200-300B might be plausible depending on how market perceives the company
How can you confidently say that? At some point they oversaturate and growth stalls. Didn’t F100 have a major pullback already this year on AI spend?
SpaceX is down, Cerebras is down, etc.
I guess it applies to everything. Universal. Like LLMs.
What is "this" referring to here?
the difference in which side of this bet people take.
of course each stock price is many bets intersecting at the same time (with all their correlations of course)
stocks are claims on future cash flows, currency is claim on future purchasing power, but all can be though of as forms of money, where the prices are based on background information [ie. priors, beliefs, experience] and people's prediction of the future supply and demand.
and the general abstract definition of said bet is asking where are you in the information flow. if you know with sufficient certainty how the people afer you will evaluate some news then you can guess which side to take to make some money.
The SOTA US labs are one chinese open model away from being abandoned overnight by the masses. Ive argued before, but argue again here, that the current hardware shortages are the main thing keeping this from being a reality currently. If more people were able to run local models for a reasonable all-in cost, we would have a much higher velocity of iteration and improvement in the open source community. But its still too niche while people cant justify the cost, and so they continue to pay the premium for inference.
If a Chinese company released an open weight model that somehow was so superior to Anthropic and OpenAI then they can just take the model and deploy it, reverse engineer it, use it to improve their models, and/or do any number of things they currently do today which are of increasing value.
Most people don’t want to or have any interest in running or managing models locally.its not just about cost. The market for running models is infinitesimal in comparison to the market for managed products. That’s why everyone uses Office and not Open Office. Nobody is going to run some Chinese open weight model no matter how good it is on their critical banking infrastructure or something.
There will be space for lots of different solutions and many products and types of models including Chinese open-weight models and American ones too once we get around to making money on them.
The game isn’t just “cheapest open-weight model at all costs” which is what you are asserting here.
> The game isn’t just “cheapest open-weight model at all costs” which is what you are asserting here.
Decide: is the claim that open weight models will be more intelligent than Anthropic/OpenAI SOTA models, or that they will be cheaper?
> Most people don’t want to or have any interest in running or managing models locally
Most people are also not the primary customers of Anthropic! 5 companies make up 25% of Anthropic's revenue, and I bet there a ton of companies spending tens of $M/month on Anthropic/OpenAI that make up much of the rest. These are all medium/big companies that have the means for self-hosting models if they decide that's more convenient. The cost is of course not the whole story here, since Anthropic/OpenAI generally make better models and that allows using them in more situations, where perhaps other models would fail, but this is also a big cost for Anthropic/OpenAI to continue getting ahead, and the moment they stop they lose this advantage. Also in my experience self hosted open weight models are _fast_, which is _very_ nice for all those tasks where you don't need SOTA intelligence to get things done.
Where is their moat? Ultimately, where is their profit? answer? There is none at least there is none that will come close to whatever imaginary valuation Wall Street comes up with, in the best of times, while Wall Street may be able to carry one such company, but two within the same year?
No one has shown a network effect for models yet. Models are commodities. Anthropic etc will make the sort of profits one makes selling commodity goods (a couple of percent), plus a perhaps a bit of extra margin from being a few months ahead of open weight models (if that profit can justify the billions in training costs).
I do agree that local models are irrelevant. But local inference and open weights are different things. You can go on openrouter and see that there are dozens of companies selling open weight inference.
For one thing, one needn't run a model locally, that's a separate--but also very damning--issue.
For another, that's fine if Anthropic and OpenAI can do that, but they will have sunk a bunch of cost into their own training already and people wouldn't pay them what they need when there is always a vastly cheaper substitute that gets the job done.
When I argue the hardware cost, I should note that I think this is the limiting factor for _improvement_ of these models. But ultimately I dont think most users will self host in that world. They will still pay for inference, but it will be for one of a dozen commoditized vendors who are able to reap the benefits of inference at scale.
It's the same reason that open source software has been a commercial failure except for a few evangelists or whoever becomes a cool "thought leader" by releasing a new library and then gets invited to conferences to speak. Software was commoditized in many ways, yet all the big software vendors made more money than GOD because they took whatever commodity parts came to exist and monetized them.
I think the models can be improved with either recursive improvement (it would be faster with better hardware but if you can do it at all you just need time) and algorithmic developments. I don't know what those would be, but usually we see improvements to both over time. The algorithms will tend to be trade secrets as well.
The problem here is that let's say it's true that the Chinese are cutting off US companies at the knees. Well the US can just do the same thing back to the Chinese. We might have some money lost and stuff but the data centers will continue to be useful, as will all the infrastructure and energy grid buildout so it's not a total loss (I couldn't care less if accredited investors like myself lose money - you know the risk). Then we just steal Chinese models, do the open-weights thing, and all that stuff. So where does that leave us? Nobody makes any money? (Also where is Europe in all this?)
> I also dont think these need to be better than SOTA, just reach a "good enough" bar.
I agree with you in some respects but I'd like to digress here a bit. Good enough for a general purpose model isn't good enough. You want the best or at least what performs to your expectations. I don't even care about old models nor how much they cost (in my personal life and of which I will be a paying customer as will tens of millions of other people) I just always use the latest and best one unless something else triggers me to use a different one. It reminds me a bit about software updates. Folks around here are like, I'm not upgrading to macOS 17.2 and I'm staying on 17.3! Regular people? Not only can they not tell the difference they don't care - just give me the latest and best thing. period.
On the other hand you can have models that are OK at general purpose tasks but are perhaps specialized for certain applications (software or hardware). Those could wind up being on a spectrum from cheaper (and integrated with lower-quality products kind of like an Android Tv model) to very expensive and high-performant.
I don't see a world in which people are by and large sifting through well today I want to use Kimi 1.1 and well later today I'm going to switch over to Claude Lite 18.7 and dang it's not giving me what I want so I'll switch to whatever model. That's not how people operate - they don't want the cognitive overhead of trying to mess around with something that's good enough. They may have to suffer that based on their economic means, but it isn't going to be an expressed preference outside of very niche categories (like home-brew hackers playing with their toys).
Similarly this is just my opinion as well. I enjoy a good discussion here and I hope you are as well. Positive intents, even if my words my be curt, at times.
I have no issues with this thread : D But back to the discussion...
Again, I elided a lot of thoughts on the whole situation, but the other factor in my belief that this is an inevitability is that the SOTA labs churn too fast in both model behavior but also harness behavior. Normie businesses of the future are not going to hitch their wagon to a moving-target of both behavior and enforcement vs just pinning the equivalent of "opus 4.5" from a run of the mill inference provider. It just doesnt make business sense to be at the whim of a model update. And the entire perpetual motion machine aspect of the SOTA financial outlook _necessitates_ constant churn. If they dont release, they die.
> So where does that leave us? Nobody makes any money?
I think this leaves us with SOTA labs pushing hard for protectionist regulation to essentially require federal blessing to be an inference provider. I also think we will see a _ton_ of dark money poured into research to try and show things like theoretical smuggling/side-channel exfil attacks on models/model weights so that the SOTA companies can paint the _open models and model weights themselves_ as a bogeyman. If they can show that theres potentially latent attack vectors in those paths then that will then prevent all the "middle class" companies from being able to use them, especially if they need to adhere to eg SOC2/HIPAA certification etc
Also, because of the three hyenas of memory, the Chinese probably eventually will have control of the memory market too in the near 3 to 5 year future.
Win on what terms, or in what sense? We've seen this play out via Android versus iOS for example. The cheapest models are more prevalent, but Apple makes all the money. I'd be perfectly happy being the Apple in this case.
It's like you can go and buy a cheap Android phone. It technically does all the same things as the latest iPhone, but it does them worse. So Apple ends up making all the money and delivering the product that people go on to use to do all sorts of other economically advantageous things that make them even more money. Where does that leave the Android phone/model? In the dust, for the most part.
What you are arguing is that all progress on models (outside of government labs or internal corporate workings) will effectively stop because there won't be funding for it. That seems unlikely to me. You just need to play it out...
US companies/investors go and invest a trillion dollars or something on models. Someone "does stuff" and reverse engineers their model. They go out of business (you said so [1]).
So now what happens?
Well, the US/DARPA/universities/whatever (all countries with interest) will keep working on this stuff and there will be some funding for it but that's about it. Military applications, physics, things like that.
Chinese companies will also fail and it will not be possible to create new SOTA ones because they'll obviously just be reverse engineered too and given away for free or at super low pricing - whatever happens in America must happen in China too.
So who will build new models? There's nobody around to pay for it besides governments or maybe corporate labs. But corporations and governments will protect the models (obviously, as OpenAI and Anthropic do today to varying extents) and keep the weights and stuff closed source. They'll have some productive use for them even if it's not in the download Claude Code and have at it style of B2C/B2B we have today.
Or
Maybe the trillions of dollars of investment going into these companies actually makes sense and the threat of open-weight models isn't what you and others seem to be suggesting it is. Whatever fate we seem to want to assign to America also happens to China here. To the extent they are behind, they're still running the same exact race. If open-weight models undercut companies they'll also undercut Chinese companies.
Just as Open Office exists today, so does Microsoft Office. Microsoft just makes all the money. I'm not sure why some seem to think that they can use the open source comparison here but not also use the open source economics that we've seen play out.
I think as a productive exercise someone should run this argument and explain why Chinese companies won't also be eaten alive by this alleged race to the bottom and have all of their stuff reverse engineered and given away. How can they invest in new AI companies with SOTA aspirations? Whatever argument you want to make for the US applies to China as well, and we can just give away freedom models to all the countries just like supposedly they will.
[1] "The important distinction is SOTA US labs - ANT and OAI - who can not economically survive reverse engineering / fast following PRC models at PRC pricing."
>whatever happens in America must happen in China too.
No, because PRC literally could not do what US did... they simply didn't have access to enough compute to dump so much $$$ capitalization for SOTA race. On the other hand, on paper US WITH MORE compute and (currently) CHEAPER compute can put forth cheaper models to undercut PRC open models... but functionally they can't for the same reason cheap PRC operators can't win in US - geopolitics. Undermining PRC AI requires giving PRC cheap US AI on cheap US compute which is sanctioned. But ultimately that is entirely irrelevant to the fact that CHEAPER US models will undercut EXPENSIVE SOTA US models. The undercutting is not symmetrical. US companies can use cheap PRC hammers (or even cheaper US hammers) to break $$$ US SOTA. US who on paper can provide even cheaper hammers to PRC... but cannot sell those hammers... so PRC stuck using cheap hammers... and since PRC never had chance to spend $$$ on SOTA, there's no equivalent valuable things to break even if they had cheaper hammers.
> trillions of dollars of investment going into these companies actually makes sense
We know the context for which those $$$ investment makes sense... if SOTA held meaningful lead to sell high volume of $$$ tokens which was hedge 1-2+ years ago when years of $$$ commitments were made. But doesn't seem to hold anymore, ergo it no longer makes sense. Microsoft had decades to build ecosystem and moat, just like adobe it was, actually hard and high friction to move away. People are switching models all the time. SOTA AI only labs has much less moat than Microsoft or other incumbent tech with decades of overlapping capture mechanisms.
> I think as a productive exercise someone should run this argument and explain why Chinese companies won't also be eaten alive
The productive answer is THEY ARE ALSO BEING eaten alive.
The even more productive answer is PRC commodity AI doesn't have to win for US SOTA AI to lose.
There's just not much meat on PRC AI bones so they are losing billions instead 100s of billions / trillions. Like there is ongoing domestic AI involution irrespective of US SOTA. Literally the primary reason PRC AI was forced to go free was because DeepSeek is billionaire side project who priced his hobby as free which set economics for all other players. Being sanctioned/compute constrained = they figured out how to close SOTA gap on the cheap but also set ceiling on how much they can lose on race to bottom. In so far their SOTA aspirations, again refer to DeepSeek, he takes some couch change from hedge fund gains or investors, and throws it at smart people and hope it's enough. If its not then they'll probably stop too. What PRC not going to do is mobilize resources equivalent to several % of GDP concentrated in few entities for SOTA AI, at least not until their compute is sorted, in which case competitive exposure / risks becomes symmetrical. In the mean time, things are not symmetrical and arguments against economics of US SOTA AI does not apply to PRC commodity AI.
I think the fundamental disagreement is you believe that cheaper models can just break the state of the art. This seems false because if it were true then the SOTA can also stop developing and spending on SOTA and just use a cheaper model to build their own better models and leverage their existing compute and energy infrastructure (or even build it out for better and more capabilities later). Anything that is released the SOTA has access to, but it also has access to things that others don't. It's ahead by definition and it's not possible to come out ahead of it because the SOTA will just absorb whatever open-weight/open-source/whatever is released and add to its own lead.
> SOTA AI only labs has much less moat than Microsoft or other incumbent tech with decades of overlapping capture mechanisms.
And those companies are worth a lot of money, continue to execute at high levels, are systematically important, and are by any definition wildly successful. That's my point. You have to apply the open source style argument fairly and equally. If you're going to suggest the cheaper models (which follow this open source software argument) are going to come into being and do all this stuff you also have to acknowledge that they made no money in comparison to the Microsofts of the world and Microsoft won because it made all the money. Whatever open-weight stuff that gets released everybody gets to use it and so tech companies get to use it too and accelerate their advantages. This is what we've seen play out, why would AI be any different when you're using the same economic paradigm as your argument?
> Literally the primary reason PRC AI was forced to go free was because DeepSeek is billionaire side project who priced his hobby as free which set economics for all other players.
American billionaires can also "just set up a side project" and reset whatever economics exist. The crux of your argument is someone can do this and then SOTA isn't economically viable, but you can't escape that this has to be true for everyone, not just the US. If it's true for everyone you also then have to admit the SOTA can't progress because there's no viable economic model - anytime you make an advance in the SOTA someone "just sets up a side project" and/or steals your model or w/e.
The US SOTA losing just means the SOTA has to be equal for everyone or the US is ahead - you can't have it both ways. Progress is practically impossible because you can't make money and you can't obtain a lead so nobody will do any work since their work will just be immediately utilized by someone else.
If what you are proposing here is true, then China is permanently behind. There isn't an economic model to obtain a "lead", yet the American companies get to just take all the work that Chinese companies do (until they go bankrupt not making money) and add it to their lead.
> The productive answer is THEY ARE ALSO BEING eaten alive.
Well you can just run them out of business then. If they spend any money you just spend more until they can't make money and close up shop. Or is someone going to subsidize them indefinitely?
The economics of these arguments that the US SOTA companies are just going to go bankrupt because of cheaper models is something I just don't find compelling. I've been trying. I debate these things so I can learn, but I'm just not seeing a coherent argument on the topic.
I don't think all of these tech companies and investors are stupid. Greedy to a large extent and some will lose their hats? Absolutely. Stupid? I don't think so. Worst case is all the infra is still useful.
No their argument is that the existence of a worse quality cheaper product will eat the majority of the market for the SOTA company, making the SOTA company less profitable (not necessarily unprofitable) because the number of people who need/want the quality difference supplied by the SOTA product aren't enough to pay to sustain the extra resources required to supply the next iteration of the better product or get a respectable ROI on the current investment.
>SOTA can also stop developing and spending on SOTA and just use a cheaper model
Yes, and they will be forced to (if the market shapes out as above) but the point is that then the company won't be getting the returns that justify the valuation they are talking about in the IPO and will likely never return the amount of money that has already been invested in it even if it is a profitable company. i.e. If I spend $20,000,000 making a lemonade stand that is a successful best in class stand and it nets me $50 profit a day then it doesn't matter how that as far as lemonade stands go that it's a pretty amazing example. I am effectively never going to see that 20 million again let alone a real profit.
They're not arguing that these SOTA companies aren't better, they're arguing that there won't be a return on investment let alone justifing further investment and means everyone will be taking a huge haircut on what is already invested (with the possible exception of early investors who take the bag with the IPO when it inevitably succeeds based on people disagreeing with this evaluation and thinking it a good investment or simply because of index funds being required to buy in)
The only examples we see, open source software, for example, shows that the major market players still make a boatload of money. Anyone off the street can download OpenOffice and it's the "worse quality cheaper product", but Microsoft makes hundreds of millions of dollars (estimate) on Office and that's what everyone uses.
What compelling evidence do you have that things will be different now?
Cheaper models break financing SOTA pace, PRC budget fast follow currently is catching up. If SOTA slows/stalls slower SOTA pace = instead of PRC 6 months behind, they're 3 months behind, or period of overlapping parity before US SOTA grabs lead again for a few months. There may not be economic model to maintain durable lead, which SOTA economic doom spiral condition. Now if PRC fast follow couldn't close gap, i.e if PRC only fast following due to distillation and US SOTA can technically prevent PRC fast follow, and extend gap then maybe exorbitant SOTA spend sustainable. If not sustainable everyone moves to cheaper cycles. This an issue if SOTA cannot be done on the cheap, which also true for PRC, but PRC not doing expensive SOTA, they're doing cheap fast follow.
>why would AI be any different
It's not, someone going to grab big share of AI... but that someone has to be competitive enough for exquisite lock in, and frankly with enterprise reducing SOTA AI spend due to SOTA AI costs, current SOTA AI ain't it. See last paragraph for hopium.
> someone going to subsidize them indefinitely
Yes. Same way Alphabet/Meta can subsidize AI, China has their Baba and tech giants with alternate revenue streams to subsidize AI, even if not SOTA, and geopolitics will make sure US cannot run PRC AI out of business in PRC and vice versa. But geopolitics not stopping PRC companies running other PRC companies out of business, same with US companies running US companies out of business. In this case host countries tech giants with high margin businesses squeezing sole play AI labs. It's just US SOTA ANT/OAI getting murdered is massive economic contagion risk for US vs Zai/kimi not for PRC.
> find compelling
This really not complicated, last couple years, US SOTA labs signed 100s of billions of compute based on projected demand that they could monopolize enough demand to finance compute capitalization. This not stupid investor prior, because their only competitor, PRC was export controlled, compute constrained and shouldn't close gap. But PRC closed gap, made models good enough and architecture efficient / cheap enough so other players, including those in US, can capture large % of demand WHILE using fraction of compute. SOTA labs now looking at margin compression, losing share to good enough competitors and obligatory (now excessive) capitalization. This bad combo for current US SOTA economics. The numbers suggest this is what is happening. There is chance demand increases past projections and extra SOTA compute will be boon, but again, we are seeing most profitable SWE enterprise curtail not increase demand.
> Worst case is all the infra is still useful
IMO this actually best case for US medium term favour, but also another strike against CURRENT US SOTA. Reality is current SOTA is grossly overpaying for compute / AI premium. Even if OAI or ANT dies, their compute can be picked up for pennies on dollar and as long as PRC behind on compute there is potential for cheaper future US SOTA economics from cheaper hardware since crash will lead hardware price regress towards mean. But that boon is for the next guy.
Here's the US hopium: post current US SOTA crash, there is potentially curve where US has so much cheap discount compute, because current SOTA taking the haircut, and where PRC has not much compute, that US will have sustainable SOTA prospects, i.e. the world where US pays $1 for shovels instead of $10 and can buy $10000 worth of shovels while PRC can only buy $100. That's feasible after current exorbitant SOTA economics bleeds it's patients to death and the next US SOTA host get the transplants and lighter balance sheet, i.e. web 2.0 for SOTA AI. Incidentally more likely "winner" of AI... if I had to guess assuming incumbent FAANG doesn't just snatch up ANT/OAI if they explode - new player/entrant who are near SOTA and eventually SOTA but is priced more competitive than AI 1.0 SOTA, that enterprise will gladly pay for since not at current exorbitant rates that's causing enterprise to bail, but it will be fine because AI 2.0 SOTA can be sustained from 3T TAM instead of 30T TAM.
No they don't.
> PRC budget fast follow currently is catching up.
If it is then your argument falls apart because by catching up on funding in any way disproves your theory about pricing and investment. China and Chinese companies can't invest because to do so is self-defeating, someone else already has a cheaper model and therefore any additional marginal investment will result in a large financial loss. You can't have it both ways. The same economic rules apply for both countries even if the CCP wants to pretend that they don't.
> and geopolitics will make sure US cannot run PRC AI out of business in PRC and vice versa
At least for my part I've never espoused that this will happen. Both countries will pursue and invest in state of the art AI and the crappy models mostly don't matter.
> This really not complicated
It actually is pretty complicated - and the points you are putting forth don't match your conclusions as I highlighted in my previous post.
> IMO this actually best case for US medium term favour,
Well it's not, it's actually just the base worst case. The best case will be something like an insurmountable lead in AI. If investors lose money, who cares, they still bought a lot of physical products and built infrastructure. For the US it's totally fine - boom/bust cycles are normal and fine.
> if I had to guess assuming incumbent FAANG doesn't just snatch up ANT/OAI if they explode
Provide some predictions then. You're just throwing things at the wall. OpenAI is going to "explode" and then Google is going to buy them or something? Based on what? Show the pricing and the economics.
Yes they do. See how easy being flippant is.
> disproves
No it just proves you're business innumerate. The same economic rules at work is PRC fast follows on the cheap which gives them net pricing advantage, which in context of everyone bleeding money just means they lose less money. Losing money is structurally fine, i.e. forever subsidies for some companies, like tech giants with alternative revenue, it's lethal for pure play AI, which US SOTA are. Economics of bleeding the same, US SOTA just bleeding much more without a tourniquet.
> crappy models mostly don't matter
Well yeah, but we're not talking about crappy models. We talk about good enough models. Good enough models that siphon SOTA demand and force SOTA discount which otherwise funded/set pace for SOTA development. Good enough competition entirely matters.
> don't match your conclusions as I highlighted
Yes because I'm arguing against your points, which are bad, hence they don't match, by design.
> insurmountable lead in AI. If investors lose money, who cares
Well that's another tier of hopium. Back to reality, if you want US AI to lead, the setup you want is for US SOTA AI to be sustainable, you want OAI, ANT dollars to stretch as far as it can. You want Nvidia, ASML, SKHynix i.e. all the component makers whose margins exploded to revert to mean so US labs can buy more compute per $ while shovel makers rub pennies. What you don't want is hardware makers soaking up generational $$$ from investors and tech company warchests in mother of all tier boom/bust cycles that would starve future booms.
> Provide some predictions then.
ANT/OAI balance sheet will be obviously non-sustainable in 1-2 years. They will slow down frontier spending to balance book. They will fail and either political capture to be too big to fail, or be bought out by other US tech majors. Based on ANT operating margin going from positive to negative (i.e. lose for every dollar spent) for SpaceX compute commitments which is just first tranch/preview of what will happen for other 800b on horizon. Based on the market they thought they were going to capture, they no longer fully capture and can't capture at price they though (i.e. 80% reduction in API pricing recently - because they got deepseeked). So their TAM is smaller than they thought, their margins are being compressed faster they they thought, but they still have to spend/pay shovel makers like it hasn't. How about you show some math, where's the $$$ going to come from. We know their largest consumers are enterprise who are cutting back. Where are they going to find 100s of billion year year. Name high value niches they'll capture... frontier math, replacing underpaid TAs?
There's a world where capabilities narrow and Anthropic is stuck selling a commodity (not like they don't have any competition now). In this world it's not unbelievable that Anthropic would trade around or below 1x revenue no?
I think calling Anthropic a tech company could be like calling an electric utility a tech company.
The $2T comes from the implied odds (perhaps 10 or 20%) that we get near term AGI or RSI and AI becomes the entire economy, crossed with the odds that Anthropic gets there first and is the runaway winner (maybe 25%).
The trouble, in my view, is that today's revenue numbers are essentially irrelevant to the outcome that drives ~all of the expected value. The market has been pretending otherwise, but fundamentally we're looking at a globally top-10 public company that boils down to a binary call option, and the probability underlying that option is inscrutable. That's never happened before, and an inability to divine a fair valuation properly demands extreme volatility, especially in reaction to capabilities progress.
Finally (total speculation): I think Anthropic knows this. I think they're sitting on something with Fable 5.5 and the surprise outperformance of Opus 5.5 is a tell. They intend to release it just before the IPO and make some big claims underwritten by its performance.
At this point every churches should IPO and raise hundreds of billions from the implied odds that judgment day is near and Paradise becomes the entire economy, that wouldn't be crazier.
Anthropic simply has a better claim, better PR
most people have been already inoculated against that kind of doomsday predictions
(and also because of this there's a huge problem with solving slow moving diasters, like climate change)
Where have you seen a trillion dollar IPO from a cult?
In addition to:
> the probability underlying that option is inscrutable
The "positive" side outcomes are also pretty inscrutable imo: even if AI completely replaced labour, how does it trade against natural resources? Even if anthropic gets there first, might open weights arrive a few months later? If AI drives the value of labour to cents per hour, does that crater global demand?
For sure, the "anthropic stock is a ticket to heaven" scenario is in there somewhere, but even taking complete AGI victory as a given, it seems like a very narrow outcome.
On the third hand: if AGI victory is assumed, even if Anthropic isn't a great bet even under that assumption, where else to put your money? While the music plays, you have to dance.
FWIW Bitcoin had the same dynamics. Its valuation between 2011-2021 basically derived from a superposition of two futures: one in which it took over the entire global economy and became the dominant reserve currency, and one in which it failed and its value went to zero. The middle-ground was effectively excluded because currencies have strong network effects and a currency that is not dominant within a market quickly ends up unacceptable to merchants and worthless.
It's still somewhat in that grey zone, used as a form of digital gold by people who believe that the dominant reserve currency (USD) is going to zero as well as a trading medium by rogue nations, but its performance has been lackluster since 2021. It's currently up about 20% since then.
Entirely possible that AI follows a similar trajectory. People assume that it's either going to go bust or take over the economy, but people are really good at overlooking excluded middle grounds. Personally I think it'll settle into a space where it's very effective as a coding assistant, effective at creating propaganda and fake forum posts, and has the effect of killing Google Search (but not really replacing it), but doesn't fulfill all the hype around completely replacing the economy.
But a big part of the reason this has been true is that (mostly-software-based) tech companies haven't historically been spending significant billions of dollars every year in capex.
I think an open question is if SOTA labs should be priced like software companies have been. Software has been so valuable because the marginal cost approaches zero as distribution increases. Profit/employee at a company like meta prior to massive AI spending was measured in the millions. Meanwhile, the balance sheets of the labs look entirely different - remaining competitive requires continuous and astronomic capital costs, and there's no indication that this goes away. They are perennially being chased by competitors offering the same performance as models mere months old at fractions of the cost.
Their business to me, at least long term, looks a lot more like an airline, railroad, or other extremely capital-heavy industry with very competitive and low margins. None of those trade at 6+ p/s. Particularly on the buyer side, the switching cost to me as a high-AI-spending employee if a cheaper model offers compelling performance is less than deciding that I should ditch my preferred business travel airline.
I think that's why you're seeing so many calls from the leaders of those companies to regulate their competition away.
The big foundational labs are not like this. Every model they release takes a ton of data infrastructure and training (pre, train, post) compute. Then inference itself requires compute. Even if the labs sell tokens, each token embodies a lot of training and inference compute. This is a much more capital heavy business, so I can see revenue multiples being smaller than pure SaaS businesses. A model release bust (like Opus 5) ends up leading to big losses because of the necessary training compute spent on the model and the labor spent on training. A bust SaaS product ends up being wound down quickly, and only costs the labor spent on building it and the compute spent running it.
A good analogy is chip companies before all of the AI related compute pressure. These were largely low margin businesses because of the fixed capital cost associated with each new chip gen, architecture, process, etc
So I think the reason for the value gap is purely in the proprietary IP. The recipe for making a good LLM is more or less public, unless you buy into the "everyone just distils claude/gpt" angle. The recipe for making a good search engine is private.
Note the huge difference in strategy and outcome!
Google published essentially nothing on how to build a search engine. Huge R&D spend, zero publications, intense culture of secrecy, extremely low attrition to startups or competitors. They did publish some papers on big data systems but only on tech that was many years old - a typical Google "research" paper pre-2015 would announce that the system it described had been running at scale in production for years already. Google also never faced any kind of split or mutiny in which lots of people with the core secrets left the firm to set up a competitor.
With AI however this changed and they published all their research immediately, also their best employees constantly ran away to join startups. The result was predictable: smaller competitors deployed their ideas faster and better than they did, and now Google has no moat anymore.
OpenAI took the classic Google approach and stopped publishing R&D after developing GPT but this didn't help them build margins because:
1. Apparently the core transformer idea is powerful enough that breakthroughs after that didn't matter much. Nobody has been able to develop some killer edge that leaves competitors unable to catch up. Most improvements to transformers have been about cost reduction, not capability improvement.
2. They had a huge split when the Anthropic guys walked out over safety, taking the knowledge of how to build a frontier LLM with them. Result was a peer competitor from day one, whereas Google had at least six years to build edge before anyone realized web search mattered.
Result: Google-style hardware costs but without any secret IP that could generate revenue commensurate with those costs.
Because we're just waiting for the world to find out that they can spend 1/10th and get virtually the same results.
Open source models are now 1/100th in price per task of US sota from this summer...
With the price of AI rapidly falling to the price of compute, with models therefore basically having zero value, and with Anthropic having rented all its compute from others then there’s little evidence that such high profits are in the future.