My company pays for a $100/month Claude seat for every developer, but every dev uses $5000+ worth of tokens a month. Explain how this is sustainable for Anthropic.
>>110230418the only thing that makes sense is they're pulling an Uber by selling at a loss until you everyone gets dependent on them at which point they'll jack up the price
>>110230431thisevery startup does this ever
>>110230418Anthropic’s revenue has already surpassed compute costs. Now imagine 5 years from now when every programmer is dependent on AI and the prices are doubled.
>>110230431>>110230450How long can you go on selling your product at a 98% loss though?
>>110230451It hasn't.
>>110230451>the prices are doubledIf you have to pay $10k a month for Claude tokens in addition to the employee's salary, why wouldn't you just hire smarter devs who can code without AI?
>>110230460until enough are dependent on the product that they will be profitable after price jackinglike he said
>>110230465hell, why use AI at all! It would be cheaper to just have humans. Dumbass.
Larger companies pay the unsubsidised API rates which allows the rest to freeload.Check out their website for enterprise plans. If a company is too big they have to pay API.
>>110230460depends on how long venture capital and investment lastslast I heard Sam Altman tried doing another raise to extend his runway at OpenAI and only got half the amount he wanted
>>110230495How do you become dependent on AI? Is every human on earth going to forget how to write code in the next few years so there will be nobody left to hire?
>every dev uses $5000+ worth of tokens a monthYes anon, I'm sure running those GPUs wastes 5k of electricty a month. The fuck are you talking about.>but my hardware costs and R&DThose are not costs consumed when you use tokens, they are purchases amortized into the price structures, and only a retard would think you can accurately measure them without making up some arbitrary statistics about ROI or the speed at which hardware will progressYou know how much it costs to generate a minute on video on Minimax H3 at home? Less than a dollar. You know how much OpenAI was charging via Sora? 70 cents a second.
>>110230418two problems with your analysis.First, you're assuming that because those tokens would have cost your company $5k at API pricing, that this reflects the actual cost of those tokens to Anthropic. In reality their profit margins aren't known. Maybe they break even at $100 and enjoy a 50x profit margin when someone pays API prices. I doubt it, but you get my point.Second, as I understand it, only small companies can do what yours does. Big companies need to pay for the enterprise plan which basically involves paying API pricing, IIUC.
>>110230460> implyingListen retard-kun, the retail price is never ever the cost the company has. If it was, no company would make money and they'd all go bankrupt.The retail price is always the literal most ridiculous price they can slap on it without either being fined or laughed at.New industries make the prices more ridiculous because there is no precedent and little to no competition.It'll be even more ridiculous still when it's venture capital backed because VCs are just retards with infinite money that will ask you about whether you have more customers or usage and not ask what you're charging or what your margins are.The CLI price being 98% cheaper is all indications that this is literally all made up. Like a 12 year old designing an economy for his make believe playground session.
>>110230545You shouldn't compare any sort of ML model to LLMs because all of these have tiny amounts of parameters in comparison. H3 is just 30B, for example
>>110230589How much money has Anthropic made so far?
>>110230418>$5000+ worth of tokensWho says those tokens are worth that much?By what metric?Anthropic can put any price on their tokens and then pretend to give you some amazing deal, so they are no viable source for that a token is really worth.
>>110230598I'm comparing a video model you can run at home to the video model OpenAI was selling you for $0.70 a second and allegedly "discontinued because it was too expensive". What the fuck are you going on about?
>>110230605Cash net profit is deeply negative and will remain so due to the $900 billion debt, though most people judge them by operating profit instead and the latest quarter is only slightly negative.
>>110230541You don't have to forget how to code.You just have to be unable to figure out how your own codebase works without having an AI explain it to you.
>>110230636>judge them by operating profit instead and the latest quarter is only slightly negative.Their operating loss/profit is a secret.They claim to make a small profit but that's by their own secret metric where any cost could be excluded for whatever reason.Right now they are just trying to create hype before their IPO so they will make the wildest claims.But once they are publicly traded they have to publish operational profits according to standards instead of their own fantasy metric.
>>110230545How long does the hardware last at the rate it's being used in the datacenters? At least year or two? Well, MI300x costs about 50k, nvidia are probably twice that amounth and I doubt that even haiku or luna fit on only GPU, it's probably 2 or 3 that have the context pre-warmed at all times.
>>110230418We have an enterprise plan and I spend $4000 on claude every month
>>110230732That's an extra $48k a year your employer could be paying you instead.
>>110230418Because the $5k "worth" is a meme number. In reality the cost of those $100 in monthly compute is way less than $100. They are printing money
>>110230499>erm you're dumb if you think AI is more expensive than humans What kind of world do you live in?
>>110230729A B300 cluster costs ~300-500k, 1k users using paying $20 (which would give them an average limit of ~2 hours of compute a month) covers it in a year or two. The amount of shit they buy and which models they run when and where varies a lot, including making models dumber as time goes on.Luckily for you and most clueless people in this thread, Anthropic has filed for IPO last week, meaning they have to actually publish their finances now. Have fun https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm You'll notice they had a quarter of operating profit (not the last one) and have closed the gap significantly.> But they have a lot of debt, 900 billion!What the fuck do you think the IPO valued at 2 trillion is for?
>>110230740Without it my output would drop to like 30% of what it is nowIts good justification to not hire more people at least
my fucktard company gives us $50 a month claude api budget (not subscriptions, a $20 sub would be 10x better) and if the job market wasn't so shit I would quit over it
>>110230965lol, i'd use it up in a day probably
>>110230418corporate socialism
>>110231004they pay me 6 figures to sit there rationing the retarded models. It really makes me have to just accept my time is being wasted by fucktards for the nice paycheck.I have a personal $200 codex plan that really is about 100 times more than this shit
>>110230885> covers it in a year or twoBusiness is not about covering costs, it's about making profit. You haven't counted in the electricity costs, too and at 6KW per hour, in some european countries this will account to over a 1k a month. So, in 2 years you will have a barely payed off GPU cluster that's already outdated and might not even survive 2 years of such harassment.80 billion direct losses in a year is very optimistic, surely, they will make money, just believe, goy.
>>110230431Because technology never improves, never becomes cheaper, never becomes more efficient. ESPECIALLY at scale.
they start a 800M token conversation, go to lunch, come back and preload the whole convo as cache miss again cause it's expiredturn on out automatic context compacting at least
>>110230450>>110230431Except that there are literally free chink models that perform 90% as well. You can change model provider like nothing, there is literally 0 lock-in and they won't ever be able to have lock-in. They must keep the prices relatively low or nobody will use their ai.
yeah i doubt it's anywhere near profitability for softdevs since they're heavy token users.They probably get the rest of money back from non tech fags who use it to write email, generate presentation or some shit.
>>110230465The $20 plan can replace 90% of programmers *right now*. Even if prices increased 10 fold, it would still be nothing compared to the salary of a developer.
>>110230418Because the US government will actually bail out these AI companies with our tax dollars when the market finally collapses. They should nationalize them in that case but they won't. Just going to write them a check.
>>110230418>Explain how this is sustainable for Anthropic.The same way a drug dealer makes money.First you create an addiction, then you profit like crazy.
>>110230956it's also a good justification to fire your jeet ass and just have the AI do 100% of your "work"
>>110230418They will jack up the prices after nobody knows how to code by hand anymore and is fully dependent on them. You really are in the golden age of vibecoding right now, it will be unaffordable after the investors demand their returns. Got to get everyone addicted to the drug first though.
>>110230418>$5000 of tokensthat's what they charge for them, but not necessarily the actual cost to anthropic
>>110230418Video games cost $60 and cost hundreds of millions to develop. How is this sustainable?
>>110231221then why are those 90% devs still employed?
>build thing>China copies it inside and out, makes a small improvement and practically gives it awaythis is not a sustainable model for for existing as a country
>>110231300the most retarded analogy i've seen today, good job /g/eet
>>110231131>80 billion direct losses in a year is very optimistic, surely, they will make money, just believe, goy.72 billion of that is fair value option accounting losses.They have financial obligations to early investors that grew because their valuation rose so high. It does not mean that they handed over 72b to someone. Basically the early investors are entitled to a bunch of shares in the company that has to be reflected in the accounting.
>>110231131no, business is about execufags pedalling fraud schemes while riding the company down into bankruptcy and paying themselves fat checks. find next fraud. repeat.
>>110231131> Business is not about covering costs, it's about making profit. Yes, and if you bother to read the IPO filing instead of making shit up, there has been operating profitYou're the kind of fool that thought Amazon was a doomed company that everyone should sell off in 2013 when their P/E ratio was 1000 because "capex bad"> in some european countries What kind of retard builds data centers in Europe for things that don't need low latency? Lmao
>>110231156>ESPECIALLY at scaleBy scale, you mean the largest scaling possible: The PC of people at home.LLM-as-a-service is a dead end. Anthropics buisness model is a model of suicide.That's why they need the government to stop development. They beg them for a pause. They need the regulatory capture because otherwise they are dead.
>>110230418>Explain how this is sustainable for Anthropic."the first sample is free"
>>110230418Their plan is to make it up on volume
>>110231440The whole model is free, if you go Chinese.
>>110231131>6KW per hour, in some european countries this will account to over a 1k a monththats what happens when you sanction the whole world in some grand virtue signaling sperg outThe EU deserves it.You have no reason to exist and whoever takes the land, when you are dead, will use it more efficiently than you do now.
>>110231476why do browns hate eu so much on this board?
>>110231489gob bless iran, god bless the houthis, god bless the resistance, death to the zionist axis
>>110231489Natozists shouldn't have bombed their countries
>>110231438> The PC of people at homeNobody is selling GPUs with 2tb of ram for a price anyone can afford any time soon. Maybe in 2040 you'll have a point and we'll have another transition from "mainframes" to small computers, but it's going to take a long while and it makes some assumptions about the scalability of hardware like we were still under Moore's law.
>>110230431My expectation over the last year's been all the frontier labs are doing this.Build your big projects while you can and find access to compute in case all this stuff gets nerfed in six months or a year.
>>110231674but that's the point, if you slop up a big enough project you won't be able to change it anymore without ai assistance. It's pretty much dead code at that point.
>>110231204Not only do the free chink models perform 90% as well, stuff like OpenAI, which is still cheaper, has been consistently only a couple months behind, and the difference between their models and Claude is negligible for 99.9% of user's needs.The Claude hype has been a meme for close to a year. You don't even see reputable people hyping it anymore (unless they literally work there), it's all students, jeets, and teenagers who clearly don't know what they're talking about exclaiming Claude is the best and nothing compares.
