Ultimately it's marginal, meaning that every $ of revenue that Google and Meta make have an associated compute cost with it, but scale your revenue down and you scale compute down also. Moreover this compute cost per revenue is fairly low. A small cloud instance equivalent's of compute can generate a ton of business value.

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

The value in Facebook is in the social network that can't be replicated by just buying servers. The value in Google is/was in the secret IP. Hardware costs are low compared to the value of those. Hardly anyone knows how to make a search engine that works as well as Google does, or how to make a social network that draws everyone away from Insta/FB (ok bytedance do...).

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.

The scales are so dramatically different that it's akin to pointing out a home office and a steel plant both consume electricity.

I dont think electricity cost is what is keeping John Doe the developer from running local models.

> 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, ...

> 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.

High profit software products have monopoly pricing power from network affects: everyone uses office for compatibility/familiarity with what everyone else uses, everyone uses Facebook because everyone else does, everyone buys advertising buys from Google because that's who everyone sells to because that's who everyone buys from.

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.

The important distinction is SOTA US labs - ANT and OAI - who can not economically survive reverse engineering / fast following PRC models at PRC pricing. Like obviously PRC operators are not going to flourish in US due to geopolitics alone, but that doesn't stop US operators running PRC models, or reverse engineered Murican open models from creaming US SOTA labs on economics eventually. The closer open models are to SOTA in capability the smaller SOTA TAM gets, and that market might not be large enough to fund/maintain SOTA. SWE is one of the highest margin sectors to capture, and open models already shrunk SOTA niche to critical infra and cybersecurity etc then what other high $$$ niche market going to replace? BioPharma maybe? The point is SOTA capitalization depends on large capability gaps that economically justifies 30T TAM - the more domains open models fast follows to compress TAM, less places SOTA has to carve sufficiently profitable moat, and less sustainable maintaining SOTA becomes, especially for dedicated AI labs with no alternative revenue. It's not cheapest open-weight wins, it's cheap and good enough open weight simply murders current SOTA economics because SOTA likely cannot finance at all costs anymore.

> Most people don’t want to or have any interest in running or managing models locally

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.

Worldwide, it is will be the cheapest model. That is going to win out. In the United States. and Europe, that will be their (OpenAI/Anthropic’s) markets, every place else probably will be a Chinese Open solution. And that will be no different than EV cars. or solar panels the Chinese dominate worldwide and the same probably will happen with AI models outside the United States, Europe, Australia, New Zealand and one or two other countries.

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.

You are misunderstanding a number of things aggressively.

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.

"Biggest" and "most important" are not given, settled facts

Isn’t the whole reason companies go public now to dump on retail? What growth is left?

SpaceX is down, Cerebras is down, etc.

Anthropic has something close to first-mover advantage but that doesn't mean they'll continue to grow revenue. Ten years from now the dominant player could be Google or DeepSeek or some random startup that doesn't even exist yet.

I didn't know that the "left behind" rhetoric also extends to investing into their stock?

I guess it applies to everything. Universal. Like LLMs.

> This is what makes a market.

What is "this" referring to here?

> 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

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.

> 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

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.

Good points.

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.

Another consideration: if we get AGI and it becomes the entire economy, what are the odds that the money you've invested still matters? It seems like both the "universal post-scarcity" or "AI company controls all wealth and power, the peasants live in squalor" scenarios are likely outcomes.

Their forward pricing prices in all expected growth... so if they don't have accelerating momentum that reduces expected price.

Mathematically they can work very similarly though because you have a revenue growth rate, and margin. The question is can that revenue/margin reach break even before you run out of runway (the expected sum of all rounds of investing) .

The main difference I would concede is that gravity, mass, forces, awind drag etc are much easier to model.

Analogies however are: they often fly over people's heads.

Do you respond to every metaphor this way?

Businesses also aren't financial statements. These are abstractions we use to understand and communicate about the world.

But they do have S curves. If you're on the shoulder of the S curve while your spending commitments are accelerating, that won't end well.

Someone needs to tell that to the AI startups.

>At this rate Anthropic will exceed AWS Revenue within a year

And in 10 years it will be worth the entire Earth's economy.

It is a lot easier to switch model providers than it is to switch cloud providers. And the open weight models from Chinese Labs make Claude a lot more of a commodity than any service provided by AWS.

> At this rate Anthropic will exceed AWS Revenue within a year

Well that’s the point. They’re not continuing at that rate of growth.

I did not suggest a downward trend. (It's a trend graph)

And yet one can argue that the "garbage businesses" you label are the ones doing their best at matching market demand. See how well things are going for other entrants, like Google, Microsoft or Meta. Putting Copilot in every app makes a whole lot less sense than Cowork, Codex and the like. They just are doing worse things while attached to traditional money fountains, like ads in social media and search.

Opus 4.6's coding ability is nowhere near as good as Opus 5.5's. Some model next year will almost certainly be significantly better at coding (and planning, and designing, and ideating) than Opus 5.5.

You're narrowly focused on writing code. What about when they can run entire firms themselves, better than the average human CEO could? What about when we have models way, way, way better than Opus 4.6 and 5.5?

The point is we will eventually very likely have AI systems that can do everything. Everything. Anything and everything a human can do with - or even without - a computer, and then better than anything or everything a human can do, including in terms of creativity and innovation and novelty and insight. Obviously I don't know when that might possibly be, and it's not happening next year or something, but you can't just pretend like we hit the cap on the technology.

Don’t bother - these knuclkleheads don’t understand nuance and despite much of their devices being manufactured in China - they struggle to accept Chinese offerings can take significant market share.

I am not trying to even make that argument. Just if we do assume it, I am probably correct.

As for the rest of your post: bizarre to see such ignorant slopulist drivel on HN of all places. Is this a YouTube comment section or an attempt at an intellectual message board?

I am actually somewhat bearish on adoption/propagation. I would be expecting much larger numbers than these if I were bullish on it.

Yes, but their balance sheet seems fine to me.

They're certainly correlated, even if one might quibble about the coefficient.

Yes.

How does it change if you’re building via subscription?

Arguably, it was cheaper.

By the way, we get way more value from our Anthropic subscriptions than any Chinese subscription.

>People PREDICTED token maxx was a fad/trend Q1 while it was underway,

No, a lot of people like Gary Marcus said it had specifically already ended in May/June. Why is this time different?

>Tokenmaxxing is over. https://finance.yahoo.com/sectors/technology/articles/tokenm...

>Tokenmaxxing is dead, long live valuemaxxing https://www.ibm.com/think/insights/tokenmaxxing-dead-long-li...

>Token Maxxing, April 2026-May 2026, RIP https://x.com/GaryMarcus/status/2057610132742156603?lang=en

What are his numbers and analysis?

So he can get all of his conclusions and predictions wrong based on his numbers and analysis, and then you think his value is in his numbers and analysis?

What is unsustainable about Anthropic's numbers?

That is true, but not even really connected at all to the comment I was replying to. Revenue is indeed an indicator of a successful business.

Additionally, the Anthropic S-1 does indeed show scalable profitable growth. This is perhaps the most shocking thing in the S-1, to me.

The dot-com era had a lot of companies with high valuations and zero revenue. That was the main problem.

Yes, revenue shows people's interest and willingness to pay for the product. That's exactly what it is.

ANT also has positive gross margins, so they aren't selling gold bars for $1.

I don't understand why people on HN treat ECON101 like some fringe theory every time this is discussed. This is basic stuff.

> Numerical high performance computing?

Any numerical high performance computing from 1995 will run on a 100$ android phone.

> You think we will eventually be able to run inference on a rasp pi?

Not today's raspberry pi, but the equivalent of it absolutely. I don't work in it, but have a number of friends who are involved in LLM development - the "breakthrough" optimizations are obvious work in hindsight. We are still in the very early innings of optimizing both hardware and software techniques. The recent market imbalance has blurred the idea that hardware simultaneously gets both cheaper and much better, but no one serious expects that to stop on a longer timeline.

Going beyond the "my watch has more computation power than NASA at the moon landing" - speech-to-text used to be unreliable tech that needed datacenters backing it as recently as the early iphone, and now runs in near-realtime on smartphones that cost about the same. A ~$50 processor running in my closet does realtime person/vehicle/package detection on my front door camera reliably.

Whatever you can buy for inflation-adjusted-45$ in ~2035, almost surely.

Possibly. 1.58-bit models already run on CPUs https://github.com/microsoft/BitNet They're not good exactly but they are surprisingly better than you might think.

https://huggingface.co/microsoft/bitnet-b1.58-2B-4T#evaluati...

This is them working within their existing market. This isn't the new market that the valuations for companies like Anthropic are operating in.

> Get rich following the bot's trade advice and never tell anyone about your bot.

The fun thing about open markets is the market will punish you for this with second order effects and it's in your best interest to get some whales on the bandwagon.

This is the existing ad market with the existing TAM. Not the AI market.