LLM providers on the cusp of an 'extinction' phase as capex realities bite
theregister.com
theregister.com
I've read somewhere that generating a single AI image draws as much power as a full smartphone charge.
In case the suspicion is true that costs are too high to be monetized, then the current scale-up phase is going to be interesting. Right now people infrequently have a chat with AI. That's quite a different scenario from having it integrated across every stack and it constantly being used in the background, by billions of people.
Late as they may be, for the consumer space I think Apple is clever to push as much as possible to the local device.
Cellphone battery charge: I have a 5000mAh cellphone battery. If we ignore charging losses (pretty low normally, but not sure at 67W fast charging)... That battery stores about 18.5 watt-hours of energy, or about 67 kilojoules.
Generating a single image at 1024x1024 resolution with Stable Diffusion on my PC takes somewhere under a minute at a maximum power draw under 500W. Lets cap that at 500*60 = 30 kilojoules.
So it seems plausible that for cellphones with smaller batteries, and/or using intense image generation settings, there could be overlap! For typical cases, I think that you could get multiple (but low single digit) of AI generated images for the power cost of a cellphone charge, maybe a bit better at scale.
So in other words, maybe "technically incorrect" but not a bad approximation to communicate power use in terms most people would understand. I've heard worse!
That's insane, holy shit. That's not even a very large image.
Apparently I was off on my estimates about how power hungry gpus are these days by an order of magnitude.
Right, but there was a point at which we could stop people from doing stupid shit because it's useless and they're bad with money. Now it seems we've embraced irrational and misanthropic spending as a core service.
We honestly just need to take money away from people who obviously have no clue what to do with it. Using AI seems like a perfect signal for people who have lost touch with an understanding of value.
1920x1080 is still, by far, the dominate desktop and laptop resolution in 2025.
Most of the energy cost is how many images I have to generate to get a satisfactory one (say 100 with a decent prompt). Looking at the broader picture, the biggest energy cost of all is hiring a human designer for layout / typography and to produce print-ready files. Then managing the manufacturer, haha.
A bit off topic, but hopefully something that will brighten your day: I make physical products, so they have to be perfect (it also means I deal in DPI, not pixels). AI speeds up the number of concepts I can generate and send to contractors. I don't try to copy anyone's specific style, that would be boring. I'm sure there are less wholesome uses of this AI thing, but I feel like I've stumbled into something I'm comfortable with.
Believe it or not, I use all this to sell antiques. Like, genuine physical artifacts made by humans hundreds of years ago. I like to tell the story of the era they are from, bring it to life a little with printed supplements. No LLMs for the writing though, I do the research and writing myself, I enjoy it too much. I don't make much money with it, but it's fun, and a way to tell my country's history.
You can generate a lot of images with the energy you would use to play a game instead for two hours; generating an image for 30 seconds uses the same amount of energy as playing a game on the same GPU for 30 seconds.
https://cloud.google.com/blog/products/compute/accelerating-...
Plus, not all these models run on optimized TPUs, but mostly on nVIDIA cards. None of them are that efficient.
Otherwise I can argue that running these models are essentially free since my camera can do face recognition and tracking at 30fps w/o a noticeable power draw since it uses a dedicated, purpose built DSP for that stuff.
Again, cell phones are just confusingly not energy intensive.
That's not a very big image, though. Maybe if this were 25 years ago
You should at least be generating 1920x1080, pretend you're making desktop backgrounds from 10 years ago
it would be insightful for competitors too, because they could use this as part of their analysis and price strategies against you.
Therefore, no company would possibly allow such data to be revealed.
And in any case, if these LLM providers burn cash to provide a service to you, then you ought to take maximal advantage of it. Just like how uber subsidized rides.
For whom would this be beneficial? The design goals of these products are to get as many users as fast as possible, using it for as long as possible. "Don't make me think" is the #1 UX principle at work here. You wouldn't expect a gas pump terminal to tut-tut about your carbon emissions.
To put that in perspective, using the 67 kJ of energy for a smartphone charge given in Saigonautica's comment you can charge a smartphone 336 times for $1 if you are paying the average US residential electricity rate of just under $0.16/kWh.
You could charge a smartphone 128 times for $1 if you were in the state with the most expensive electricity (Hawaii) and paying the average rate there of around $0.42.
Saigonautica's battery is on the large size. It's a little bigger than the battery of an iPhone 16 Pro Max. A plain iPhone 16 could be charged 470 times for $1 at average US residential electricity prices.
For most people energy used to charge a smartphone is in the "this is too small to ever care about" category.
We can do a similar calculation for AA rechargeable batteries, and the results might be surprising.
$1 of electricity at the US average residential rate is enough to recharge an AA Eneloop nearly 2300 times. Of course there are inefficiencies in the charger and charging, but if we can get even 75% efficiency that's good enough for more then 1700 charges.
That really surprised me when I first learned it. I knew it wasn't going to be a lot...but 1700 charges is I think more than the number of times I'll swap out an AA battery over my entire lifetime. I hadn't expected that all my AA battery use for my whole life would be less than $1 worth of electricity.
Maybe they didn't fall behind in anything, maybe they just did an analysis of what it would cost to train transformer models with hundreds of billions of parameters, to run inferencing on them, and then decided that there was no way to actually be profitable doing this.
People have mentioned on hacker news that there seems to kind of "weather patterns" with how hard the various llms think, like during business hours they get stupid. But of course there is some disagreement about what "business hours" are. It's one of those "vibes".
Imagine scheduling your life around the moods of AIs.
That's the business model. If you don't want a surly and moody AI with a hangover and bad attitude, you gotta pay more!
Like isitdownrightnow.com for crowd sourcing web site availability, there should be a isitdumbrightnow.ai site!
I’m having a hard time squaring the number $644 billion and the phrase “extinction phase.”
I don’t believe their actual estimate of GenAI spending but if it’s even in the same ballpark as the real value, that’s not an extinction.
Pushing towards a trillion bucks a year for what LLMs are mostly currently used for does not seem like a sustainable system.
Of the "real" categories, they expect: Service 27bn (+162% y/y) Software 37bn (+93% y/y) Servers 180bn (+33% y/y) for a total of $245bn (+58% y/y)
That's not shabby numbers, but way more reasonable. Hyperscaler total capex [2] is expected to be around $330bn in 2025 (up +32% y/y) so that'll most likely include a good chunk of the server spend.
[1] https://www.gartner.com/en/newsroom/press-releases/2025-03-3...
[2] https://www.marvin-labs.com/blog/deepseek-impact-of-high-qua...
Altman loudly hyping "look you can ghibli-fy yourself", stating inflammatory things like "we are the death of the graphic designer"; a desparate ploy to rapidly consume the market before the bubble bursts.
I don't think OpenAI has that option.
The business model was you could sell books over the internet at a much cheaper cost compared to Barnes and Noble or Borders because they weren't paying for physical locations and there was no sales tax on the transactions because it was on the internet.
A key difference from OpenAI is that Amazon was cash flow positive from very early on and before that first profitable quarter. They only needed one funding round Series A of $8 million instead of repeatedly trying to raise extra rounds of funding from new VC investors.
The Amazon startup already had enough free cash from operations to internally fund their warehouse expansions. The "Amazon had no profits" was an accounting side-effect because of re-investment. Anybody seriously studying Amazon's financial statements in the late 1990s would have paid more attention to their cash flow rather than "accounting profits".
On the other hand, OpenAI doesn't have the same positive cash flow situation as early Amazon. They are truly burning more money than they take in. They have to get billions from new investors to buy GPUs and pay salaries. ($40 billion raised in latest investment round.) They are cash flow negative. The cash flow from ChatGPT subscription fees is not enough to internally fund their growth.
ChatGPT was released in Nov 2022. They are expecting $12B in revenue this year[2].
[1] https://adainsights.com/blog/when-did-amazon-start-making-mo...
[2] https://www.cnbc.com/2025/03/26/openai-expects-revenue-will-...
Thank you for the correction. My memory was faulty and Jeff Bezos actually said, "we always had positive gross margins". Deep link: https://www.youtube.com/watch?v=zN1PyNwjHpc&t=36m11s
The positive gross margins allowed enough discretionary use of cash to take out loans and service that debt. I just looked at the 1998 10k filing and the page on "Consolidated Statements of Cash Flows" has "Net cash provided by (used in) operating activities" of positive $31 million compared to negative -$6 million in 1996.
To invest in OpenAI is to bet that they won't need to keep investing more in hardware than they bring in. To me that doesn't seem a sure thing, but it isn't obviously wrong either. They are growing revenue very quickly and already have significant cashflow, and it isn't clear to me that they'll need to sustain the CapEx forever.
Amazon out-priced everybody when it arrived, because it didn't charge any sales tax for years, until the laws had to be re-written to close the loophole. It didn't have the eye-watering sums poured into it that AI has had, nor did it have any significant competition internationally. Things couldn't be more different for OpenAI.
The more interesting question to me is how gpu vs tpu plays out. Plus the other npu like approaches. Sambanova cerebras groq etc
Yes, there are differences between the models, and yes some may work better.
But picking the model at this point is just picking the cheapest option. For most use cases any model will do.
Isn't that where the cost lies? Data, annotation, and model generation all have mostly linear responses to changes in spending.
> For most use cases any model will do.
They'll operate. They will not produce reliable results. Adoption is one metric, but intentional avoidance should be another.
Models are still leapfrogging each other every month in e.g. coding or research capability, or even in more mundane tasks such as summerizing long multi topic texts.
Depending on which side of the issue you fall, you're hoping this will go on for a long time to come, or praying that it will end asap.
I'm not using the cheapest in neither my own support, nor in my production systems.
If the choice is between something that costs $10 a month or $20 a month, and both solve those use cases, it's rational to pick the cheap one.
They all seems to racing to the plateau... It doesn't look like there will ever be a "stand out" leader and the product that each company presents to the market appears to be essentially the same product that everyone else presents to the market. Maybe with some slight twist to it that is easily recruitable or exceedable within a few months.
This is the issue really. at some point the investors are all going to realize that non of their investments are going to be market leaders. When they get to that stage the bubble will well and truly pop.
To me it feels there is no plateau and the models are already very useful and impactful.
I believe there is no plateau because there is nothing objectively special or magical about the human mind and it all can and will be eventually solved, one hack at a time.
Claude 3.7 is a great example of a model clearly beating 3.5 in all benchmarks, but slowly destroying my code base by adding lots of extra lines or hacking around my instructions (adding ,,if'' statements when I want it to change the code to handle a case instead of understanding what change is really needed to be done for it).
I still prefer o1 pro and a lot of those leapfrogging in benchmarks don't translate to being smarter anymore.
Being close to the edge of AI usage, it's important to realize that most AI use cases are not "fully autonomous AI software engineer" or "deep research into a niche topic" but way more innocuous: Improve my blog post, what's the capital of France, what are some nice tourist sites to see around my next vacation destination.
For those non-edge use cases, costs are an issue, but so are inertia and switching costs. A big reason OpenAI and ChatGPT are so huge is that it's still their go-to model for all of these non-edge use cases as it's well known, well adopted, and quite frankly very efficiently priced.