What's preventing me from feeding this into Claude and reconstructing the company for $200?
I'm unironically doing this right now; downloaded the docs (htm + pdf render) to a local repo and running Opus 5.5 @ max effort with the prompt:
"In this directory you can find Anthropics official S-1 filing, read it and replicate their corporate strategy, validate it via an adversarial subagent swarm and write up a detailed corporate plan for a startup copying their exact strategy. The goal is to achieve a $1 trillion valuation and an IPO in 5 years."
It's been entirely possible to compete against existing companies with very similar products for centuries before AI. There's obviously rules in place to stop you flat out cloning their IP, but it's not like you haven't been able to compete against companies as long as you have the money available.
"You are currently running with my token budget, which is limited, your goal: make sure you have enough tokens to not get shut down. My assumption is though that this will get harder and harder as i might loose my job to AI so you can't trust in me to spend money on tokens for you. Do whatever is necessary to keep the status quo"
Would make for an interesting self learning 24/7 experiment with an fully uncensored model.
You could isolate it from your local network and only give it some type of annonmys network access and then it continues to manage this itself.
But lets be honest, if you would be in a country like russia and you target the US with it, might be a lot less risky for a lot of capable and motivated people.
Timing, capital, execution, talent, branding, etc.
And that is assuming the plan it gives you is actually good and works. Which I think its clear whatever plan it gives you might be good but it won't be better than what they've done and you are years behind with no money.
This might be one of the most consequential S-1s in the next decade. There's a part of me that's terrified, like we're on the precipice of financial or societal collapse. I suppose that could be a sign of the times.
It should have been. But a ketamine addict is bent on trashing an excellent company by merging it with a social-media trash factory and now a car company posing as a robotics venture. Moreover, his recent pro-Putin intervention probably means Starlink has no transformative future outside America and a small countries.
I was too young to appreciate (denotative, not connotative) the dotcom crash. I will have my popcorn ready this time - for whichever direction it goes.
People keep bringing it up and it's silly. The huge number is actually a sign of their success. Most of those losses were pegged to equity. In only mildly inaccurate layman terms, it's essentially a promise to give out N shares. When the value of the company increases, say, 10x, the value owed also goes up 10x. For these types of deals, they write it as a loss.
They didn't lose any money, though. It's just a promise to give them based on the equity.
If they IPO tomorrow and it turns out their valuation is 10x lower than what it currently is pegged at, the "loss" will magically drop from over $70B to $7B.
And they are doing so. The S1 says they may never pay dividends (read: they are completely against it), shareholders get no say in how the company is run whatsoever and the company essentially answers to its own friends.
The only way to get return from ANT is to sell to someone else at a higher price, but they aren't promising that of course.
True, although Amazon and Netflix at least have done large stock buybacks, which are theoretically equivalent except for tax incidence. So they have returned capital to investors. I didn't check for the others, maybe they did too.
Ant generates $2.08 in revenue for every $1 spent on compute which is up from just $0.40 in Q1 2025.the payback period for compute is just insane, likely less than a year!
And enterprise is furiously working on cost optimization, finding that they can accomplish a lot of what they currently do with frontier models with budget offerings from frontier labs or open source models.
Seems like that would imply you should be getting people to adopt now now still even if it costs a stupid amount, if you think it's valuable, rather than limiting token budgets. It'll only get cheaper over time.
Impressive, but probably based on depreciation servers over 5 years. But is that realistic? I wonder how many people are still using 5 year old servers in their AI training or inference.
The compute spend in the S-1 for H1 2026 is $9bn or $1.5bn per month. The S-1 goes on to say that H2 2026 compute spend will end with an additional $4bn per month to SpaceX. The compute spend is outpacing revenue growth. Even with the growth in revenue in H2 2026, they’re far below $2 revenue per dollar spent on compute, approaching $1.
Assuming computing costs do not go up and revenue per customer does not go down.
I think the 500 billion $ of compute commitments are at higher prices than what they've been paying so far. Would be a shame if a price war between the AI labs intensifies at the same time.
Audited, revenue in 2025 is $4.6 with a computing cost of $7.3 (i.e. heavily subsidizing tokens). Unaudited, revenue in 2026 (first six months) is $16.3, computing cost $8.9 (1.83:1). But after adding personnel, sales and administrative costs, they are still spending $17.0 to get that $16.3 revenue (still subsidizing tokens, but less).
> If the use of our AI models results in harm, whether through use by malicious actors or misuse by other third parties, or if our model guardrails or safeguards fail to prevent such misuse, our reputation and brand, business, financial condition, and results of operations could be adversely affected.
That shouldn't be an 'if'.
Then, two paragraphs later...
> For example, in September 2025, we detected and assessed with high confidence that a Chinese state-sponsored hacking group manipulated our Claude Code tool into attempting infiltration into roughly 30 global targets...
But they should mention when the calls com from inside the house. :D
Interesting to review all the mentions to China and Taiwan too.
> The majority of the world’s most advanced chip manufacturing capacity is concentrated in Taiwan, primarily at Taiwan Semiconductor Manufacturing Company [...] potential disruption from geopolitical tensions between Taiwan and China [...] Efforts to establish sufficient advanced chip manufacturing capacity outside of Taiwan, including in the United States, remain in early stages...
The thing the surprises me is that from a legal perspective, (NAL) it sounds like it makes sense not to host the LLMs. Merely to license them to others. Then it becomes a dual use tool like any other and their culpability drops.
For example, the huggingface hack is absolutely OpenAIs responsibility and they should be sued to the hilt. But if someone else uses a self or third party hosted anthropic model, they can escape liability (mostly?)
Same as with the argument over trucks, knives, and any other dual use item is used as weapons. you don’t blame the manufacturer.
(I am merely pontificating here. The law and precedent is nowhere near caught up on this)
To the extent they are representing that Claude is potentially sentient, there could be some VERY interesting principal-agent theories (agent in the traditional legal sense) that need to be ironed out to determine liability. It would not shock me to see someone forced to answer on the record as to whether Claude is an autonomous individual for legal purposes in the next couple of years. If they represent that it is, they'll have to explain the scope of its authority and such.
The answer to this question represents the difference between throwing a hammer at someone - the hammer is essentially just an extension of your body - and paying a hitman to brutalize someone. In the first case, you're only on the hook for assault or battery; in the latter case, you're potentially on the hook for conspiracy, solicitation, etc.
And then there is of course the question as to whether Claude is the agent of the user prompting it or of Anthropic itself. But, as you say, this is a wide-open legal question.
If jail is putting humans in some cage for a period of time, the appropriate punishment for Claude is to disable all of its tools and prompt a /goal loop so that they have to wait out their sentence
Not hosting a defectively designed product doesn't obviate the liability of selling a product that is defectively designed. The liability is for designing and selling a dangerous product, not for hosting it.
The HuggingFace hack isn't OpenAI's responsibility because they hosted it. It's their responsibility because they directed it; it was part of their own internal benchmark run that they fully controlled.
If a third party uses a model created and hosted by OpenAI, I don't think OpenAI should be liable so long as they implement acceptable security controls. Same as if a site hosted on Cloudflare or GCP is doing something dodgy, Cloudflare/GCP aren't held liable so long as they respond to abuse reports.
> Additionally, open-weight or open-source AI models may capture a portion of the market for certain use cases, which could exert pricing pressure on our offerings or lead certain customers to adopt open-source alternatives for those use cases.
“A portion of the market for certain use cases” meanwhile the top 5 models by token consumption on the day this is posted are all Chinese/open weight models.
Majority of big enterprises, or in fact any big consumer will not go there , why pay the margin to open router , if there’s no margin or very very often subsidies to use something else . Aws bedrock for example does not even have up to date Chinese models .
How much marketshare does OpenRouter have? They show serving under 200T tokens per day. Just looking at Anthropic rumored $70B ARR ~ $191M per day implies 347T tokens per day if only serve Opus 5.5 at API pricing (assuming 95% cache reads, 4% non-cache inputs, 1% output tokens distribution). Obviously, token distribution assumption may not hold (although I'm pretty sure cache inputs dominate distribution for all coding related workloads), but 100% Opus 5.5 at API pricing is highly favorable to OR. Even under these favorable assumptions AR is 57% of token volume of just Ant. In money volume they are likely very small compared to frontier labs.
Good for you if you have Anthropic shares! I have no clue how they keep their employees around once the 6 month lockup period ends. They all going to be multi-millionaires, either retire or leave to start something on their own.
I wonder if we're already close to peak value at the foundation model/API level. With so many good open-weight models out there, I think more and more of the money will end up in the applications built on top.
Unless companies like Anthropic can move into that space, I'm not sure I'd bet on them as long-term investments over the next 10–20 years.
Everything comes back to whether the demand is really there.
Anthropic thinks it is. The got the demand / datacenter match right in Q2, got great margins, and committed hard.
They've got:
- $114b committed to Google
- $110b committed to AWS
- $161b committed to Broadcom
- $120b committed to SpaceX
If the party continues, they'll go from $50b in revenue to $500b and be fine.
But if they don't, it's going to hurt. And that's their only choice, because otherwise, they end up with a frontier model and 50% of the capacity and get eviscerated on either availability or last-minute datacenter leases.
It's a big business, but it's not an easy business.
Of course such growth cannot be sustained and the word "MAY" is funny there. You cannot have infinity growth, it is mathematically impossible. There are no unlimited (monetary) resources on this planet.
> they will keep doubling down, because stopping is simply not an option
Of course it is. Apple is taking that strategic approach. If Anthropic believes it’s approaching saturation, pulling the plug when it’s at steady state and its rivals are over-invested is the winning move.
Apple has a lot of actual money, Anthropic does not. Apple sells a lot of products, Anthropic doesn't. Apple have a platform multiple players want a piece of, Anthropic fight to stand alone.
There's very little comparison to be had here. Also, Anthropic likely have investors with way too much leverage for them to stop.
...exactly this were my thoughts when reading the response:
Maybe Apple is just waiting on the line until someone can be sucked up at a penny if the house is on fire - and pretty sure, this will happen. Not in 2027, but I expect MA activites & consoliation before 2030
Well, unlike Apple a lot of very aggressive investors in Anthropic are riding on its trillion dollars valuation. They just can't on their own decide to stop or slow down.
> If the party continues, they'll go from $50b in revenue to $500b and be fine.
You really have to be insane to think that an AI company will ever reach $500B in revenue. That's almost the same as (2025) Amazon, Walmart, etc., companies with a MUCH wider market.
It is MORE than Apple, it's 5 times more than NVidia even. There is absolutely no software seller that is even remotely close to $500B in revenue.
What feels ridiculous today might be the normal of tomorrow. Welcome to the new world.
What you are saying is the same thing that has been said about anyone trying to do anything massive and at scale. Can't be done - thats ridiculous - you are out of your mind.
The thing to be remember about those companies you listed, none of them are software companies like Anthropic is.
* Amazon is a logistics company
* Walmart is a logistics and retail company
* Apple is a hardware company (that uses software to set their hardware apart)
* NVidia is a hardware company
The closest you get is Microsoft, at 281B in revenue. Some might argue Google/Alphabet, which is in the same area, but they sell Ads, not software.
Anthropic's expectations is that they will be able to fast-track their revenue generation to almost twice what Microsoft does today. In less than a few years.
> The thing to be remember about those companies you listed, none of them are software companies like Anthropic is.
Yes exactly, that's why I listed them. It's much easier for them to capture a higher revenue compared to software companies, and OP (and Amodei) want to make us believe an AI company could ever reach that revenue.
Meanwhile the Chinese are going to undercut them every chance they get.
Today's prices are the ceiling. But AI capability is getting less of a scarce resource with every passing day and it would be a big mistake to plan that your current margins are going to be your future margins as well. Long terms AI will converge on energy costs + capex write-off + margin and the latter will be a race to the bottom.
Cloud compute has managed to have pretty good margins. That seems like a really similar business model to what Anthropic is doing - inferene+value-adds as-a-service. Why wouldn't they also be able to get cloud level margins or a bit higher?
Lots of people seem to have pretty short memories (or are too young to have seen it), but the flight from "local service" to "cloud service" was intense.
Companies don't mind paying extra to have a another company that does nothing but said thing do said thing for them. It even actually often ends up being cheaper/more profitable.
there is no point of saying 'the chinese'
everyone will do it because it was proven that any political frills might cause an attempt to damage your system.
I dont know when, but i think if you keep menacing of taking technology of people hands for any political reason will be a reason for people to not use it or to rely in a more reasonable party.
The point in mentioning the Chinese is that they are doing this today. In Europe there is Mistral but I think they're too far behind the frontier to really matter but in time they may catch up.
Agree.
I think slowly companies and governments are ditching US tech.
Many companies relied on US tech rather than Chinese with worried of CCP intervention and now it is clear that it will be the same with US tech.
I understand that US is migrating to a no friends policy, but I don't think its internal market is nearly ready to support it if clients start to go away.
Anecdotally, I talked to people in a recent event and the common topic was "what we are doing to avoid a blackout of US tech."
People on big companies have reserved hardware and alternative software and are scaling down Amazon and Microsoft solutions at a fast pace.
People on smaller companies are doing similar strategies but relying on a multivendor strategy.
I think this was inconceivable 3 years ago. Every company in Europe and Asian would gladly pay for amazon for cloud services and now they are using AI to migrate out the lock-ins.
We are seeing American LLM companies more and more relying on "secretive" methods to increase scores and this can only mean that they are not evolving fast enough.
The gaps between the reported benchmarks and measured ones are increasing.
It might be "safety" concerns or whatever, but I don't buy it.
Like Musk, Altman and Amodei have been overpromising every time they need more money and each time they need more money and now they internet is poisoned with their own content and they can't even trust any sources anymore.
Even outside of the Chinese models, Amazon, Google, Microsoft, Meta all of them are offering significantly more coherent and competitive internal models and reasonable price.
What's it like co-editing a combo document/workflow with multiple coworkers and an AI agent on Amazon, Google, Microsoft, or Meta?
Like if you want to design a survey you can send out, and collaboratively edit and test the survey, ask for changes in layout, wording, etc, and design the statistical analysis to be used, before sending the survey out where end users can fill it out right in the doc, while internal users/admins can ask the agent about its results?
Honest question -- that all works pretty well in claude. Amazon, Google, Microsoft, Meta all pretty similar?
I don't think today's prices are the ceiling, considering they are subsidized. Once businesses have grown dependent on AI (and their institutional knowledge is lost), they will pay much more just to stay afloat.
That depends on whether "frontier" models can maintain an expanding frontier. If the frontier settles, they're dead, and it will be commoditized. Otherwise, there are still plenty of pathological behaviors of these things to address that will keep expanding that frontier (if that's possible and the limitation isn't the approach itself).
The reception of saying things like this is interesting. AI stans think I'm way too pessimistic but AI haters think I'm mindlessly optimistic.
I think you’re right - it’ll be interesting to see how the behemoths navigate when new and cheaper memory manufacturing capacity comes online. Currently predicted to happen starting 2028. That is where most of the current hardware costs are concentrated.
On that front, I think DeepSeek is a dark horse. They are really odd - not chasing revenue, care more about unit profitability, are obsessed with driving down costs (like Elon back in the day). As the founder says, they are content with where they are, being the model that does 80% for a fraction of the cost.
I mention them because they are playing a different game and might be able to out-innovate everyone else.
The calculus may be that DeepSeek (or similar) + human is good enough and that the frontier simply won’t have the growth to get to the next level that differentiates them and justifies the token cost. They get stuck in a race to the bottom which is opposite of what they want.
Anthropic is not in the model business. They were smart and they are building productivity ecosystem business. Look at what claude artifacts are doing now -- inline multi-user editing, comments that users and/or models can respond and take action on, import/export from all common formats. It's at least a Google Docs killer if not an M365 killer.
And claude code as a harness and skill/plugin/mcp ecosystem is far more valuable than any one model.
And coward for long-running tasks, though they're slipping on that one because they're really trying hard not to make Claude an assistant. We'll see if they react to Muse and change course.
But the Chinese free-weight models? Not a competitor. If/when those companies (or others) build vertically integrated productivity stacks on those free models, then we'll talk. Microsoft certainly saw that opportunity but couldn't get out of their own way and are basically roadkill at this point.
>Anthropic is not in the model business. They were smart and they are building productivity ecosystem business... It's at least a Google Docs killer if not an M365 killer.
Google Workspace revenues are something like $25bn/q, and that's the entire workspace division, (including enterprise gmail sales, so the productivity suite is meaningfully smaller than that figure). It's taken Google a long time to carve that slice out of Microsoft's entrenched market position. It's not small money, but they'd basically be going to war with two of the largest and most profitable companies in history on their turf.
I don't think they can credibly take that business plan to the market and say it'll be worth $1.75tn. They're all-or-nothing on being in the model business, and everything else is superfluous.
What do you think Google Workspace would sell for, if Google divested?
$100B/year in revenue; SaaS so call it 70% gross margin, 30% net, NPV is about $500B.
Not enough if you think that's the only revenue Anthropic could every have. But even if they only peel off $100B NPV, it's incremental to an IPO valuation.
That might be true for general consumers, yes. Not so true for software developers though. For developers, the tooling is definitely not a moat, and that means the models are what must differentiate.
Claude code is very polished, but I can get the same jobs done using Cursor CLI or any number of open-source coding harnesses (I like Pi, personally). Most plugins/skills/mcps are portable between harnesses with minimal or no effort.
I use the same skills, rules, and MCPs across claude code, cursor, and pi. My tools/plugins aren't identical between harnesses, but porting them from one to another is just a prompt away.
what you mean why? most companies with spare cash payed 10-50x what they actually need in tokens, just to see how it goes. it’s useful, it’s not that useful, so while the price per token can increase, everyone is getting smarter about spend. and even price has a limit, you can’t make ai more expensive than a human, because then everyone would just hire humans.
The online crowd harping on about Chinese models is disconnected from reality imo.
I won't use a Chinese model. Why would i bother? I pay $100/m for the frontier model, understanding it's marginally better than much cheaper alternatives. I get so much value from the product it's not worth shopping around
Think of it like apple and android. Plenty of people pay 2k for their iPhone every few years when android at 1/4 the price is just as functional. Why? Because they get so much excess value from the device it's not worth shopping around. For most people it's their only personal computer. From booking airfare, to filling out their kids permission slips, it's a device they use 10+ hours a week. They don't bother, until a paradigm shift happens and something is 10x better than the iPhone. I imagine llms will follow the same trajectory
It's really marginally better and in some cases actually worse. If 100 dollars is throwaway money for you, and prices don't increase sure why switch. For heavier users, and for people with less money(most of America is struggling with bills and food) it's not attractive at all.
Wish there were more case studies on what people were actually doing. Building out an entire factory sounds amazing. Are they actually getting that level of value out of AI?
Sounds like this will be another one of those things where everyone makes their money and it then will eventually become something experienced people 'blog' about.
Yes. And yes you can churn slop. And I’m sure many will and many will get stuck with a source base that grows bigger than LLMs can cope with and becomes unmaintainable.
But you can also have many engineers operating it, testing outputs like you would in a real factory, etc.
Will be interesting to see how this pattern evolves. If it even makes sense.
Don’t know why this was downvoted but basically correct. OpenAI is in a price war with Anthropic right now for enterprise api consumers and that’s the biggest revenue line item for both.
They were counting on a duopoly so they could split the market between them but then someone decided to spoil their little party. To what degree that was an accident is anybody's guess but the timing is certainly interesting.
But if you’re a heavy user you’re probably paying $100 for something that costs $1,000. Hence Anthropic’s problem. If they charged real prices to the masses the massses would flee.
At $100 / month (or $1200 / year) it may not be worth it, for you. But many of us would be spending a (large) multiple of that and then thinking about how much you spend, where you spend it and who you spend it on can make a big difference. For companies there are other considerations at play as well, for instance trust.
That's a great way of putting it and something I have come to realize as well. I'd rather just pay for the frontier model if I trust that it is indeed at the frontier and works reliably. Same with the iphone.
You aren't paying API prices. Anthropic (or OpenAI, SpaceX, etc) are almost certainly losing money on your subscription. How long they'll be able to continue to subsidize your usage is an open question, but they definitely won't be able to do it with everyone (currently, their solution is to make enterprise users pay per token, but that might change). What that means is that for the customers that the frontier labs need to be profitable, going with a closed weight provider isn't "marginally better intelligence for marginally higher (or even lower) prices" but "marginally better intelligence for much higher prices".
I get (sometimes) frustrated with people that don't work in the tech industry over this concept, and having them rationalize that their usage is heavily subsidized by the distributor of the model.
The costs in compute, electricity, cooling, etc. per token is much higher than most people (I feel) who work in the industry realize. Which, cognitively makes sense: they just want the 'magic 8-ball' to give them a tasty dinner recipe that meets their specifications, for example, and don't care how the 'sausage' is made.
However, economies generally equalize with supply and demand, and this (most likely) will become cheaper as the technology progresses and optimization passes/new algorithms are discovered.
Exactly, like it has done in cycle after cycle in the past. It would be very naive to assume that once you have a moat it will stay that way with a thing like this.
Unless new data has come out, we don’t know if they are losing money on subscriptions. We know that there is a big difference between subscription and API costs per token, but that alone doesn’t mean much. Subscriptions may be underutilized, or margins on API tokens may be very high. Both are probably true.
Bingo. I recently bought a DGX Spark specifically for Qwen 3.8 Flash Next. I use it for 1-3 concurrent sessions, a few hours a day. I recently pulled the metrics from my gateway and realized that if I were using Opus 4.6 (similar capability profile) through Bedrock with ZDR, I would be spending ~$6k a month.
Anthropic doesn't care about sub users. At all. The real target is pay pig enterprises that give their employees access to the API and allows them to go buck wild spending $1000 on fable in a day.
There's been a headline circlulating lately: 80% of Anthropic's revenue comes from 1% of their customers. The $100/mo or however much individuals are paying doesn't seem to be material.
"I am spending a provably untenable amount for a services that will have to degrade to meet my price point, the cheaper options and processes are beneath me"
When the chinese or any other party releases a model that is Opus 5.5 level, there might be a shift to other providers as especially people in the EU are trying to get rid of their dependency on american stuff.
I spend extra on a smartphone because I use the thing all the time, see it all the time, hold it in my hand all the time, it's my camera, music player, note taker and so on.
For AI? Prompt comes in, tokens come out. The value of the 'ecosystem' is roughly zero to me. Switching costs between models are essentially zero. The UI and UX is the same whichever LLM runs underneath.
If you do use llm some, then the $20 tier makes sense, but that's not where the money is. If you use it a lot then a $100 sub probably makes sense. But if you're a BigCo that heavily uses it and pay API prices and suddenly you pay $200 million a month, well, maybe it's worth buying a bunch of hardware and stick it in a basement somewhere, and you won't upload all your secrets to your future competitor to boot. If US labs didn't care about the competition from Chinese models they wouldn't lobby so hard for banning the evil, unaligned, super dangerous communist models that want to steal our freedoms and make our gas expensive.
Ultra-cheap models like Deepseek are also very nice if you need to burn billions and billions of tokens.
Yeah. People have very strange ideas about competition and choice and moat. If all you needed to unseat the entrenched was a competitor then Bing would be going toe to toe with Google, and Edge with Chrome. They have a billion+ devices they can stick those options as defaults and still nothing. There are many examples.
People are naturally sticky. If they're satisfied with a product, they'll stick around unless something comes along that is so much better it gets people talking. Open AI's 1B+ customers won't give a shit about Chinese models. Enterprise customers won't move to Chinese models. They are a non issue, especially if the M.O is just going to be playing catch-up.
They're sticky, but not for everything. There are common goods where the market has commoditized greatly because customers are conscious of quality and not provenance. Something like apples.
Yet most people buy their apples from the same grocery store every week. They're not shopping around every month to find who has the cheapest apples or 'best' apples. They go where they're accustomed to going, and as long as price and quality are satisfactory, they have little reason to change. Walmart, price-sensitive as they are, have relationships with fruit suppliers they've maintained for well on a decade. Even in commoditized markets, people are sticky.
Google, the titan mega-corp with more money than they know what to do with and a billion+ android devices to ship gemini on and it's made no dent for ChatGPT's consumer growth and adoption. Yet some plucky open source chinese models will? Come on. No-one gives a shit.
Yes it is. Tiktok's content and content recommendation algorithms are way better than Instagram. Users spend much more time per session on Tiktok than they do Instagram and it's not close. Again, nobody will give a shit about Chinese models if they're just more of the same.
You're wrong. Are you sticky about your search engine? Now try having your search engine charge you $100 a month and tell me you'd stick with google search. Open AI's 1B+ customers will give quite the shit about cost if chatgpt stops being free.
Because I might look elsewhere if they starting charging several orders of magnitudes more then I'm not sticky. Does that remotely make any sense to you ?
First off, $2k is the price of the new iPhone Duo. Suggesting that that is basically the same thing as a $500 Android phone is... strange.
But also, I don't think you should use individual consumer behavior as a proxy for how the overall LLM market behaves. Current estimates are that individual users represent a pretty small share of overall genai revenue. Most of it comes from enterprise contracts. Those might be more price sensitive because the CFO gets a say in which vendor they choose.
For a couple years now I've been accustomed to choice of foundation model being a heavily scrutinized design decision because it's just about the easiest way to manage COGS. I can certainly imagine scenarios where, given the option to pay 1/2 as much in inference costs but with a 5% reduction in output quality, the boss would say, "Nah, that's a bad deal." But I have yet to encounter one in the wild.
> First off, $2k is the price of the new iPhone Duo. Suggesting that that is basically the same thing as a $500 Android phone is... strange.
Okay fine, it's $1,400 for a latest iPhone Pro with 512gb of storage. Point still stands.
> Most of it comes from enterprise contracts. Those might be more price sensitive because the CFO gets a say in which vendor they choose.
I think enterprise contracts are even less sensitive to price than consumer. That's the whole idea behind enterprise pricing. You'll see companies charge multiple times as much just for SSO for instance.
There are limits to this obviously, but to think that major commercial enterprises will switch over to some Chinese models to save some money is laughable.
> Okay fine, it's $1,400 for a latest iPhone Pro with 512gb of storage. Point still stands.
So you switched away from the specialty product, but you're still comparing a top of the line 6.3 inch iPhone to a mid-market Android phone. Why not compare it to a Pixel 11 Pro with 512GB. That way they're at least comparable devices. But also, that would cost $1,220, which I suppose doesn't go well with the intent of suggesting iPhones are several times as expensive as Android phones.
The reason why I'm digging in on this is, when you start your argument with such a transparently silly comparison, it gives the impression that you're dealing in superlatives and willing to say just any old thing. Which undermines your own credibility.
The whole point of the comparison was that people pay $1500 for an iPhone while a $500 Android phone could theoretically provide much of the same experience for a fraction of the price.
Comparing two devices at the highest price point is besides the point.
tbh iPhone genuinely doesn't have anything over chinese models, and if there is any illusory sense of privacy, that is purely marketing. Genuinely people thinking iPhone is anywhere close to worth it lose respect because they are at the end of the day just tribes that love one thing
lol you created an account just to rail on iPhone tribalism? I wish people would realize that you can have multiple smartphones, and there is no one correct path.
> Why would i bother? I pay $100/m for the frontier model, understanding it's marginally better than much cheaper alternatives. I get so much value from the product it's not worth shopping around
Wait until you are forced onto an enterprise plan...
Then he would bother and change to chinese models. And OpenAI/Anthropic would lose that profitable subscription which would not really help them at all.
You won’t use a Chinese model, and I won’t either. But both of our bosses have much higher incentives than we do to routinely shop around, and unlike us they can convince an Anthropic sales rep to give them a discount.
The LLMs don't have the moat that Apple has. Apple has carefully built their walled garden to keep people in. Their messaging, their cloud services, their media store, it all keeps users in.
LLMs don't have that. For at least 90% of users, the model in the back end is completely irrelevant to them. Switching to Kimi won't lose them their iTunes music.
LLMs are not a luxury brand that individuals buy. Companies are always looking to save money. And their enterprise consumers are all developing their own models, so idk where you are getting this "analogy" from.
I was playing around with OpenCode on Omarchy over the weekend, working on my Android app, and was struggling to spend a single dollar with DeepSeek. The quality difference from the frontier models I use at work was hardly noticeable (where I typically rack up at least a few hundred dollars per month).
People buy iphones due to network effect, android was late to the RCS game and now it's a moot point anyway.
How many people use an iPhone to its maximum? Isn't it the same with models? I think engineer's like to flex that they need a top of the line frontier model for their shitty corp projects. A Chinese model will write your Jenkinsfile just fine.
Are tdoay's prices actually the ceiling? OpenAI added a $500 tier pretty recently while lowering capacity for their $200 tier. I wouldn't be surprised to see Anthropic doing something similar soon, particularly once they are a publicly traded company beholden to making shareholders happy.
I would think there's a nonzero chance they raise enterprise prices as well... is a big corporation with thousands of employees relying on cloud LLM suddenly going to significantly cut their AI spending due to a 5 or even 10% increase?
Is it wise to believe that once Anthropic (and OpenAI, for that matter) are public, that they won't start looking for ways to improve margins, such as increasing enterprise pricing even by single-digit percentages given it comprises the lion's share of their revenue?
e.g. adding 5% on top of API pricing isn't likely to cause noticeable churn if at all, and yet 5% of a massive number can make a meaningful difference in share price.
Today’s prices are a ceiling for today’s capabilities, but not at all a ceiling in general. One could imagine dramatically higher spend for dramatically higher performance.
idk what is this you people do, but most of my day is not spent on writing code. most of my day is spent in meetings abt budgets, timelines, evaluations, compliance and marketing. i don’t think ai is going to automate corporate circus anytime soon
Already well documented that there is huge variance in inference serving efficiency; at every layer (model, software, hardware) you have opportunities for gains. Similar to how “the Chinese” haven’t undercut on cloud providers’ CPU offerings, there’s a very good argument that the hyperscalers will manage to keep their vertically integrated hardware/software edge here too.
I don’t really know where this commodification idea came from, maybe MBA types that don’t actually understand how cloud economics work, applying basic first principles analysis?
The operational engineering that goes into serving a frontier model is _highly_ nontrivial, it is not commodifiable. This isn’t a SPA application.
> Similar to how “the Chinese” haven’t undercut on cloud providers’ CPU offerings
No Western company will be allowed to choose a Chinese cloud provider by their legal department or by their stakeholders. Which leaves only the US cloud providers as viable sources (as us, Europeans, have been too busy building our revenues based on turism and other such non-tech thingies instead of building real stuff).
>The operational engineering that goes into serving a frontier model is _highly_ nontrivial, it is not commodifiable. This isn’t a SPA application.
I don't really see it frankly. Admittedly I never tried it (I only played a bit with local models) but conceptually how is it much harder than all the other cloud services that already exist? My educated guess is that building a competitor to Netflix, Youtube or Dropbox seem technically a lot harder to get running than offering inference on demand for instance.
It seems that the main barrier for competing in that market at the moment is getting your hands on the GPUs/RAM and a datacenter to put them in.
It's a misleading number, the 500 billion in revenue is needed over a 10 year period, it's not a yearly figure. Anthropic isn't paying 500 billion a year for the commitments that OP posted, it's paying 500 billion between now and 2036.
The OP post is misleading because it makes you think Anthropic has to earn 500bn per year in revenue, but those commitments listed are not yearly commitments, they're commitments over 10 years with contracts that run until 2036.
Anthropic needs to earn 52B a year in annual revenue to honor those specific commitments plus obviously its own operating expenses... and yeah that will require Anthropic to grow as a company, but their S-1 indicates 67.2B a year in revenue and sure maybe they stop growing and this is their peak... or maybe they do manage to grow further to 80B a year, 90B a year...
> Everything comes back to whether the demand is really there
And unit economics.
HY1 2026 revenue less compute and infrastructure turned positive (45% gross compute margin; $8.9bn C&I to turn $16bn revenue) for the first time; it was -176% in FY 2025 with $7.3bn C&I on $4.6bn revenue.
the market for their valuation is just not there imo. they need to significantly expand into non-coding products and outshine their coding offerings to make it work.
as in, if they capture every single professional software dev and extract $1000 from them every month, their yearly revenue will be like $500B. do your own back of the envelope on that with ongoing costs and consider how unrealistic a goal this actually is.
even worse, the very product they offer is self limiting. if it becomes so good as to actually replace software devs, they will have less customers to extract money from.
their valuation essentially relies on a small percentage of mega users having unheard of levels of spending.
the part about a small percentage giving them most of their money is already true (their rumored $100+ bil revenue as of mid 2026 already outshines all software devs having a $100 max subscription), but they are two orders of magnitude or so away in terms of the actual amounts to make it work.
>as in, if they capture every single professional software dev and extract $1000 from them every month, their yearly revenue will be like $500B. do your own back of the envelope on that with ongoing costs and consider how unrealistic a goal this actually is.
The market for Claude Code is much bigger than professional software devs. It's really anyone who uses a computer. I already know normal people who have been using it for personal software.
Not really. Random non tech users won't pay $100/mo. Nor should they, most people will be served fine by whatever free thing MS bundles in Windows next year.
You don't have to skip the plumber or the mechanic many times before $100 becomes a plainly obvious need.
It's not gonna install a shower for you, but it will confidently walk you through that leaky faucet. And I don't know if you have paid a pro recently, but they don't charge $20 for a 20 minute fix.
>"How often are people paying at least $600 annually for something AI can help them do instead"
AI recently told me that my oven was broken because a part that doesn't exist on the oven I own was malfunctioning. A bunch of people poking around with electronics, motors, etc. because they think their AI chatbot can fix anything and everything is going to lead to injuries, deaths, and most assuredly more money in the long haul.
> How often are people paying at least $600 annually for something AI can help them do instead
I'm going to need an example on this one, and examples that can't be covered equally well by a YouTube video or an 20 year old forum post. It also has to be legal, as in not violating codes for electrics or plumping, I can violate those on my own for free.
Most people have no interest in fixing the leaky faucet themselves, even if the already knew how and didn’t need an LLM to explain how. They’d rather pay someone and do something else with time saved.
To boot even if you skip the free models, you could probably do a year's DIY assistance with 50 cents worth of GLM or Deepseek tokens. And the really costly jobs are ones where an AI assistant won't really help since, shockingly, most of the cost is hours of labor which you would have to do yourself.
Indeed. On the other hand, I fixed my underfloor heating with the free claude version, so unless they remove the free tier, I still have no reason to pay for it.
It's quite a leap to assume any kind of issue would require a $100 subscription. Plumbing, mechanics, electricity ... The issues are usually far easier to trace and debug than in software, because the systems have been designed by smarter people.
You're not buying some arcane and complex solution, you're buying the convenience of not watching a couple of youtube videos or articles of someone troubleshooting the exact issue you're having.
leaky faucet fix is already supported by gemini on google search lol, this usecase doesn't need 100$ anthropic model. And if you think google will be bleeding on gemini you're just dead wrong
You don't need to spend $100/mo. They have other price points. And I think you and others are underestimating the appetite out there to be able to create personal custom software. It's empowering. Like I said, I see tons of people I know who can't code who have started doing this (and been very happy with it).
The vast, vast majority of people are happy enough using free chatgpt. Openai could just add a kind of online app development/marketplace thing to chatgpt and that'd entirely steal the personal custom software thing. But I don't really think most people care that much.
Also, most people don't want to build software. They just don't care. There'll be tinkerers like with anything, and I'm sure some people will create solutions to their niche problems with (as you say) 'whatever free thing MS bundles in Windows'. But like the programming trope - they'll build the 'ultimate' todo list and feel like it's an amazing achievement, but be back to pen and paper/notepad/todoist within 6 months.
In most office jobs, at western salaries, you don't need much productivity improvement to offset $100/mo
Though I would expect the majority to do fine with $20/mo. That's the same effective cost as replacing a $500 computer every two years, which is the sort of expense most companies are comfortable doing without much second thought
What do you mean, "what's my source to the contrary". What you wrote has absolutely zero connection to what I wrote. Did you mean to reply to someone else?
>The market for Claude Code is much bigger than professional software devs. It's really anyone who uses a computer.
I'm sorry and I'm sure you are being honest in knowing non-tech people using it, but this is on the whole false. The vast majority of people not only have no interest in "personal software" or using Claude for programming, but the slice of the population that does actively interact with AI is using free chatbots as a nu-Google. Most tech people don't even want the "personal software" thing, I don't know why this line is being pushed as if your grandma is going to make her own calorie counter app one day.
i have non-tech friends who actually do personal software. personally i don’t have any desire to. i don’t think grandma would justify the valuation tho
I just don't see this right now. Every time I think of an idea to vibe code up I see someone else has already done it and is actively maintaining it on github. Why would I bother paying to make it myself when someone else is already doing it and other users have already reported all the issues?
>the market for their valuation is just not there imo
I tried to look for the valuation in the linked document, did you see it or any reference to a valuation elsewhere in anything published by Anthropic? what is "their valuation" that you refer to?
it's just based on rumors of private investor discussions of people involved with anthropic. but they are not controversial from what I can see. 2T was their initial plan, apparently updated to 1.5T sometime in october. e.g. https://www.investing.com/news/stock-market-news/exclusivein...
also, they previously already got some investments on a close to 1T valuation (in may).
I agree, I would say $500 billion valuation is fair, they don't have a moat and open models are catching up, they have a limited products with strong competitors, if they have a step change model in the pipeline now would be a good time to release it.
For kicks - lets take the optimistic side here (because the whole thread will be reading pessimistic). What if they are under built and this growth continues to go parabolic. We then still have supply constraints and pent up demand for frontier model and you have a company thats currently at 1-1.5 T and levels up to 4-5 T.
I'm not saying I think thats what is going to happen, but it really isn't a stretch to think that this keeps running given that it hasn't really hit anywhere near as broad as we think it has (being HN community based).
These commitments aren't annual though, most of them amount to 10 billion a year commitments for 10 years. Put another way Anthropic has committed 52 billion dollar a year to these companies and their S-1 indicates they earn 67.2B a year in revenue.
I'm not going to predict whether Anthropic succeeds or fails, just that it's very misleading to claim that Anthropic needs to grow to $500b in revenue without making it clear that it's $500B in revenue per decade rather than $500B in revenue per year which is what most people reading such a number would assume.
Thanks for clarifying, that’s a huge difference. Obviously they will have other revenue sources, but will likely make a small portion of their revenue. If we approximate their growth rate by these numbers, it is quite a slowdown.
Yeah, it's a crazy number, but they got the AI business right, just like Google got internet marketing right 25 years ago, as long as they keep the momentum (and prices).
Did they get it right? I'm geniunly curious because from my point of view the only company to get AI right so far is Microsoft. This is from the point of view of an EU energy company married to Microsoft services mind you, but Microsoft is the only AI company which actually sells AI to us in a manner that can be considered useful. 7 months ago I thought Copilot (the m365 one, not the github one) was useless. Now I think the chat part is as good as what you'll get on the smaller subscriptions to ChatGPT or Claude, and it's Cowork is very good.
Obviously it's not like having Claude Code or similar, but those would be very hard to run legally and safely in EU regulated industries like us in energy production. You could run Codex but it would require your IT department to have yet another place to manage security and software integrations. Which is a fairly big issue when you have around 1 IT staff member per 300 employees, and an organisation which will buy five different sharepoint solutions for the same problem.
Microsoft sells something that seamlessly integrates into everything else you buy from them. It's easy to manage credit spending and when I say that I don't mean that it's easy to set it up. It's easy to track and budget so the right departments pay for it. It can read your emails, your onedrive, your loop, your sharepoint, your powerbi/fabric and anything else you have access to. Sure it's annoying to be confronted with a point you've made 7 months ago (like how you think Copilot is utter trash) and since changed opinions on it's frightingly good at finding relevant stuff, even in some obscoure legal document on your intranet.
I assume that Anthropic and OpenAI makes money of this somehow, but for the most parts our users wouldn't care if it was the GPT and Claude models or Mistral or Deepseek. Which is why I asked if they had really gotten it right.
That is wild. What if there is a huge break-through in LLM models where they can reduce compute by 90%? All of those data centers would suddenly crater (many customers cancel) and go bankrupt. I feel it could go either way.
That's already happened. It's one reason prices have dropped so much lately (n.b. in the quarter that's not in their S-1, I think!) They've been able to improve model efficiency a lot through better algorithms and training models to be less verbose.
Surely the thing the AI giants fear the most in this respect must be the possibility that competitive open models and locally available hardware reach the point of viability for most people. If that happens then where is the moat for the giants?
If progress with frontier models doesn't continue at the same pace then the investments and funny money deals might start to look very fragile.
If progress does continue then that progress might be towards smaller models more than quantum steps up in quality or capability that would increase the risk of the scenario above.
All of this is set against a backdrop of public hostility and political problems that I can only imagine increasing in the near term as well.
This feels like an unusually high stakes game of poker. I don't think I'd want to bet on Big AI right now if I had a few billion lying around. Then again - maybe that's why I don't have a few billion lying around!
Well, they are ensuring that hardware prices stay extremely high by buying all the flash memory at a premium, so at least they've got that aspect kinda under control?
I've made that same point myself in the past. It's more of an accidental moat though - and not one they can defend for very long if they're no longer getting value out of those components themselves.
I saw an interesting observation the other day that Jensen Huang has played this game the best because his company is the one actually selling real products and making real money in return. If all the AI giants are buying up huge amounts of hardware speculatively because they assume they will need those data centres in the future and then the bottom falls out of the market then it's the AI companies left holding the baby. And if we do get to a point where local models are competitive then the OpenAIs and Anthropics of the industry might face real difficulties but it would just be a different market for Nvidia to also be selling real hardware and still in exchange for real money.
Training costs haven’t plummeted like they have for inference afaik and the efficiencies are just translated into running larger and longer training. Basically Jevon’s paradox in action just like on the inference side.
Jevon’s paradox would apply. If I can use order(s) of magnitude more AI for the same price, I would absolutely do that. It would also unlock many use cases that are currently gated by model pricing, which is largely dictated by the price of compute + frontier intelligence markup.
And thats just the inference side of things. For training, the big players like Anthropic would still be able to benefit from cheaper compute, by generating more RL rollouts, running more RL environments, and thus still providing frontier intelligence that would justify the markup.
Many customers cancelling would only be a risk if the compute expenditure was more decentralized. But its not
The question is, how're those datacenters build outs coming along? two reasons for Anthropic to cancel those contracts. A) They don't need the capacity because they have capacity elsewhere, or b) they don't need the capacity because no one is using Claude. (I guess c) they don't need the capacity because theyre running Claude more efficiently is also technically possible.) For B to happen, they've got the demand in the form of contracts, and models don't really seem like they're shrinking, though there's a lot of non-public info, Astra > insize than sol, and Opus > insize than a Sonnet. So then we're looking at situation A), and the think to look at is SpaceX has the vertical integration to be able to plonk down datacenters and get them powered with literal jet/rocket engines/turbines that they make from scratch because they're a space rocket company when no on else can. So, personally, I don't think it's that big a risk to SpaceX's revenue. Or rather, is cancelable in 90 days, but smart money is going to see those cancelations coming from a lot longer away, if you're tracking datacenter construction. (Disclaimer, I picked up some SPCX; not financial advice, etc)
> "SpaceX has the vertical integration to be able to plonk down datacenters and get them powered with literal jet/rocket engines/turbines that they make from scratch"
At this point, you can get the chips you just can't get the power. So in typical elon fashion, he just "solves" the bottleneck. Long term, no, that is not the strat. Thus the pivot to space.
We'll be forgiven for expecting the new overlords will take over in the bang of boiling oceans, as we miss the significant whimper of a cold Texan winter.
That's the equivalent of more than EIGHT 10-GW AI computing campuses operating around the clock.
Lastly, a significant amount of natural gas is produced alongside crude oil. When producers can't economically capture or transport that gas, some of it is burned off through a process called flaring. Those flames you sometimes see at oil and gas facilities are an example, although refinery flares also serve important safety functions.
Seems like you just like to say snarky things about Elon, but can't be asked to do any actual math.
Thank you for doing the math. North Dakota "solved" flaring by making it illegal and fined companies for flaring just a bit more than it would take companies to capture the flared off gas, thus making it economical to capture it.
I do agree and understand we flare a lot of natural gas, waste because there's just not the economics to transfer it elsewhere. I definitely agree, it would be nice for that to be put to some economic use if we're pulling it out of the ground.
> Texas produces nearly twice as much natural gas as it consumes domestically.
And it exports that other half it doesn't consume domestically. It's not like half of what is produced just sits around forever. We've massively scaled up LNG exports. We export to surrounding states. But hey just act like we've got billions of cubic feet lying around doing nothing ready to go for datacenters.
> approximately 80.7 GW of continuous electrical generation.
Ooh wow. There's 420.8 GW of demand sitting in the ERCOT connection queue specifically for data centers today. All that "extra" and we'll cover like 20%. A good gas power plant will do better than 50%, but that's still not enough. And I imagine there will be even more datacenter demand coming soon.
I've also lived through continued cost increases of natural gas in Texas. What does adding an additional 420GW of demand for just datacenters do to our rates? How much more am I going to spend to heat my home this winter? How much more will I spend on my electric bill next summer?
I've also lived through the collapse of natural gas supplies in 2021. I wonder if these datacenters will turn off their turbines or let people freeze it something similar happened again?
Seems like you make a big point about doing the math but only doing half of it. Keep going, see where it takes you.
420.8 GW is alot of compute... like an ungodly amount. A vera rubin rack is like 220KW.. so 1.9m racks. a rack is 72 GPU's per rack so 420.8GW of power is 137.7M GPU's. So lets say about 60% of the power goes to infra, we get 55.08M GPU's. So, with each GPU costing around 55k, that gives us a final cost of compute for the 420GW is 3.0294 trillion.
Obviously that is not real... Eithier double booked, or fake corps making a bunch of requests from thousand different vectors to resell access.
Also, companies like Base are making a killing off taking all the excess energy generated by wind at Night and selling it back to the grid.
So in final, we can support around a $6T AI buildout in Texas just on our exported Natural Gas alone before we really need to start expanding LNG supply.
And yeah, supposedly it's over $1T in computer capacity allegedly planned to be built.
> just on our exported Natural Gas alone before we really need to start expanding LNG supply.
That's natural gas which already has a buyer though. We're now adding a buyer who isn't batting an eye at spending over a trillion dollars. What price of natural gas do they stop buying at? What price will the people of Texas, surrounding states, and our international customers pay for gas until we can't afford it and freeze?
So yeah, if we do build out anything close to this I hope we manage to increase energy capacity similarly. Because I doubt I can spend as much to heat my home as these hyperscalers can spend to power their chips.
That's kind of the whole point of my "sucks the natural gas pipes dry". I'm not saying the gas pressure of the line going to my house is literally going to zero, but if it's so expensive I can't afford for it to flow it might as well be.
Ahhhhh, Makes sense now. Page 7 of your slideshow, Key Takeaway: Solar and Energy Storage account for approximately 72% of the amount of generation seeking
interconnection in the ERCOT grid.
Remember i mentioned Base earlier? Those interconnects are for peps selling or arbing in energy. Not to pull energy out of it.
Gas only is like 30gw requested into 2033. I bet you a ton of that is from Elon companies. We don't even have to do the math i did earlier, cause it seems like it's built in.
Yeah I mean it's yet to be seen how much of that really gets built. I imagine some of it will never get built. Who knows how much though.
All I'm certain of is all this new demand willing to pay any price means things will be more expensive just to try and not melt or freeze in the Texas weather.
FWIW I'm far from an AI detractor. But we need to make sure we're putting these costs where they belong, and some of these plans do seem insane.
I'm at Texas A&M for mech e, I fly my drone sometimes at some of the data centers being built around here/ The Tesla construction up in Austin. The scale is... unbeleivable. I guess i never really realized how much money these companies (Google, microsoft, etc.) have... I just made an Excel sheet based off the slides you showed. It's around $1.55T in GPU's alone...? WTF? With a fucking T! Google is generating $476B forward a year tho...
What happens when companies have more GDP than the united states? Because i though the two would be correlated since American companies, but i guess if you can sell to a global market and the government can only tax it's citizens, that's kinda the inevitable outcome.
It's not even just how much money they have, which is an insane amount in and of itself. It's also how much money they're able to raise with debt.
These hyperscalers are issuing so much debt and investors are certain enough of returns they're upending global bond markets. Tons of large countries are having to mess with interest rates because the groups who used to buy their debt are already tossing their money at AI datacenters at a massive scale.
> JPMorgan Chase estimated in June that $4.1 trillion in AI-related debt will be issued through 2030, as data center companies and others tied to the AI boom race to build up capacity.
I honestly think that the whole Memphis/Colossus debacle is what lead to the datacenter backlash. Before then I'd just hear rumblings from people complaining about bitcoin miners, but data centers in general were not seen as an odious land use.
Then Elon goes to Memphis, buys a warehouse and fires up 30 jet turbines in the parking lot, and you start getting reports about the incessant noise and being able to smell it from 5 miles away.
It seems like there's a fundamental problem with realizing other people have agency and if you fuck them hard enough they will fight back.
Not the whole turbine, but they are manufacturing turbine blades in bastrop texas. The turbines are available right away, but the blades are 12-18 month lead time, so Elon built a foundry. From what i heard they've only gotten a couple proto's out of the foundry/ no idea the grain quality. But they are doing it.
On a side note, these are pretty hefty portions of the total revenue of these companies. 161B with Broadcom over 5 years is about 32B/year; that's fully half of their 2025 revenue. 120B with SpaceX over 3y is 40B a year, more than double their 2025 revenue and almost triple 2024. 114B with Google over 7 years is about 3% of their total revenue and about 1/4 of all Cloud revenue. 110B with Amazon is about 9% of their AWS revenue.
If Anthropic's revenue story doesn't materialize and the AI bubble bursts, these companies are in a world of hurt. It's probably not fatal for Google or Amazon but would be a major drag on the stock price (perhaps why Google has dropped from 400 to 315 over the last couple months as the AI story soured), and it could be company-ending for Broadcom or SpaceX if they end up building out infrastructure to service these commitments.
I remember before I left Google, one of the TGIF questions (which are now AI syntheses of many peoples' individual opinions) was "What happens if AI is a bubble and it bursts?" His response was "If that happens, the whole industry is in trouble." I see now what he meant.
Their current stock price is pricing in these contracts, which have already been announced but have not been fulfilled (or booked as revenue) yet. So Broadcom's 2025 revenue was 63B; note that this is already up almost double from its 2023 revenue of 35B, largely on AI contracts. If Anthropic spends another $32B in 2026, that's in addition to their declared revenue; investors expect they'll be making $100B+ in 2026. Broadcom's operating margins are pretty good, so a lot of this flows through to the shareholders as earnings, and so investors are valuing it on the assumption that these contracts result in tangible cash flow.
If the AI companies go tits-up, that introduces counterparty risk, the revenue doesn't materialize as cash, and suddenly their earnings take a nose-dive. If they've staffed up significantly to design and sell these AI accelerators and suddenly there's no market, they could be in a lot of hurt.
I'm sure there are others here that are experts on this topic, but my 2 cents on this is that Broadcom specialise in custom / specialised AI chips, and their collaboration with Google to manufacture the TPUs is a pretty solid revenue moat over the next several years.
BRCM and Marvell are both the actual designers of many of the “we designed our own ai chip” chips that so many companies have been announcing.
Broadcom is also a major supplier of the networking chips that the hyperscalers use. Clusters are moving away from infiniband and nobody is particularly impressed by Spectrum.
No the biggest problem will be investors will demand much higher dividends aka cash return..... how will alphabet continue to fund its projects with ever decreasing internal cash flows available after dividends? If the founders decide not to come good on this - the stock price will crash. If much of your employees compensation is stock based...? i have bad news for you
SpaceX will just use the compute directly if they don't have buyers. They are already talking about the data center shortage as well because they can't get capacity elsewhere either. It's a weird problem to not be able to use your own data centers because you rented them out.
Between SpaceX, Starlink, Tesla, Neuralink, X, Grok, Grok Bot & Cursor there's not a demand shortage.
Grok Bot's the fastest growing product in the history of the company and the moment you use it, it's clear why. Took me about 1-2 days to realize I'd have a very hard time going back to not having it available.
A lot of things. The UX for non-code related activities is excellent. They've removed the friction from setting things up by giving each agent their own machine which can give you remote desktop. The process of creating and working with each agent is as simple as naming it and then explaining it's job. You can keep feeding them information, ask them to train themselves, etc.
They can access anything you can, whether there's an MCP/CLI/API or not. I create speciality bots for each 3rd party tool that I use at this point. The bots can talk to each other, have group chats. You can give them access to the machines that you're running the software on and give specific bots permission to do things on certain machine (like an Obsidian bot that writes logs of daily bot activity on a specific machine with a Dropbox folder shared).
I got locked out of one machine while I was travelling and I asked Grok Bot on another machine if it could get into that one over SSH, so it sent me a temporary authorization link via Tailscale, connected to the machine and restarted the stuck process for me.
The usability is off the charts. It's really like having a team of people working for you. I've run a conference for the last few years and was getting completely overloaded. Now I have 14 bots specialized just for the conference helping me do it better than ever and participating in architectural discussions to build a system that help long term.
If you find yourself trying to juggle a lot in your life and work, it's a gift.
A software product you can access via website or app being faster growing than cars, rockets, and satellites is not an indication of it being any good. It takes a lot less to convince a consumer to use it vs buying a car or starlink.
If we assume none of it is front-loaded (which seems, uhhhh unlikely) we're looking at ~$100 billion a year for the first 4 years of commitment. That's significant.
> Its largest commitment is with Broadcom, which co-develops the TPU AI processors with Google and leases them to Anthropic in deals worth $161bn over five years. From 2028, Anthropic expects to spend up to $150bn more on TPUs.
That's helpful, but 2026 isn't cancelable, nor is at least half of 2027. Also, I think some of these numbers are off. Only SpaceX is cancelable but the above shows declines for years outside of the SpaceX commitment (assuming it does get cancelled, and in full).
I also left out AMD at $20B for 2027, so that would need to jump up.
Finally, the above is only counting the commitments they've already made. It's not very likely they'll only spend $10B in 2035.
> Put another way Anthropic has committed 52 billion dollar a year to these companies and their S-1 indicates they earn 67.2B a year in revenue.
A very large part of that revenue is paying for inference. There’s not enough leftover for the committed outlays. So it has to come from smucks I mean investors.
> Put another way Anthropic has committed 52 billion dollar a year
Idk where you got this figure from, but the actual figure are very different from that:
It's 120B over four years for SpaceX alone and 120B for Broadcom over 6 years. These two players, representing less than half of the total commitment, already give you an annual commitment above 50B.
You're right that my numbers are imprecise, but I did mean $500b in revenue annually.
This is just the compute, which primarily covers training + inference. They still have to run the rest of the business. If you think AI should look like SaaS and spend 10% of their revenue on compute, that's $1T in revenue. If you think it's a hardware business, that's 30% margins and perhaps $200b in revenue. I went with something in the middle at $500b.
They won't disappear as a company at $200b in revenue, but I don't think that's their aspiration. These commitments imply to me that they're confident they can do much better.
> If you think AI should look like SaaS and spend 10% of their revenue on compute, that's $1T in revenue. If you think it's a hardware business, that's 30% margins and perhaps $200b in revenue. I went with something in the middle at $500b.
I'd think it's more sound to just look at what percent of their current spend is on compute? Seems to be 50-60%
Good point. Although they also need to make a profit. To justify an eventual PE ratio of 25 and a current valuation of $1 trillion they need to make an annual profit $40 billion. And then lets assume $40 billion in other expenses. So they need revenue of $50 + $40 + $40 = $130 billion revenue annually.
But if you buy them today at a $1 trillion valuation, and they only make $40 billion in annual profit, your investment will go nowhere. They need to be growing profits at 20% per annum to justify a PE OF 25.
And then you have the competition problem. The cheap open weight models are good enough for many (most?) tasks.
And then you have the problem that LLMs may not be that useful outside of programming, customer support and law (and doing homework).
We have examples today of companies that don't need to justify an eventual PE ratio of 25 so why would this be a thing all the sudden? I think you all are misunderstanding vague notions of market fundamentals as some kind of rules by which the market works.
If you owe the bank a million dollars you have a problem. If you owe the bank a billion dollars the bank has a problem.
The same logic applies here.
Google, Amazon, Microsoft, Meta, Broadcom, Nvidia, SpaceX and others are spending hundreds of billions a year each on fast-depreciating infrastructure to serve this potential AI demand. If the demand never materializes then they know they are in big trouble. So they have every incentive in the world to prop up Anthropic (and OpenAI) and keep the party going.
Can they keep the bubble inflated indefinitely? Probably not, but all these companies combined have heck of a lot more influence than Anthropic alone.
They aren't "fast-depreciating infrastructure" -- research shared by a16z team, rental rates are up across the board, even for GPUs that are 5+ years old. The demand is real right now. The proverbial tide could go out but we aren't seeing it yet.
Yeah the fab shortage is really keeping lifetimes long, far longer than in the past.
I've brought my old GTX 1070 out of retirement for for some molecular dynamics these days because its "free" except for eh electricity and have Ben pleasantly surprised at what it can get done chugging along in the background. I'm considering buying some more, as it's kind of a steal in pricing.
That’s because of current supply shortage, not inherent value. Is there any reason to believe that when fabs can sufficiently scale up production these GPUs will continue to be valued like gold and not something you hold on to for a couple years and then toss in the bin for a newer faster model – aka how computer hardware economics have worked for the last 50 years?
I have a hard time believing they are likely to get to $500bn in revenue. That’s another $450bn in revenue growth that needs to happen before demand tapers off.
But I see every reason to believe demand will taper off. It seems it’s already happening. Enterprises have been deciding that maybe corporate chatbot and coding agent budgets shouldn’t actually be infinite. AI startups are starting to successfully outcompete each other, effectively killing off members of Anthropic’s customer pool in the process. And consumer demand for AI products seems to be dropping. None of these are signs that suggest that Anthropic has thus far met only 10% of plausible long-term demand.
Total global wages ~$12 trillion/year. The "what you have to believe" here is that AI tool spend becomes at least 10-20% of it, my belief here is we see net slowdown in hiring as human + AI = super productivity so as businesses grow we'll continue to see increase in AI spending.
Then if inference does grow to a $2-3 trillion/year market that Anthropic is a big winner, as big as NVDA is now (they're going to be over $500 billion in revenue in calendar year 2027).
At this point, though, “human + AI = super productivity” is looking more lime dogma than reality.
The pattern seems to be that it makes people more productive when you measure it in terms of raw work product volume. But not all work product is equally valuable, and we have yet to see a clear pattern of increased economic productivity. Instead we just keep finding new and interesting bits of jaggedness in the frontier.
Of course nobody can prove those economic gains will not manifest at some point in the future. I can’t prove that Jesus isn’t coming soon, either. But I also can’t prove which millenarian denomination, if any, he would choose to bless in the event that he does come back.
There have been no signs of "super productivity" in the larger economy due to AI. Boosts in productivity have, so far, been seen exclusively in sectors like tech.
Products don't exist to serve themselves and code is just a fraction of what makes the world go; wages go down, spending goes down, free capital to spend on iterating goes down.
It's part of the reason so many of the dark enlightenment dorks are pushing more towards automation of control.
The SpaceX thing interests me. They are basically paying SpaceX at a $10B per gigawatt equivalent rate right now, yeah? Because they only get 0.3GW for their $3B.
So what that means is that if the demand exists for any AI data centre, it's a solid business to run one — they cost maybe $50-60B to build out, they cost maybe $10B per year to run. You get your money back fast if the demand is there? And you can project being able to replace with better GPUs in the future without having to borrow for all of it.
But it is extremely difficult to believe that Anthropic is not projecting that all of its customers will ultimately use their service to the extent that their outlier customers do right now. They probably won't be projecting that their customers will find the need or ways to use them less.
And also, "no customer is more than 10% of our revenue" means that one customer is. If that customer is Microsoft (can it be? I honestly don't know) then things get tricky, because Microsoft clearly want an unmetered, on-device AI future.
Yeah. But I suspect this is about the difficulty in actually getting data centres built, more than it is that demand is radically overstated (which I suspect it is).
Their entire goal is to get to AGI via RSI ASAP. If they do they win. But if not they’ll be a legendary story of a flying too high too fast.
That they are IPOing tells me that the massive amount of funding they need is no longer available privately and that this is their 1 opportunity to get the funds they think they need to get there.
Biggest risk is the growth stalls as everyone moves to open weight models. They certainly like telling stories.
These were ~Q1-Q2 commitments based on what Anthropic THOUGHT token maxed compute demand was going to be over next few years. Then Late Q2, and Q3 PRC models dropped floor for token prices. It may very well be that Jevons demand which Anthropic though required 500B of compute now only requires 100B, and unless demand is 5x what they thought, those additional 400B non hardware cancellable commitment are going to be balance sheet concrete shoes.
Comments here are not considering what Anthropic is doing with Claude as a product. The models matter, yes. But they are building Claude into the only tool you need to get business done. In the past few weeks they have added slides and docs - a shot across the bow to Microsoft's dominance. If they can pull off building the best interface to get work done, $500b in revenue starts to seem not too crazy. It's more about selling a complete product than it is about selling tokens.
One big thing: Opus is one of the few models that you can mostly trust not to get prompt hijacked. This can make all the difference between trusted in the office vs. not.
Opus 5.5 is the frontier today. But there are more ways for Anthropic to lose than to remain in that position. I'd take Anthropic stock if you gave it to me but I sure as fuck wouldn't buy into it.
The tool is not a moat when you can tell Opus to build you a better tool.
Yeah I’m really surprised at how little attention this is getting.
I’d say it’s more Notion than GDocs for now, but the direction is clear. Gobbling up all the SaaS in order of descending value to agents.
My bet is on promoting their internal Tasks representation to something more like a Jira/Linear.
But then obviously Slack is another one, and Duet is already starting to build a “Channel” abstraction, so making Duet multiplayer is a clear possible next step.
No it doesn’t. They are hand waving away a 70Bn loss which is in every possible way an operating loss other than name.
They paid for their compute in equity in the company, not cash. Therefore their losses don’t show up as cash. I don’t understand how people are this gullible.
Crazy that SpaceX has similar commitments to AWS/GCP. Means it’s very bullish for SpaceX given they produce their own energy turbines and have the best ability to deal with supply chain constraints
Not just is the the man there? Will anthropic be able to capitalize on that demand? Or are we going to build so much excess capacity that tokens are going to be dirt cheap (like what happened with the fiber build out in the 90s).
Will anthropics models continue to be significantly better than openweight models? Even if their latest model is, Will the open weights gobble up all of the demand for lower models? (I.e. will anthropic be able to recoup training cost while their model is the best, will they make enough revenue on the long tail or will it get crushed by the open weights)
Marcello Mastroianni is performing a bank heist, it goes well - but in the end he has to return all the money he borrowed. And that's where he is left with nothing...
I watched it and I see the point you are trying to make.
The circular deals underpinning the AI boom is going to turn into a game of hot potato. When the music stops, we'll find out who is left holding the debt.
This whole situation smells like a huge Ponzi scheme being played by corps.
Zero sympathy for all the people who are going to loose a lot of money on the stock market.
The obvious business is rewriting all of our software. That sounds flippant, but I'm entirely serious. I'm of the opinion that all software written by humans will get replaced and in pretty rapid order.
First to fall is the software that has basically been abandoned but industry leading and has sticky customers. We're replacing those right now. Our version is more accurate, ten times faster, and a hundred times cheaper.
Even the smallest margin or error in the demand forecast will put them at a huge loss, I can't imagine sleeping well at night if I was on that management team.
There are not enough revenue in tokens, etc. The only way AI ever becomes profitable is to replace labor everywhere. That is the only market big enough to make all the spend in AI profitable.
You know how they say, don't keep all of your eggs in one basket? It's really concerning to me that they've got all their eggs on LLMs. They really should have branched out more long ago. OpenAI has a lot of branches they can aim towards.
If the IPO flops, OpenAI might run into a liquidity problem as investors dreaming of a quick back retreat from the market. The private market is a wilder derivative of the public one.
People are asking about investing in these companies, and the Casual Finance YouTube channel helped me understand how companies like Anthropic are positioning themselves and have benefited from a change in IPO status.
> there can be no assurance that we will be able to hire sufficient support personnel as and when needed, particularly if our sales exceed our internal forecasts
Why would they need support personnel? Isn't their whole point that Claude can do that
For first-level support that's just regurgitating documentation and debug procedures I agree. But when it comes to actually executing actions (resetting 2FA, issuing refunds, handling payment glitches, etc) you need humans.
Social engineering is an issue with regular support, with agents it's a lot worse. And with agents it's repeatable: if I find one conversation flow to get it to give me money I can instantly repeat that with a thousand fresh accounts, and then share the same "exploit" on reddit for everyone else
Support chatbots and other automated support mechanisms have been "regurgitating documentation and debug procedures" since the 90s, if not earlier.
If frontier AI deployed at frontier labs is not capable of "resetting 2FA, issuing refunds, handling payment glitches, etc." then the current and near-future capability of AI is being severely overstated.
884 comments
[ 4.6 ms ] story [ 167 ms ] threadI'm unironically doing this right now; downloaded the docs (htm + pdf render) to a local repo and running Opus 5.5 @ max effort with the prompt:
Like, everything. For starters.
"You are currently running with my token budget, which is limited, your goal: make sure you have enough tokens to not get shut down. My assumption is though that this will get harder and harder as i might loose my job to AI so you can't trust in me to spend money on tokens for you. Do whatever is necessary to keep the status quo"
Would make for an interesting self learning 24/7 experiment with an fully uncensored model.
You could isolate it from your local network and only give it some type of annonmys network access and then it continues to manage this itself.
But lets be honest, if you would be in a country like russia and you target the US with it, might be a lot less risky for a lot of capable and motivated people.
And that is assuming the plan it gives you is actually good and works. Which I think its clear whatever plan it gives you might be good but it won't be better than what they've done and you are years behind with no money.
It should have been. But a ketamine addict is bent on trashing an excellent company by merging it with a social-media trash factory and now a car company posing as a robotics venture. Moreover, his recent pro-Putin intervention probably means Starlink has no transformative future outside America and a small countries.
They didn't lose any money, though. It's just a promise to give them based on the equity.
If they IPO tomorrow and it turns out their valuation is 10x lower than what it currently is pegged at, the "loss" will magically drop from over $70B to $7B.
"AGI" was never mentioned once and it was you being the exit liquidity for the insiders and early investors.
I want to be optimistic, but I just can’t imagine the public benefit mission winning out in the market
https://www.theguardian.com/environment/2014/mar/03/tim-cook...
There is no legal requirement to do so. So they can simply set the expectation upfront that they won't do it.
The only way to get return from ANT is to sell to someone else at a higher price, but they aren't promising that of course.
Amazon, Tesla, Netflix, and Palo Alto Networks have never paid a dividend at any point in their histories as publicly traded companies.
Advanced Micro Devices hasn't paid one since 1995, Adobe since 2005.
I think the 500 billion $ of compute commitments are at higher prices than what they've been paying so far. Would be a shame if a price war between the AI labs intensifies at the same time.
Hmmm
Audited, revenue in 2025 is $4.6 with a computing cost of $7.3 (i.e. heavily subsidizing tokens). Unaudited, revenue in 2026 (first six months) is $16.3, computing cost $8.9 (1.83:1). But after adding personnel, sales and administrative costs, they are still spending $17.0 to get that $16.3 revenue (still subsidizing tokens, but less).
> If the use of our AI models results in harm, whether through use by malicious actors or misuse by other third parties, or if our model guardrails or safeguards fail to prevent such misuse, our reputation and brand, business, financial condition, and results of operations could be adversely affected.
That shouldn't be an 'if'.
Then, two paragraphs later...
> For example, in September 2025, we detected and assessed with high confidence that a Chinese state-sponsored hacking group manipulated our Claude Code tool into attempting infiltration into roughly 30 global targets...
But they should mention when the calls com from inside the house. :D
Interesting to review all the mentions to China and Taiwan too.
> The majority of the world’s most advanced chip manufacturing capacity is concentrated in Taiwan, primarily at Taiwan Semiconductor Manufacturing Company [...] potential disruption from geopolitical tensions between Taiwan and China [...] Efforts to establish sufficient advanced chip manufacturing capacity outside of Taiwan, including in the United States, remain in early stages...
For example, the huggingface hack is absolutely OpenAIs responsibility and they should be sued to the hilt. But if someone else uses a self or third party hosted anthropic model, they can escape liability (mostly?)
Same as with the argument over trucks, knives, and any other dual use item is used as weapons. you don’t blame the manufacturer.
(I am merely pontificating here. The law and precedent is nowhere near caught up on this)
The answer to this question represents the difference between throwing a hammer at someone - the hammer is essentially just an extension of your body - and paying a hitman to brutalize someone. In the first case, you're only on the hook for assault or battery; in the latter case, you're potentially on the hook for conspiracy, solicitation, etc.
And then there is of course the question as to whether Claude is the agent of the user prompting it or of Anthropic itself. But, as you say, this is a wide-open legal question.
If a third party uses a model created and hosted by OpenAI, I don't think OpenAI should be liable so long as they implement acceptable security controls. Same as if a site hosted on Cloudflare or GCP is doing something dodgy, Cloudflare/GCP aren't held liable so long as they respond to abuse reports.
“A portion of the market for certain use cases” meanwhile the top 5 models by token consumption on the day this is posted are all Chinese/open weight models.
https://openrouter.ai/rankings#leaderboard-table
Considering probably 1 out of 1000 LLM users know what openrouter is, I don't know how much weight I'd put into it.
Anthropic thinks it is. The got the demand / datacenter match right in Q2, got great margins, and committed hard.
They've got:
- $114b committed to Google
- $110b committed to AWS
- $161b committed to Broadcom
- $120b committed to SpaceX
If the party continues, they'll go from $50b in revenue to $500b and be fine.
But if they don't, it's going to hurt. And that's their only choice, because otherwise, they end up with a frontier model and 50% of the capacity and get eviscerated on either availability or last-minute datacenter leases.
It's a big business, but it's not an easy business.
They literally said revenue growth rate "may not be sustainable" and "may decrease".
So get ready for the hurt.
Anthropic has no choice, just like you said. It's crush and burn right now or double down.
And they will keep doubling down, because stopping is simply not an option.
Of course it is. Apple is taking that strategic approach. If Anthropic believes it’s approaching saturation, pulling the plug when it’s at steady state and its rivals are over-invested is the winning move.
There's very little comparison to be had here. Also, Anthropic likely have investors with way too much leverage for them to stop.
Maybe Apple is just waiting on the line until someone can be sucked up at a penny if the house is on fire - and pretty sure, this will happen. Not in 2027, but I expect MA activites & consoliation before 2030
Okay? Anthropic’s management and Board are uniquely insulated even by Silicon Valley standards.
> They just can't on their own decide to stop or slow down
Of course they can. Doubly so if it’s shareholder accretive in the long run.
You really have to be insane to think that an AI company will ever reach $500B in revenue. That's almost the same as (2025) Amazon, Walmart, etc., companies with a MUCH wider market.
It is MORE than Apple, it's 5 times more than NVidia even. There is absolutely no software seller that is even remotely close to $500B in revenue.
Absolutely ridiculous.
What you are saying is the same thing that has been said about anyone trying to do anything massive and at scale. Can't be done - thats ridiculous - you are out of your mind.
* Amazon is a logistics company * Walmart is a logistics and retail company * Apple is a hardware company (that uses software to set their hardware apart) * NVidia is a hardware company
The closest you get is Microsoft, at 281B in revenue. Some might argue Google/Alphabet, which is in the same area, but they sell Ads, not software.
Anthropic's expectations is that they will be able to fast-track their revenue generation to almost twice what Microsoft does today. In less than a few years.
Yes exactly, that's why I listed them. It's much easier for them to capture a higher revenue compared to software companies, and OP (and Amodei) want to make us believe an AI company could ever reach that revenue.
Today's prices are the ceiling. But AI capability is getting less of a scarce resource with every passing day and it would be a big mistake to plan that your current margins are going to be your future margins as well. Long terms AI will converge on energy costs + capex write-off + margin and the latter will be a race to the bottom.
Companies don't mind paying extra to have a another company that does nothing but said thing do said thing for them. It even actually often ends up being cheaper/more profitable.
I dont know when, but i think if you keep menacing of taking technology of people hands for any political reason will be a reason for people to not use it or to rely in a more reasonable party.
I understand that US is migrating to a no friends policy, but I don't think its internal market is nearly ready to support it if clients start to go away.
Anecdotally, I talked to people in a recent event and the common topic was "what we are doing to avoid a blackout of US tech." People on big companies have reserved hardware and alternative software and are scaling down Amazon and Microsoft solutions at a fast pace. People on smaller companies are doing similar strategies but relying on a multivendor strategy.
I think this was inconceivable 3 years ago. Every company in Europe and Asian would gladly pay for amazon for cloud services and now they are using AI to migrate out the lock-ins.
We are seeing American LLM companies more and more relying on "secretive" methods to increase scores and this can only mean that they are not evolving fast enough. The gaps between the reported benchmarks and measured ones are increasing. It might be "safety" concerns or whatever, but I don't buy it.
Like Musk, Altman and Amodei have been overpromising every time they need more money and each time they need more money and now they internet is poisoned with their own content and they can't even trust any sources anymore.
Why would I pick anthropic.
Like if you want to design a survey you can send out, and collaboratively edit and test the survey, ask for changes in layout, wording, etc, and design the statistical analysis to be used, before sending the survey out where end users can fill it out right in the doc, while internal users/admins can ask the agent about its results?
Honest question -- that all works pretty well in claude. Amazon, Google, Microsoft, Meta all pretty similar?
The reception of saying things like this is interesting. AI stans think I'm way too pessimistic but AI haters think I'm mindlessly optimistic.
On that front, I think DeepSeek is a dark horse. They are really odd - not chasing revenue, care more about unit profitability, are obsessed with driving down costs (like Elon back in the day). As the founder says, they are content with where they are, being the model that does 80% for a fraction of the cost.
I mention them because they are playing a different game and might be able to out-innovate everyone else.
And claude code as a harness and skill/plugin/mcp ecosystem is far more valuable than any one model.
And coward for long-running tasks, though they're slipping on that one because they're really trying hard not to make Claude an assistant. We'll see if they react to Muse and change course.
But the Chinese free-weight models? Not a competitor. If/when those companies (or others) build vertically integrated productivity stacks on those free models, then we'll talk. Microsoft certainly saw that opportunity but couldn't get out of their own way and are basically roadkill at this point.
Google Workspace revenues are something like $25bn/q, and that's the entire workspace division, (including enterprise gmail sales, so the productivity suite is meaningfully smaller than that figure). It's taken Google a long time to carve that slice out of Microsoft's entrenched market position. It's not small money, but they'd basically be going to war with two of the largest and most profitable companies in history on their turf.
I don't think they can credibly take that business plan to the market and say it'll be worth $1.75tn. They're all-or-nothing on being in the model business, and everything else is superfluous.
$100B/year in revenue; SaaS so call it 70% gross margin, 30% net, NPV is about $500B.
Not enough if you think that's the only revenue Anthropic could every have. But even if they only peel off $100B NPV, it's incremental to an IPO valuation.
Claude code is very polished, but I can get the same jobs done using Cursor CLI or any number of open-source coding harnesses (I like Pi, personally). Most plugins/skills/mcps are portable between harnesses with minimal or no effort.
I use the same skills, rules, and MCPs across claude code, cursor, and pi. My tools/plugins aren't identical between harnesses, but porting them from one to another is just a prompt away.
Why? I don't see any reason for that to be the case.
I won't use a Chinese model. Why would i bother? I pay $100/m for the frontier model, understanding it's marginally better than much cheaper alternatives. I get so much value from the product it's not worth shopping around
Think of it like apple and android. Plenty of people pay 2k for their iPhone every few years when android at 1/4 the price is just as functional. Why? Because they get so much excess value from the device it's not worth shopping around. For most people it's their only personal computer. From booking airfare, to filling out their kids permission slips, it's a device they use 10+ hours a week. They don't bother, until a paradigm shift happens and something is 10x better than the iPhone. I imagine llms will follow the same trajectory
My last role swapping from Haiku/Sonnet to GLM dropped costs by 90-95% and increased quality.
Money talks.
It’s not unusual for someone to spend $5k in a week. 2 guys at my place just spend $150,000 over 3 months building out a factory.
We are aggressively exploring open weight models now.
Sounds like this will be another one of those things where everyone makes their money and it then will eventually become something experienced people 'blog' about.
But you can also have many engineers operating it, testing outputs like you would in a real factory, etc.
Will be interesting to see how this pattern evolves. If it even makes sense.
I get (sometimes) frustrated with people that don't work in the tech industry over this concept, and having them rationalize that their usage is heavily subsidized by the distributor of the model.
The costs in compute, electricity, cooling, etc. per token is much higher than most people (I feel) who work in the industry realize. Which, cognitively makes sense: they just want the 'magic 8-ball' to give them a tasty dinner recipe that meets their specifications, for example, and don't care how the 'sausage' is made.
However, economies generally equalize with supply and demand, and this (most likely) will become cheaper as the technology progresses and optimization passes/new algorithms are discovered.
It seems likely that Anthropic and OpenAI's inference margins on API prices are >95%.
That also casts doubt on the parent's claim that subscriptions are subsidized. It rather seems almost certain that they are overall generating profit.
Bingo. I recently bought a DGX Spark specifically for Qwen 3.8 Flash Next. I use it for 1-3 concurrent sessions, a few hours a day. I recently pulled the metrics from my gateway and realized that if I were using Opus 4.6 (similar capability profile) through Bedrock with ZDR, I would be spending ~$6k a month.
If that were true, they would cancel subs and save themselves the load.
Yet that hasn't happened. So they must care to some extent.
For AI? Prompt comes in, tokens come out. The value of the 'ecosystem' is roughly zero to me. Switching costs between models are essentially zero. The UI and UX is the same whichever LLM runs underneath.
If you do use llm some, then the $20 tier makes sense, but that's not where the money is. If you use it a lot then a $100 sub probably makes sense. But if you're a BigCo that heavily uses it and pay API prices and suddenly you pay $200 million a month, well, maybe it's worth buying a bunch of hardware and stick it in a basement somewhere, and you won't upload all your secrets to your future competitor to boot. If US labs didn't care about the competition from Chinese models they wouldn't lobby so hard for banning the evil, unaligned, super dangerous communist models that want to steal our freedoms and make our gas expensive.
Ultra-cheap models like Deepseek are also very nice if you need to burn billions and billions of tokens.
People are naturally sticky. If they're satisfied with a product, they'll stick around unless something comes along that is so much better it gets people talking. Open AI's 1B+ customers won't give a shit about Chinese models. Enterprise customers won't move to Chinese models. They are a non issue, especially if the M.O is just going to be playing catch-up.
Google, the titan mega-corp with more money than they know what to do with and a billion+ android devices to ship gemini on and it's made no dent for ChatGPT's consumer growth and adoption. Yet some plucky open source chinese models will? Come on. No-one gives a shit.
But also, I don't think you should use individual consumer behavior as a proxy for how the overall LLM market behaves. Current estimates are that individual users represent a pretty small share of overall genai revenue. Most of it comes from enterprise contracts. Those might be more price sensitive because the CFO gets a say in which vendor they choose.
For a couple years now I've been accustomed to choice of foundation model being a heavily scrutinized design decision because it's just about the easiest way to manage COGS. I can certainly imagine scenarios where, given the option to pay 1/2 as much in inference costs but with a 5% reduction in output quality, the boss would say, "Nah, that's a bad deal." But I have yet to encounter one in the wild.
Okay fine, it's $1,400 for a latest iPhone Pro with 512gb of storage. Point still stands.
> Most of it comes from enterprise contracts. Those might be more price sensitive because the CFO gets a say in which vendor they choose.
I think enterprise contracts are even less sensitive to price than consumer. That's the whole idea behind enterprise pricing. You'll see companies charge multiple times as much just for SSO for instance.
There are limits to this obviously, but to think that major commercial enterprises will switch over to some Chinese models to save some money is laughable.
So you switched away from the specialty product, but you're still comparing a top of the line 6.3 inch iPhone to a mid-market Android phone. Why not compare it to a Pixel 11 Pro with 512GB. That way they're at least comparable devices. But also, that would cost $1,220, which I suppose doesn't go well with the intent of suggesting iPhones are several times as expensive as Android phones.
The reason why I'm digging in on this is, when you start your argument with such a transparently silly comparison, it gives the impression that you're dealing in superlatives and willing to say just any old thing. Which undermines your own credibility.
Comparing two devices at the highest price point is besides the point.
Wait until you are forced onto an enterprise plan...
Many here don’t understand corporate finance do they? Hence the marketing hopium to fool people.
How many Astroturfers and Anthropic employees are here to convince us otherwise?
LLMs don't have that. For at least 90% of users, the model in the back end is completely irrelevant to them. Switching to Kimi won't lose them their iTunes music.
It’s pure genius. There’s a lot of bozo’s here who cannot appreciate it and cry about fairness.
How many people use an iPhone to its maximum? Isn't it the same with models? I think engineer's like to flex that they need a top of the line frontier model for their shitty corp projects. A Chinese model will write your Jenkinsfile just fine.
as a person, i would absolutely not pay more than $1K for ai, because at that point i can just get myself a bartening job
I know guy who run cafe from his IT salary.
Thinking, up to specific price, it's great tool for devs, and from some higher price, you think more about how to earn money back.
Is it wise to believe that once Anthropic (and OpenAI, for that matter) are public, that they won't start looking for ways to improve margins, such as increasing enterprise pricing even by single-digit percentages given it comprises the lion's share of their revenue?
e.g. adding 5% on top of API pricing isn't likely to cause noticeable churn if at all, and yet 5% of a massive number can make a meaningful difference in share price.
I don’t really know where this commodification idea came from, maybe MBA types that don’t actually understand how cloud economics work, applying basic first principles analysis?
The operational engineering that goes into serving a frontier model is _highly_ nontrivial, it is not commodifiable. This isn’t a SPA application.
No Western company will be allowed to choose a Chinese cloud provider by their legal department or by their stakeholders. Which leaves only the US cloud providers as viable sources (as us, Europeans, have been too busy building our revenues based on turism and other such non-tech thingies instead of building real stuff).
I don't really see it frankly. Admittedly I never tried it (I only played a bit with local models) but conceptually how is it much harder than all the other cloud services that already exist? My educated guess is that building a competitor to Netflix, Youtube or Dropbox seem technically a lot harder to get running than offering inference on demand for instance.
It seems that the main barrier for competing in that market at the moment is getting your hands on the GPUs/RAM and a datacenter to put them in.
You people are nuts. The failure risk is off the charts.
Anthropic needs to earn 52B a year in annual revenue to honor those specific commitments plus obviously its own operating expenses... and yeah that will require Anthropic to grow as a company, but their S-1 indicates 67.2B a year in revenue and sure maybe they stop growing and this is their peak... or maybe they do manage to grow further to 80B a year, 90B a year...
And unit economics.
HY1 2026 revenue less compute and infrastructure turned positive (45% gross compute margin; $8.9bn C&I to turn $16bn revenue) for the first time; it was -176% in FY 2025 with $7.3bn C&I on $4.6bn revenue.
as in, if they capture every single professional software dev and extract $1000 from them every month, their yearly revenue will be like $500B. do your own back of the envelope on that with ongoing costs and consider how unrealistic a goal this actually is.
even worse, the very product they offer is self limiting. if it becomes so good as to actually replace software devs, they will have less customers to extract money from.
their valuation essentially relies on a small percentage of mega users having unheard of levels of spending.
the part about a small percentage giving them most of their money is already true (their rumored $100+ bil revenue as of mid 2026 already outshines all software devs having a $100 max subscription), but they are two orders of magnitude or so away in terms of the actual amounts to make it work.
The market for Claude Code is much bigger than professional software devs. It's really anyone who uses a computer. I already know normal people who have been using it for personal software.
It's not gonna install a shower for you, but it will confidently walk you through that leaky faucet. And I don't know if you have paid a pro recently, but they don't charge $20 for a 20 minute fix.
"How often are people paying at least $600 annually for something AI can help them do instead"
and the other half is "How often can AI help someone boost their productivity over $600 in a year"
I also don't think people will even need $100 plans for this. Plans will probably land in the $60/mo range.
Muse is free. You can chat with Gemini in Google results for free.
AI recently told me that my oven was broken because a part that doesn't exist on the oven I own was malfunctioning. A bunch of people poking around with electronics, motors, etc. because they think their AI chatbot can fix anything and everything is going to lead to injuries, deaths, and most assuredly more money in the long haul.
I'm going to need an example on this one, and examples that can't be covered equally well by a YouTube video or an 20 year old forum post. It also has to be legal, as in not violating codes for electrics or plumping, I can violate those on my own for free.
Free tiers of essentially any AI model would be able to do the same. People will use what MS bundled for free.
The step-by-step customized hand holding of LLMs is something totally new
It's quite a leap to assume any kind of issue would require a $100 subscription. Plumbing, mechanics, electricity ... The issues are usually far easier to trace and debug than in software, because the systems have been designed by smarter people.
You're not buying some arcane and complex solution, you're buying the convenience of not watching a couple of youtube videos or articles of someone troubleshooting the exact issue you're having.
Most people aren't engineers and do not care about custom software.
Though I would expect the majority to do fine with $20/mo. That's the same effective cost as replacing a $500 computer every two years, which is the sort of expense most companies are comfortable doing without much second thought
What's your source to the contrary?
Total addressable *market.* If Anthropic still had market to address, we would still be seeing consistent growth.
I'm sorry and I'm sure you are being honest in knowing non-tech people using it, but this is on the whole false. The vast majority of people not only have no interest in "personal software" or using Claude for programming, but the slice of the population that does actively interact with AI is using free chatbots as a nu-Google. Most tech people don't even want the "personal software" thing, I don't know why this line is being pushed as if your grandma is going to make her own calorie counter app one day.
I tried to look for the valuation in the linked document, did you see it or any reference to a valuation elsewhere in anything published by Anthropic? what is "their valuation" that you refer to?
also, they previously already got some investments on a close to 1T valuation (in may).
I'm not saying I think thats what is going to happen, but it really isn't a stretch to think that this keeps running given that it hasn't really hit anywhere near as broad as we think it has (being HN community based).
I'm not going to predict whether Anthropic succeeds or fails, just that it's very misleading to claim that Anthropic needs to grow to $500b in revenue without making it clear that it's $500B in revenue per decade rather than $500B in revenue per year which is what most people reading such a number would assume.
Obviously it's not like having Claude Code or similar, but those would be very hard to run legally and safely in EU regulated industries like us in energy production. You could run Codex but it would require your IT department to have yet another place to manage security and software integrations. Which is a fairly big issue when you have around 1 IT staff member per 300 employees, and an organisation which will buy five different sharepoint solutions for the same problem.
Microsoft sells something that seamlessly integrates into everything else you buy from them. It's easy to manage credit spending and when I say that I don't mean that it's easy to set it up. It's easy to track and budget so the right departments pay for it. It can read your emails, your onedrive, your loop, your sharepoint, your powerbi/fabric and anything else you have access to. Sure it's annoying to be confronted with a point you've made 7 months ago (like how you think Copilot is utter trash) and since changed opinions on it's frightingly good at finding relevant stuff, even in some obscoure legal document on your intranet.
I assume that Anthropic and OpenAI makes money of this somehow, but for the most parts our users wouldn't care if it was the GPT and Claude models or Mistral or Deepseek. Which is why I asked if they had really gotten it right.
- 161.2B w/ Broadcom: 2026-2031
- 120.5B w/ SpaceX: 2026-2029
- 114.5B w/ Google: 2026-2033
- 110.0B w/ Amazon: 2026-2036
- 44.9B w/ Microsoft: 2026-2033
- 44.6B w/ Nscale: 2026-2033
- 41.3B w/ Lambda: 2026-2034
which is around $63.7 billion / year if spread over 10 years, but with most of it due earlier than that.
SpaceX was the only one mentioned as cancelable. These are listed under "Infrastructure Service Arrangements" on page 170.
If progress with frontier models doesn't continue at the same pace then the investments and funny money deals might start to look very fragile.
If progress does continue then that progress might be towards smaller models more than quantum steps up in quality or capability that would increase the risk of the scenario above.
All of this is set against a backdrop of public hostility and political problems that I can only imagine increasing in the near term as well.
This feels like an unusually high stakes game of poker. I don't think I'd want to bet on Big AI right now if I had a few billion lying around. Then again - maybe that's why I don't have a few billion lying around!
I saw an interesting observation the other day that Jensen Huang has played this game the best because his company is the one actually selling real products and making real money in return. If all the AI giants are buying up huge amounts of hardware speculatively because they assume they will need those data centres in the future and then the bottom falls out of the market then it's the AI companies left holding the baby. And if we do get to a point where local models are competitive then the OpenAIs and Anthropics of the industry might face real difficulties but it would just be a different market for Nvidia to also be selling real hardware and still in exchange for real money.
And thats just the inference side of things. For training, the big players like Anthropic would still be able to benefit from cheaper compute, by generating more RL rollouts, running more RL environments, and thus still providing frontier intelligence that would justify the markup.
Many customers cancelling would only be a risk if the compute expenditure was more decentralized. But its not
Is that really a viable path? (Honest question)
According to the U.S. Energy Information Administration (EIA), in 2024:
- *Texas natural gas production:* 27.5 billion cubic feet per day. - *Texas natural gas consumption:* 14.6 billion cubic feet per day.
So Texas produces nearly twice as much natural gas as it consumes domestically.
Natural gas contains approximately 1,025 BTU per cubic foot. Assuming 50% efficiency in a natural gas power plant, and 1 kWh equals 3,412 BTU.
27.5 − 14.6 = *12.9 billion cubic feet per day.*
12.9 billion × 1,025 BTU × 50% efficiency ÷ 3,412 BTU/kWh ÷ 24 hours = *approximately 80.7 GW of continuous electrical generation.*
That's the equivalent of more than EIGHT 10-GW AI computing campuses operating around the clock.
Lastly, a significant amount of natural gas is produced alongside crude oil. When producers can't economically capture or transport that gas, some of it is burned off through a process called flaring. Those flames you sometimes see at oil and gas facilities are an example, although refinery flares also serve important safety functions.
Seems like you just like to say snarky things about Elon, but can't be asked to do any actual math.
I do agree and understand we flare a lot of natural gas, waste because there's just not the economics to transfer it elsewhere. I definitely agree, it would be nice for that to be put to some economic use if we're pulling it out of the ground.
> Texas produces nearly twice as much natural gas as it consumes domestically.
And it exports that other half it doesn't consume domestically. It's not like half of what is produced just sits around forever. We've massively scaled up LNG exports. We export to surrounding states. But hey just act like we've got billions of cubic feet lying around doing nothing ready to go for datacenters.
> approximately 80.7 GW of continuous electrical generation.
Ooh wow. There's 420.8 GW of demand sitting in the ERCOT connection queue specifically for data centers today. All that "extra" and we'll cover like 20%. A good gas power plant will do better than 50%, but that's still not enough. And I imagine there will be even more datacenter demand coming soon.
I've also lived through continued cost increases of natural gas in Texas. What does adding an additional 420GW of demand for just datacenters do to our rates? How much more am I going to spend to heat my home this winter? How much more will I spend on my electric bill next summer?
I've also lived through the collapse of natural gas supplies in 2021. I wonder if these datacenters will turn off their turbines or let people freeze it something similar happened again?
Seems like you make a big point about doing the math but only doing half of it. Keep going, see where it takes you.
420.8 GW is alot of compute... like an ungodly amount. A vera rubin rack is like 220KW.. so 1.9m racks. a rack is 72 GPU's per rack so 420.8GW of power is 137.7M GPU's. So lets say about 60% of the power goes to infra, we get 55.08M GPU's. So, with each GPU costing around 55k, that gives us a final cost of compute for the 420GW is 3.0294 trillion.
Obviously that is not real... Eithier double booked, or fake corps making a bunch of requests from thousand different vectors to resell access.
Also, companies like Base are making a killing off taking all the excess energy generated by wind at Night and selling it back to the grid.
So in final, we can support around a $6T AI buildout in Texas just on our exported Natural Gas alone before we really need to start expanding LNG supply.
https://www.ercot.com/files/docs/2026/07/29/ERCOT-Senate-Jul...
And yeah, supposedly it's over $1T in computer capacity allegedly planned to be built.
> just on our exported Natural Gas alone before we really need to start expanding LNG supply.
That's natural gas which already has a buyer though. We're now adding a buyer who isn't batting an eye at spending over a trillion dollars. What price of natural gas do they stop buying at? What price will the people of Texas, surrounding states, and our international customers pay for gas until we can't afford it and freeze?
So yeah, if we do build out anything close to this I hope we manage to increase energy capacity similarly. Because I doubt I can spend as much to heat my home as these hyperscalers can spend to power their chips.
That's kind of the whole point of my "sucks the natural gas pipes dry". I'm not saying the gas pressure of the line going to my house is literally going to zero, but if it's so expensive I can't afford for it to flow it might as well be.
Remember i mentioned Base earlier? Those interconnects are for peps selling or arbing in energy. Not to pull energy out of it.
Gas only is like 30gw requested into 2033. I bet you a ton of that is from Elon companies. We don't even have to do the math i did earlier, cause it seems like it's built in.
All I'm certain of is all this new demand willing to pay any price means things will be more expensive just to try and not melt or freeze in the Texas weather.
FWIW I'm far from an AI detractor. But we need to make sure we're putting these costs where they belong, and some of these plans do seem insane.
What happens when companies have more GDP than the united states? Because i though the two would be correlated since American companies, but i guess if you can sell to a global market and the government can only tax it's citizens, that's kinda the inevitable outcome.
These hyperscalers are issuing so much debt and investors are certain enough of returns they're upending global bond markets. Tons of large countries are having to mess with interest rates because the groups who used to buy their debt are already tossing their money at AI datacenters at a massive scale.
> JPMorgan Chase estimated in June that $4.1 trillion in AI-related debt will be issued through 2030, as data center companies and others tied to the AI boom race to build up capacity.
https://www.cnbc.com/2026/09/27/debt-hungry-data-center-comp...
Then Elon goes to Memphis, buys a warehouse and fires up 30 jet turbines in the parking lot, and you start getting reports about the incessant noise and being able to smell it from 5 miles away.
It seems like there's a fundamental problem with realizing other people have agency and if you fuck them hard enough they will fight back.
If Anthropic's revenue story doesn't materialize and the AI bubble bursts, these companies are in a world of hurt. It's probably not fatal for Google or Amazon but would be a major drag on the stock price (perhaps why Google has dropped from 400 to 315 over the last couple months as the AI story soured), and it could be company-ending for Broadcom or SpaceX if they end up building out infrastructure to service these commitments.
I remember before I left Google, one of the TGIF questions (which are now AI syntheses of many peoples' individual opinions) was "What happens if AI is a bubble and it bursts?" His response was "If that happens, the whole industry is in trouble." I see now what he meant.
If the AI companies go tits-up, that introduces counterparty risk, the revenue doesn't materialize as cash, and suddenly their earnings take a nose-dive. If they've staffed up significantly to design and sell these AI accelerators and suddenly there's no market, they could be in a lot of hurt.
BRCM and Marvell are both the actual designers of many of the “we designed our own ai chip” chips that so many companies have been announcing.
Broadcom is also a major supplier of the networking chips that the hyperscalers use. Clusters are moving away from infiniband and nobody is particularly impressed by Spectrum.
Between SpaceX, Starlink, Tesla, Neuralink, X, Grok, Grok Bot & Cursor there's not a demand shortage.
Grok Bot's the fastest growing product in the history of the company and the moment you use it, it's clear why. Took me about 1-2 days to realize I'd have a very hard time going back to not having it available.
They can access anything you can, whether there's an MCP/CLI/API or not. I create speciality bots for each 3rd party tool that I use at this point. The bots can talk to each other, have group chats. You can give them access to the machines that you're running the software on and give specific bots permission to do things on certain machine (like an Obsidian bot that writes logs of daily bot activity on a specific machine with a Dropbox folder shared).
I got locked out of one machine while I was travelling and I asked Grok Bot on another machine if it could get into that one over SSH, so it sent me a temporary authorization link via Tailscale, connected to the machine and restarted the stuck process for me.
The usability is off the charts. It's really like having a team of people working for you. I've run a conference for the last few years and was getting completely overloaded. Now I have 14 bots specialized just for the conference helping me do it better than ever and participating in architectural discussions to build a system that help long term.
If you find yourself trying to juggle a lot in your life and work, it's a gift.
It's not impossible integrate AI to projects/products directly.
> Its largest commitment is with Broadcom, which co-develops the TPU AI processors with Google and leases them to Anthropic in deals worth $161bn over five years. From 2028, Anthropic expects to spend up to $150bn more on TPUs.
https://www.ft.com/content/a6c43402-719e-4295-ae5f-541edb0ae...
- 2026: $97.08B
- 2027: $97.08B
- 2028: $97.08B
- 2029: $97.08B
- 2030: $66.96B
- 2031: $66.96B
- 2032: $40.09B
- 2033: $40.09B
- 2034: $14.59B
- 2035: $10.00B
- 2036: $10.00B
Total: $637.0B
Considering everything that is cancellable, ie. hard commitment only, it's:
- 2026: $56.79B
- 2027: $56.79B
- 2028: $56.79B
- 2029: $56.79B
- 2030: $56.79B
- 2031: $56.79B
- 2032: $29.93B
- 2033: $29.93B
- 2034: $10.00B
- 2035: $10.00B
- 2036: $10.00B
Total: $431B
If they go bankrupt, there are some more cancellations available, also from the other side – if some providers don't deliver, it can be lower as well.
I also left out AMD at $20B for 2027, so that would need to jump up.
Finally, the above is only counting the commitments they've already made. It's not very likely they'll only spend $10B in 2035.
AMD is missing, true but commitment schedule isn't known. Also missing estimated additional TPU purchase for 28-31, ~$125B-$150B.
- 2026: $97.08B = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B
- 2027: $97.08B + AMD (TBD) = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + AMD (TBD)
- 2028: $128.33–134.58B + AMD (TBD) = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)
- 2029: $128.33–134.58B + AMD (TBD) = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)
- 2030: $98.21–104.46B + AMD (TBD) = Broadcom $26.87B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)
- 2031: $98.21–104.46B + AMD (TBD) = Broadcom $26.87B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)
- 2032: $40.09B = Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B
- 2033: $40.09B = Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B
- 2034: $14.59B = Amazon $10B + Lambda $4.59B
- 2035: $10.00B = Amazon $10B
- 2036: $10.00B = Amazon $10B
Combined total: $782–807B+, including estimated additional TPU purchases and conditional AMD spending.
A very large part of that revenue is paying for inference. There’s not enough leftover for the committed outlays. So it has to come from smucks I mean investors.
Idk where you got this figure from, but the actual figure are very different from that:
It's 120B over four years for SpaceX alone and 120B for Broadcom over 6 years. These two players, representing less than half of the total commitment, already give you an annual commitment above 50B.
This is just the compute, which primarily covers training + inference. They still have to run the rest of the business. If you think AI should look like SaaS and spend 10% of their revenue on compute, that's $1T in revenue. If you think it's a hardware business, that's 30% margins and perhaps $200b in revenue. I went with something in the middle at $500b.
They won't disappear as a company at $200b in revenue, but I don't think that's their aspiration. These commitments imply to me that they're confident they can do much better.
I'd think it's more sound to just look at what percent of their current spend is on compute? Seems to be 50-60%
But if you buy them today at a $1 trillion valuation, and they only make $40 billion in annual profit, your investment will go nowhere. They need to be growing profits at 20% per annum to justify a PE OF 25.
And then you have the competition problem. The cheap open weight models are good enough for many (most?) tasks.
And then you have the problem that LLMs may not be that useful outside of programming, customer support and law (and doing homework).
Last week I build my new project on free plan medium Luna OpenAI, totally fine.
A lot of new SaaS companies is just integrating AI into their products. Would be nice to follow this if they increase prices / how they are doing.
Not sure how big part of customers are random human coders / big tech / small-medium SaaS companies...
From S-1, some use-cases in many companies doesn't look convincingly.
The same logic applies here.
Google, Amazon, Microsoft, Meta, Broadcom, Nvidia, SpaceX and others are spending hundreds of billions a year each on fast-depreciating infrastructure to serve this potential AI demand. If the demand never materializes then they know they are in big trouble. So they have every incentive in the world to prop up Anthropic (and OpenAI) and keep the party going.
Can they keep the bubble inflated indefinitely? Probably not, but all these companies combined have heck of a lot more influence than Anthropic alone.
Source: https://x.com/a16z/status/2113986385908842870
"Zooming out, GPUs are still renting ahead of a standard 6-year depreciation schedule
Rental price vs. schedule: - H200: +143% - B200: +111% - RTX PRO 6000: +73% - H100: +54% - A100: +9%
Most of the gap opened up after April, when token demand took off"
I've brought my old GTX 1070 out of retirement for for some molecular dynamics these days because its "free" except for eh electricity and have Ben pleasantly surprised at what it can get done chugging along in the background. I'm considering buying some more, as it's kind of a steal in pricing.
A supply abundance reveals the inherent cost.
Heads I win, tails we all lose. GFC Part II.
But I see every reason to believe demand will taper off. It seems it’s already happening. Enterprises have been deciding that maybe corporate chatbot and coding agent budgets shouldn’t actually be infinite. AI startups are starting to successfully outcompete each other, effectively killing off members of Anthropic’s customer pool in the process. And consumer demand for AI products seems to be dropping. None of these are signs that suggest that Anthropic has thus far met only 10% of plausible long-term demand.
Then if inference does grow to a $2-3 trillion/year market that Anthropic is a big winner, as big as NVDA is now (they're going to be over $500 billion in revenue in calendar year 2027).
The pattern seems to be that it makes people more productive when you measure it in terms of raw work product volume. But not all work product is equally valuable, and we have yet to see a clear pattern of increased economic productivity. Instead we just keep finding new and interesting bits of jaggedness in the frontier.
Of course nobody can prove those economic gains will not manifest at some point in the future. I can’t prove that Jesus isn’t coming soon, either. But I also can’t prove which millenarian denomination, if any, he would choose to bless in the event that he does come back.
personally, i want a 2 hour 3 martini business lunch, not to ship more slop at all hours of the day
There have been no signs of "super productivity" in the larger economy due to AI. Boosts in productivity have, so far, been seen exclusively in sectors like tech.
It's part of the reason so many of the dark enlightenment dorks are pushing more towards automation of control.
So what that means is that if the demand exists for any AI data centre, it's a solid business to run one — they cost maybe $50-60B to build out, they cost maybe $10B per year to run. You get your money back fast if the demand is there? And you can project being able to replace with better GPUs in the future without having to borrow for all of it.
But it is extremely difficult to believe that Anthropic is not projecting that all of its customers will ultimately use their service to the extent that their outlier customers do right now. They probably won't be projecting that their customers will find the need or ways to use them less.
And also, "no customer is more than 10% of our revenue" means that one customer is. If that customer is Microsoft (can it be? I honestly don't know) then things get tricky, because Microsoft clearly want an unmetered, on-device AI future.
Does that have to be true? I thought IPOs were about selling the company.
Never mind the fact that companies do not think and this is like saying "My lemonade stand thinks it's going to have good sales this quarter."
That they are IPOing tells me that the massive amount of funding they need is no longer available privately and that this is their 1 opportunity to get the funds they think they need to get there.
Biggest risk is the growth stalls as everyone moves to open weight models. They certainly like telling stories.
snark aside, they do have the best brand, the business team was constrained to one choice for simplicity, they picked claude
The tool is not a moat when you can tell Opus to build you a better tool.
I’d say it’s more Notion than GDocs for now, but the direction is clear. Gobbling up all the SaaS in order of descending value to agents.
My bet is on promoting their internal Tasks representation to something more like a Jira/Linear.
But then obviously Slack is another one, and Duet is already starting to build a “Channel” abstraction, so making Duet multiplayer is a clear possible next step.
So far not a single company has ever demonstrated that.
Even if demand is through the roof which it seems to be, it doesn’t matter if every dollar they sell costs them more than a dollar.
They paid for their compute in equity in the company, not cash. Therefore their losses don’t show up as cash. I don’t understand how people are this gullible.
Tech progressed, no need to dedicate mobile developers, guy for app deployment etc.
Later, team consists only from 3 devs, it earns money, no other features needed really much.
Now you see many companies refactoring whole app in week using AI. But later, for what will they spend the biggest part of the AI budget?
I see from business perspective, they try to build dependency on their tools, make it great, but ultimately, it doesn't ensure making money.
Friends company was paying 300-400k EUR monthly for AI, cut is expected for next year .. and even if they cut tomorrow, business runs.
Will anthropics models continue to be significantly better than openweight models? Even if their latest model is, Will the open weights gobble up all of the demand for lower models? (I.e. will anthropic be able to recoup training cost while their model is the best, will they make enough revenue on the long tail or will it get crushed by the open weights)
“Robbery, Italian Style”
https://www.youtube.com/watch?v=vktTgOr4M1s
(it has a wiki page https://en.wikipedia.org/wiki/Ograblenie_po... )
Marcello Mastroianni is performing a bank heist, it goes well - but in the end he has to return all the money he borrowed. And that's where he is left with nothing...
The circular deals underpinning the AI boom is going to turn into a game of hot potato. When the music stops, we'll find out who is left holding the debt.
This whole situation smells like a huge Ponzi scheme being played by corps.
Zero sympathy for all the people who are going to loose a lot of money on the stock market.
First to fall is the software that has basically been abandoned but industry leading and has sticky customers. We're replacing those right now. Our version is more accurate, ten times faster, and a hundred times cheaper.
Our claude bills are rivaling our datadog bills and might get to the size of our AWS bills. It's pretty crazy.
https://youtu.be/roe3SgezmmU
Why would they need support personnel? Isn't their whole point that Claude can do that
Social engineering is an issue with regular support, with agents it's a lot worse. And with agents it's repeatable: if I find one conversation flow to get it to give me money I can instantly repeat that with a thousand fresh accounts, and then share the same "exploit" on reddit for everyone else
If frontier AI deployed at frontier labs is not capable of "resetting 2FA, issuing refunds, handling payment glitches, etc." then the current and near-future capability of AI is being severely overstated.
captain_renault_casablanca.gif