will take that bypass, because to resist it takes willpower. It makes it harder to follow the path you need to.
It's sort of like exercise: it used to be required and unavoidable (you had to walk to get anywhere), now it's optional so most people don't do it with anywhere near enough regularity or intensity.
Could we stop with the Luddite label? It’s such a lazy, vague way to gesture that you find some people dumb instead of using a more descriptive term to communicate who you’re referring to. Virtually nobody who would define as AI skeptic or anti-AI identifies as a Luddite
Actually, it's an apt label because the machine looms were brought in by factory owners so they could employ less skilled operators who commanded a much lower wage, and the factory owners pocketed the difference. Skilled weavers were up in arms and smashed the machines.
Today, the purpose of AI is not to get a machine to do your job, it's to look good enough to persuade your boss to sack you and split your salary between the boss and the AI company.
> Skilled weavers were up in arms and smashed the machines.
And how did that work out of them? People conveniently ignore that part of history which seems to clearly indicate that outright destruction rarely leads to progress. With AI, I'd argue that we have much more opportunity to steer than to destroy, and the latter is not even a viable option so - why even entertain it at all? It's just raging for raging sake.
Tell that to entire towns and villages of unemployed weavers.
Ultimately, the factory owners asked their rich and powerful friends in government to help them out, and they started having show trials and executing anyone who impinged on factory owners' rights to make out like bandits.
If the Luddites made a mistake, it was in breaking the machines when they could've broken the owners. That would make the next set of owners a little more considerate of the human cost of their enrichment.
I think you answered the problem in the 2nd paragraph yourself - then why are we breaking LLMs and not the owners? We are capable of organizing our society more like a scalpel than a hammer. We can keep LLMs and have them work for all of us and anyone who's saying otherwise is either too afraid or too stupid.
This Luddite AI prohibitionist movement appears to directly harm all of us. It's much better to steer than to destroy.
Exactly. In the end, it doesn't matter. We used to value someone having physical strength, but we automated away the need for that kind of work. Now, people lift as a hobby, but strength isn't particularly valued.
If we manage to automate thought so it can atrophy (and, I think we're on the way), intelligence will go the same way as physical strength.
Knowing the way in which bureaucracies work, I suspect there is zero (or likely negative) money available for new humans or old-school systems, but heaps available for anything with AI in the title. A manager who increases human staff will be blamed for causing bloat, but any manager who manages to replace humans with AI is an efficiency guru and promising future leader (even if the AI sucks and winds up costing more than what it replaces).
> This smells like a bandaid over an underfunded system, and a way to sneak in further cost-cutting in the future where the real calls will be actually answered by AI.
I think there’s also an aspect of accountability-washing: when regular 911 makes mistakes, managers and politicians get hauled into the spotlight to explain the failure (my city has that now due to the outgoing mayor’s decision to put an unqualified crony in charge). I will bet that the first time this fails badly it’ll be blame shifting and assurances that some huge tech company has fixed the problem and it’ll never happen again, with as little talk about oversight responsibility as they can get away with.
The other thing with the AI debate is that it's nothing new- these kinds of failures are the same kind that have been going on for a long time, for the exact same reasons.
Any time a local government fails people these same kinds of incentive misalignments are at the real heart of it.
The fact that it has the word AI in it just gives it that cool cyberpunk dystopian sheen, where you call 911 and the AI says something about mechahitler instead of trying to help you.
What about cases where e.g. there's an earthquake with thousands of incidents and you need to triage help to the worst? It's not common, but there are absolutely cases where small amounts of intelligence applied in parallel is a genuine advantage.
Im a bit daft today, getting over strep throat, and this went over my head: is an equivalent to this “Yes, but it’s scary because it might turn from automating voice to no more humans answering ever?”
Having someone answer and provide updates or screen for new information is valuable. For instance someone calling can know exactly where the shooter is. Or it could be something related. Information is valuable. What's the alternative? Not have someone pick up the phones? No matter how much you over staff there will be times where peak volume will be more than you can support
This assumes that LLMs can do those things. Given day to day experience with "chat bots", they can't. The LLM eii be circular do nothing of value then route your call to a number where nobody picks up.
All party in the deal understands this. They are just convinced they will no longer be around when things go wrong, hence won't be held accountable for this particular decision.
> This smells like a bandaid over an underfunded system, and a way to sneak in further cost-cutting in the future where the real calls will be actually answered by AI.
I don't get why are people so against cost cutting and use of AI in government services?
So much of my tax money is already wasted on bullshit. If you are in Europe - it's majority of your work, being set on fire by bureaucrats.
Most people mean this to say that 1 trillion is a lot of money, but it still comes back to what you believe AI is- in hindsight, does 1 trillion dollars to build the internet sound like a lot or a little? (That is, spending 1 year of USA's defense budget to get the entire internet)
It comes back to your perception of what AI is because to people who say AI is glorified auto-complete won't believe that the money is worth it.
The AGI-pilled true believers who say it will end all money and result in a post-scarcity world believe literally any amount is justifiable.
Most people, me included, land somewhere in the middle- it seems like AI is a humanity-level sea change in technology and how computers work and serve us. It seems plausible that a few trillion is a reasonable amount.
I'm not going to make a prediction of what will happen with AI whether it will autocomplete / productivity or AGI. I will say it seems to be trending towards former than the latter just by how scaled down the promises have become over the last year (we went from curing all disease and cancer / post-scarcity to productivity and code.) The amounts being spent on this can only really justified by some paradigm shifting returns and within the timeframe investors expect. This isn't something like Apollo / Manhattan project - those were taken on by the government with public money. This is explicitly a profit making enterprise funded by markets.
I have personal experience using it at work so I come down on the “it’ll be a big productivity boost” side rather than Deep Thought. At my job we are still evaluating ROI so it is mandated we all use it to help us, via copilot. For example I recently did a code port that one team said would take a month and sonnet did it for 18000 tokens in a seven hours. Thats 180.00$ vs tens of thousands. Thats a huge productivity boost. Now with the rise of token cost the question is about ROI for my company. What bothers me the most are the hidden costs to the environment that my company does not factor into the equation. From that perspective it seems more than a profit making exercise but I can’t help wonder if it could be done cheaper, and really is an over inflated circle jerk to mint a million new billionaires. It’s like if Martin Shkrelli discovered fire. A great product from an absolute scum bag.
> Most people mean this to say that 1 trillion is a lot of money, but it still comes back to what you believe AI is- in hindsight, does 1 trillion dollars to build the internet sound like a lot or a little?
You are talking about the Value of AI, but the key is the Revenue of AI.
If AI companies cannot get the Revenue to pay for all those investments, somebody is going bankrupt.
what's the threshold for model routing where you're willing to trust the router?
For coding my own work I don't trust the model router, and it would have to be shown to be to save a real dollar amount.
From a buying perspective it's a hard sell to save x but lose out on bugs you are probably introducing at an unquantifiable severity and frequency. How much is it worth to hedge your bets by doing every single inference request on the frontier model?
How much will it cost to go back later and fix things, but also the meta question of how to be able to decide on a hypothetical unknowable? (You'll never know how much better or worse your code was gonna be, it's untestable at a project level)
> LLMs are really only capable of combining different things into output
In terms of how writers think about creativity this doesn’t reflect reality- there’s the old writer’s saying that there are only 7 stories (and other variations of this idea) and that writing is about creatively remixing these well-worn ideas.
The LLM’s capability to write acceptable fiction and nonfiction is coming soon if it’s not already here. (I think right now it still needs some high level input from humans, depending on length and topic)
It’s clear to me that, especially as LLMs get better and better, there’s only one real difference- it’s that I don’t want to hear what an LLM has to say. I’m only interested in what other humans have to say. I feel the same way about AI music- even if it sounds ok, even as good as the kind of unoriginal pop music that’s not AI made (but is a kind of it’s own slop) or a bad committee made hollywood movie, it’s still fundamentally more interesting than something that’s been generated. Even if it’s 100% fiction it’s still based on a real human’s life experiences.
This is a bit like saying “there are only five fundamental flavors” and everything is some combination of them.
Maybe, but that has basically zero effect on the experience of actually eating food. Just because the form may be the same doesn’t mean an LLM will be able to supply original content.
I’m not sure I understand the analogy. You could say that the reason why artificial flavors work in food is because we know how to recreate flavors from first principle chemistry. The dislike of the artificialness of cheetos also doesn’t stop many many people from eating them.
I would say human made slop writing isn’t any better or worse quality than LLM writing.
Knowing if it was made by a real person makes a difference to me, but I also know it won’t to everyone.
Just because the form of a story is the same as another story does not imply that the content is interchangeable.
I could write you a story right now about being an expat living in Poland in 2026. It’s unclear to me how an LLM would have the cultural knowledge or experience to write such a story that is actually accurate and up to date, without already having something like my story already in its dataset.
Yeah but you would just be doing the same thing: mashing up some known information about humans and Poland. You would pick plausible names like Marek or Piotr, plausible places like Warsaw or Gdansk, and some random flavour of story that is either completely general across cultures (hero's journey) or some local tale from Poland.
Shakespeare did the same, so I wouldn't blame you for it.
I am an expat in Poland, that’s why I used it as an example. The point being that I have knowledge of that experience that is not written down in an LLM dataset.
Superhero Wars Trek 42 - the latest franchise installment, billion dollar Hollywood budget, financed and produced by billionaires, audience tested committee-made script, lots of people involved, designed primarily to push merchandise.
Personal Story - a movie about someone's personal reality or fantasy, something typically not made by studios because the story isn't profitable to a wide audience, painstakingly generated and edited by one person using AI, designed primarily to tell a story.
I put up a different strawman because your strawman is too beatable. As humans, of course we're all interested in what other humans have to say. But you haven't explained why you feel people can't use AI to say things, or why AI output can't be based on a real human's life experiences.
Personally, it seems like movie studios are the ones with only 7 stories to tell, and I've seen the movie and its sequel, the prequel, the requel, the remake, the gritty reboot, the musical, and the animated series. I'm excited that AI will let people skip the studio gatekeepers to tell some new stories for a change.
It seems like there are some credible rumors that Google is actually winning in terms of actually building models that work and don't lose money- between how they're able to price them, the TPU advantage and their capex advantage (being able to raise debt + just having a lot of cash - well I said not lose money... more like not go bankrupt).
From the outside they look like they're behind in terms of frontier models, but I think they might be the best positioned to not go out of business when the bubble pops.
Also look at the fact that they've been able to deploy AI-assisted search at google scale. It must be another order of magnitude larger (at least) than the model deployments for OpenAI and Anthropic.
Of course unless you're inside Google it's impossible to know for sure.
In terms of open models, Gemma 4 beats the pants off everything else to the point that paying for APIs becomes hard to justify. Qwen has the meme-share for coding, but it feels much less well rounded. I have no doubt that Google have both the infrastructure and the expertise to curb stomp everyone else, should they resolve in earnest to do so.
Lest we forget, "Attention is All You Need" came from Google.
It also came directly from the university of Toronto, and the university of Toronto seeded all American frontier labs (including Grok (why do you think they could start so fast))
Interesting, glad to hear. We have gemma4 at work, and I was considering localhosting qwen, but gemma4 is so far behind the Opus and Fable I have at home that I've decided to hold off for another model release.
"We have a company provided Toyota at work but it is so far behind the Ferrari I rent at home that I've decided to hold off for another model release."
At 20x the parameter count I should hope GLM beats Gemma! But is it 20x better? Expertise is demonstrated, not by making big models, but by making small ones. Bigger isn't better if you can't run it at all.
It's rumored that Gemini 3.5 flash has a >50% margin, and I'd imagine 3.6 flash is even higher.
I do not think OpenAI or Anthropic are actively chasing margins - though, Anthropic is supposed to be profitable on some form of non-GAAP accounting...
I suspect Google isn't really interested in seeing how far it can get dragged into a race of selling dollars for $0.25, and is more interested to see if it can stay in the race selling $0.50 for a dollar - when everyone else is losing or barely breaking even.
It kind of doesn't make sense though, because typically a large org like Google can afford to crush competitors on pricing. They could probably even go toe to toe with chinese model pricing for years without feeling it.
Maybe they don't want to price war with the other labs so they can comfortably maintain healthy margins on selling them compute?
google has to make money. flash is awesome. you can run it free on their infra and the performance and latency is excellent for what you wait and pay for right now, with great perf per watt. every person in the world going to google.com runs it. every query. its far larger than free gpt, localhost qween and what not.
it's their pro that isn't awesome at all. in fact, their pro kinda suck now that everyone else woke up.
But maybe the TPU advantage is in inference? That's what I assume because the number of compute cycles are going to be all in inference vs training. So they could train on GPUs if they want.
"As of October [2025], OpenAI's compute margins reached 70%, up from 52% at the end of 2024 and double the rate in January 2024, [The Information] said, citing a person familiar with the figures."
As for Anthropic, the rumors I remember seeing for their API margins were more like 85-90%, but I don't have a reference at hand for those. But once you know the API is wildly profitable and the subscriptions are roughly break-even and not even a big slice of their income, all of the investment makes a lot more sense.
It says the original report was in the Information, which I can't see, but I'm skeptical that they includes the training cost? And how much that changes the figure?
That's the profit margin on inference, not overall. Each model does end up being profitable over its lifetime, but the money they're making is being immediately churned into buying more data centers & the training for the next giant model up, so they're not profitable overall at the moment. It's a bit like how Amazon kept churning their profits into more growth instead of taking the profit early.
That said, Anthropic has supposedly crossed over into profitability and made $1 Billion in profit so far this year, in the lead up to their IPO. Being profitable sounds good for launching on the stock market! But as a customer, that's noticeable in the downtime due to lack of compute, and only getting 50% access to Fable.
OpenAI might not be profitable, but they've got so much compute access that they've been able to give their customers full access to Sol, and as a result they've almost doubled their Codex subscriber base in the last two weeks (6 million on July 12, 10 million on July 21 - that would be an extra $1-$10 Billion in Annual Recurring Revenue that they've gained in just these 2 weeks). Doing the unprofitable thing in the short term can result in outsized rewards in the long term.
On the other hand, the consumer side of the market seems to be less competitive right now.
OpenAI's new Mac app doesn't even have a normal "Chat" option now. OpenAI might be chasing coding and b2b sales more now that they realise very few regular consumers pay for subscriptions.
It has nothing to do with npm. However, a binary could be configured to extract your git/npm secrets using this exploit, which could then lead to a npm supply chain attack (or pip, etc. etc.).
It's interesting that all models seems to be unbiased out of the box so far- that is, mostly reflecting the training data (the internet).
The whole mecha-hitler thing doesn't seem to reflect fine-tuning, it was just a prompt change.
There's been some studies that suggest that certain usage of LLMs reduces political bias, which seems reasonable. Like, how credible is climate change, are Haitians eating pets, etc. THings that have a basis in fact.
I don't put it past Elon to train a model with political bias, just that it hasn't happened yet.
> It's interesting that all models seems to be unbiased out of the box so far
This is really begging the question. If something relies on the perception of a human, it has bias. The data (or lack thereof) used to train models is per se a bias.
The mistake is assuming bias-removal is some virtuous goal to be achieved. It can't, and shouldn't. Alignment, while equally impossible, is at least a goal worth aiming towards.
I think the issue here is that there's no objective difference between "alignment" and "actively trying to shape its replies to fit a political narrative", beside the fact that the latter refers to the kind of political narratives that you don't like.
> But your brain works the other way around. First there is a concept, a feeling, an image, and then the words come out to describe it.
After this I can't take the essay seriously- this sort of blanket statement about the one true hierarchy of consciousness and knowledge is BS.
While it seems to have been disproven that words in other languages cause people to speak and think differently, that also doesn't mean that words don't have any effect on the way we think, or that the concept of a word always has to come before the word itself.
The reason why we experience the uncanny valley of LLMs is because they don't represent a true consciousness, BUT it's also clear that the architecture represents certain qualities of consciousness- as the models have scaled we can see that it has some other non-word related understanding.
The evidence points to consciousness as a set of interlocking systems- attention, long term memory, short term memory, emotions, etc.
reply