Discussion

Volpe's Blog

franze: .... so what? the technology exists, the models exist. Even when the bubble bursts things will not go to the state "before AI". Even if model development would stop today (not the worst thing to happen) it would still be the most impactful invention since the printing press

hk__2: Yes, that’s what the author wrote in the second sentence of the post: "AI is here to stay."

nopinsight: > nobody is sure if even their metered pricing is profitableThis is most likely wrong. Lab executives insist that serving tokens is profitable. It's the cost of training next-gen models that requires them to keep raising ever larger rounds. More importantly, many independent providers price tokens of open-weight models at a fraction of Anthropic's prices.

atwrk: But are they actually profitable, or do they employ creative accounting where only parts of overhead expenses are counted against all of inference revenue, similar to what Uber did?

qoez: History doesn't have to repeat. There's barely anything else going on in terms of innovation, and AI is a real step function technology. We might be overspending but there's no way we're getting another AI winter like last time (remember last time investment in 90s AI had to compete for resources with the internet boom).

hk__2: Isn’t that covered at the top of the post?> AI is here to stay. If used right, chances are it will make us all more productive. That, on the other hand, does not mean it will be a good investment.

Chance-Device: From the beginning of this I’ve wondered the same question: how do these companies justify spending such massive amounts now (and 3 or 4 years ago) when software and hardware efficiencies will bring down the cost dramatically fairly soon?They basically decided that scaling at any cost was the way to go. This only works as a strategy if efficiency can’t work, not if you simply haven’t tried. Otherwise, a few breakthroughs and order of magnitude improvements and people are running equivalent models on their desktops, then their laptops, then their phones.Arguably the costs involved means that our existing hardware and software is simply non viable for what they were and are trying to do, and a few iterations later the money will simply have been wasted. If you consider funnelling everything to nvidia shareholders wasting it, which I do.

sunaurus: The point is that you can’t just serve tokens without also training the next models. It’s an inseparable part of your costs, so naturally you can’t be profitable unless the price you are charging ALSO covers training.

256BitChris: I could see OpenAI hitting financial issues which triggers some media induced panic and for people to claim the AI bubble has popped.However, the core utility of the best AI (read: Anthropic's ATM, by miles), will still exist and be leveraged by those who have learned to use it well.I could also see the exponentially declining power requirements offsetting the exponential-but-slower rate of AI compute demand, which then renders a lot of unused capacity in these massive data centers.I think of it like the old mainframes in the 70s which would take an entire city block to run, and now we have the equivalent of millions, if not billions of them in our pockets.

eieje: It’s pretty much undeniable at this point that the sentiment has changed.About 2 months ago this place was unbearable - filled with doom and hype AI posts. I welcome the calming and eventual slow release of the bubble.

general_reveal: HN is no longer a reliable place for the truth. Quite frankly, unless you are utterly self educated, you are terribly vulnerable to this place.At this rate, I’d almost prefer to talk on a private mailing list with vetted resumes.

myspy: Why?

general_reveal: You have to be uneducated to even read an “AI is bubble article”. Anyone working this stuff knows how much more compute we need.

monegator: > How this affects you?> checks list ...nope, nothing will either directly or indirectly affect me. Let it happen sooner, rather than later, and unleash the mobs at the tech bros that set the world on course to make everybody's life more miserable. We'll still be here to get the scrapped RAM and GPUs to train and infere local models thank you very much.

coffeebeqn: The current best models are already very capable of disrupting the job of millions of people. I don’t think a scenario where we just go back to pre-Claude Code exists and I’m sure the same models can be tuned for much of other white collar work at similar capability

positron26: When will this concern farm end? Internet is ant-milling harder than a model gone psychotic on synthetic data. Call me when it's over.Back to the mines. The Vulkan only writes itself when prompted with well-conditioned problem statements.

jqpabc123: Another possibility not really addressed here --- local LLMs.AI on hardware you own and control --- instead of a metered service provider. In other words, a repeat of the "personal computing" revolution but this time focused on AI.TurboQuant could be a key step in this direction.

elorant: I feel that even if the bubble bursts hardware prices will still take years to normalize. So no clear benefit for the average consumer here.

joshstrange: > RAM prices are crashing because new models won’t need as muchReality begs to differ [0] and following the link for that text goes to an article [1] where they talk about Google's TurboQuant which supposedly will lower the RAM requirements. Now if that means RAM prices come down (as speculated, not reported on, in the link) or the AI companies just do more things with their extra ram is yet to be determined. The fact this article links there with text "RAM prices are crashing" throws the entire rest of the article into doubt for me.RAM prices are most certainly not crashing (yet) and treating it as a forgone conclusion because _one_ lab found gains could be made and hasn't even reported on the efficiency of their method is just irresponsible. It's almost as bad as when LLMs link things to prove their point, you visit the link, and find it says nothing of the sort or even the opposite.[0] https://pcpartpicker.com/trends/price/memory/[1] https://tech.sportskeeda.com/gaming-news/how-google-s-new-tu...

infecto: It’s incredible how polarizing the AI rush is. I keep the perspective that the technology is an absolute step change but I have no idea where the cards will fall. I take a lot of issue with these style of articles. I get a sense that the authors are being overly defensive.The cost to serve tokens is absolutely profitable today and that’s been true for at least a year. What’s unclear is how R&D and capex fit into the picture. I am not that pessimistic on this front either though. For the data center build outs, demand for tokens is still exceeding supply. On the R&D front, well most of us here on HN have benefited from decades of overinflated engineering salaries being paid by often companies that were not profitable and not only unprofitable, usually without a plan for success. In this current rush, companies cannot keep up with supply, it’s a much easier math problem when you have something that people want (tokens) and you need to figure out profitability when including R&D.

malfist: > The cost to serve tokens is absolutely profitable todayHow can you possibly say that? Everyone knows that's not the case, these companies are losing money every day selling tokens. Revenue is not the same thing as profit.

A_D_E_P_T: Two things can be true at the same time:- AI is a genuinely transformative technology on par with the internet and on track to probably surpass the smartphone- The inflated valuations, the circular flows of money (or "money"), and the financial cup-shell game mean that the players of the game are all a few bad weeks away from catastrophe. This is, of course, nothing new for SV -- but the scale this time is new. Some believe it will soon collapse -- "bubble," thus.

thereitgoes456: > The cost to serve tokens is absolutely profitableCan you explain why you know better than the analyst at Cursor cited in this article?

ap99: They're not just betting on the current tech, they're building out infra like this because probably any future tech currently being researched will also require massive data centers.Like how the gpt llms were kind of a side project at openai until someone showed how powerful they could be if you threw a lot more parameters at it.There could be some other architecture in the works that makes gpts look old - first to build and train that new ai will be the winner.

boriskourt: > The cost to serve tokens is absolutely profitable today and that’s been true for at least a year.> For the data center build outs, demand for tokens is still exceeding supply.Can you provide any numbers for this?

baq: If they shut down all training today they’d be absolutely printing money for the next couple quarters and then die with a bang once the other lab releases the next frontier to the public.

shimman: How? They're already burning $2 bills to make $1, court documents shown that Anthropic has already been lying around revenue (claimed to have made $19 billion when it's actually $5 billion to date [1]).Not hard to believe they're lying about other things when they've been lying about the capability of their products since inception.[1] https://www.reuters.com/commentary/breakingviews/anthropic-g...

irusensei: I guess the point is that without the hype subsiding it enshitification will ensue.

schnitzelstoat: Yeah, I don't think local LLM's will keep up with what the massive corporations put out. But they might get to a level of performance where it just doesn't matter for most users.And people would prefer to run a model locally for 'free' (not counting the energy cost) rather than paying for an LLM subscription.

nickphx: step change? how? profitable? where did you read that? people want tokens? really? who are these people?

dash2: Is that right? I think that you can serve tokens without training the next models. It would be bad strategy, but it would work. So it's an important question, are they covering their operating expenditure? If they are the business has legs (and it will be worth spending a lot to train the next models). If not, maybe not.

camdenreslink: If a major model provider were to just halt progress on developing new and improved models, the open weight alternatives would catch up in a couple years.They would have a period of great margin, followed by possibly zero margin as enterprises move to free options.They would have to come up with a lot of great products around the inferior models to justify charging at that point.

netdevphoenix: Local LLMs don't sound profitable at all for those building them. If you really wanted a SOTA model, you would be paying eye watering amounts to own it unless you got an open sourced one.

phito: I think their current goal is to capture as much market as they can while they still have the best models, their only moat. Look at Anthropic, they are clearly trying to lock their users in their ecosystem by refusing to follow conventions (AGENT.md etc) and restricting their tools exclusively to their own services.

baggachipz: Consumers and retail investors will bear most of the brunt from this bubble. Even taxpayers, as the government will most likely bail out the "too big to fail" ai companies in the "race against China". All based on bullshit, hype, and greed.

techpression: I wouldn't trust those claims from any private companies, even public ones play the most insane tricks in earnings calls to inflate numbers or heck, just make up new ones.I'm not saying they're wrong, but I don't take much stock in their words.

Aperocky: The decision is the right one. Scaling at any cost is the right way to go.You cannot find the efficiency if you haven't been experimenting at scale, this is true personally as well.If someone haven't been burning a few B tokens per month, everything coming out of their mouth about AI is largely theory. It could be right or wrong, but they don't have the practice to validate what they're talking about.Not everyone scaling to that degree would have the right answer or outcome, many would be wrong and go bust. But everyone who didn't will not have the right answer.

fotcorn: Also, there is zero reason to think that the big labs did not have anything similar to TurboQuant for a long time already.TurboQuant itself is already a year old! So even smaller labs have probably seen and implemented it.

nyeah: [delayed]

MattRix: This is not lying, that is just what run rate revenue means! It makes sense to use as a metric when a company’s user base is growing as fast as Anthropic’s is.

rvz: > HN is no longer a reliable place for the truth."No longer?" It never was.Especially with AI boosters being allowed to degrade the comments section and shilling their paid blogs and violate the guidelines.

keybored: > On the R&D front, well most of us here on HN have benefited from decades of overinflated engineering salaries being paid by often companies that were not profitable and not only unprofitable, usually without a plan for success.I feel like giving a Richard Nixon lecture now.

Aurornis: This article tries to build upon a lot of half-truths or incorrect facts, like this:> OpenAI is struggling to monetize. They turned to showing ads in ChatGPT,The ads aren’t going into your paid plans (except maybe a highly discounted tier, depending on the market). The ads are a play to offer a free version. Having an ad-supported free tier isn’t new.The discussion about being unprofitable also repeats the reductionist view that these companies are losing money and therefore the business model doesn’t work. This happens with every VC cycle where writers don’t understand that funded companies are supposed to lose money while they grow. That’s what the investment money is for.We have very strong indicators that inference is not a money loser for these companies and is likely very profitable. They should be spending large amounts of money on R&D to get ahead and try new things while they’re serving up tokens.The “but they’re losing money” argument never seems to be brought out against competitors that literally give away their models for free and for which we can calculate the cost of serving 400B-1T parameter open weight models.

Izkata: > The ads aren’t going into your paid plans (except maybe a highly discounted tier, depending on the market). The ads are a play to offer a free version. Having an ad-supported free tier isn’t new.Sounds like it is new for ChatGPT though. That's also how it started with TV and Youtube, first on the free tier then expanding to the paid ones.

faangguyindia: If the gains are real why the limits are so bad? Google can barely serve Anti-gravity.

agentultra: It sounds like most of the data centers promised in 2025 and 2026 are not even built yet and most of the GPUs bought haven't even been installed.If it does all go down in flames, even floor value is not going to be that valuable.I can't predict the future but it's smelling a lot like a recession already under way that is bigger than the sub-prime crash.

throwaway27448: > We have very strong indicators that inference is not a money loser for these companies and is likely very profitable.Why is OpenAI specifically losing money hand over fist then?

aurareturn: Training. But training costs are a smaller and smaller percentage of revenue as inference revenue grows faster than training costs.

schnitzelstoat: It's a winner-takes-all market and everyone wants to be the next Google and not the next Lycos or AskJeeves etc.It'd be interesting to see what they spend all the money on though as we seem to be hitting diminishing returns and I'm not sure if the typical enterprise user really cares about small improvements on benchmarks.It seems like it'd probably be better to spend all that on marketing, free trials, exclusivity/bundle deals etc. ChatGPT already has a strong advantage there as it has so much brand recognition. I've seen lay people refer to all LLM's as ChatGPT like my grandparents did with Nintendo and all video game consoles.

delecti: I don't think it's winner-takes-all. Google is Google in 2026 because Lycos and AskJeeves were bad in comparison. The average user doesn't care whose LLM they're using because they're all close enough. It's hard to see past the bubble bursting, but I expect most people will use multiple of them depending on context (Copilot via the integration in windows, Gemini via Siri on their phone, etc), likely without paying.

Eridrus: The article is just helpfully illustrating how artisanal you can make your slop if you really try!

shubhamjain: > OpenAI is struggling to monetize. They turned to showing ads in ChatGPT, something Sam Altman once called a “last resort”, while Anthropic is crushing them with the more profitable corporate customers and software engineers. Their shopping feature flopped and they shut down Sora, both supposed to be revenue drivers.I don't think Sora ever thought of as a "revenue driver" considering how notoriously expensive and unpredictable video generation via inference is. OpenAI is just a repeat of Uber—minus the scandals—in a different decade. Uber got itself into tons of businesses related to transportation on the assumption that it would all be viable "one day." Same stuff that OpenAI is going.I would say, once the bubble bursts—which is likely, considering the geopolitical environment—OpenAI, Anthropic, and Alphabet are likely to be the winners, with a lot of small players at the tail end. Anthropic won over programmers and OpenAI on everyone else. For millions of people, AI = ChatGPT, so I would bet that OpenAI can still become profitable, once they cut down their expenses.

JohnTHaller: > minus the scandalsGiven the tech bros involved, we just don't know about them yet. Also was this comment generated using AI? Look at all the em dashes.

richard___: Complete bs.

thereitgoes456: That is not what the article says, it says $19B ARR.I don’t necessarily see a contradiction. $19B run rate, achieved very recently, is actually consistent with $5B lifetime earnings, because their growth curve is so sharp. Zitron is not good at math.

Havoc: Gov bailout seems like the only way out.

dgb23: Aren't you conflating the technical side of it with the economic one?A bubble doesn't necessarily mean that the the underlying tech/innovation isn't useful. It's a financial and economic phenomenon that is pretty well understood and researched:- During the hype cycle, investors tend to overestimate the short to mid term effects and underestimate the long term effects.- It's near impossible to pick the winners in advance, and research has shown that investors underestimate how many losers there will be.- The financial system/market works very well when there are localized issues with debt. Those get seemingly automatically detected and repaired. But broad increases in credit not so much. Those spread into the whole system in non-obvious and complex ways and destabilize the whole system, which can lead to very large corrections.etc.

sandworm101: There is also demand for ram in others areas of data centers. As we are all pushed deeper into clouds, i can see the rise of ram for data storage (ram drives) continue to eat into the supply. A module of ddr5 will be more useful in a netflix rack streaming movies 24/7 than in a gaming PC where it may only be used an hour or two every day.

monegator: might be: there is too much busy work as it is, but we need people to work in order to make money in order to spend it in order to keep the circus from going under. It's the circle of lifeLet me remind you that you are not paying the full price for the service and all the value of those company is out of thin air. More or less the premise of the article. *when* you will be asked the real price, we'll see if the company will prefer a human or a bot it can't pass blame to

joefourier: It’s absolutely not winner take all. LLMs have become a commodity and the cost of switching models is essentially nil.Even if ChatGPT has brand recognition amongst lay people, your grandparents aren’t the ones shelling out $200/mo for a Claude code subscription and paying for extra Opus tokens on top of that. Anthropic’s revenue is now neck and neck with OpenAI, but if tomorrow they increased the price of Opus by 5x without increasing its capabilities, many would switch to Gemini, GPT 5.4, Cursor, or any cheap Chinese model. In fact I know many engineers that have multiple subscriptions active and switch when they hit the limits of one, precisely the tools are so interchangeable.At some point it could even become cheaper to just buy 8x H100s and host Qwen/Deepseek/Kimi/etc yourself if you’re one of those companies paying $3k/mo per engineers in tokens.

piker: > They lose a big customer for their cloud services. Even worse considering that now, using the AI they helped fund, everyone can compete with their sub-par products. GitHub is a good candidate for disruption, and that’d be just the start.Look, I'm a Microsoft hater like the rest of us, but calling Microsoft's products sub-par discredits the author a good bit. I invite anyone who thinks this to try and compete with them. Go after something like Word, for example. Then prepare to be awed by what some of the most brilliant programming minds ever can produce after grinding for four decades.

karolist: You can have an opinion about a tool as a user, without ever having ability to create such a tool yourself, that's literally what every tech and auto reviewer does.

nickcageinacage: AI is shit. I just want this to be over. Can we move on

thebeardredis: Hopefully soon. My new unwords are f.e. "agentic".

jeromegv: Yep, especially if we look at what happened just last week, both Google and Anthropic have dropped how much you get out of your existing plans.

bob1029: https://www.cerebras.ai/blog/cerebras-cs-3-vs-nvidia-dgx-b20...

chrisweekly: > "decades of overinflated engineering salaries"'Overinflated' relative to what? You make some good points but I don't accept this as a premise.

pier25: So these companies will be profitable if training stops? Is that even a real possibility?

amelius: > Now if that means RAM prices come down (as speculated, not reported on, in the link) or the AI companies just do more things with their extra ram is yet to be determined.I think it is determined:https://en.wikipedia.org/wiki/Jevons_paradox

hirako2000: Not crashing yet. The article is looking 1 to 5 years to come.Given Nvidia's CEO's agitation I would give credit to the prediction, and if it's correct the price will go back to what it was, or even lower of investment in capacity are made today.

baq: > It's a winner-takes-all market and everyone wants to be the next Googleabsolutely isn't! if billed per token, there is no reason to be married to a single model family provider at all. the models have very different strengths and weaknesses, you should be taking advantage of this at all times.

paulddraper: Anthropic has said inference is profitable. That’s a biased source, but the math pencils.This is why switching to local open weight models saves a lot of money. (Even though it’s not apples to apples.)

nyeah: Can you give a few penciled numbers?

iterateoften: According to open router token demand is growing at something like 10% a weekIt’s insane

aurareturn: This is an awful article.Bottom line is that H100 prices are near 3 year highs, A100s are still profitable to run, B200 prices are increasing, no one has enough compute. Google, OpenAI, Anthropic, Meta, AWS, Azure are all compute constrained. Every single one of them said so publicly. OpenAI is struggling to monetize. They turned to showing ads in ChatGPT, something Sam Altman once called a “last resort”, while Anthropic is crushing them with the more profitable corporate customers and software engineers. AI bubble is bursting because OpenAI is trying to monetize free users on ChatGPT with ads but Anthropic is kicking butt in AI. What kind of logic is that? So it seems like AI can be monetized as Anthropic shows. I wouldn’t be surprised at all if in the next couple of quarters we see OpenAI looking for an exit. It will be interesting because the sizes are now so big that we will probably know all the details. The most likely buyer is Microsoft, they already own a lot of it, and because of that, they are the most interested in showing a win. I'll take the opposite stance. I think OpenAI is going to be bigger than Microsoft in market cap within the next 3 years. I think Anthropic and OpenAI are going to run laps around current big tech except maybe Google. Independent reports state that Claude metered models are priced 5x more expensive than their subscribers pay Already dispelled. It isn't 5x more expensive than their subscribers pay. Inference has a gross margin of 50%+. It's been repeated over and over again by Anthropic CEO, OpenAI CEO, and just about anyone who's down deep analysis on token profitability.

albinn: I would think that we are going to see RAM prices increase even more, given, among other things, pure helium disruptions and increased electricity prices.I haven't looked closely into TurboQuant, but perhaps it will revolutionize just as much as the 1-bit llm did...

H8crilA: Where to go next? I don't think anyone has gotten close to automating everyday PC usage, likely via screen capture and raw keyboard+mouse inputs. Imagine how much bigger would that market be than vibecoding.

Tade0: My main worry is - once this is all over, the market consolidates and using LLMs will become a requirement in job listings, what's the highest price per million tokens companies will be able to charge us?Currently on a given day I'm chewing through approximately the equivalent of my lunch money, but where there's opportunity to extract wealth, someone will find a way to do it.

adjejmxbdjdn: I’m not disagreeing with you, but consumer RAM prices are lagging indicators. If commercial RAM prices are dropping then consumers will see those price drops last, especially given the fact that several consumer manufacturers turned to commercial only.

drakythe: Is there a source that says commercial RAM prices are dropping? I was recently told (without a source, so I am not sure if it is true or not) that OpenAI never even bought any of the RAM they signed deals on last year, and that those deals were just letters of intent. So if prices are coming down I wouldn't be shocked but the economy is pretty well vibe coded these days so who even knows.

ajay-b: I would be very sad to lose services like ChatGPT. It has significantly improved my workflow by digesting and analyzing huge documents, and helping me to synthesize and respond better. May be I am part of a minority.

raincole: Don't worry lol. It's not going anywhere. The article is just ragebaitng. Verbatim:> Anthropic is already in a push to reduce costs and increase revenueYeah, it's totally a bad sign when a company tries to... reduce costs and increase revenue.

mattmanser: Try doing some inference with local models.I'd be surprised if they're making money on inference just from that. There's no way someone paying $20 p/m and using it all day is not spending way more on even just the electricity for tokens, let alone the capex.

smt88: YouTube, Spotify, and most video steamers have zero ads on paid tiers. I never see video ads.

adjejmxbdjdn: YT has a Premium Lite paid tier (at least in the U.S.) that does show ads on music and in certain other areas of the app, such as shorts, searching, browsing, etc.

14113: > companies are supposed to lose money while they growAt what point do we declare that a company has "grown" and now must make money? OpenAI is a multi-billion dollar company right now, surely that's a point at which they should be profitable, instead of propped up by further investment and borrowing.> We have very strong indicators that inference is not a money loser for these companiesAll of the economic analysis that I've read strongly states the opposite. Running a GPU is a net loss /even for the data centre operators/. For them to break even, they currently charge OpenAI/Anthropic/Etc more than OpenAI/Anthropic/Etc make per-token.

curtisblaine: I'm sure Word is full of arcane backwards compatible tricks that 20% of users use, but I find it hard to differentiate the Pareto 80% of the product from Google Docs or any other competitor (LibreOffice?) Adding rich text, tables, headings and colors is pretty much a solved problem for all of these softwares. Adding images or handling more complex layouts sucks everywhere, it's not like that Word has a great user experience and the other don't. All of them are bad. IMHO, if we had any of the competitors being the de-facto standard for word processing, the vast majority of users wouldn't feel the difference. Power users would for sure, but I'm not sure they're many or they use existential features. If Word didn't have a near monopoly in office settings due to aggressive marketing, OS presence and a proprietary file format that constantly changes and never renders well outside of Microsoft products, it could disappear without anyone (save Microsoft) losing much.

piker: Yes. That 80% you find useful is served fine by Google Docs, but there’s a good reason the enterprise overwhelmingly goes for Word, and it lives deep in that 20% and a lot of the time has zero overlap with others.

SirensOfTitan: This is a classic HN mistaking the map for the territory. R&D and capex absolutely figure into de-facto profitability and sustainability for AI labs, despite their separate treatment in accounting.> well most of us here on HN have benefited from decades of overinflated engineering salaries being paid by often companies that were not profitable and not only unprofitableThis is a really concerning perspective: people were paid what they were worth. Software is or was one of the few remaining arenas wherein a person can find a middle or upper middle class lifestyle consistently.I will also note: a startup raising an 8 MM series A and eventually fizzling out is not the same at the hundreds of billions invested in these AI companies without a path to profitability. It is utterly absurd to pretend these are the same thing: any company ingesting that much cash needs to justify its capacity to survive.

9rx: > This is a really concerning perspective: people were paid what they were worth.The parent comment doesn't discount that, only pointing out that "what they were worth" was inflated due to a speculative environment. Wherein lies your concern?

pydry: Jevons paradox only applies if demand hasnt already been saturated.The fact that public LLM usage is leveling off at a price of $0 and Jensen "we make the shovels in this gold rush" Huang is rather desperately claiming that you need to spend $250k/year in tokens to be taken seriously suggests that demand saturation may not be that far off.Whether it applies to software engineers I think is another open question. Im constantly being told that it doesnt and that LLMs make half of us redundant now, but Im skeptical.

chasd00: i don't think it will work, it's too easy to switch models. When google comes out with a new model people will just switch. I think Google wins in the long run, they have the money to just wait until everyone else goes bankrupt and they also have the Apple contract and therefore the mobile market.

leoc: And apparently the most efficient training and inference thanks to their TPUs, IIUC?

Aperocky: Demand of tokens is absolutely skyrocketing.And unlike the traditional "this will replace humans right away", I think what this introduce is a lot of incentive to spread those token in places where there was never any incentive to hire a software engineer for previously. In turn, that will drive a lot of business activity in those area that will potentially fail given the current quality of the output.This feels like a boom before bust scenario, and I'm not even sure if it will bust.

hirako2000: Tulips sales also skyrocketed.Seriously, what value are tokens providing other than justifying layoffs. Concretely. Today. Not in the speculating scenario that cardiologist could be replaced with models.We see this new trend of agentic coding, again a promise software will be written that way going forward, despite the number of fiasco already experienced when trusting a model turned bad. The use case may provide value, but right now all it does is fullfil the push for token consumption all these AI leaders are advocating for.

gruez: >Seriously, what value are tokens providing other than justifying layoffs. Concretely. Today.It's adding tests for me and doing medium complexity refactors that I'd otherwise have to spend hours on

owlmirror: Isn't that at the moment still a free product? Of course they will not prioritize serving those requests. That tells you nothing.

butlike: It tells you there's no clear path to monetization.

mrbungie: > Lab executives insist that serving tokens is profitable.Maybe marginally profitable, but right now they need to give out subsidies for people to use their products (Antigravity, Codex, Claude Code et al) in an actually useful manner that prevents churn and at the scale they need to justify usage growth forecasts, which they need to keep the wheel turning.Probably if you look at the users who exclusively use the simple chat box interfaces (i.e. ChatGPT, Gemini in UI, Claude in UI) plans it is actually profitable, but I'd also say that's not where most of the usage comes from.I'd love to actually look at both usage + profitability from each user segment to see if their PxQ growth expectations from non-enterprise usage make any sense.> Many independent providers price tokens of open-weight models at a fraction of Anthropic's prices.Are those open-weight models as good as Anthropic? Are they the same parameter class?

est31: It's a loss leader but this is normal. Same has happened with Uber, Airbnb, Amazon, etc. Using VC money to buy marketshare and once you have it, you can milk it.The question is more around the moats that these companies have and it seems to me while their models are amazing technology, they don't really have a moat. The open/chinese models still continuously catch up to the american ones.

hirako2000: And what possible moat. It isn't hard to foresee that in just a couple of years, models outpacing the latest frontier tech we have today will run on consumer hardware. With open source workflows anyone can pull in to run, providers won't see a penny.Another scenario is that dense models get replaced entirely, in which case the likelyhood of OpenAI and co pioneering the concept is pretty slim. They will be left with billions worth of infrastructure which cost them 10 times that 2 years earlier, faced with the reality touched by the article: liquidate.

slfnflctd: > almost as bad as when LLMs link things to prove their point, you visit the link, and find it says nothing of the sort or even the oppositeTo be fair, they got it from us. This happened to me plenty of times long before modern LLMs.

piker: Sure, and the less you understand about the tool’s fundamental capabilities, the less useful your opinion is. The best reviewers have deep knowledge.

Zardoz84: I don't paid anything to YouTube and I don't see any ads. Because I block ads.

fcarraldo: > Software is or was one of the few remaining arenas wherein a person can find a consistentlySoftware salary inflation and expansion has made this the case. Tech’s accessibility to the educated has accelerated gentrification massively, rising up prices on rent and food. While the statement is correct, tech’s contribution to income inequality is part of the issue. If you’ve lived in Austin or Chicago (especially Austin) prior to ~2010 you’ll have seen this first hand.

drakythe: Anthropic also recently tweaked their usage limits to discourage use during peak hours. Why would they do that if inference was profitable?

wongarsu: I can get Kimi K2.5 inference on openrouter for about $0.5/MTok input + $2.5/MTok output, from six providers that have no moat besides efficiently selling GPU time. We can assume they are doing so at a profit (they have no incentive to do this at a loss), giving us those numbers as the cost to serve a 1T-a32b model at scale.Now we don't know the true size of any of the proprietary models, but my educated guess is that Sonnet is in about the same parameter range, just with better training and much better fine tuning and RLHF. Yet API pricing for Sonnet is $3/MTok input + $15/MTok output, exactly six times as expensive. Even Haiku is twice as expensive as Kimi K2.5.I find it difficult to believe in a world where those API prices aren't profitable. For subscription pricing it's harder to tell. We hear about those that get insane value out of their subscription, but there has to be a large mass who never reaches their limits. With company-wide rollouts there might even be a lot of subscription users who consume virtually no tokens at all.

jerojero: Companies doing foundational models need to cover the cost of training which is much more expensive than training something like kimi.

gruez: >Companies doing foundational models need to cover the cost of training [...]But that's moving the goalposts? The original claim was on inference itself, not the whole company.> The cost to serve tokens is absolutely profitable today and that’s been true for at least a year.

infecto: Most/all private labs have cited inference is profitable. This was happening before the large push to scrap plans and largely charge folks the underlying api rates. Second take a look at the pricing of open models. Now certainly it’s not direct 1-1 comparison but we can use it as a baseline. Now of course folks might not be telling the truth but one of those situations where I see too many markers on the true side.For supply look at outages and growth rates at companies like openrouter. The demand is growing every week.

michaelcampbell: Same, and constructing at least drafts of huge documents that I can iteratively fine-tune that have (at least last week) saved me 10's of hours.And based on reality (code) rather than my feelz of what I vaguely remember the code to have been doing in some long past.

keybored: > This is a really concerning perspective: people were paid what they were worth.Even interpreting what-they-were-worth in the usual sense, I’m not so sure about this. We have seen wage collusion reported by the usual US West Coast-based companies. And some news on here[1] have reported that some engineer with a salary of $100K[2] might be producing $1M of value. And even factoring in the usual “but benefits and overhead” comes out to a solid factor of profit per programmer/engineer.Despite that the sense I get (only from this site since that is my only reference) is that the so-called overpaid engineers are incredibly content to just have this happen to them. As long as they are paid well compared to other workers, it’s fine. No matter the profit factor. In fact, the discourse is very much focused on how “privileged” they were if the tide ever changes. Instead of realizing how much value they provided, collectively.Outlets for capturing more of the value they create is entrepreneurship (Hello HN). Never any collective organizing. And entrepenurship is easily bought via aqcuisition.Collective bargaining would have been relevant in case they ever get automated... by the very software they co-created.One could imagine that this “privileged” collection of programmers could have served as a vanguard for the collective good of programming professionals as well as collective ownership of software goods, using their privilege to that end. The former never happened, and the latter is partly realized in people’s free time (see the OSS maintainer in Nebraska meme).[3][1] All from recollection since this is just news from the Frontier to me[2] Of course the pay might be much higher now; this might have been a while ago[3] when it isn’t simply exploited by corporations just using OSS without giving any back; a logical turn of events when no license or law forces them to contribute back

guzfip: > As long as they are paid well compared to other workers, it’s fine.Well I’m sure they’ll be thrilled to know they can collect $100 a week more in unemployment benefits than their neighbor.

joefourier: The dotcom bubble burst and 26 years later we’re all hopelessly addicted to the internet and the top companies on the stock market are almost all what would have been called “dotcoms” then.The railroad bubble burst in 1846 not because trains were a dead end - passenger number would increase more than 10x in the UK in the following 50 years.

fcarraldo: Well, not GP, but I do. Let’s look at the numbers:Median senior SWE salaries in SF: https://www.levels.fyi/t/software-engineer/levels/senior/loc...Median income in metro areas: https://www.cnbc.com/2024/07/11/the-median-salary-for-the-25...Engineering salaries are significantly higher than nearly every other industry on average and on median. Much of this is driven by VC funding rather than sound, profitable, bootstrapped businesses with sustainable profit margins.Engineering salaries have also been driven upwards significantly the past ~10 years (since the post-2008 crash recovery), while wage growth in the US is mostly stagnant. I don’t have a source handy for that, but there are plentiful studies.Outside of the US this may be less true, but I took GP’s “most of us on HN” to mean people who work in US tech companies which are primarily concentrated in high COI areas.

rileymichael: > Engineering salaries are significantly higher than nearly every other industry on average and on mediannow compare the profit per employee at tech (software engineering) companies and those industries..

hnthrow0287345: I don't see this bubble really popping as-in sinking the economy. Some circular investing and enough write offs will happen to avoid the largest recession indicators from informing the general population that there's actually a recession. You also have a government willing to do shady shit for their own benefit at the expense of responsible governing and ethics, and we have already seen the business leaders of the biggest tech companies cozy up to the administration.My guess is that cloud companies will scoop up the data centers for pennies on the dollar and the GPUs get written off or fire-sold to enthusiasts still wanting to run local models. Then they can offer exceptionally low initial prices to new customers and get more people to be locked in. Or maybe we see a couple of new cloud companies start up but that would likely need lower interest rates.

lstodd: DC infra will be scooped up by cloud guys, that's a given. As for GPUs.. well low-precision tflops have other uses besides inference. You can run Doom for example.

mattmanser: Their point is it is a bad sign at this stage in the game, there's a lot of competition still.Usually in a land grab like this you spend, spend, spend.Uber was still paying to subsidize customer's rides until fairly recently to kill off the competition.

nexos: I think ultimately the AI bubble is bound to burst solely based on the fact that no AI company has turned a profit. A business model consisting of pure speculation on profitability when profit has not come in for 4 years now indicates that the tech industry is over-betting on AI. That plus consumer backlash at the way AI is jacking up consumer prices on RAM and etc means that the bubble is bound to burst. To paraphrase Linus Torvalds, AI is a helpful tool but I look forward to the day it’s a regular part of life and the hype cycle ends

butlike: They clearly have some vested interest/skin in the game. Not sure it's worth retorting that one.

infecto: I wish this was higher up. I have been tracking the same since Thanksgiving ‘25 and the growth is unreal. Again I don’t know where the cards fall maybe the industry overspent on capex but it’s at least easier to see why they are spending based on demand. The risk of being left out is greater than overbuilding.

red_admiral: Microsoft's AI, on the other hand, is underwhelming at the moment and might well go the way of Windows Phone. Plus enough people hate the copilot icons everywhere that Microsoft is hinting at dialing down a bit.MS Office should last a while if they stop calling it "Copilot 365 Office" or whatever it was.

michaelcampbell: My take is new capabilities will consume any price reductions, making them moot. At least in the medium term.A RAM price drop due to some magic efficiencies assumes everything else doesn't change, which I doubt anyone honestly thinks will be the case.

ajross: > Reality begs to differHonestly you're both wrong. RAM prices spiked speculatively, and they're going down for the same reason. Market people always want to argue in fundamentals, when in practice *ALL* the high frequency components of the signal are down to a bunch of traders trying to guess where it's going in the short term.At best those guesses are informed by ground truth ("AI needs a lot of RAM!" "Sam cornered the marked!" "TurboQuant needs less RAM!"), but they remain guesses, and even then you can't tell the difference between that and random motion.

Forgeties79: I’ll believe they’re going down when it doesn’t cost $550 for the $105 ram I purchased 1 year ago. Yes consumer prices lag commercial prices yada yada, I think any hot takes are pointless until we see lower prices or far more convincing evidence it’s coming. When it costs basically a MacBook neo for 32gb of DDR5 ram it’s hard to hear “ram is coming down for sure”

hyperpape: > Magnificent 7 companies are increasing capex to their biggest ever to differentiate their tech from each other and the big AI labs, but the key realization is that they don’t have to spend it to win. It’s a defensive move for them, if they commit $50B, OpenAI and Anthropic need to go raise $100B each to stay competitive, which makes them reliant on investors’ money.Stay competitive how? If the Magnificent 7 aren't spending the money, then how could it possibly hurt OpenAI/Anthropic to not raise equal amounts of money? Maybe you can pull together an explanation, but this author didn't even try to do so.This piece seems poorly thought-out, but well designed to get shared.Promote writers who will actually explain their claims carefully.

dist-epoch: There are private companies which rent/buy GPUs, run open-source LLMs on them and sell the tokens. They absolutely make profit, and their clients think they get a good deal and are buying the tokens.

guzfip: > Software is or was one of the few remaining arenas wherein a person can find a consistently.I want to add something additional to this: it is one of the few fields that can afford middle or upper middle class lifestyle and is accessible.I have no doubt if I could redo my life with the necessary resources I’d be more than capable of putting myself through med school and gone with a secure career that paid more than I ever made in software.But at this stage of life? I don’t have the time or money to spend a decade+ paying some institution tens of thousands of dollars to hopefully maybe have a real career.Once software as a career dies, I suspect many will find themselves locked out the middle class for generations.

WarmWash: It was kind of a flash in the pan moment where you could leave your retail floor manager job, crash course this thing called "javascript" in a 3 month class, and then get hired for a six figure remote job if you could choke out a mildly competent github repo.

rhaen: Tulip futures skyrocketed, it was economic speculation on a useless asset, not supply and demand. Crypto is the analogy, not AI. Given that the major AI labs other than GDM are private, this is even more true.Agentic coding absolutely blew up from demand, users are not being tricked into paying $200 a month, and they’re not complaining about hitting rate limits because it’s useless.

aurareturn: users are not being tricked into paying $200 a month I can't believe people actually believe that people and companies are tricked into paying for tokens. My $20 Codex subscription is so useful, I can easily see myself paying $200 for it.

skeeter2020: >> Building a datacenter is supposed to be a “safe” investment in normal times, so banks give private credit and mortgages to finance them.Except the investment is more like a railway or utility. It generates like 3% return, which is definitely not good enough for the people providing the money, or (in the case of the profitable companies) anywhere near the double-digit returns they make on their technology products. I won't be surprised when we see consolidation of marginal players and abandonment of the losers, just like you can find rail lines to nowhere, and fiber that's never been used.

cma: > RAM prices spiked speculativelyDidn't OpenAI buy up 40% of the capacity all at once?

ajross: No, they signed a bunch of contracts for future deliveries. That's not a supply constraint. The factories making RAM continued operating and serving their existing deliveries, and in fact they still are.Freshman economics would say that supply is fine and that prices shouldn't move. But they did anyway. And the reason is speculation.

leoc: I don't get it tbh. What market participants were speculating here? There aren't futures markets in RAM as far I know, though I certainly don't know much. And the supply constraints appear to have been pretty real (though maybe not immediate) if eg. Valve was begging publicly for RAM consignments. Were there pure-play speculators filling warehouses with DDR5?

SirensOfTitan: I think calling it inflated is to play to a narrative that labor was overvalued broadly in tech.Salaries across industries in the US have remained flat since the 1970s. Calling the one sector that can provide access a middle class lifestyle inflated s to play into a narrative capital is eager to tell, even if OP didn't intend that.

WarmWash: >potentially fail given the current quality of the output.The question is how big the fail is if you measure it in 3 month increments going back to late 2022.

Aperocky: fails are beneficial to an economy. If there are no fails, you end up with Soviet Union.As long as there are more amount of success, then it should be net positive.

dist-epoch: Jensen is already talking about $1000/mil tokens soon.But there is no real higher limit. Imagine a LLM which could answer the question "what does my company need to do to beat the competition?". And then realize that the competition asks their LLM the same question. So now everybody is bidding the price up or using more tokens to get a better answer

noelsusman: That analyst was talking about subsidizing tokens through the subscription plans, which is a different claim.

infecto: Ty for sharing and agree. I think there is confusion with some folks in the comments for this post confusing inference profitability and plan profitability. Most plans as we can tell are probably teetering the line of profitability and that’s why we have seen some like Cursor really tighten how many tokens you get.

martinvol: RAM prices haven't crashed yet and it'll take time because it has to propagate within the supply chain. Micron is +20% from the top already https://www.investing.com/equities/micron-techStock price is the best forward indicator I can think of

titzer: The article says "...and RAM prices are crashing because new models won’t need as much," and I went and read the link. The link was a puff piece for a very specific compression mechanism that...no one is using?I do hope that RAM prices come down but this was just wishful thinking.

lotsofpulp: That prices change from one point in time to another is a trivial fact.“Inflated due to a speculative environment” is not an accurate way to frame labor prices that held for many years. At that point, the prices were simply high due to high demand relative to supply (compared to other types of labor).

surajrmal: That's not a necessarily profitability thing as much as a demand thing. The only way to improve the supply for those willing to pay more is to take it away from those paying less. Once supply catches up to demand things will change

Throaway1975123: Tulips had literally no economic value. LLM's do.

drakythe: I say this as someone who has used them to boilerplate/scaffold a bit of code by this point: Economic Value of LLMs is debatable, if only because they're being too broadly applied.

Throaway1975123: Debatable sure. Not 0. Tulips are 0. They add nothing to anyone's output. LLM's are not. LLM's are not tulips.

martinvol: they have to fight to stay competitive because mag7 can outspend them, but my hypothesis is that they wont need to ultimately.

skeeter2020: This is changing the narative. Nobody really cares about tulips and some dumb throwaway comparison. Unless LLMs are worth an awful lot the math here does not make sense. That is both debatable and important.

h14h: My (potentially naive) take is that open models will save us. The biggest markets for LLMs (e.g. coding) are narrow-enough to be served well by smaller models with proper RL. Cursor's Composer 2 (created from a Kimi K2.5 base) is a great example, and I expect it to be the first of many.The wealth of great open models provide an excellent base for fine-tuning, distillation, and RL. I see a lot of untapped potential in the field of bespoke, purpose-built models that can be served far more cheaply than the frontier competition. I would not be surprised if we see frontier-adjacent experiences running comfortably on a Mac Mini by year end.With frontier models seemingly hitting diminishing returns in quality, I struggle to see a world in which gigantic, expensive, general-purpose models don't become increasingly niche.

naravara: I think they’re losing money because they have to amortize the costs of training the models in the first place, which is where most of the resource sink is.This is why they were freaking out about DeepSeek just taking the trained model weights and slapping an interface on it.

malfist: Thats like saying a restaurant is profitable because they're making money selling meals if you ignore the costs of ingredients.Of course they are profitable if you ignore their cost to bring a product to market.

infecto: The problem with that comparison is restaurants largely don’t have much room to adjust price or optimize cost. The AI industry is too new with many unknowns right now so investors are willing to take risk. For the hyperscalers the bet is that being left out is going to be a greater loss than overbuilding.

ZitchDog: > they have no incentive to do this at a lossAre you sure? Surely there is a lot of interesting data in those LLM interactions.

wongarsu: Many of them are promising not to store any of this. Of course we have to trust them, for all we know they are all funded by various spy agencies

sempron64: It's ridiculous to call this tulips, in the sense of a speculative asset whose price depends on resale. A more similar recent example is the dotcom boom and bust based on building internet infrastructure, or the 2008 crash which was based on cyclical infrastructure overinvestment. These crashes were characterized by demand growth not keeping up with investment because the target markets were tapped out. Not clear when we'll get there with AI. The consumer market seems saturated on chatbots but we're not even close to saturated for b2b or self driving for example. And this discounts other new technological offerings which may unlock larger consumer markets (products where people are willing to pay $100 a month instead of 10 or 20)All that said the dotcom boom is extremely analogous and that crash was quite bad.

skeeter2020: dotcom was maybe 100B a year focused on the US and mostly VCs. AI is perhaps 250B global VC (with more than half of ALL VCs concentrated in one sector) and another 800B+ from non-VC. These numbers are basically a guess but structurally we are set up for something much, much worse.

naravara: That’s the wrong analogy. Model training is more like the setup costs of developing the menu and training staff. What’s driving the costs is important when talking about financial sustainability. If it’s mostly coming from optional R&D investments instead of the direct costs of producing the food then you can simply not exercise the option and be profitable. If it’s more coming as a variable cost that scales with each meal served that’s a very different situation.Yeah it should be factored in, but it’s a different set of implications for long term sustainability. They don’t actually need to test and optimize a new menu every day or week. If they decide to just stick to the same one longer they can get way more return from each dollar spent on development. It’s just that right now the rate of improvement you get with training is really high and nobody can afford to fall behind their competition.

schmidtleonard: Overinflated relative to the wet dreams of the ownership class.

gruez: It's not exactly stuff of "wet dreams of the ownership class" to say that of the possible white collar careers, software engineering is pretty hard to beat in terms of salary vs work you need to put in.

scw: TurboQuant has a specific benefit by compressing the KV cache at a negligible cost to quality. That mainly means that the context lengths can go up in models for the same amount of memory, however the KV cache only accounts for something like 20% of the overall model size, and this will not dramatically decrease memory demands in the way that some of the more sensationalist reporting has stated.

wavemode: people used to say this about search engines and web browsers, as wellregardless, eventually Google became the universal default for both. When it comes to software, the average person doesn't shop around for the technologically optimal choice, they just use what everyone else is using.

baq: my point is today there is no clear winner. opus, gpt 5.4 and gemini have different strengths. google search was running circles around competition in basically all use cases.

Throaway1975123: No it isn't changing the narrative. Tulip bulbs were a huge bubble based on speculation, completely. No one ever used a tulip to create a piece of software, or anything else. Their economic value was precisely 0. The whole thing was based on a bubble. LLM's may be IN a bubble, but they aren't tulips.

lnfromx: Okay lets suppose all those companies are profitable if training would stop today. What if token demand is shrinking ? I think big parts of the current demand is artificially build by e.g. FOMO and marketing without real value generated by them. There is no indication in economic data about some productivity boom resulting from AI usage. Next thing is Energy costs - that will soon eat into profitability too. I don't see how this bubble can't burst.

martinvol: I don't think token demand will shrink because we're still just learning how to use it, demand will skyrocket. The problem is what price we'll be willing to pay for it, specially if competition keeps soaring.

infecto: Exactly. I don’t know why folks take so much offense to it. You could absolutely do just as you describe. Spend 3-4 hours truly working while enjoying the lunch today sessions and in-house lunch and barista. I definitely benefited from this and I am not ashamed of it but absolutely it was this weird moment in time.

gmerc: consumer ram is starved by production capacity shifting to HBMs. Hbms dropping in price would not affect consumer RAM on any immediate timeline. Also, as pointed out by many, Jevons Paradox

infecto: Don’t confuse what I say. Bottom line these companies are not profitable yet but it is profitable to serve a token via the API. They have increasing demand, not enough supply, models are getting better on quick timelines. For sure there may be some losers but it’s not hard to see that that token serving can be a profitable activity.

wavemode: tbh I don't think this use case is going to be as big as people seem to thinkthere are a lot of reasons, but in brief - I think AI desktop use is a product that the average person isn't going to get much value out of. to make an analogy - the creators of Segway thought people would buy them in large numbers, but it turned out most people don't mind walking manually (or at least, don't mind it enough to spend money on a scooter). I think makers of AI Desktop Use products are going to find out the same thing as it relates to everyday tasks like checking email and shopping.

H8crilA: I was thinking more remotely managing the computer in a warehouse, replacing the mouse of an architect, or some physical object engineer.

matheusmoreira: > Seriously, what value are tokens providing other than justifying layoffs. Concretely. Today.Claude helped me implement a ridiculous amount of features in my programming language. It's helped me migrate the heap to an easily moveable index-based object space. It's helped me implement generators. It's helped me implement a new memory allocator. It's helped me fix a ridiculous amounts of bugs and make a huge number of small improvements everywhere. Its ability to provide me repository wide code review was a game changer for a solo developer like me. And it's doing so much more than that.It's actually addictive to build things with Claude. The usage limits are starting to make me anxious, just like withdrawal syndrome. I applied for their open source max subscription program even though I'm too small for it because who knows, I might get in anyway and it costs nothing.AI is quite literally a world changing technology. I hope the open models keep steadily progressing and that hardware remains available to all so we can run our own models on our own computers one day.

fsloth: [delayed]

vekker: > Seriously, what value are tokens providing other than justifying layoffsLike the OP said, it's incredible how polarizing this debate is. When I read comments like yours, I feel like a significant part of the global workforce in IT must be living on another planet? Or they never really used Claude Code, Codex, OpenCode, ... intensively before because of company policies?I legitimately am at least 10x more productive than a year ago, and I can prove it in number of commits and finished monetizable features developed per day. Obviously my workflows still very much require an active, constantly context-switching human-in-the-loop, but to me there's absolutely no question both output volume & quality have skyrocketed.

Throaway1975123: I created 5 websites this year and am working on 3 prototype games. For free. Without any knowledge of coding beforehand.

sigmoid10: Yeah, I also stopped reading at that point. If I want a bunch of random, made up facts to sell lukewarm opinions or steer the uneducated masses, I'll tune in on a Trump press conference. Why does this feel like someone is desperately trying to make reality mirror his flailing market bets?

Forgeties79: Sometimes it's real easy to see who has risky short positions right?

paulddraper: You can rent a H100 GPU for $4/hour. [1]300k tokens for that hour.OpenAI charges $6.Those are pessimistic assumptions.[1] https://lambda.ai/instances

drakythe: 3.99 at 8x instances, with a minimum 2 week commitment. Good luck getting 70% usage average during that time. Useful when you're running a training round and can properly gauge demand, not so great when you're offering an API.

infecto: Is it not a good penciled number? It helps set the directional tone that at inference cost is being covered.

drakythe: It says the numbers are theoretically possible. Requiring a 66% usage to break even when 100% usage will piss off customers by invoking a queue means it’s a balancing act.“Technically correct. The best kind of correct”. So inference may technically be _capable_ of being profitable, but I have question’s about them being profitable in _practice_.

aurareturn: Even if TurboQuant, which was released a year ago, drastically lower RAM requirements, AI labs will just release bigger models.Jevons Paradox. When are we going to learn that efficiency gains in AI does not decrease hardware usage?

functional_dev: valid point, it reminds me of video games. GPUs got faster, devs pushed higher resolutions, more complex lighting instead of saving power :)

marcyb5st: Isn't salary a proxy of how hard to replace one person or a group of persons is or how valuable they are?There was a surge in demand for SWEs and scarcity brought salaries up. Are them too high? Hell no. On average, my colleagues and me generated ~2M$ each in 2025 for our company, while we get payed a fraction of that (grants and bonuses included). If you look at net income per employee we are at around 700k each in 2025.Additionally, employers try their hardest to drive costs down (eg. offshoring as much as possible, everyone doing layoffs at the same time, ...) and average/median salaries remained high. If the salaries were overinflated those numbers should have came down I believe. The fact that they didn't makes me think that it still is a scarcity problem not an overinflation one.

coldpie: I do wonder how much of the apparent demand is driven by companies automatically running these things when users didn't actually ask for it. For example every web search I make now has an AI response that I scroll right past. I'm sure that counts for someone's token usage data, but I got zero value from it. This is happening in almost every software product now.

raincole: It is quite hard to imagine how the demand is saturated now. I think any company that uses a sliver of AI will happily increase their token consumption 100x if it's free.

fcarraldo: At the top end (say, top 100 tech companies) it’s pretty high indeed. Public companies, for sure, as otherwise their stock price would tank. It’s not uncommon in this industry to have margins above 70-80%.But there are thousands if not tens of thousands where the profit per employee is minimal or negative.I can’t find a source for all tech (the data wouldn’t exist for private firms anyway) but I think it’s telling to look at this list, scroll down to about the middle and look around at salaries you or your colleagues are pulling. Software revenues are certainly high but the industry is afloat because of these high margin businesses creating returns so that low margin businesses can exist. Without the massive infusion in upfront capital, very uncommon in other industries, it’s simply not sustainable.Typically a market that’s buoyed by its top performers but has significant amounts of capital tied up in under performers is called “a bubble”.https://www.trueup.io/revenue-per-employee

beepbooptheory: I don't really get the last bit. It's hard to imagine what a new fangled "frontier model" could do that would blow anyone out of the water. Like what does this look like? Really good benchmarks? Who cares about that anymore?

layer8: Not hallucinating anymore would be a good start.