Anthropic raises $13B Series F
anthropic.com
anthropic.com
What gets me is that this isn't even a software moat anymore - it's literally just whoever can get their hands on enough GPUs and power infrastructure. TSMC and the power companies are the real kingmakers here. You can have all the talent in the world but if you can't get 100k H100s and a dedicated power plant, you're out.
Wonder how much of this $13B is just prepaying for compute vs actual opex. If it's mostly compute, we're watching something weird happen - like the privatization of Manhattan Project-scale infrastructure. Except instead of enriching uranium we're computing gradient descents lol
The wildest part is we might look back at this as cheap. GPT-4 training was what, $100M? GPT-5/Opus-4 class probably $1B+? At this rate GPT-7 will need its own sovereign wealth fund
Anecdotally moving from model to model I'm not seeing huge changes in many use cases. I can just pick an older model and often I can't tell the difference...
Video seems to be moving forward fast from what I can tell, but it sounds like the back end cost of compute there is skyrocketing with it raising other questions.
Called efficient compute frontier
They don't even burn it on on AI all the time either: https://openai.com/sam-and-jony/
This is an extraordinary moment.
Computers are now seeing, thinking and understanding.
Despite this unprecedented capability, our experience remains shaped by traditional products and interfaces."
I don't even want to learn about them every line is so exhausting
"We would like to introduce you to the spawn of Johnny Ive and Sam Altman, we're naming him Damien Thorn."
I don't think we're hitting peak of what LLMs can do, at all, yet. Raw performance for one-shot responses, maybe; but there's a ton of room to improve "frameworks of thought", which are what agents and other LLM based workflows are best conceptualized as.
The real question in my mind is whether we will continue to see really good open-source model releases for people to run on their own hardware, or if the companies will become increasingly proprietary as their revenue becomes more clearly tied up in selling inference as a service vs. raising massive amounts of money to pursue AGI.
Initially, first party proprietary solutions are in front.
Then, as the second-party ecosystem matures, they build on highest-performance proprietary solutions.
Then, as second parties monetize, they begin switching to OSS/commodity solutions to lower COGS. And with wider use, these begin to outcompete proprietary solutions on ergonomics and stability (even if not absolute performance).
While Anthropic and OpenAi are incinerating money, why not build on their platforms? As soon as they stop, scales tilt towards an apache/nginx type commoditized backend.
That's still a pretty good deal for an investor: if I give you $15B, you will probably make a lot more than $15B with it. But it does raise questions about when it will simply become infeasible to train the subsequent model generation due to the costs going up so much (even if, in all likelihood, that model would eventually turn a profit).
"probably" is the key word here, this feels like a ponzi scheme to me. What happens when the next model isn't a big enough jump over the last one to repay the investment?
It seems like this already happened with GPT-5. They've hit a wall, so how can they be confident enough to invest ever more money into this?
If model training has truly turned out to be profitable at the end of each cycle, then this company is going to make money hand over fist, and investing money to out compete the competition is the right thing to do.
Most mega corps started out wildly unprofitable due to investing into the core business... until they aren't. It's almost as if people forget the days of Facebook being seen as continually unprofitable. This is how basically all huge tech companies you know today started.
Having experienced Anthropic as a customer, I have a hard time thinking that their inevitable failure (something i'd bet on) will be model/capability-based, that's how bad they suck at every other customer-facing metric.
You think Amazon is frustrating to deal with? Get into a CSR-chat-loop with an uncaring LLM followed up on by an uncaring CSR.
My minimum response time with their customer service is 14 days -- 2 weeks -- while paying 200usd a month.
An LLM could be 'The Great Kreskin' and I would still try to avoid paying for that level of abuse.
https://arstechnica.com/tech-policy/2024/02/air-canada-must-...
“There Is No AI Revolution” - Feb ‘25:
Putting aside the hype and bluster, OpenAI — as with all generative AI model developers — loses money on every single prompt and output. Its products do not scale like traditional software, in that the more users it gets, the more expensive its services are to run because its models are so compute-intensive.
While OpenAI's numbers aren't public, this seems very unlikely. Given open-source models can be profitably run for cents per million input tokens at FP8 — and OpenAI is already training (and thus certainly running) in FP4 — even if the closed-source models are many times bigger than the largest open-source models, OpenAI is still making money hand over fist on inference. The GPT-5 API costs $1.25/million input tokens: that's a lot more than it takes in compute to run it. And unless you're using the API, it's incredibly unlikely you're burning through millions of tokens in a week... And yet, subscribers to the chat UI are paying $20/month (at minimum!), which is much higher than a few million tokens a week cost.
Ed Zitron repeats his claim many, many, excruciatingly many times throughout the article, and it seems quite central to the point he's trying to make. But he's wrong, and wrong enough that I think you should doubt that he knows much about what he's talking about.
(His entire blog seems to be a series of anti-tech screeds, so in general I'm pretty dubious he has deep insight into much of anything in the industry. But he quite obviously doesn't know about the economics of LLM inference.)
This was the power of Moore's Law, it gave the semiconductor engineers an argument they could use to convince the money-guys to let them raise the capital to build the next fab- see, it's right here in this chart, it says that if we don't do it our competitors will, because this chart shows that it is inevitable. Moore's Law had more of a financial impact than a technological one.
And now we're down to a point where only TSMC is for sure going through with the next fab (as a rough estimate of cost, think 40 billion dollars)- Samsung and Intel are both hemming and hawing and trying to get others to go in with them, because that is an awful lot of money to get the next frontier node. Is Apple (and Nvidia, AMZ, Google, etc.) willing to pay the costs (in delivery delays, higher costs, etc.) to continue to have a second potential supplier around or just bite the bullet and commit to TSMC being the only company that can build a frontier node?
And even if they can make it to the next node (1.4nm/14A), can they get to the one after that?
The implication for AI models is that they can end up like Intel (or AMD, selling off their fab) if they misstep badly enough on one or two nodes in a row. This was the real threat of Deepseek: if they could get frontier models for an order of magnitude cheaper, then the entire economics of this doesn't work. If they can't keep up, then the economics of it might, so long as people are willing to pay more for the value produced by the new models.
He says "You paid $100 million and then it made $200 million of revenue. There's some cost to inference with the model, but let's just assume in this cartoonish cartoon example that even if you add those two up, you're kind of in a good state. So, if every model was a company, the model is actually, in this example is actually profitable. What's going on is that at the same time"
notice those are hypothetical numbers and he just asks you to assume that inference is (sufficiently) profitable.
He doesn't actually say they made money by the EoL of some model.
If not, this seems like a recipe for bankruptcy. You are always investing more than you're making, right up until the day you don't make it back. Whether that's next year or in ten or twenty years. It's basically impossible to do it forever - there simply isn't enough profit to be had in the world if you go forward enough orders of magnitude. How will they know when to hop off the train?
Surely the Anthropic CEO will have no incentive to lie.
The companies doing foundational video models have stakeholders that don’t want to be associated with what people really want to generate
But they are pushing the space forward and the uncensored and unrestricted video model is coming
When it comes to sexually explicit content in general with adults, all of our laws rely on the human actor existing
FOSTA and SESTA is related to user generated content of humans, for example. They rely on making sure an actual human isnt being exploited and burdening everyone with that enforcement. When everyone can just say “thats AI” nobody’s going to care and platforms will be willing to take that risk of it being true again - or a new hit platform will. That kind of content currently Doesnt exist in large quantities yet, until a video model ungimped can generate it.
Concerns about trafficking only rely on actual humans not entirely new avatars
regarding children there are more restrictions that may already cover this, there is a large market for just adult looking characters though and worries about underage can be tackled independently. or be found entirely futile. not my problem, focus on what you can control. this is whats coming though.
people already dont mind parasocial relationships with generative AI and already pay for that, just add nudity
that said, I'm sure you can imagine that the really illegal, truly, positively sickening and immoral stuff is children-adjacent and you can be 100% sure there are sociopaths doing training runs for the broken people who'll buy the weights.
Additionally, the entire "payment processors leaning on Steam" thing shows that it might be very difficult to monetize a model that's known for generating extremely controversial content. Without monetization, it would be hard for any company to support the training (and potential release) of an unshackled enterprise-grade model.
That article is nothing to do with AI, really.
In the meanwhile, "better data", "better training methods" and "more training compute" are the main ways you can squeeze out more performance juice without increasing the scale. And there are obvious gains to be had there.
Does this apply to Google that is using custom built TPUs while everyone else uses stock Nvidia?
If Google wants anything better than that? They, too, have to wait for the new hardware to arrive. Chips have a lead time - they may be your own designs, but you can't just wish them into existence.
All of the big AI players have profited from Wikipedia, but have they given anything back, or are they just parasites on FOSS and free data?
Probably because you're doing things that are hitting mostly the "well-established" behaviors of these models — the ones that have been stable for at least a full model-generation now, that the AI bigcorps are currently happy keeping stable (since they achieved 100% on some previous benchmark for those behaviors, and changing them now would be a regression per those benchmarks.)
Meanwhile, the AI bigcorps are focusing on extending these models' capabilities at the edge/frontier, to get them to do things they can't currently do. (Mostly this is inside-baseball stuff to "make the model better as a tool for enhancing the model": ever-better domain-specific analysis capabilities, to "logic out" whether training data belongs in the training corpus for some fine-tune; and domain-specific synthesis capabilities, to procedurally generate unbounded amounts of useful fine-tuning corpus for specific tasks, ala AlphaZero playing unbounded amounts of Go games against itself to learn on.)
This means that the models are getting constantly bigger. And this is unsustainable. So, obviously, the goal here is to go through this as a transitionary bootstrap phase, to reach some goal that allows the size of the models to be reduced.
IMHO these models will mostly stay stable-looking for their established consumer-facing use-cases, while slowly expanding TAM "in the background" into new domain-specific use-cases (e.g. constructing novel math proofs in iterative cooperation with a prover) — until eventually, the sum of those added domain-specific capabilities will turn out to have all along doubled as a toolkit these companies were slowly building to "use models to analyze models" — allowing the AI bigcorps to apply models to the task of optimizing models down to something that run with positive-margin OpEx on whatever hardware that would be available at that time 5+ years down the line.
And then we'll see them turn to genuinely improving the model behavior for consumer use-cases again; because only at that point will they genuinely be making money by scaling consumer usage — rather than treating consumer usage purely as a marketing loss-leader paid for by the professional usage + ongoing capital investment that that consumer usage inspires.
Instead, I mean that these later-generation models will be able to be fine-tuned to do things like e.g. recognizing and discretizing "feature circuits" out of the larger model NN into algorithms, such that humans can then simplify these algorithms (representing the fuzzy / incomplete understanding a model learned of a regular digital-logic algorithm) into regular code; expose this code as primitives/intrinsics the inference kernel has access to (e.g. by having output vectors where every odd position represents a primitive operation to be applied before the next attention pass, and every even position represents a parameter for the preceding operation to take); cut out the original circuits recognized by the discretization model, substituting simple layer passthrough with calls to these operations; continue training from there, to collect new, higher-level circuits that use these operations; extract + burn in + reference those; and so on; and then, after some amount of this, go back and re-train the model from the beginning with all these gained operations already being available from the start, "for effect."
Note that human ingenuity is still required at several places in this loop; you can't make a model do this kind of recursive accelerator derivation to itself without any cross-checking, and still expect to get a good result out the other end. (You could, if you could take the accumulated intuition and experience of an ISA designer that guides them to pick the set of CISC instructions to actually increase FLOPS-per-watt rather than just "pushing food around on the plate" — but long explanations or arguments about ISA design, aren't the type of thing that makes it onto the public Internet; and even if they did, there just aren't enough ISAs that have ever been designed for a brute-force learner like an LLM to actually learn any lessons from such discussions. You'd need a type of agent that can make good inferences from far less training data — which is, for now, a human.)
Last week I put GPT-5 and Gemini 2.5 in a conversation with each other about a topic of GPT-5's choosing. What did it pick?
Improving LLMs.
The conversation was far over my head, but the two seemed to be readily able to get deep into the weeds on it.
I took it as a pretty strong signal that they have an extensive training set of transformer/LLM tech.
Model specialization. For example a model with legal knowledge based on [private] sources not used until now.
Or, as in the case of a leading North American LLM provider, I would love to be able to choose an older model but it chooses it for me instead.
Doesn’t explain Deepseek.
The problem is that in the meantime, they're going to nuke our existing powergrid, created in the 1920's to 1950's to serve our population as it was in the 1970's, and for the most part not expanded since. All of the delta is in price-mediated "demand reduction" of existing users.
Edit: for the curious, no. An H100 costs about ~25k and produces $1.2/day mining bitcoin. Without factoring in electricity.
Labs can just step up the way they track signs of prompts meant for model distillation. Distillation requires a fairly large number of prompt/response tuples, and I am quite certain that all of the main labs have the capability to detect and impede that type of use if they put their backs into it.
Distillation doesn't make the compute moat irrelevant. You can get good results from distillation, but (intuitively, maybe I'm wrong here because I haven't done evals on this myself) you can't beat the upstream model in performance. That means that most (albeit obviously not all) customers will simply gravitate toward the better performing model if the cost/token ratio is aligned for them.
Are there always going to be smaller labs? Sure, yes. Is the compute mote real, and does it matter? Absolutely.
....while degrading their service for paying customers.
This is the same problem as law-enforcement-agency forwarding threats and training LLMs to avoid user-harm -- it's great if it works as intended, but more often than not it throws a lot more prompt cancellations at actual users by mistake, refuses queries erroneously -- and just ruins user experience.
i'm not convinced any of the groups can avoid distillation without ruining customer experience.
And if LLMs don't keep getting qualitatively more capable every few months, that means that all this investment won't pay off and people will soon just use some open weights for everything.
And honestly I don't think a lot of these companies would turn a profit on pure utility -- the electric and water company doesn't advertise like these groups do; I think that probably means something.
> the electric and water company doesn't advertise like these groups do
I'm trying to understand what you mean here. In the US these utilities usually operate in a monopoly so there's no point in advertising. Cell service has plenty of advertising though.
Not the GP (in fact I just replied to GP, disagreeing with them), but I think that economies of scale kick in when you are provisioning M GPUs for N users and both M and N are large.
When you are provisioning for N=1 (a single user), then M=1 is the minimum you need, which makes it very expensive per user. When N=5 and M is still 1, then the cost per user is roughly a fifth of the original single-user cost.
This doesn't seem correct. I run legacy models with only slightly reduced performance on 32GB RAM with a 12GB VRAM GPU right now. BTW, that's not an expensive setup.
> You think anyone is going to make money on that vs $20 a month to anthropic?
Why does it have to be run as a profit-making machine for other users? It can run as a useful service for the entire household, when running at home. After all, we're not talking about specialised coding agents using this[1], just normal user requests.
====================================
[1] For an outlay of $1k for a new GPU I can run a reduced-performance coding LLM. Once again, when it's only myself using it, the economics work out. I don't need the agent to be fully autonomous because I'm not vibe coding - I can take the reduced-performance output, fix it and use it.
It's closer to $3k to build a machine that you can reasonable use which is 12 whole years of subscription. It's not hard to see why no one is doing it.
With my existing setup for non-coding tasks (GPU is a 3060 12GB which I bought prior to wanting local LLM inference, but use it now for that purpose anyway) the GPU alone was a once-off ~$350 cost (https://www.newegg.com/gigabyte-windforce-oc-gv-n3060wf2oc-1...).
It gives me literally unlimited requests, not pseudo-unlimited as I get from ChatGPT, Claude and Gemini.
> and with the sub you get the new models where your existing hardware will likely not be able to run it at all.
I'm not sure about that. Why wouldn't the new LLM models run on a 4yo GPU? Wasn't a primary selling point of the newer models being "They use less computation for inference"?
Now, of course there are limitations, but for non-coding usage (of which there is a lot) this cheap setup appears to be fine.
> It's closer to $3k to build a machine that you can reasonable use which is 12 whole years of subscription. It's not hard to see why no one is doing it.
But there are people doing it. Lots, actually, and not just for research purposes. With the costs apparently still falling, with each passing month it gets more viable to self-host, not less.
The calculus looks even better when you have a small group (say 3 - 5 developers) needing inference for an agent; then you can get a 5060ti with 16GB RAM for slightly over $1000. The limited RAM means it won't perform as well, but at that performance the agent will still capable of writing 90% of boilerplate, making edits, etc.
These companies (Anthropic, OpenAI, etc) are at the bottom of the value chain, because they are selling tokens, not solutions. When you can generate your own tokens continuously 24x7, does it matter if you generate at half the speed?
Yes, massively it's not even linear 1/2 speed is probably 1/8 or less the value of "full speed". It's going to be even more pronounced as "full speed" gets faster.
I don't think that's true for most use-cases (content generation, including artwork, code/software, reading material, summarising, etc). Something that takes a day without an LLM might take only 30m with GPT5 (artwork), or maybe one hour with Claude Code.
Does the user really care that their full-day artwork task is now one hour and not 30m? Or that their full-day coding task is now only two hours, and not one hour?
After all, from day one of the ChatGPT release, literally no one complained that it was too slow (and it was much slower than it is now).
Right now no one is asking for faster token generation, everyone is asking for more accurate solutions, even at the expense of speed.
Well who does the inference at the scale we're talking about here? That's (a key part of) the moat.
Each new commercial model needs to not just be better than the previous version, it needs to be significantly better than the SOTA open models for the bread-and-butter generation that I'm willing to pay the developer a premium to use their resources for generation.
And for the newcomers, the scale needs to be bigger than what the incumbents (Google and Microsoft) have as discretionary spending - which is at least a few billion per year. Because at that rate, those companies can sustain it forever and would be default winners. So I think yearly expenditure is going to be 20B year++
Taxi apps are a commodity today.
The only people that really benefited from Uber are:
- Uber executives
- early investors that saw the share price go up
- early customers that got VC subsidized rides
1. opportunity costs are a thing.
2. if you add Uber's financial numbers since creation, the crazy amount of VC that was invested Uber would have provided better returns by investing it in the S&P 500.
3. Uber will settle in as a boring, profitable company that's going to be a side note in both the history of tech and also of transportation and will primarily be remembered for eroding worker rights.
Uber IPO May 2019: market cap $82bn. Uber now: $193bn. 2.35x multiplier.
S&P 500 May 2019: $2750. S&P 500 now: $6460. 2.35x multiplier.
So the much, much riskier Uber investment has barely matched a passive S&P 500 investment over the same time frame. And the business itself has lost money, more money was put into it than has been gotten back so far.
I'm not even sure why I'm in this conversation as it seems ideological. I bring up facts and you bring up... vibes?
I was replying to this: "So far Lyft seems to be doing okay, which proves the business plan doesn't really work." when I said Uber is profitable
Your retort to that was S&P grew more than Uber, which is a nonsensical argument. Our standard for what is a good business is if it grows faster than S&P after going public?
Edit: I dug up some research related to this, most companies do worse than S&P after becoming public. What's your point then?
The same is currently true of Anthropic.
Or the last investor. When this type of money is raised, you can be sure the earlier investors are looking for ways to have a soft landing.
Like sure it saves me a bit of time here and there but will scaling up really solve the reliability issues that is the real bottleneck.
For what it is worth, $13 billion is about the GDP of Somalia (about 150th in nomimal GDP) with a population of 15 million people.
The GDP of the Netherlands is about $1.2 trillion with a population of 18 million people.
I understand that that’s not quite what’s meant with ‘small country’ but in both population and size it doesn’t necessarily seem accurate.
California (where Anthropic is headquartered) has over twice as many people as all of Somalia.
The state of California has a GDP of $4.1 Trillion. $13 billion is a rounding error at that scale.
Even the San Francisco Bay Area alone has around half as many people as Somalia.
That’s the known minimum cost. We have a lot of room to get costs down if we can figure out how.
I'm reminded of how insanely complex the human brain is: ~100 trillion connections. The Nvidia H100 has just 0.08 trillion transistors.
When you consider where most of that money ends up (Jensen &co), it's bizarre nobody can really challenge their monopoly - still.
You think any of these clusters large enough to be interesting, aren't authorized under a contractual obligation to run any/all submitted state military/intelligence workloads alongside their commercial workloads? And perhaps even to prioritize those state-submitted workloads, when tagged with flash priority, to the point of evicting their own workloads?
(This is, after all, the main reason that the US "Framework for Artificial Intelligence Diffusion" was created: America believed China would steal time on any private Chinese GPU cluster for Chinese military/intelligence purposes. Why would they believe that? Probably because it's what the US thought any reasonable actor would do, because it's what they were doing.)
These clusters might make private profits for private shareholders... but so do defense subcontractors.
Nice timing? I am sure they have scored a deal with the selling of personal data
I really have to wonder, how long will it be before the competition moves into who has the most wafer-scale engines. I mean, surely the GPU is a more inefficient packaging form factor than large dies with on-board HBM, with a massive single block cooler?
But I do believe that their cost per compute is still far more than disparate chips.
how so? deepseek and others do models on par with previous generation for a tiny fraction of a cost. Where is the moat?
That's just pure insanity to me.
It's not even Internet speed or hardware. It's literally not having enough electricity. What is going on with the world...
So we can at least assume that whoever is deciding to move the capacity does so at some business risk elsewhere.
Understanding scale is a weird thing, that I guess is psychological. I think the different experiences and travels I have made have an impact on this. Despite living for over a decade in Texas I live in one of the most densely populated places. But I recently got to visit Colorado, which is far less populated and has lots of weird places.
You can drive up to the base of the Great Sand Dunes and walk up the first few hills quite easily if you're in good shape. Here's some photos
https://www.hydrogen18.com/p/2024-great-sand-dunes-national-...
If you pull out your smartphone and look at Google Maps, it becomes pretty obvious how insane the scale of the place is. There's no real prohibition on where you can walk there because it isn't necessary. It's so large and the environment is so harsh, you aren't going to cross much of it on foot. Ever.
> GPT-4 training was what, $100M? GPT-5/Opus-4 class probably $1B+?
Your brain? Basically free *(not counting time + food)
Disruption in this space will come from whomever can replicate analog neurons in a better way.
Maybe one day you'll be able to Matrix information directly into your brain and know kung-fu in an instant. Maybe we'll even have a Mentat social class.
Fifty years ago, we were starting to see the very beginning of workstations (not quite the personal computer of modern days), something like this: https://en.wikipedia.org/wiki/Xerox_Alto, which cost ~$100k in inflation-adjusted money.
I mean, it's like the djin giving you three whishes, and not a single character will ask "what's the two best wishes I can do to (ensure mankind will reach perpetually best peaceful harmonious flourishing social dynamics forever| whatever goal the character might have as greatest hope)". When you have a instant perfect knowledge acquisition machine at disposal, the first thing to obviously understand is what the most important things to do to reach your goal.
The film didn't mention everything Neo learned like that though, just that he accumulate straight forward for many hours. Wouldn't be an action movie, certainly you would hope the character first words after such an impressive feat wouldn't be "I know kung fu".
Is there room for a smaller team to beat Anthropic/OpenAI/etc. at a single subject matter?
It still amazes me that Uber, a taxi company, is worth however many billions.
I guess for the bet to work out, it kinda needs to end in AGI for the costs to be worth it. LLMs are amazing but I'm not sure they justify the astronomical training capex, other than as a stepping stone.
> LLMs are amazing but I'm not sure they justify the astronomical training capex, other than as a stepping stone.
They can just... stop training today and quickly recuperate the costs because inference is mostly profitable.
Indeed, LLM companies likely turn operating profits, but I'm not sure that alone justifies their valuations. It's one thing to make money, it's another to make a return for investors.
And sure, valuations are growing faster than you can blink. Time will show if this in turn is justifiable or a bubble.
Since before AWS was even a thing, Amazon was already turning up great revenue and could’ve easily just stopped expanding and investing into the company growth, and they would be profitable easily. Instead, Amazon decided to reinvest all their potential profits into growth/expansion (with the favorable tax treatment on top) at the expense of keeping the cash profits. At any given point, Amazon could’ve stopped reinvesting all potential profits into their growth, and they would be instantly profitable.
This is not the same as Uber, which ran their core service operations at a net loss (and was only cheap due to their investors eating the difference and hoping that Uber will eventually figure out how to not lose money on operating their core service).
Today's LLM providers have the same choice as Amazon had back then. They could just stop advancing/expanding, and instead try harnessing that profit. They choose not to, they instead bet on fast expansion/advancement, and it makes total sense to me (especially as the competition rn between all those LLM providers seems to be very heated, as opposed to what Amazon dealt with at the time).
Uber is different, though. There was no point at which they could just call it a day and go "alright, we are fully self-sufficient, and we can just start collecting profits from our services as soon as tomorrow." Expansion or not, unless I am missing something, Uber wasn't in a position where they were profitable (and just chose to forego profits in favor of spending the money on expanding). To be clear, they indeed spent money on expanding, I am just questioning the part about their expansion being done using the funds that could've otherwise made the difference between them being profitable and not.
How do you know models are expensive to run? They have gone down in price repeatedly in the last 2 years. Why do you assume it has to run in the cloud when open source models can perform well?
> The hype is insane, and so usage is being pushed by C-suite folks who have no idea whether it's actually benefiting someone "on the ground" and decisions around which AI to use are often being made on the basis of existing vendor relationships
There are hundreds of millions of chatgpt users weekly. They didn't need a C suite to push the usage.
Because cloud monetization was awful. It's either endless subscription pricing or ads (or both). Cloud is a terrible counter-example because it started many awful trends that strip consumer rights. For example "forever" plans that get yoinked when the vendor decides they don't like their old business model and want to charge more.
i’d rather have a subscription than no service at all
oh, and one can always just not buy something if it’s not valuable enough
Citation needed
I think those actually using "AI" have a lot better idea of which are which than the C-suite folk.
I'd wager the personal failure rate when using LLMs is probably even higher than the 95% in enterprise, but will wait to see the numbers.
Definitely not. That came years later but in the late 2000s to mid-2010s it was often engineers pushing for cloud services over the executives’ preferred in-house services because it turned a bunch of helpdesk tickets and weeks to months of delays into an AWS API call. Pretty soon CTOs were backing it because those teams shipped faster.
The consultants picked it up, yes, but they push a lot of things and usually it’s only the ones which actual users want which succeed.
But there was also cool stuff happening at smaller places like Joyent, Heroku, Slicehost, Linode, Backblaze, iron.io, etc.
A lot of my awareness started in the academic HPC world which was a bit ahead in needing high capacity of generic resources but it felt like this came from the edges rather than the major IT giants. Companies like IBM, Microsoft, or HP weren’t doing it, and some companies like Oracle or Cisco appeared to thought that infrastructure complexity was part of their lock on enterprise IT departments since places with complex hand run books weren’t quick to switch vendors.
Amazon at the time wasn’t seen as a big tech company - they were where you bought CDs – and companies like Joyent or Rackspace had a lot of mindshare as well before AWS started offering virtual compute in 2006. One big factor in all of this was that x86 virtualization wasn’t cheap until the mid-to-late 2000s so a lot of people weren’t willing to pay high virtualization costs, but without that you’re talking services like Bingodisk or S3 rather than companies migrating compute loads.
The argument is something like that is not really possible anymore given the absurd upfront investments we're seeing existing AI companies need in order to further their offerings.
But yes, there was a window of opportunity when it was possible to do cutting-edge work without billions of investment. That window of opportunity is now past, at least for LLMs. Many new technologies follow a similar pattern.
– Hank Rutherford Hill
Today the only way to scale compute is to throw more power at it or settle for the 5% per year real single core performance improvement.
- Running up single-core performance through increasingly sophisticated core design and clock speed (which is now at the 5% per year point mentioned)
- Going wider by throwing more SMT, more cores, and larger caches at the problem.
Assuming here that x86 was the last major architecture that was going for high single-thread performance at all costs, the first phase lasted us a good 30 years-- from the 4004 to the flameout of Netburst.
We could consider the second phase starting when they started delivering the P4 with Hyperthreading, and its true-dual-core predecessors shortly thereafter, so we're now about 20 years into that era.
Do we have another 10 left in it?
What I always thought was exceptional is that it turns out it wasn't the incumbents who have the obvious advantage.
Take away the fact that everyone involved is already at the top 0.00001% echelon of the space (Sam Altman and everyone involved with the creation of OpenAI), but if you had asked me 10 years ago who will have the leg up creating advanced AI I would have said all the big companies hoarding data.
Turns out just having that data wasn't a starting requirement for the generation of models we have now.
A lot of the top players in the space are not the giant companies with unlimited resources.
Of course this isn't the web or web 2.0 era where to start something huge the starting capital was comparatively tiny, but it's interesting to see that the space allows for brand new companies to come out and be competitive against Google and Meta.
There are many models that call themselves open source, but the source is nowhere to be found, only the weights.
Wouldn't it be the same for the hardware companies? Not everyone could build CPUs as Intel/Motorola/IBM did, not everyone could build mainframes like IBM did, and not everyone could build smart phones like Apple or Samsung did. I'd assume it boils down the value of the LLMs instead of who has the moat. Of course, personally I really wish everyone can participate in the innovation like the internet era, like training and serving large models on a laptop. I guess that day will come, like PC over mainframes, but just not now.
The model leaders here are OpenAI and Anthropic, two new companies. In the programming space, the next leaders are Qwen and DeepSeek. The one incumbent is Google who trails all four for my workloads.
In the DevTools space, a new startup, Cursor, has muscled in on Microsoft's space.
This is all capital heavy, yes, because models are capital heavy to build. But the Innovator's Dilemma persists. Startups lead the way.
And for the record I really wish more money was being thrown outside of LLM.
Any stall in progress either on chips or smartness/FLOP means there's a lot of surplus previous generation gear that can hang and commoditize it all out to open models.
Just like how the "dot com bust" brought about an ISP renaissance on all the surplus, cheap-but-slightly-off-leading-edge gear.
IMO that's the opportunity for a vibrant AI ecosystem.
Of course, if they get to cheap AGI, we're cooked: both from vendors having so much control and the destabilization that will come to labor markets, etc.
> no idea whether it's actually benefiting someone "on the ground"
I really don't get it. Before, we were farmers plowing by hand, and now we are using tractors.
I do totally agree with your sentiment that it's still a horrible development though! Before Claude Code, I ran everything offline, all FOSS, owned all my machines, servers etc. Now I'm a subscription user. Zero control, zero privacy. That is the downside of it all.
Actually, it's just like the mechanisation of farming! Collectivization in some countries was a nightmare for small land owners who cultivated the land (probably with animals). They went from that to a more efficient, government controlled collective farm, where they were just a farm worker, with the land reclaimed through land reform. That was an upgrade for the efficiency of farming, needing fewer humans for it. But a huge downgrade for the individual small-scale land owners.
I'm curious to hear from experts how much this is true if interpreted literally. I definitely see that having hardware is a necessary condition. But is it also a sufficient condition these days? ... as in is there currently no measurable advantage to having in-house AI training and research expertise?
Not to say that OP meant it literally. It's just a good segue to a question I've been wondering about.
FCFF = EBIT(1-t) - Reinvestment
If the hardware needs constant replacement, that Reinvestment number will always remain higher than what most people think.
In fact, it seems none of these investments are fixed. Therefore there are no economies of scale (as it stands right now).
I don't know the efficiency gains per generation, but let's just say to get the same compute with this 1.25+delta system requires 2x energy. My impression is that while energy is a substantial cost, the total cost for a training run is still dominated by the actual hardware+infrastructure.
It seems like there must be some break even point where you could use older generation servers and come out ahead. Probably everyone has this figured out and consequently the resale value of previous gen chips is quite high?
What's the lifespan at full load of these servers? I think I read coreweave deprecates them (somewhat controversially) over 4 years.
Assuming the chips last long enough, even if they're not usable for LLM training/serving inference, can't they be reused for scientific loads? I'm not exactly old, but back in my PhD days people were building our own little GPU clusters for MD simulations. I don't think long MD simulations are the best use of compute these days, but there's many similar problems like weather modeling, high dimensional optimization problems, materials/radiation studies, and generic simulations like FEA or simply large systems of ODEs.
Are these big clusters being turned into hand-me-downs for other scientific/engineering problems like above, or do they simply burn them out? What's a realistic expected lifespan for a B200? Or maybe it's as simple as they immediately turn their last gen servers over to serve inference?
Lot of questions, but my main question is just how much the hardware is devalued once it becomes previous gen. Any guidance/references appreciated!
Also, anyone still in the academic computing world, do people like de shaw still exist trying to run massive MD simulations or similar? Do the big national computing centers use the latest greatest big Nvidia AI servers or something a little more modest? Or maybe even they're still just massive CPU servers?
While I have anyone who might know, whatever happened to that fad from 10+ years ago saying a lot of compute/algorithms would be shifting toward more memory-heavy models(2). Seems like it kind of happened in AI at least.
(1) Yes I know it's complicated, especially with memory stuff.
(2) I wanna say it was ibm Almaden championing the idea.
These AI data centers are chewing up unimaginable amounts of power, so if nvidia releases a new chip that does the same work in half the power consumption. That whole datacenter of GPUs is massively devalued.
The whole AI industry is looking like there won't be a first movers advantage, and if anything there will be a late mover advantage when you can buy the better chips and skip burning money on the old generations.
That makes me wonder if it's more a performance thing than an energy cost thing. I guess you also have to factor the finite supply of reliable energy hookups for these things, so if you're constrained on total kwH consumption your only way to more TOPs is upgrade. Probably ties in with real estate/permitting difficulty too. I guess what I'm picturing is if energy availability(not cost) and real estate availability/permitting timelines weren't issues, that %20 of cost probably wouldn't look too bad. So it's probably those factors combined. Market pricing dynamics are hard :/
I didn't know the recycle time of the Asics was that fast! That's an interesting point about the first-mover. I would counter that a large part of the first move value in this case is ai engineer experience and grabbing exception talent early. But with Facebook and these guys paying for them like athletes, I'd guess that experience build up and talent retention aren't as robust. Openai lost most of it's best, but maybe that's an exceptional example..
All of that to say, yeah, first mover advantage seems dulled in this situation.
Who knows, maybe apple is playing the long game and betting on just that.
One more unimpressive release of ChatGPT or Claude, another 2 Billion spent by Zuckerberg on subpar AI offers, and the final realization by CNBC that all of AI right now...Is just code generators, will do it.
You will have ghost data centers in excess like you have ghost cities in China.
You'd be competing with ASIC miners, which are 100x more cost effective per MH/s. You don't need 100,000GB of VRAM when mining GPU, therefore its waste.
From Dario’s interview on Cheeky Pint: https://podcasts.apple.com/gb/podcast/cheeky-pint/id18210553...
Is it?
Seems like theres a tiny performance gain between "This runs fine on my laptop" and "This required a 10B dollar data centre"
I dont see any moat, just crazy investment hoping to crack the next thing and moat that.
So long as there is competition it’ll be available at marginal cost. And there is plenty of innovation that can be done on the edges, and not all of machine learning is LLMs.
Most things are not perfect competition, so you get MR=MC not P=MC.
We're talking about massive capital costs. Another name for massive capital costs are "barriers to entry".
There is a natural monopoly aspect given the ability to train and data mine on private usage data but in general improvements in the algorithms and training seem to be dominating advancements. Microsoft's search engine Bing paid an absolute fortune for access to usage data and they were unable to capitalize on it. LLMs have the unusual property that a lot of value can be extracted out of fine tuning for a specialized purposes which opens the door to a million little niches providing fertile ground for future competitors. This is one area where being a fast follower makes a lot of sense.
> There is a natural monopoly aspect given the ability to train and data mine on private usage data but in general improvements in the algorithms and training seem to be dominating advancements.
There's pretty big economies of scale with inference-- the magic of how to route correctly with experts to conduct batching while keeping latency low. It's an expensive technology to create, and there's a large minimum scale where it works well.
The problem is, the arguments you're making could be used for almost any industry, including ones that we've seen sustained excess profits.
Being an unusually effective and well managed company can certainly yield sustained above average returns but that hardly means that is a meaningful barrier to entry - the question at hand.
Ozempic is marked up 200x (20,000%) because they're able to extract monopolistic rents, that's a completely different ballpark of 2x or even 5x markups.
In a fully commoditized industry participants yield average ROIs, in a mature market it's industry dependent but I would consider ranges from 10% to 100% above average ROIs to be reasonably normal. It's when things get to 10x to > 100x that I would consider to be able to extract monopolistic rents. I know I'm mixing up profits (ROIs) with markups on COGs with the issue being that companies extracting monopolistic rents tend to obscure the fact with padded expenses and the financial details needed to calculate Total Production Cost per Unit are generally not available.
In the medium term China has so much spare capacity that they maybe be the only game in town for highend models, while the US will be trying to fix a grid with 50 years of deferred maintenance.
This article is a good example of the bear case https://www.honest-broker.com/p/is-the-bubble-bursting
And then we’ll realize we wasted an entire Apollo space program to build an over-complicated autocompleter.
If the diminishing returns that we see now continue to prove true, ChatGPT6 will already be financially not viable so I doubt there will be GPT7 that can live up to the big version bump.
Many folks already consider GPT5 to be more like GTP4.1. I personally am very bearish on Anthropic and OpenAI.
It's quite likely that an order of magnitude improvement can be had. This is an enormous incentive signal for someone to follow.
Does this really describe a "most" ør are you just describing capital?
The capitalization is getting insane. Were basically at the point where you ned more capital than a small nations GDP.
That sounds mich more accurate to my ears, and much more troubling
I really hope we can get to a point where modest hardware will achieve similar results for most tasks and these insane amount of hardware will only be required for the most complex requests only, which will be rarer, thereby killing the business case.
I would dance the Schadenfreude Opus in C major if that became the case.
Whatever it is, the signal it's sending of Anthropic insiders is negative for AI investors.
Other comments having read a few hundred comments here:
- there is so much confusion, uncertainty, and fanciful thinking that it reminds me of the other bubbles that existed when people had to stretch their imaginations to justify valuations
- there is increasing spend on training models, and decreasing improvements in new models. This does not bode well
- wealth is an extremely difficult thing to define. It's defined vaguely through things like cooperation and trade. Ultimately these llms actually do need to create "wealth" to justify the massive investments made. If they don't do this fast this house of cards is going to fall, fast.
- having worked in finance and spoken to finance types for a long time: they are not geniuses. They are far from it. Most people went into finance because of an interest in money. Just because these people have $13bn of other people's money at their disposal doesn't mean they are any smarter than people orders of magnitude poorer. Don't assume they know what they are doing.
This place is better than much of the internet but still. Ah the dream would be to have this place somehow be filled with the experts on all topics and let them duel it out.
With all these models converging, the big players aren’t demonstrating a real technical innovation moat. Everyone knows how to build these models now, it just takes a ton of cash to do it.
This whole thing is turning into an expensive race to the bottom. Cool tech, but bad business. A lot of VC folks gonna lose their shirt in this space.
It's going to rock the market like we've never seen before.
This ignores differential quality, efficiency, partnerships, and lots more.
But in the consumer market segment, for most cases, its all about who is cheapest (free preferably) - aside from the few loonies who care about personality.
The true lasting economic benefits within enterprise are yet to play out. The trade off between faster code production vs poorer maintained code is yet to play out.
On what basis do you know this? Or more like your personal impression — based on asking how many people? Your friends?
Just informal conversations? Or surveys of some kind? I’m genuinely curious. After studying statistics and research design, I have learned to be very cautious at what I conclude.
I often talk with various people outside of my usual peer and work groups, but I wouldn’t know how to get anything close to an unbiased take given the social constraints. Maybe I would notice large shifts over time.
Observation: LLMs are pretty new in the scheme of things and information diffusion seems pretty different across age groups, geographies, and professions. This makes it pretty hard to even be confident that one knows how to collect good data.
I think of it a bit like the Windows vs. macOS comparison. Obviously there will be many players that will build their own scaffolding around open or API-based models. But there is still a significant benefit to a single company being able to build both the model itself as well as the scaffolding and offering it as a unit.
0 - https://x.com/thisritchie/status/1944038132665454841
1- https://docs.anthropic.com/en/docs/agents-and-tools/tool-use...
Claude Code is just much _pleasant_ to use than most other tools, and I think people are overly discounting that aspect of it.
I'd rather use CC with slightly dumber model, than Cursor with a slightly better one; and I suspect I'm far from being the only one.
They can afford to burn a good chunk of global wealth so that they can have even more global wealth.
Even at the current rates of insanity, the wealthy have spent a tiny fraction of their wealth on AI.
Bezos could put up this $13 billion himself and remain a top five richest man in the world.
(Remember Elon cost himself $40 billion because of a tweet and still was fine!)
This is a technology that could replace a sizable fraction of humamkind as a labor input.
I'm sure the rich can dig much deeper than this.
And if it does? What happens when a sizable fraction of humamkind is hungry and can't find work? It usually doesn't turn out so well for the rich.
But it's pretty obvious wealth can be created and destroyed. The creation of wealth comes from trade, which generally comes from a vibrant middle class which not only earns a fair bit but also spends it. Wars and revolutions are effective at destroying wealth and (sometimes) equitably redistributing what's left.
Both the modern left and modern right seem to have arrived at a consensus that trade frictions are a good way to generate (or at least preserve) wealth, while the history of economics indicates quite the contrary. This was recently best pilloried by a comic that showed a town under siege and the besieging army commenting that this was likely to make the city residents wealthy by encouraging self-reliance.
We need abundant education and broad prosperity for stability - even (and maybe especially) for the ultra wealthy. Most things we enjoy require absolute and not relative wealth. Would you rather be the richest person in a poor country or the poorest of the upper class in a developed economy?
I'm not sure to what extent you meant this, but I don't know that I'd agree with it. Trade allows specialization which does increase wealth massively, no doubt. And because of how useful specialization is, all wealth creation involves trade somewhere. But specialization is just one component of wealth creation. It stands alongside labor, innovation, and probably others.
> I'm not sure to what extent you meant this, but I don't know that I'd agree with it.
At a very foundational level, all wealth comes from trade, even when there is no currency involved.
When two parties voluntarily make a trade, each party gets more value out of the trade than they had before, so the sum total of value after the trade is, by definition alone, greater than the sum total of value before the trade.
Small example: I offer to trade you a bag of potatoes for 2 hours of your time to fix my tractor, and you accept.
This trade only happens because:
1. I value a running tractor more than I value my bag of potatoes
2. You value a bag of potatoes more than you value 2 hours of your time.
After the trade is done, I have more value (running tractor) and you have more value (a bag of potatoes), hence the total value after the trade is more than the total value before the trade.
The only thing that creates value is trade. It's the source of value.
All we have done is become more elaborate and sophisticated in this stuff but at the core, its been the same throughout much of time.
It's not a deficit if the value they assign to what they get is higher than the value they assign to what they give.
If they are giving away something they value highly for something they value less highly, then it's not a voluntary trade, now is it?
That's still a value assignment: they value eating and paying rent more highly than job satisfaction.
The value comes from doing the thing. Fixing the tractor creates value. Growing and harvesting the potatoes creates value.
Trading the potatoes for something you value more, yes, also creates additional value. But notice that potatoes have value to you even if you don't trade them for anything.
If you choke on a grape and a helpful volunteer saves your life, they create value. Nothing was exchanged, but they sure as hell generated wealth right there.
Labour and voluntary trade both create value. I wonder what sort of mindset one must have to forget that labour exists.
The value from value doesn't come by default; it's not a given.
Neither of those things create value unless they are actually used.
Fixing a tractor you neither use nor trade does not create value.
Growing potatoes you will never eat does not create value.
So, sure, labour can create value, but it's not a guarantee. Trade, OTOH, always creates value.
Value can be created either way: through labour or trade.
Regardless, your statement that all wealth is created by trade is false.
Scenario: You are a medium level engineer, who got laid off from a company betting on AI to replace a significant portion of their junior/medium level developers. You were also employing a middle-aged woman, to help with the kids after school and around the house, until you and your wife come back from work. She now needed to be let go as well, as you can't afford her anymore. The same thing happened to a large portion of your peers and work in the same industry/profession is practically no longer available. This has ripple effects on your local market (restaurants, caffes, clothing stores etc).
How do you see this as empowering and a net positive thing for these people individually, and for the society? What do they do that replaces their previous income and empowers them to get back to the same level at least?
More likely automatic stabilizers and additional stimulative spending would have to happen in order to fully utilize all the new productive capacity (or reduce it, as people start to work less). It's politically hard to sustain double digit unemployment, and ultimately the government can always spend enough or cut enough taxes to get everyone employed or get enough people to leave the labor force.
On a larger level, I would just ask your fictitious medium-level engineer what are they able to do today, with an AI/LLM, that they were unable to do before? As a very basic example, and one that is already true with existing LLM's, a mid-level engineer who wanted to build an app might've formerly struggled with building a UI for their app. Now, sans designer, a mid-level engineer can spin up an app UI much more quickly, and without the labor of finding and actually paying a designer. That's not to say there's no value left in design, but if you're starting out it's similar to how bootstrap (dating myself here) was an enabler because you were no longer in need of a designer to build a website (was still a huge time suck and pain in the ass though). You can multiple that by a bunch of roles and tasks today because LLM's make it possible to do things you just formerly wouldn't have been able to do on your own.
Last thing is the much more high level. Every time some new tech is introduced there's a lot of concern about displacement. I think, again, that's valid and perhaps moreso with AI. But it does seem to me like major new tech always seems to create a lot of opportunity. It might not be for the exact same people like your mid-level engineer (although I think it might for him/her), but I stay hopeful that the amount of opportunity created will offset the amount of suffering it will cause. And I don't say that in some kind of "suffering is ok" way, but just like revenue growth is the be all end all for so many companies, tech brings change and some suffering is a part of that. Prior skills become less important, new skills are preferred. Some folks adapt. Others thrive. Some are left behind.
If you're still checking in on this thread, and you actually read my diatribe, do you think I'm totally full of it? Again, I don't know that I would bet it would work out this way. Actually I probably would bet on that. But I'm definitely hopeful it will.
Narrow point: In general, one person’s impression of what is crazy does not fare well against market-generated information.
Broader point: If you think you know more than the market, all other things equal, you’re probably wrong.
Lesson: Only searching for reasons why you are right is a fishing expedition.
If the investment levels are irrational, to what degree are they? How and why? How will it play out specifically? Predicting these accurately is hard.
Example: I could also make up dozens of reasons why something permitted by the laws of physics seems ridiculous.
Google also bought Motorola for 12 billion and Microsoft bought Nokia for 7 billion. Those weren't success cases.
Or more similarly, WeWork got 12B from investor and isn't doing well (hell, bankrupt, according to Wikipedia).
A lot of that was patent acquisition rather than trying to run those businesses so it's hard to say a success or not.
I’ve explained various points at length in other comments: (i) why I selected this example (simply to show that folk wisdom or common sense is less reliable than market-driven valuations) (ii) how a funding round is influenced by markets even though it isn’t directly driven by a classic full market mechanism.
Something I haven’t said yet would be a question: how can an outsider rigorously assess the error in a funding round or acquisition? To phrase the question a different way: what price or valuation would an oracle assign based on known information?
One might call this ex-ante rationality. Framing it this way helps remove hindsight bias; for example, a subsequent failure doesn’t necessarily mean it was mispriced (sp?) at the time.
Not to be mean, but aren't you being a little hypercritical here, bringing up your bespoke example of YouTube?
In general, a market synthesizes more information than any one individual, and when they operate well it is unlikely for an individual is going to beat them.
This is a well known general pattern, so if someone wants to argue in the other direction, they need to be ready to offer very strong evidence and reasoning why the market is wrong — and even when they do, they’re still probably going to be wrong.
To answer: no, and even if it was a “yes” it wouldn’t affect the argument I was making. I’ll explain.
I was wondering how long it would take for this kind of meta-critique would pop up. Meta critiques are interesting: some people use them as zingers, hoping to dismantle someone else’s entire position. But they almost never accomplish that because they are at a different level of argument: they aren’t engaging with the argument itself.
Meta-critiques are more like an argument against the person crafting the argument. In this sense, they function not unlike ad hominem attacks while sneakily remaining fair game.
Lastly, even if I was a hypocrite, it wouldn’t necessarily mean that I was wrong — it would simply make me inconsistent in the application of a principle.
Yeah, I meant hypocritical. For some reason I couldn't find the right word.
It's at least possible that the investment pays off. These investors almost certainly aren't insane or stupid.
We may still be in a bubble, but before you declare money doesn't mean anything any more and start buying put options I'd probably look for more compelling evidence than this.
I'm sure this exact sentence was said before every bubble burst.
I would assume the majority of investors in AI are playing a game of estimating how much more these AI valuations can run before crashing, and whether that crash will matter in the long-run if the growth of these companies lives up to their estimates.
I think that's one possible interpretation but another is that these funds choose to allocate a controlled portion of their capital toward high risk investments with the expectation that many will fail but some will pay off. It's far from clear that they are crazy or stupid.
If you don't have hindsight then passing on FTX probably implies passing on some successful opportunities too. So another opportunity cost and possibly a larger one.
Answer: It's easy to pick and choose to prove one's point.
Softbank has been doing well lately by the way:
https://www.ebc.com/forex/softbank-stock-price-jumps-13-afte...
The WeWork investment proved Son never has an investment thesis - other than spray and pray.
What a minute. Isn't this the very definition of a bubble?
[1]: It was $3B at the end of May (so likely $250M in May alone), and $5B at end of july (so $400M that month).
and we continue to pretend that market generates any semblance of value.
Doing proper intrinsic valuation with technology firms is nigh-on impossible to do.
They they achieve AGI or a close approximation, and end up wealthier than god.
That's basically the bet here. Invest in OpenAI and Anthropic, and hope one of them reached near AGI.
This guy's analysis says they are bleeding out despite massive revenue
They earned $250M in May based on ARR, and about $400M in july. Model training is going to be amortized over multiple years anyway. I am not privy to how much they spent, not going to comment on that. GM was public news, and hence I got that.
Re Zitron's analysis, I don't find them to be reliable or compelling.
Claude 4 launch was not even fifteen months after the launch of Claude 3 (which is discontinued). The “multiple” is 1.2 - I wouldn’t call that “multiple years”.
I expect the next breakthroughs to be all about efficiency. Granted, that could be tomorrow, or in 5 years, and the AI companies have to stay all at in the meantime.
If there's a step-function breakthrough in efficiency, it's far more likely to be on the model side than on the semiconductor side. Even then, investing in the model companies only makes sense if you think one of them is going to be able to keep that innovation within their walls. Otherwise, you run into the same moat-draining problem.
That said innovation on the model side is more likely to come from a 10B-funded startup that still has some money to spare on the brightest researchers on top of giving them all the data and compute they want to play with.
- Buy an old warehouse and a bunch of GPUs
- Hire your local tech dude to set up the machines and install some open-source LLMs
- Connect your machines to a routing service that matches customers who want LLM inference with providers
If the service goes down for a day, the owner just loses a day's worth of income, nobody else cares (it's not like customers are going to be screaming at you to find their data). This kind of passive, turn-key business is a dream for many investors. Comparable passive investments like car washes, real estate, laundromats, self-storage, etc are messier.
I think once the sheen of Microsoft Copilot and the like wear off and people realise LLMs are really good at creating deterministic tools but not very good at being one, not only will the volume of LLM usage decline, but the urgency will too.
($100-plan, no agents, no mcp, one session at a time)
Are they putting Canadian public funds into Anthropic?
On the lawyer fees...do you expect someone to do it for free? Would you rather hire someone without the expertise? Do you think the crimes are the estate lawyers' fault?
Process costs money. You can sneer at the professionals that work hard to keep a lawful society in order, if you like. SBF sure did. Look where that got him.
---
[1] https://www.crunchbase.com/organization/ontario-teachers-pen...
[2] https://www.otpp.com/en-ca/investments/our-investments/teach...
Although I admit that the government may be on the hook to replenish any spectacular failures in such a pension plan so in that way, it is somewhat fair -- though I doubt any one investment is weighted so heavily in any pension fund as to precipitate such an event.
[1]https://www.reuters.com/technology/openai-tells-investor-not...
people don't even remember the era before the current brands. like the time a bell offshoot almost crashed canada because they siphoned all the telephone money into bad routers.
Now he's in AI investments.
A fascinating investor. I just finished re-reading Microserfs. The buzzwords may have changed between 1993 and 2025, but the human behaviors certainly have not.
On a long enough timeframe, the open source models will catch up to the proprietary models and inference providers will beat these proprietary companies on price.
Hindsight says, don't do fraud.
Things working out in the end doesn't make what he did not a crime at the time. He was a common paper hanger, albeit with billions instead.
Morally speaking, no. Practically speaking, it does. He would not have seen jail time.
It's literally exactly what Shkreli got 7 years for, even after repaying investors. If you defraud money from someone and put it back before they find out, it's still a crime. Fraud is about intent more than anything else, and they proved it for SBF.
Some who take on unreasonable risk will be among the most successful people alive. Most will lose eventually, long before you hear about them if they keep too many taking crazy risks.
Who is a great genius, and is who is just winning at "The Martingale entrepreneurial strategy"?
Filipinos have a more predictable low-loss version of this call Paluwagan.
It's not just about surviving a downtown and unforseen circumstances with some luck (like the sibling talking about FedEx barely making it). Tesla, for example, was famously extremely close to bankruptcy.
But SBF got into the situation he was in due to his egregious fraud. The accounting at FTX was a criminal joke, with multiple sets of books, bypassable controls, outright fake numbers. My guess is that if SBF had survived that particular BTC downturn that his extreme hubris and willingness to commit fraud would have eventually done him in - downturns always happen at some point, and his brazenness in his criminal enterprise showed no signs of learning from mistakes.
Sure, all hugely successful companies have a ton of luck involved. But I think it's a mistake to pretend that SBF was just done in by bad timing, or that all companies do what he did. His empire collapse was pretty inevitable IMO if you look at what a clown show FTX was under the covers.
Whether by negligence or intent, FTX was arranged so that they couldn't go bust without stealing.
That number isn’t 0
Does no one still remember that tether continually stalled audits FOR YEARS in the face of increasing scrutiny?
Right, but parent didn't say anything about an all-in bet, just double-or-nothing on a positive EV bet.
Frankly, I'd repeatedly bet on a positive EV bet too; it's a guaranteed win if you're allowed to go on for as long as you want to.
We should all try to remember this the next time we vote to cut taxes on billionaires.
"what do you call a rouge trader that makes money?"
"Managing director"
If someone makes money on time, everything is forgiven. Money blinds us.
That applies to individuals, but it probably also applies to companies. We're in an AI boom? Raise some money while it's easy.
I got the impression that some people were reselling access and adding layers of fees to profit from the hype.
More importantly, we should ask who will be left holding the bag when this bubble bursts. For now, investors are getting their money back through acquisitions. Founders with desirable, traditional credentials are doing well, as are early employees at large AI startups who are cashing out on the secondary market. It appears the late-stage employees will be the ones who lose the most.
My feeble uncle isn't allowed to buy a single lightbulb in his state yet , but burning terawatts for useless porn generators is where we are investing our engineering efforts.
5 minutes into my first opus prompt on Claude Code on an empty repo, I've already been warned by Claude Code that I'm about to hit my opus limit despite not using it in 12 days.
Intellectual engagement goes down, users get dumber and only look at quantity. China is taking first steps to continue its excellence. In the New York Post of all places:
https://nypost.com/2025/08/19/world-news/china-restricts-ai-...
"It’s just one of the ways China protects their youth, while we feed ours into the jaws of Big Tech in the name of progress."
Iconiq (Mark Zuckerberg's family office) was one of the lead investors in the round
I think one key question is can Anthropic replicate this on some other segment. Like with people working with financials.
What a fantastic amount of money flying around though, to support my inane queries to Claude.
chat gpt 5 in codex is really good
so much that i stopped used claude code altogether
cheaper too
made me realize nobody has moat, coders especially will just go to whoever provides best bang for their buck.
But who knows what will be to best tool/model to use in October.
I am very curious about the GAAP numbers here.
Step 2: achieve AGI.
Step 3: ?
Step 4: transcend money.
We're in a VC bubble; any project that mentions AI gets tons of money.
also if your founder has to use dozens of buzzwords when asked to describe what their app does and that still doesn't even explain it, its obviously just bs.
"Arcarae’s mission is to help humanity remember and unlock the power each individual holds within themself so they can bring into reality their unique, authentic expression of self without fear or compromise.
Our research endeavors are designed to support this mission via computationally modeling higher-order cognition and subjective internal world models."
lol
What do you mean lol? Isn't that awesome? Feel free to share if you think that isn't awesome. I personally don't think there is enough information here to tell if that is awesome or satire, but it is interesting how usually things like this are considered awesome, but this particular one is deemed satire.
What does the product do?
I think this is like ChatGPT, but it generates "inner monologue" in the background, and the "inner monologue" is then added to the context, and this "addresses" "sycophancy, attention deficits, and inconsistent prioritization"
Unreasonable doesn’t even start to capture it. Anthropic being worth 10% of Alphabet is beyond insane.
So 10% of valuation for 1.5% of revenue, which grew 5x in last 6 months. Doesn't seem as unrealistic as you put it, if it has good gross margin which some expects to be 60%.
Also Google was valued at $350B when it had $5B revenue.[1]
[1]: https://companiesmarketcap.com/alphabet-google/marketcap/
Investors are forward looking, and market conditions can change abruptly. If Anthropic actually displaces Google, it's amazingly cheap at 10% of Alphabet's market cap. (Ironically, I even knew that NVidia was displacing Intel at the time I invested, but figured that the magnitude of the transition couldn't possibly be worth the price differential. News flash: companies can go to zero, and be completely replaced by others, and when that happens their market caps just swap.)
Anthropic have several similiar competitors with actual real distribution and tech. Ones that can go 10x are underdogs like Google before IPO or Amazon, or Shopify etc. Anthropic current stock is beyond that. Investors no longer give any big opp. to public. They gain it via private funding
Right now nobody wants to be the first to offer advertising in LLM services, but LLM conversation history provides a wealth of data for ad targeting. And in more permissive jurisdictions you can have the LLM deliver ads organically in the conversation or just shift the opinions and biases of the model through a short mention in the system message
As I said, insane. And that’s not even considering the 10 to 15% shares of Anthropic actually owned by Alphabet.
You may not agree with the market's estimation of that, but comparing just present revenue isn't really the right comparison.
Basically, 5x-ing revenue in 8 months off of a billion dollars starting revenue is insane. Growing this quickly at this scale breaks every traditional valuation metric.
(And no - this doesn't include margins or COGS).
Somebody above said that Anthropic might reach $9 billion ARR by the end of this year.
How much was google revenue in 2003? It was 1.5 billions (2.6 in today's USD)
Not saying the price is justified, but the comparison is not very fair.
I could see two or three percent, but this seems like a pretty big stretch. Then again, I'm not a VC.
Machine ice became competitive in India and Australia in the 1850s, but it took until the start of World War 1 (1914) for artificial ice production to surpass natural in America. And the industry only disappeared when every household could buy a refrigerator.
Self-driving doesn't have to scale globally to be economically viable as a technology. It could already be viable at $400k in HCOL areas with perfect weather (i.e. California, Austin, and other places they operate).
If AI is winner take all, then the value is effectively infinite. Obviously insane, but maybe it's winner take most?
Convincing billions of users to make a new account and do all their e-mail on a new domain? A new YouTube channel with all new subscribers? Migrate all their google drive and AdSense accounts to another company, etc?
This is trivially simple and creates no moat?
I know you aren't asserting this but rather just putting the argument out there, but to me at least it's interesting comparing a company that has vendor lock-in and monopoly or duopoly status in various markets vs one that doesn't.
I'd argue that Google's products themselves haven't been their moat for decades -- their moat is "default search engine status" in the tiny number of Browsers That Matter (Arguably just Chrome and Mobile Safari), being entrenched as the main display ad network, duopoly status as an OS vendor (Android), and monopoly status on OS vendor for low-end education laptops (ChromeOS). If somehow those were all suddenly eliminated, I think Google would be orders of magnitude less valuable.
When their sales have nosedived, new products have flopped, their CEO is the most disliked man in America, and their self driving still requires someone in the car at all times?
Tesla is a GameStop level meme stock.
From a technical perspective, they manage to attract top talent - Google / OpenAI lose a lot of good people to Anthropic. This is important since there are few people who can transform a business (e.g., the guy who built Claude Code). Being attractive for top talent means you're more likely stumble upon them.
Edit: After looking it up, normal P/Sales ratios are on the order of about 1. They vary from like .2 to 8 depending on industry.
Its not internally consistent, at all.
I do think this is important. Many of the best researchers are also religious AGIists and Anthropic is the most welcoming to them. This is a field where the competence of researchers really matters.
Impressive in the valuation, terrifying in the fact that they need to keep raising and these valuations might not prove justifiable
However, I remembered when Youtube was young. It was burning money every month on bandwidth.
After selling out to Google, it took another decade to turned profit. But it did. And it achieved its end game. As the winner, it took all of the video hosting market. And Google reaped the entirety of that win.
This AI race is playing out the same way. The winner has the ability to disrupt several FAANGs and FAANG neighbors (eg. Adobe). And that’s 1-2 trillion dollar market, combined.
I often think about that when trying to evaluate forward looking tech. Even though 99% of the time logic like that proves to be correct, it's also true that most of the time the winners in a race did that exactly because they defined some piece of the standard framework of logic that everybody else played by. Uber is similar - they shouldn't exist, they basically broke the law in most countries they moved into, brazenly violated all kinds of barriers that kept taxi industry completely entrenched for decades. But now they are dominating in most of these countries.
"Uber is similar - they shouldn't exist, they basically broke the law in most countries they moved into, brazenly violated all kinds of barriers that kept taxi industry completely entrenched for decades"
And here's a simple way to demonstrate my point - backed by VC - Uber accelerated its growth and got to the point it was so widely adopted nobody could stop them from operating.
Between OpenAI, Anthropic, Google, Facebook, xai, Microsoft, Mistral, Alibaba, DeepSeek, z.ai, Falcon, and many others, AI feels a lot more competitive.
And it's cash from asset managers. Its not 10Bn worth of compute time from Microsoft or Google.
Much like any other investment. What do you think makes this more speculative than any other investment?
Deficit spending doesn't create new money. Deficit spending borrows existing money from the population and institutions in exchange for a promise of future government revenues. The Fed does not participate in treasury primary auctions and does not monetize the debt as a means of funding government operations.
If you printed new money to pay for the government, you wouldn't have a debt. That's double-counting. Not to mention the debt is twice as large as the entire money supply so what you're suggesting isn't even physically possible. It would be inflationary to simply print new money to finance spending, which is exactly why it's not done.
[edit] Also the debt limit is a stupid concept that's likely unconstitutional. Congress authorizes spending, meaningful debate over paying for it by adjusting the debt limit likely falls afoul of the 14th amendment's public debt clause. But yeah I mean the debt limit goes up because the government spends more money than it takes in, so it needs to borrow more each year.
The Fed doesn't have nearly as much control as folks think.
The Fed directly created money during QE and they are directly destroying it during QT. There's a net add, but that's mostly because the economy is growing, which creates new demand for money as expressed by demand for debt.
The money supply staying fixed or shrinking is a non-goal anyways. It's irrelevant. What matters is inflation as measured from the change in actual prices.
The conversation always goes like this.
You: "The government is lying about inflation!"
Me: "Ok, what rate do you think it's actually been?"
You: "10%!"
Me: "So you're telling me inflation over the last 30 years was 1700%? So prices are now 17X higher than in 1995? You sure?"
Then we look up historical prices like this.
https://www.tasteofhome.com/collection/this-is-what-grocerie...
In 1995 ground beef was $1.49/lb.
Bread was $.89/loaf.
Eggs were $0.92/doz.
Milk was $2.50/gal.
idk if you're shopping at Erewhon but where I shop ground beef isn't $25/lb, bread isn't $15/loaf, eggs, well, you got me there lol, and milk isn't $42.50/gal.
Unless the conspiracy is far bigger than we think, or "they" are everywhere, whoever "they" are, I think it's safe to assume that inflation numbers have been pretty accurate.
That new money is different from the new money the central bank creates to push interest rates down. That later one the US has been destroying. But both do many of the same things (but not all).
Think about it this way: if money were just created to fund the deficit why would we have a debt? That's double-counting. You can invalidate your hypothesis very easily: the M2 money supply is about half the size of the debt. It's not possible to square that circle unless deficit spending was re-pledging existing money.
1) Will I (and others) be able to get a H100 (or similar) when the bubble pops, and would that lead to new innovations from the GPU poor?
2) Will China take the lead in AI as they are less "capitalistic" with the demands for outsized returns on their investment compared to US companies, and they may be more willing to continue to sink money into AI despite possible market returns?
Some will be used a lot will be written off and tossed away.
H100s will not age this well. It's not like owning old railroad tracks, it's like owning a fleet of 1992 Ford Taurus's. They'll be quickly obsolete and uneconomical in just a few years as semiconductor manufacturing continues to improve.
[Voted down by the cash cabal! Arise! Knowledge workers of the world, you have nothing to lose but your SPARE CHANGE!]
A man looks at economics. Understands nothing. Thinks it must be all fake and made up. He must be so smart for seeing it through!
Btw there's a decentish board game called Modern Art based around the pricing of art with no intrinsic value.
A company has agency; it seeks to add economic value to itself over time including changing people’s perceptions.
I don’t see how your comments have any bearing to the point I was making. What am I missing?
>>> me: Perhaps there are salient differences between art on a wall and a company.
>> you: At heart, not really. The whole point of all of this is to motivate humans to get off their butt and reduce entropy.
> me: A painting on a wall is merely an inanimate object. / A company has agency; it seeks to add economic value to itself over time including changing people’s perceptions.
The Horror! Just look at the disjointed conversational history above. It seems like some sort of drunken history episode where people aren’t paying attention to each other.
Should I assume you are trying to understand what I’m saying? It is becoming less plausible with every comment. (I’m referring to the “be charitable” part of HN guidelines.)
Additionally, there is another anti-pattern at work here: this seems like a pretty inane definitional argument. You’re claiming there’s no difference between art on a wall and a corporation entity? By what definition? What is the utility of your definition; meaning, what can you do with your definition that provides differential predictive power?
My claim: when it comes to valuation, an agent is sufficiently different from a non-agent (yes, even if it appreciates!) What is the criteria for “sufficiently different”? To explain: if you get more benefit out of a distinction than it costs you to make the distinction, it is a net benefit.
In this case about valuing things, someone who makes a living building predictive valuation models is going to distinguish wall art from corporate entities because doing so is useful for prediction.
Of course they have some things in common. This is irrelevant to the question of “is making this distinction worth it?” As long as predicting the difference between them is valuable paying attention to the distinction is valuable.
This kind of talking past each other is one of many reasons “why we can’t have nice things” such as useful discussion. Shameful.
If you propose some grand unified theory that says two things ultimately derive from the same thing, that’s fine, but if you’re going to use it for prediction you’ll have to explain how to apply it.
How? The market is the one that made the decision to invest. They are not playing musical chairs.
The laws of economics have the kind of inevitability you expect from the laws of physics. Disrespect them at your own peril.
Absolutely not- that is ridiculous.
Let's take "supply and demand" for example. Supply and demand only applies when you assume greed and capitalism. In a different social construct, the traditional supply & demand completely falls apart. But the problem is economics is presented as some sort of fact of nature. It just reinforces survival of the fittest instead of society that helps everyone.
Increasing prices because of demand is not the law of the land- it's a greed of humans that you normalize and make acceptable.
All of whom have a real world standardized thing to exchange for this already
Why do you think this discussion even needs to include the people who don’t have that standardized thing to exchange? If thats what you think
* I’m an equal opportunity critic of comments that are indistinguishable from people yelling into the void with whatever pops into their head. So yes, I’m extremely critical of this very human tendency that isn’t helpful.
Remember, every technology you use today followed this pattern, with winners emerging that absolutely did go on to be extremely profitable for decades.
Most of us remember the .com era. But in the early 1900s there was literally hundreds of automotive startups (actual car companies, and tens of thousands of supplier startups) in the metro-detroit area: https://en.wikipedia.org/wiki/List_of_defunct_automobile_man...
Some of these went on to be absolutely fantastic investments, most didn't. All VCs and people who invest in venture know this pattern.
Everybody involved knows exactly the high risk level of the bets they are making. This is not "dumb" money detached from reality, and the pension funds with a 3% allocation to venture are going to be just fine if all these companies implode, this is just uncorrelated diversification for them. The point of these VC funds is to lose most of the time and win big very rarely.
There will be crashes, and more bubbles in the future. Humans will human. Everything is fine.
Too many normies betting their life savings without understanding this risk in prior bubbles, so we regulated away the ability for non-institutional investors to take venture risk at all.
Some institutions try to achieve this by launching their own cryptocurrencies, but by and large, the market isn't biting.
Sure, but in my view, I think we are on the downtrend now and this line of thinking has been taken way too far.
It'll take a solid year and about 30k.
Any chance of even talking to a VC as an outsider?
I'm basically hiring a part time front end junior assistant to fill in some gaps.
I'm hiring out of a cheaper county where 30k can actually do something. The idea is the first 30 gets an MVP done.
I don't think I could finish anything worth selling for that.
That’s just about the most tangible benefit I see this AI breakthrough delivering. What an asset to have too, socially and civilly, especially when compared to the west’s primary adversary: the CCCP and its communist message of ‘equality’ for the people when they’re still working six days a week!
> Headline: OpenAI raises 400 Trillion, proclaims dominion over the delta quadrant
> Top comment: This just proves that it's a bubble. No AI company has been profitable, we're in the era of diminishing returns. I don't know one real use case for AI
It's hilarious how routinely bearish this site is about AI. I guess it makes sense given how much AI devalues siloed tech expertise.
Because Nvidia is making actual profit selling hardware to those who do, not hoping for a big payout sometime in the future. Different risk/reward model, different goals.
It's hard to escape the conclusion this is dumb money jumping on a bandwagon. To justify the expected returns here requires someone to make a transformer like leap again, and that doesn't take spending huge amounts in one place, but funding a lot more speculative thinkers.
Because of the legal uncertainty about what they were doing. There was no fundamental technological impediment.
Here the technology simply doesn't exist and this is a giant bet that it can be magically created by throwing (a lot) more money at the existing idea. This is why it's "dumb money" because they don't seem to understand the dynamics of what they're investing in.
I made a new top-level comment mentioning the 2006 YouTube acquisition only to show that many people were shocked, but -surprise- markets are usually better predictors than individual hunches.
It is very far from a situation where the price discovery mechanism is allowed to work.
1. How much an organization is willing to invest in X competes against other market opportunities.
2. The effective price per share (as part of the latest round of financing) is an implicit negotiation.
It is a matter of degree, sure, but my point still stands: there is a lot of collective information going into this valuation. So an individual should be intellectually humble relative to that. How many people have more information than even an imperfect market-derived quantity?
No, there isn't. For example, I would like to legally bet against Anthropic existing as a going concern in five years. Where can I do this? All the information against them is discarded and hidden.
> How many people have more information than even an imperfect market-derived quantity?
I’ll restate the point because I don’t think you’re understanding what I mean.
Do you think this funding round was irrational from the point of view of the investors? If so, how can you make such a claim? Do you have information they do not?
It is possible you have some bit of knowledge they don’t, but on balance it is unlikely that you are operating from a position of having more relevant information.
Sure it was overcome, but not because of YT or Google, but because of external forces causing those people fighting it to converge on hosting their content on the platform.
Clear, testable predictions are possible if you try.
(if it hadn't been liquidated)
Do you own any Amazon, Alphabet, or Salesforce, perhaps through some index fund? Congratulations, you own some Anthropic. This matters to you.
And market conditions matter to you, too. Every deal is a comparable mark that factors into every other deal. Where this tech is going, and whether we're in a bubble or just getting started... these are forces that are interested in you, even if you're not interested in them.
The issue wasn't that crypto markets in general were down at that point; the issue was they were doing frauds.
If the liquidators had perfect hindsight, they'd be trading their own money. Not cleaning up other people's messes.
Their job is to be responsible and follow procedure.