Comment by simias
1 day ago
The problem for me is that it's really hard to project what this industry is going to look like 6 months from now, let alone 6 years.
What's the endgame once the dust settles? Is it going to be a utility like water or electricity and the cheapest offer will prevail, in which case margins will be crushed? Or is Anthropic (or some other big AI lab) going to swallow everything and become the Google Search of LLMs, with everybody else fighting for scraps?
Or is it local compute, models becoming more efficient and hardware becoming cheaper (again) and companies hosting their own AI clusters like they would run SAP?
I think there is a plausible path where this valuation actually works out, but it's a narrow one IMO. I also want to acknowledge that I personally would prefer if it wasn't the case and a more distributed, open model prevailed.
I just look at technology over the last 50 years and see: decreasing prices, increasing capabilities. Barring temporary hiccups. Altman & co did a very smart thing by locking up the GPUs and the RAM, they effectively created a temporary near monopoly. At the same time: that incentivized the rest of the world to look more closely at efficiency negating some of the gains of that expense.
It's just like a utility with the slight difference that it can be sold at a distance without significant cost for transportation, in many ways it is the ultimate digital good. Unlike Spotify copyright is not an issue, unlike your electricity, gas, internet or water company there is no last mile problem. Unlike Google there are a ton of qualified contenders already (Google had to be better than Altavista, which wasn't hard, but they were so much better they managed to lock up search for long enough to cement their position).
Anthropic has none of the advantages and all of the disadvantages. On top of that all AI companies seem to be run by megalomaniacs.
> they were so much better they managed to lock up search for long enough to cement their position
There is still a chance for something like this to happen with AI, no?
> On top of that all AI companies seem to be run by megalomaniacs.
Not all of them. Only those who outspend all the others and then you never really hear of any others. It's like the most basic requirement for the job.
Fair. I know of at least one other though.
> On top of that all AI companies seem to be run by megalomaniacs.
Isn’t that what it takes though? Doesn’t startup culture celebrate this attitude?
> Doesn’t startup culture celebrate this attitude?
Some do, unfortunately. But that's not all of them.
Apple certainly did.
3 replies →
> On top of that all AI companies seem to be run by megalomaniacs.
I think this might genuinely going to become a problem for them in the long run. People just absolutely hate their guts. That's a very strong motivator to do everything possible to get rid of them.
We'll of course see in 5-10 years if the doom PR will work out, but that always was a rather high-risk gamble, and it already looks like it might not.
There is a significant amount of businesses and people that are very willing to switch to literally anyone else just out of spite alone.
It's not necessarily spite as well, it goes without saying that international relations between the US and its allies have deteriorated to the point yet another critical dependency Washington can turn off for geopolitical reasons isn't a very attractive idea for many. While I dislike a lot of Anthropic's behaviour, it was not their actions but the fact their models were subject to export controls which really motivated me to not be dependent on US providers alone.
My country has no frontier models; since I've no national grounds to support one provider over another, a strong degree of provider diversity is an obvious benefit to me. I also think a highly competitive global market is a good thing for mankind in general, any hegemonic power having a monopolistic or oligopolistic position in AI provision is probably the 'bad ending' for most of us since it represents an enormous concentration of democratically unaccountable power regardless of the power which 'wins'. I'd argue it's rational, not spiteful to favour provider diversity on those grounds.
2 replies →
>doom PR
https://www.astralcodexten.com/p/ai-safety-is-not-a-corporat...
As a French person myself, I think this antipathy is less strong in non-English speaking countries, where the population (even if they understand English) is less exposed to the antisocial rants of these megalomaniac idiots.
Someone like me who spends a lot of time in the anglosphere puts faces on companies like OpenAI or Anthropic. I doubt that's true for most people, especially outside of the USA. They're just big companies, like 3M, DHL or Toyota. Who knows what the management of these companies think of the future of humankind?
I would expect than personality-driven rejection of American tech from abroad nowadays is driven mainly by Trump and, to a much lesser extent, Musk. Everybody else is just "some guy" to 99% of the population.
5 replies →
I know I'm just one person, but I absolutely detest all of the rhetoric coming from these companies and what it means for software freedom - the FSF kind. I ABHOR the idea that free software developers have to either use the hosted (controlled) service, or go way slower than proprietary software devs.
So I'm doing what I can - I only use local models for 100% of my personal coding. (And their capability is amazing! Qwen 3.8 Flash Next leaves me wanting for nothing at this point.)
And I know open weights aren't enough. I'm ready to support open training if/when we get around to that. I hope once hardware prices ease, we can see homelabbers building small AI hosts (small by homelab standards yknow. So let's say 512GB host ram and 1-2 32GB cards, though i hope we also start to see AI cards with lots of vram but less compute cores). Then maybe we can do folding@home style distributed training.
In a time when software freedom is perhaps more threatened than ever, I'm happy to see that information still WANTS to be free, perhaps even more than ever. Look at how easy it is to distill claude! Despite, presumably, anthropic working very hard to prevent it.
10 replies →
I don't hate them. They're doing amazing work, and I'd like to believe they're genuinely trying to do good for humanity. I just think it's dangerous for any group of people to think they can impose their morality on everyone (only Dario knows what's right and wrong in military actions, only one company in the world, welfare to the current models, humanity has 10% of chance of getting destroyed by AI, and etc). For that, I'd rather see open-weighted and hopefully even open-sourced models win
The only thing protecting America's AI industry is the cartel of hardware and embargoes against sharing resources. Once those are cracked, they have zero moat. We're talking 1's and 0's at the end of the day, not nuclear engineering or nuclear fusion or some other revolutionary way to increase the value of real property.
>not nuclear engineering or nuclear fusion or some other revolutionary way to increase the value of real property.
Is engineering the factories that make the hardware not nuclear engineering? Even if not, considering there exists only 1 supply chain in the world capable of making them, it seems pretty revolutionary. And that moat seems pretty big, given that it hasn't been crossed for many years.
1 reply →
[dead]
This is a great analysis
The long-term is considerably easier because in the short-term there are limits on supply so prices aren't really able to move in a way where balance will be found.
In the long-term, the marginal cost of a request is zero so prices will plunge and a lot of the fixed investment will get stuck. The difference between the cost of training a model/infra and serving a request from that model is a clear problem for economics.
What happens with Anthropic specifically is less clear because they, as every company in the same position in the last 150 years has done, are going to push for regulation. A lot of the risk has also been shifted onto neoclouds. The value of the customer relationship is, however, effectively zero. A core part of their strategy has been to build stuff out like CC to differentiate that but, ultimately, that won't work.
The other strategy they have employed is trying to get other companies to stop training new models. Again, seems unlikely to work.
There is no scenario in which they become Google search outside of government intervention. Obviously this is a key part of their strategy now but at best, this will only work in the US and Europe (I would actually say they are much further along in Europe where there is a strong tendency towards corporatism and centralisation of power, they could be "Europe's AI").
The only thing regulatory capture in the USA will do is to improve the situation for the rest of the world to compete. I wonder how long it will take them to figure that one out, for all their smarts they seem to be incapable of reading a history book.
This "regulatory capture" stuff mostly just exists in the imaginations of HN users anyway. This is just one reason why it never made sense as a strategy.
https://www.astralcodexten.com/p/ai-safety-is-not-a-corporat...
> improve the situation for the rest of the world to compete
The risk is that USA and China will come together on a joint "AI non-proliferation" agreement in the same way that nuclear technology became restricted by the countries that already had the capability. This will lead to a de facto techno-colonization of the rest of the world just between themselves.
This was a topic of discussion at Trump and Xi's AI summit just last month.
9 replies →
Although I agree inference will become cheap (probably not zero), it's quite possible that there is a moat they can build around understanding customer context.
With each query OpenAI and Anthropic get more context about their users. When I compare answers between AIs the ChatGPT answer often includes useful things it inferred from earlier conversations. Losing that context could be a meaningful barrier for a customer to switch to a cheaper service.
> With each query OpenAI and Anthropic get more context about their users.
Not unless they break their own TOS. The temptation must be very large to do so and is probably why those settings keep mysteriously shifting back to 'yes, I give all my data to you to do with as you please'.
2 replies →
Wild how even when humanity achieves a level of abundance there’s somehow a need from a few on wall st to create artificial scarcity.
They can't help themselves. They all saw that movie and figure they can be stars too.
The evidence we have so far points to LLMs becoming commodities. I can’t think of a technology in the last 30 years where so many companies are within ankle biting distance of the leader. There seems to be no data aggregation moat like with Google.
Distilling just works too well. It costs much more to aggregate the data in the first place and train the first model, but once that exists and you are selling access, the second guy can train their model by using the first model to generate the training data.
The only ways to fight this is trying to either: 1. Block it by cutting off access, which is pretty tough because it basically means you cannot offer unmonitored access which is a no-no to a lot of customers, and even then the workload looks very similar to actual real workloads and if the company doing the distilling is a bit sneaky and using intermediaries etc it's almost impossible to stop, or:
2. Try to use IP law to block the practice, which is pretty gross given how they got the data in the first place, and so far no court has supported them on this. And even if they can make it illegal, it won't actually help because there are countries that see having top-tier AI as a national priority and will actually support their companies breaking laws to get there.
People misunderstand why Anthropic is being so stingy about granting access to their top models. It's not some kind of marketing ploy, they know that if they release them fully openly then there will be a Chinese open-weights model in 3 months that is just as good. Also the reason why they try to hide chain-of-thought and the probability distributions.
> Distilling just works too well.
I don't think the evidence supports this claim. All the important capability gains now come from RL rollouts in real environments. You can't fake that by generating text from a teacher model. Anthropic makes a lot of noise about how China is distilling their models, but all Chinese labs can access is text outputs. They can't even get the logits. Now, compare all of the recent US open-weights models that are far behind the Chinese frontier. Why haven't they simply "distilled" all of those powerful open weight models that are available to them?
Given Anthropic is basically just “distilling” the IP of others they’re in no position to cry foul when others run the same playbook on Anthropic IP. They can pound sand on that front.
2 replies →
> the second guy can train their model by using the first model to generate the training data
Isn't this somewhat akin to a model learning from its own inputs? Doesn't that behavior lead to model collapse? Either the model is trained on pristine data, in which case it stagnates, or it's distilled from another model, which leads to model collapse.
1 reply →
Anthropic has made a lot of noise about this but has not offered a shred of proof. CEO about to do IPO is not a trustworthy source.
3 replies →
The interface being arbitrary text makes the moat even looser. The cost of experimentation with alternatives is getting lower and lower.
It’s going to be hard for companies to justify paying $1000s per month to frontier providers when they see what’s doable with open versions.
multiple thousands per employee sure.... but you will be surprised how much non-compensation employee expenses can be
I think people often underestimate how far behind most of the competition is.
Isn't the first open model that likely surpasses original Mythos from February this week's Kimi K3.1 Pro?
Nothing compares to September's Opus 5.5 yet.
Now the question is whether Anthropic is preparing another phenomenal upgrade with Fable 5.5, or if the progress eventually slows down and competitors catch up.
It means nothing for the companies to compete in benchmarks if the consumers can't tell the difference.
1 reply →
I have a little different take..
Top models, capable of discovering and inventing stuff, held back from public, and available only as for-rent research service (or whatever the relevant bureaucratic term is) to big corporations and government agencies like NASA, DARPA, NSF, NIST, DOE, etc. only.
I have a feeling that we, the public, will stop getting access to more intelligent models starting next year.. i mean, not the incremental ones, but, major ones.. like the one you got going from opus 5 to opus 5.5.
The economics of that don't work out. Government is a small market and can't drive frontier innovation at the scale it's happening today, let alone in the future.
Look at the defense industry: The Big 5 contractors (Boeing, RTX, General Dynamics, Lockheed Martin, Northrop Grumman) capture most of the discretionary military budget, and the largest of them is in the $300B valuation range (Boeing, which has substantial non-government revenue). Fine businesses but nowhere near Anthropic's valuation.
Bottom line, the government doesn't have deep enough pockets.
I disagree. I think it does make sense the way it was described; top models for governments or industrial players with deep pockets, and "normal" models for the general public. None of the big 5 defense contractors sells directly to the wider public.
In fact, I'd assume this direction has been widely discussed and considered at Anthropic. I think Fable's release was an experiment to test this.
The biggest risks though are China and Moore's law. In the short term, if you don't offer a competitive model as the market grows, you fall behind in marketshare. Anthropic's edge is that it's offering is still better than Chinese offerings. If they offer anything less than that, something not better but just comparable for example, they lose on price. In the long term, as computing capacity continues to increase, local models potentially pose the same challenge.
So there's plenty of pressure for Anthropic to keep widely offering the best model they can.
1 reply →
this suffers from the same prisoner's dilemma we already see at play. Any player can defect and in theory capture a "trillion dollar market". Once they IPO, then their service is to maximize (short term) shareholder value. Saying "we reject all the billions in recurring subscription revenue because we are hoping all AI providers will do the same, therefore creating a moat" isn't going to go over well.
The SOTA model dilemma is that you have to pay more and more to train models that focus on less and less of the AI task market, because 95% of tasks can be done by cheap small models.
1 reply →
I think GP's thesis is that selling only to the US government would preserve their moat by preventing distillation.
3 replies →
There is a good chance of that. This will incentivize the rest of the world. Keep in mind that the world is much larger than the USA, that the USA has alienated a lot of parties that were its allies and friends not that long ago and that there are far more talented people outside of the USA than in it. Silicon Valley had the world by the short and curlies and then Trump came along and pretty much burned it all down. It will take another decade to know how big the damage is but it is probably in the trillions of $ of lost income and that's before we get into intangibles.
I don't want to make this discussion politicized.. but man... Trump truly burned everything down that the US built over decades of hard work and careful planning. And these things are extremely difficult to build back, especially in a world like today's.
11 replies →
How is it not clear that LLMs are being increasingly commoditized and that the foundation model producers are in a bad race to the bottom with each other AND cheap open-weight competition?
Not even just from Chinese models, but from companies/labs like Mercury
Stop spreading lies - for most people they are.
E.g I know people in high finance who switched - they don’t need frontier for many tasks - Anthropic is getting jack all revenue from them.
If the tech improves massively - Anthropic still gets nothing - since firms will want on-prem or cloud offerings that ensure zero data leakage / IP.
You lot haven’t played this out.
I think you either grossly misread my comment or are responding to someone else, because my position seems to be pretty much the same as yours
I wouldn't bet on local compute. Even if it may end up cheaper in the long run, the convenience of just another SaaS subscription sure beats the high upfront const and continuous maintenance required with local deployments.
We've seen businesses moving everything to the cloud before AI, we'll see more of that as regulatory requirements are increasing.
The difference is you can get the same model from different providers so they compete on price and features (ZDR, security, fast inference). It's less like Netflix and more like Apple Music.
I agree that the SaaS solution will be a winner, but the price diff cannot be too big or it would be too economical to run it yourself.
I agree it's hard to predict.
It's very possible given the current environment that LLMs go the way of airlines. Low margin businesses that are essential to world economics with heavy price competition.
and airlines are still finding ways to enshittify. flights included meals, now you're (un)lucky if they include biscoff. and that'll be $50 to check a bag. I can't imagine how frustrating an enshittified LLM would be. hey - I noticed this is a finance-related query. those cost 100% more.
Variable token cost depending on context is definitely something that's coming if they manage to lock up the market.
It's going to be all about the training set. Theres no moat in models trained on internet data.
There are tons of private datasets where a model will be fine tuned to provide useful domain specific responses.
And this is why I think companies will wise up and stop sending their information to the likes of Anthropic and OpenAI: they will not be able to maintain their edge if either one of them decides to defect and start using all of their uploaded user data to improve the model. So they have a massive incentive to break their own TOS.
That's the push factor. The pull is, say there's a giant repo like windows or Adobe. LLMs produce spaghetti code because the repo doesnt fit in the context and it wasn't in the training set. If Adobe fine tuned it the LLM would be much more useful.
4 replies →
The idea that noisy usage data is worth that much for training as to balance the fatal risk of the crime you describe, imo borders on a conspiracy theory.
I am under the impression that curated training pipelines are driving the gains, not utilizing random usage data in nefarious ways.
>The problem for me is that it's really hard to project what this industry is going to look like 6 months from now, let alone 6 years.
i keep thinking how crazy it is we went from ChatGPT releasing in 2022 to multi-trillion IPOs 4 years later. the industry still hasn't agreed on how to generally use anything. crazy times ahead.
It's going to be "all of the above" because AI has a whole lot of use cases.
For certain AI use cases, AI will be like a utility. For other user cases you need to pay a premium for the frontier. For other use cases AI will be local.
Even in the industries you mention (water and electricity) you have all sorts of businesses. Like people selling water in fancy cans for a large markup or batteries vs wall power.
When google became the king of search the incumbents were so far behind that catching up wasnt possible and they thought their position was entrenched anyway. With LLMs openAI's competitors are at or close to their capability already and everyone knows they need to keep spending to remain relevant. So unless the cash pipeline dries up for competitors that is unlikely to happen. The reason for that is anyone who has invested knows that either they give up and hand everything to their competitors or someone else does. Which means you need to keep spending to outlast your competitors. These companies are too rich and there is too much VC money around and enough billionaires with more money than they and their descendants could spent in many lifetimes, and a huge amount of greed. If the money runs out that is because some very bad shit has occurred and LLMs will be the leats of our worries.
Google is a bad comparison because they arent charging money. People swap streaming services all the time, analogously.
>Is it going to be a utility like water or electricity and the cheapest offer will prevail
This
Like ISPs, right?
In the short term perhaps, but I like to think that unlike ISPs, this kind service isn't bound by network effects. You can't just "generate your own internet"; like it or not you have to connect to something. With time though I think it'll increasingly be possible and cost-effective to generate inference locally.
So like electricity, where most people will probably just use the network, but a few will opt to invest upfront and generate their own.
>> What's the endgame once the dust settles?
Given the trillion-dollar spend, the theory is that every many woman and child in the western world will devote a substantial portion of their income (like 5-10%) towards some sort of AI subscription.
or...
That doesn't happen. All the AI investments go south and every man woman and child in the western world will devote a substantial portion of their income (like 5-10%) to bailing out an economic system that went all-in on AI. Either way, the money has already been spent.
You left out one...
The already shaky economic system (and associated political systems) can't take the strain of every man woman and child in the western world devoting a substantial portion of their income (like 5-10%) to LLM fallout, and we end up with a collapse scenario of some sort. (Which, to be clear, is near-certain to be worse than either of the above.)
I don't see a path forward for valuations like Anthropic, either the hardware requirements fall off a cliff (and why would people pay Anthropic so much if it runs on a cheap laptop?) or either they don't and they will never have the capabilities to serve all their customers (and why would they be valued so much then?).
It's a lose/lose situation here for Anthropic (and OpenAI) whatever happens.
> Or is it local compute, models becoming more efficient and [...] companies hosting their own AI clusters like they would run SAP?
Already happening, which is requiring a lot of enterprise DC power retrofits.
> Or is Anthropic (or some other big AI lab) going to swallow everything and become the Google Search of LLMs, with everybody else fighting for scraps?
Of all scenarios, based on the data we have now, this seems to be the least likely one.