So if the idea is to unseat Google, and make LLMs that are monetized by ads -- well that would be a lot of revenue!
The problem is obviously that Google knows this, and they made huge investments in AI before anyone else
---
I guess someone wants to do to Google what Apple did to Microsoft in the mobile era -- take over the operating system that matters by building something new (mobile), not by directly trying to unseat Microsoft
The problem seems to be that no one has figured out what the network effect in LLMs is. Google has a few network effects, but the bidder / ad buyer network is very strong -- they can afford to pay out a bigger rev share than anybody else
Google also had very few competitors early on -- Yahoo was the most credible competitor for a long time. And employees didn't leave to start competitors. Whereas OpenAI has splintered into 5 or more companies, fairly early in its life
[1] at least according to the Acquired podcast, which is reputable
edit: oops, it was profit, not revenue
https://www.acquired.fm/episodes/google
Google with this business model makes more profits than any other company, ergo tautologically, is the most magical business model ever discovered.
By yearly revenue, the highest revenue company is Walmart, followed by Amazon, which make somewhere near twice the revenue of Alphabet (around 11th place, per https://en.wikipedia.org/wiki/List_of_largest_companies_by_r...). Especially if you account the inflation, the total lifetime revenues of the major oil companies will easily dwarf Google.
Google is nowhere close to earning the most revenue of any business ever.
Companies like Uber and Amazon operated at a loss, true. But they had an actual product. And they didn't come close to the money Google, Meta, OpenAI and Microsoft are losing.
As people pointed, this is wrong.
But anyway, Google's revenue last year was enough to satisfy the smallest point of the interval the article points out. And barely so.
So if everything goes perfectly for the next 5 years capital-wise, and AI manages to capture Google's revenue, at the most optimist conditions, they will be able to break even with depreciation.
Honestly, that is better than what I was expecting. But it completely different from the picture you will see in any media.
Google is very large, and I’m sure Acquired framed the statement is such a way that it’s true, but this statement as you presented it is false.
Other publicly traded companies have reported more lifetime revenue. Other product categories besides internet search have generated more revenue.
At the very least, the exact same network effects with respect to advertising that search has. The vast majority of frequent ChatGPT users I know mostly use it like a search engine.
That said, those network effects will be massive. Ads in LLMs are going to be unprecedentedly lucrative, irrespective of the platform. Google/Meta currently charge so much for ads because they have such enormous proprietary profiles on users based on their search/communication history that they can offer advertisers the ability to target users with extraordinary granularity. But at the end of the day, the ad itself is static and obviously an ad. LLMs will make these ads dynamic and insidious, subtly injected into chats in the way a real-life conversation might happen to discuss products. LLMs will become the ultimate word-of-mouth advertisers, the final form of astroturfing.
We are not seeing that (currently) with GPUs. Perf/watt has basically completely stalled out recently while tokens per user has easily increased in many use cases has went up 100x+ (take Claude code usage vs normal chat usage). It's very very unlikely we will get breakthroughs in compute efficiency in the same way we did in the late 90s/2000s for fiber optic capacity.
Secondly, I'm not convinced the capex has increased that much. From some brief research the major tech firms (hyperscalers + meta) were spending something like $10-15bn a month in capex in 2019. Now if we assume that spend has all been rebadged AI, and adjust for inflation it's a big ramp but not quite as big as it seems, especially when you consider construction inflation has been horrendous virtually everywhere post covid.
What I really think is going on is some sort of prisoners dilemma with capex. If you don't build then you are at serious risk of shortages assuming demand does continue in even the short and medium term. This then potentially means you start churning major non AI workloads along with the AI work from eg AWS. So everyone is booking up all the capacity they can get, and let's keep in mind a small fraction of these giant trillion dollar numbers being thrown around from especially OpenAI are actually hard commitments.
To be honest if it wasn't for Claude code I would be extremely skeptical of the demand story but given I now get through millions of tokens a day, if even a small percentage of knowledge workers globally adopt similar tooling it's sort of a given we are in for a very large shortage of compute. I'm sure there will be various market corrections along the way, but I do think we are going to require a shedload more data centres.
At least for gaming, GPU performance per dollar has gotten a lot better in the last decade. It hasn't gotten much better in the past couple of years specifically, but I assume a lot of that is due to the increased demand for AI use driving up the price for consumers.
Why wouldn't Moore's Law continue?
With GPUs and CPUs. You need to replace entire thing. And now they are the most expensive part of the system.
Other option is doing more with same computing power, but we have utterly failed with that in general...
A few thoughts:
1. The comparison to previous tech bubbles is apt - we're seeing massive capex without clear paths to profitability for many use cases.
2. The "build it and they will come" mentality might work for foundational models, but the application layer needs more concrete business cases.
3. Enterprise adoption is happening, but at a much slower pace than the investment would suggest. Most companies are still in pilot phases.
4. The real value might come from productivity gains rather than direct revenue - harder to measure but potentially more impactful long-term.
What's your take on which AI applications will actually generate enough value to justify the current spending levels?
1) the labor angle: it’s been stated plainly by many execs that the goal is to replace double percent digits of their workforce with AI of some sort. Human wages being what they are, the savings there are meaningful and seemingly worth the gamble.
2) the military angle: the future of warfare seems to be autonomous weapons/vehicles of all sorts. Given the winner takes all nature of warfare, any edge you can get there is worth it. If not investing enough in AI means the US gets steamrolled by China in the Pacific (and other countries getting steamrolled by whomever China wants to sell/lend its tech to), then it seems to justify most any investment, no matter how ridiculous the current returns seem.
First of all, warfare is not winner take all. That's a sort of video game naive conception. The famous quote "War is politics by other means" is much more accurate.
When armed conflicts happen, it's because the belligerents have specific objectives, and very rarely is that objective "the total obliteration of the enemy" vs something more specific and concrete like territory, access to natural resources, the creation of a vassal state that can be exploited, or sometimes purely ideological (nationalist notions growing into the idea that a people are entitled to an empire).
Anyhow, the point is warfare is not a winner takes all game of obliteration.
But also the idea that the future of warfare will be all autonomous weapons is massively overweighting on drone hype, and ignoring that a lot of the fundamentals haven't changed since the days of Bismark, despite the rise of drones, computer vision algorithms, etc.
A simple example is Ukraine, where the battlefield is essentially defined by the combination of traditional artillery, mines and similar fortifications, and simple observation drones that don't have any particularly complex AI. The combination creates a 20 km "no go zone" that has nothing really to do with autonomy.
In fact, the more AI centric loitering munitions provided by US/EU firms have performed quite poorly in Ukraine, which is why they're favoring much more simple implementations like using hobby FPV drone components, or remote piloting via GSM modems, etc.
Will these technologies play an increasing role in future conflicts? Of course. But they're not going to completely upend things, or obliviate more traditional platforms.
Heck, another example would be simple hand coded AIs have been better than humans in dogfights for decades now. And it matters exactly zero for real world conflicts, because what fighter pilots actually do isn't a Top Gun movie.
If those companies replace a low 2-digits percentages of the developers, and capture their entire salary, it's still not enough to reach the depreciation numbers on the article.
On 2, that could justify it... Except that we are talking about fucking LLMs. What do anybody expect LLMs to do in a war that will completely obliterate some country?
Even if we grant that this is possible, have any of these execs actually thought through what happens when their competitors also replace large chunks of their workforce with AI and then begin undercutting them on price? The idea that "our prices will stay exactly the same, but our salary costs will go to zero and become pure profit instead!" is delusional even if AI can actually replace large numbers of people, which itself is quite doubtful.
Unless you're a monopoly, I don't see how AI will lead to these massive cost savings everyone is hoping for.
Has anyone said this?
The point is such a shift would transfer spending from labour to these AI companies. That satisfies their investors' thesis requirement.
That about sums it up.
I think the author is wrong on this point however: > Today’s tech just cannot do what will be required of it (AI shouldn’t be dispensing medication when it can’t even count to 7).
The failures of AI are thought provoking, and more so when considered together with other results where AI performs at near expert level on challenging benchmarks. However, perfect reasoning is hardly a requirement. Most humans are not particularly good at reasoning, and most jobs do not need it. Both humans and AI can use calculators and other tools. All that's needed is that the AI is more or less as good as a human, while requiring much less pay.
A good exercise to appreciate the current state of AI might be to ask AI to write an essay about this topic ("how much revenue is needed to justify current AI spend, and draw parallels to the dotcom boom and building the transcontinental railroad"). Try it with two different models, using the deep research mode. I expect the results would be humbling.
... So, in summary: We likely need on the order of hundreds of billions to low-trillions of dollars annually in AI revenue to justify the present level of infrastructure and model investment. Current realized revenues are many orders of magnitude below that.
But that’s the cold math. History suggests that such math often overlooks strategic externalities, spillover effects, hype, and speculative capital flows.
I don’t know if it’s that surreal or unexpected. There’s a reason “The Emperors Clothes” is such a classic, enduring, fable. It’s happened before. It’ll happen again.
Not shading the article. All good points, just was surprised the author threw this bit in.
Buy more tulips.
Railroads and fibre are better examples. Tulips are actually fucking useless as a productive asset. Railroads, fibre-optic cables, power production and datacentres are not.
Between 1900 and 1929 the total capital invested in rail stocks and bonds in the US grew from $10.8bn (against GDP of $21.2bn) to $21.4bn [1] (1929 GDP was $104.6 billion). [2]
So the current AI capex doesn’t really seem too far fetched by comparison.
Apple’s revenue went from $13.9bn in 2005 to $391bn in 2024.
Google’s revenue went from $6.1bn to $348bn over the same period.
Microsoft’s revenue was $197.5m in 1986, $345.9m in 1987, $591m in 1988, $804.5m in 1989, $1183m in 1990, $1843m in 1991, $14.5bn by 1998 and $19.75bn in 1999.
So revenue 10x’d from ‘86 to ‘91 and again from ‘91 to ‘99.
Those are all nominal unadjusted figures, but all three companies are arguably “category defining”.
For OpenAI to go from public launch to what looks like ~$12 billion projected annualised revenue so quickly says quite a lot. If it only follows the Microsoft trajectory that would be $120bn by 2030.
But going back to the railways point: the impact of railways wasn’t measured in the revenue of just one company, but rather across the whole physical supply chain.
If you assume that AI could be as transformative to the digital supply chain (and everything it touches) then you could argue that investments of 20% of global GDP wouldn’t be crazy.
(Though, railway mania, 1929 crash, etc…)
[0] https://www-users.cse.umn.edu/~odlyzko/doc/mania18.pdf
[1] https://www.researchgate.net/figure/Composition-of-capital-r...
[2] https://www.measuringworth.com/datasets/usgdp/result.php?use...
We are in geopolitically fraught times. Money alone is not capital.
We have been living in an era where financial capital has dominated.
We are entering an era where computing capital, intellectual capital, and military capital will dominate.
The people in control of those when the game changes are the ones writing the rules.
These are bullshit terms. Capital is capital. Military production, IP production and yes, AIs running in datacentres and on the grid, are all subject to economic forces. (Folks argued railroads were a different form of capital in the 19th century, too. And fibre optics. And tulips. And dot-com companies. And computer-assembled American mortgage instruments.)
We might be investing for a golden future. We might be the Soviet Union baited into unsustainable spending commitments. The answer to these questions isn't in pretending this time is different, or that economics can be suspended when it comes to certain questions of production and return.
We could probably debate how holding onto piles of green paper doesn’t provide much advantage in certain contexts, but I suspect you’d agree with that.
My proposal is that there’s a high likelihood the bet is that green paper matters less than high powered AI systems, and as far as I can tell, that’s a reasonable bet.
yes, true, but it would probably be better to pour money into shipyards than data centers
They think they are building an AI god.
If you think of it in religious terms it suddenly makes sense. Expected rate of return? One scenario has has infinite expected return (some kind of pascals wager/mugging)!
Of course there will be no AGI. Just a planet we'll have to live on where those deluded idiots wasted our resources on some boondoggle. Maybe this kind of concentration of power is a bad thing? I think we are going to get to those kind of questions once the party is over.
AI data centers have an impossibly short runway to achieve profitability
2) llm-provided tax/legal advice alone was worth hundreds of euros for me
3) there is no easy way to capture all this value. I still think edge intelligence/autonomy has a way to pay for the rest of the party, because those will be physical things. People eagerly part with money for things. Governments and enterprises will maybe pay for cloud services, if the price will be right
That's called a "bubble". Obviously, this time it is different until it isn't.
I own several books of trig and other tables, three slide rules and a couple of calculators, a working Commodore 64 and an IT consultancy company.
We are fiddling with LLMs as yet another tool. We are getting some great results but not earth shattering.
Tulips are very pretty flowers. I have several dozen in my garden. I have some plants that are way more valuable than tulips in my garden too.
Oh, AI.
It is artificial but it is not intelligent. A LLM (int al) is a marvelous thing. I find sheer joy in conversing with a "gpt-oss-20b F16" that runs on a £600 GPU and a slack handful of CPU and RAM because so little gives so much.
Let's use either $17 billion per launch or $32 b per year. We could compare each launch to training a whole new model, though OpenAI hasn't released official numbers of that, so any comparison would a bit specious. There are some public guesses so I'll reproduce them here and link to my source, but again, they're guesses. GPT-2 is estimated at costing roughly $100,000 to train. OpenAI did say it cost $257/hour of training, but hasn't said how many hours it took to train, and the assumption is that's just the compute cost, that's not fully loaded (eg researcher salary and office rent and etc not included) [1]. The guess is by Karpathy though, so I'd give it a lot of weight in being directionally accurate.
GPT-4 is guessed at costing "more than $100 million" according to sama [2] and each run for a train of something GPT-5 size at $1 billion or more (according to sama)(of which you'd need to make several runs before getting it right, so to speak) [3]. Again though these numbers numbers should include copious amounts of NaCl.
That's just for OpenAI, but we can assume the other labs are using roughly as much. It's a bit different than the $400 billion/year that the linked article we're posting on gives, but it's a differently shaped number.
If the whole of the Apollo program cost $350 billion (2025 dollars) and we're doing $400 b per year, the comparison is an entire 21 launch Apollo space program in a year, or a single Apollo launch per month. Which is still insane, mind you.
The other stat to look at though, because these are unfathomably large numbers, is the how much Apple invested in China, according to Patrick McGee, author of the book Apple in China. according to his book, Apple invested $275 billion into China over 5 years, or $55 billion/year in workers training and infrastructure building. We're comparing years to months here, but Apple is a singular company in the broader tech industry, and their $275 billion investment arguably paid off. Selling iPhone is a rather more proven area than LLMs though, so there's that.
[1] https://x.com/karpathy/status/1811488645175738409?s=46
[2] https://en.wikipedia.org/wiki/GPT-4
[3] https://fortune.com/2025/02/25/what-happened-gpt-5-openai-or...
There might be a slump.
But today's AI companies (both those you've heard of and those you haven't) are here to stay.
It'll be okay, don't panic.
Phase 1) Spend $400 bn / yr less a bit of income
Phase 2) ? - AGI aound 2030 - replace workforce or double output. Worth >1x world GDP or ~$100 tn ?
Phase 3) Profit!
(some discussion in "The case for AGI by 2030" https://80000hours.org/agi/guide/when-will-agi-arrive/)
Fund phase 1 by hyping it to investors
Imagine at the start of the electrification era people went "We'd need to build loads of cables and power plants and stuff that's expensive, lets just stick to steam power".
It's not a bet on this making sense via pedestrian business economics but rather that it'll be a game changer.
...whether that pans out is a technological and societal question, not an economic one in my mind
False dichotomy. There are literally infinite options between ignoring AI and spending a quarter of a trillion on it annually.
The people investing in AI companies (and the big players spending in AI) are seeking Artificial General Intelligence (AGI). It's the only way they get a return on their capital.
They are investing so they can get there first. Money basically becomes meaningless at that point, whoever owns the AGI owns the world. That's the only way to get a return on that investment.
Then why fuck around with ads?
I suspect that this revenue number is a vast underestimation, even today, ignoring the reality of untapped revenue streams like ChatGPT's 800M advertising eyeballs.
1. Google has stated that Gemini is processing 1.3 quadrillion tokens per month. Its hard to convert this into raw revenue; its spread across different models, much of it is likely internal usage, or usage more tied to a workspace subscription rather than per-token API billing. But to give a sense of this scale, this is what that annualized revenue looks like priced at per-token API pricing for their different models, assuming a 50/50 input/output: Gemini 2.5 Flash Lite: ~$9B/year, Gemini 2.5 Flash: ~$22.8B/year, Gemini 2.5 Pro: ~$110B/year.
2. ChatGPT has 800M weekly active users. If 10% of these users are on the paid plan, this is $19.2B/year. Adjust this value depending on what percentage of users you believe pay for ChatGPT. Sam has announced that they're processing 6B API tokens per minute, which, again depending on the model, puts their annualized API revenue between $1B-$31B.
3. Anthropic has directly stated that their annualized revenue, as of August, was $5B [2]. Given their growth, and the success of Claude 4.5, its likely this number is more around $6B-$7B right now.
So, just with these three companies, which are the three biggest involved in infrastructure rollouts, we're likely somewhere in the realm of ~$30B/year? Very fuzzy and hard, but at the very least I think its weird to guess that the number is closer to like $12B. Its possible the article is basing its estimates on numbers from earlier in 2025, but to be frank: If you're not refreshing your knowledge on this stuff every week, you're out of date. Its moving so fast.
[1] https://www.reddit.com/r/Bard/comments/1o3ex1v/gemini_is_pro...
[2] https://www.anthropic.com/news/anthropic-raises-series-f-at-...
OpenAI announced a few months ago that it had finally cracked $1B in monthly revenue (intriguingly, it did so twice, which makes me wonder how much fibbing there is in these statements).
I'll also say this: the fact that AI companies prefer to tout their usage numbers rather than their revenue numbers is a sign that their revenue numbers isn't stellar (especially given that several of the Big Tech companies have stopped reporting AI revenue as separate call-outs).
I believe this is incorrect; as far as I've heard, an anonymous source leaked that OpenAI had hit $12B in annualized revenue a few months ago [1]. I do not personally put any weight in leaks, and prefer to operate on data that has been officially announced.
[1] https://www.reuters.com/business/openai-hits-12-billion-annu...
Yes, I moved from Yahoo to Google in multiple apps. And yet, once I subscribed to ChatGPT, it satisfies what I need, despite all the noise I hear about the alternatives.
I haven't yet begun to absorb https://news.ycombinator.com/item?id=45513234. Presumably the median AI user won't either.
I wouldn't believe it if you told me even 1% of those users are paying. 10% is simply ridiculous.
Respectfully, the idea of sticking ads in LLMs is just copium. It's never going to work.
LLMs' unfixable inclination for hallucinations makes this an infinite lawsuit machine. Either the regulators will tear OpenAI to shreds over it, or the advertisers seeing their trademarks hijacked by scammers will do it in their stead. LLMs just cannot be controlled enough for this idea to make sense, even with RAG.
And if we step away from the idea of putting ads in the LLM response, we're left with "stick a banner ad on chatgpt dot com". The exact same scheme as the Dotcom Bubble. Worked real well that time, I hear. "Stick a banner ad on it" was a shit idea in 2000. It's not going to bail out AI in 2025.
The original content that LLMs paraphrase is itself struggling to support itself on ads. The idea that you can steal all those impressions through a service that is orders and orders of magnitude more expensive and somehow turn a profit on those very same ads is ludicrous.
Also ads only make sense for the use case of direct chatting with a user, any type of automation (the big promise of AI) ads don't matter.
Don't you think that people would have felt similarly about early Google search?
> any type of automation (the big promise of AI) ads don't matter
I would have thought that ads have no place in an OS, but they proved me wrong on that. How sure are you they won't prove you wrong on automation ("summarize this text for me - here you go, shall I run it through Grammarly now?", "build this app for me - [uses freemium tool instead of equally capable FOSS]",...)? Never underestimate the potential for enshittification.
