The world is using 300T token a day. Barely anyone is using AI. This number would grow 30x. If there is anything, Google, OpenAI, Anthropic are undervalued by a great deal.
The world is using 300T token a day. Barely anyone is using AI. This number would grow 30x. If there is anything, Google, OpenAI, Anthropic are undervalued by a great deal.
There’s no business in tokens. Zero. The only business is selling compute and Anthropic rented all its compute.
Structurally they’re in a terrible position on what’s unfolding now with the completely collapse of model valuations. With truly “open” models gaining more and more momentum models are basically worth nothing and “token” prices are collapsing to the cost of compute. That doesn’t leave anything on the table for the Anthropic and Open AI’s of the world.
but Opus 5.5 is the best product ever made in the history of the world (yes I'm serious). There's gotta be some consideration for that.
It has zero network effects and barely any economies of scale. It's a good product without a moat which will get replaced by any company with their own compute and cheaper energy.
I think the only way out for Anthropic is to buy a few neoclouds as they go under, but even then it's a coin flip as to if they survive the next 3 years.
Or they partially crash and get bought by Oracle. The Sun route
The market is a cold beast and doesn’t care. If things hit the fan it will happily take ownership Opus 5.5 in bankruptcy court and slit everything else out on the street.
If people paid a profitable price for Opus 5.5 it would not be the best product in the world and would not be slotted into so many things where it's benefiting from ubiquity. That circle needs squaring. There will be an Opus 5.5 that's a 10th of the price in a few months.
I personally don't really believe the meme that Anthropic and OpenAI are charging less than the marginal cost (cost of inference on requests). Open weight models have surpassed the abilities of their closed weight models when such claims started, and the inference costs on said open weights models are provably not so high.
Where they quite reasonably might be making a loss is their marginal costs + amortised fixed costs (training).
I predict that eventually the training expenditure is going to slow down. At that point, the large commercial providers are likely hoping they can keep inference demand up, and recoup their investment.
The driver for the training is an arms race between providers and open weights labs. But several dynamics could play out:
* Models will be seen as good enough, with only fine-tuning required, for most use cases, and demand for expensive training of foundation models drops off, stopping the arms race. * Most of the participants cannot raise capital due to lack of market confidence in them, and there is a 'winner takes all' dynamic with the remaining provider, who can then raise costs.
I personally think the first end game is more likely than the second.
You are absolutely entitled to call it a meme but that is also their own internal reporting/panicked emails. Subscriptions are offered at a large subsidy with the benefit of more training data and adoption.
> I predict that eventually the training expenditure is going to slow down. At that point, the large commercial providers are likely hoping they can keep inference demand up, and recoup their investment.
The big labs can't do this that's their edge over open models. The stickiness is critical for them to keep up. They might try again but it's going to finish badly.
openai has a billion users. Ads and they know it.
The math doesn’t math here. There isn’t enough spent on online ads to dig them out of this hole.
Is that potential source of revenue? Maybe but folks are verbally very negative on AI ads.
Is that a viable business plan for Open AI? No.
What’s with the low quality posts?
Google search ads are highly converting because they are camouflaged.
OpenAI ads are nothing of the sort!
Google search ads used to also convert highly because they were well-targeted and positioned at the point of consumption.
If I'm searching for red widgets, it's a fairly strong indicator I'm way more likely than the average to want to BUY red widgets, so a well-placed ad for redwidgets-cheap.com didn't feel hostile.
Now when it's "I am clearly searching for red widgets but you want to railroad me into blue doohickeys because they're paying more per click", you're one step above a Taboola chumbox.
Google search ads are just highly optimised over now 26 years. There's no reason that OpenAI won't do the same. How successful they are going to be is not clear though.
No clue how it's low effort to point out ads. It's not my fault you and everyone here lacks the product creativity to see the reality that they WILL be embedded in the answers and you will not be able to tell it's an ad. This will result in a literally the best ad system the world has ever seen. They simply need to wait a few years for the general pop to be reliant on them (which if you work with kids surprise they believe everything it says) and then slowly shift this over. They will be the most valuable company in the world bar none.
Do I think it's dystopian and shitty? Yes. Doesn't mean it ain't gonna happen and doesn't mean it's not the right move for the business.
> surprise they believe everything it says
It's not just kids in my observation.
The equivalent model for chatgpt would be randomly getting replies that are generated from a prompt injected with advertiser copy and indicated by a small "sponsored reply" in the corner.
Honestly, having written that out... I'm not sure that's an outcome that's generally avoidable.
Their ads aren’t converting like search ads, more like social
Even social is charitable..
Google, I hear, is very very good at Ads
Cost per token is going to tank; it's a race to the bottom. The frontier model providers have no moat between each other and eventually even the open weight models will be "good enough" for almost every purpose.
Right now the market is expecting something a lot more than "sells a lot of tokens" out of these things to justify their valuation.
The future is in private data, private models, and private APIs, which can't be distilled as easily as public APIs.
Wouldn't that depend on how much data? I assume those who own the data would rather buy the hardware and create models from this data themselves, as opposed to sending their data off to someone else, or allowing it to be pirated.
For most of those Google is much better positioned.
I disagree, the moat will be of those holding rights for valuable problem domain specific training data. Examples I can think of:
Adobe sits on a treasure trove of:
- Copyrighted multimedia
- Years of telemetry on content creation
Zeiss sits on decades of data:
- Medical imaging, and surgical footage.
- Silicon waffer inspection.
These embody a training corpus that is de facto (lawfully) out of reach of the crawler scrapper approaches of Google/Meta/OpenAI/Anthropic (even if they pirate the seas for torrents like before).
Neither Adobe or Zeiss are an AI power house but the barrier to entry has also dropped with all the movement in the field.
Google owns Google Books and a myriad other data sub businesses beside the index of the entire web.
Excellent point, and one that’s not raised enough in these conversations.
Well openai and anthropic are sitting on absolute treasure troves of private data that is their users' interaction with their models.
personally thats why i think OAI is in a way better place with them targeting ads.
Yeah, nobody is doubting that people want to consume tokens. We're doubting that that Google, OpenAI, or Anthropic has a significant enough moat to keep people buying their tokens there. Over the past couple years the moat has been the model. As open models are getting better, that moat is drying up. Eventually the market for "tokens" will converge with the market for compute, which is a commodity with little to no moat.
I think even if the moat dries up, you could think of tokens/inference similar to compute. There's all the open source tooling available in the world for you to rent your own bare metal servers/racks and use a white glove service, yet there is demand for cloud.
At current prices, it makes sense to self host, but if the price goes down and the demand grows across different verticals, there is room for this business.
CFOs looking at tons of 2026 token spend but with no clear ROI have joined chat.
ROI in the space is not complicated, their revenue is already insane. You just drive down costs with fixed capex (GPUs), that will continue to pay out over 5-10 years. These are growing pains, they need to make it past the GPU scarcity/ramp up phase.
The CFOs he is talking about are the enterprises who are spending big money on the tokens ; they are not seeing RoI ; not hyperscalers or frontier labs
Yep, i can't disclose specific numbers but we went from $X/month/person budgets to $X-100/month recently, and this is a pretty big company with ~5000 people. From what I can tell the pursestrings are only tightening further too, too little actual real-world benefit for the obscene spend
% would be helpful, is it -1% or -100% per month per person?
Correct. The people that control the pot of money these labs ultimately need to survive seem increasingly unconvinced that giving these labs money is a good idea. At least at the levels the labs need to not go bankrupt.
What GPU scarcity? Unconnected GPUs are being hoarded by the oligarchy:
https://www.businessinsider.com/companies-hoarding-unused-ai...
Reports indicate GPUs are increasingly being put on shelves in warehouses. There’s a scarcity of power and data centers to host them, but data increasingly suggests the world has an over supply of chips at the moment.
Like most “scarcity” scenarios things flip from under supply to massive oversupply real quick. Remember during COVID when lumber got stupidly expensive and then almost overnight folks couldn’t give wood away.
If everyone is using AI the ROI could just keeping up with everyone else. If anything Intelligence is now becoming a tax or the cost to use the toll road.
If models got smarter and an average prompt response used let's say 5k tokens instead of 10k tokens, would that change your prediction?
Point is, the number of tokens is a silly metric to proxy these companies' growth.
Everyone I know and their mom is using AI, technical or not. The question is, is it profitable to run a frontier AI lab and sell inference while spending billions on R&D.
Anthropic could be super profitable right now if they stopped developing new models. But then they'd quickly fall behind because there's not much of a moat. Maybe the only ones who will survive this are huge companies like Google with separate incomes who can swallow the cost of staying on top.
I saw a (highly reputable of course) tweet a few days back suggesting all of OpenRouter's DS4 traffic would fit on a single rack of modern Nvidia kit.
Hm, that’s doubtful. I think even one instance of DS 4.1 Flash alone at ~500B parameters needs a full node (8 GPUs).
4xh200 seem to be enough if you cpu offload the engrams (which is pretty typical and even expected).[1]
And these are quite old cards by now. But yeah "all of OpenRouters traffic" is gonna take more than a rack
https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hard...
[dead]
high token usage but where is the value? How come the other s&p companies aren't killing it?
Well they just need to use it properly and build a bunch of software systems for them to lever..oh hang on a minute...
It's like Agile, if it doesn't work, they must be doing it wrong.
And price could drop 100x.