ah my bad, I’ve switched the visual to A index by cost per task. Note the blog post’s plot is filtered further than AA’s, based on what models are available with inference providers we can actually work with (Europe only). Vibe-coded filtering is here (very rough): https://pareto-eco-sum.netlify.app/
The high usage was due to omp in vibe mode overnight, probably working way too hard through things. The high cost, yes we’d rather pay extra to work with providers that provide other benefits than just lowest cost possible (open models, no training, EU DC, etc)
For what it’s worth the energy numbers I get from Neuralwatt are within the ballpark of what is available elsewhere like https://cleerdash.sustainableaigroup.com/. What’s your source that there is profit / capex in there? Their published methodology seems pretty transparent
I agree but it does worry me how fast my usage is increasing. Two months ago I was using 10x less tokens and probably not much more than 5kWh on inference. This month about 30kWh on inference. If it becomes more affordable, is there going to be another jump? Not quite sure
To put that in perspective my house uses an average of 1300kWh per month. Adding 30kWh per month would be an increase of 2.3% at a cost of around $4.20, not enough to notice.
Do you live in the US? You should know in that case that the US is already one of the most wasteful energy users on the planet, far ahead of most other developed countries, even.
So potentially increasing energy usage even more, especially with likely token usage acceleration, is hardly reason for celebration.
No one is celebrating and using LLMs isn't necessarily a waste of energy. Where I live isn't relevant, LLMs use the same amount of power regardless of where their users are. You also know nothing about how I use electricity in my home, how I monitor my overall energy use and carbon output, and how I offset it.
> You also know nothing about how I use electricity in my home
I don't
> Where I live isn't relevant,
But where you live does matter. Average monthly household energy usage in the EU is 200-400 kWh. In the US it's more like 800-900kWh. So what you consider "normal energy usage" may vary considerably by where you live.
At 200-400kWh total usage, 30kwh starts to look a lot more significant. So you should at least consider the possibility that it's not that energy usage from AI is insignificant, but that your existing energy usage is wasteful.
You're American and the same 2.3% of energy you use would probably represents 5-10% of what the average Japanese person uses, and likely 15-20% of what the average Indian or Ethiopian uses.
My point isn't to attack you personally (I'd rather attack your country's disastrous energy policies) but to point out that LLMs will constitute a huge increase in our energy consumption at a time where we haven't transitioned to renewable energy and we're probably 20-30 away from it.
for me it’s more an environmental consideration than financial? Passivhaus Standard for houses is about 50kWh/m²/year, so 30kWh is getting close to 10% of total
It's a very very very rough approximation, but scaling laws (admittedly weaking over time) yield ~x^2 from a size, ~x^2 from a dynamic power, and more of the data movement is moving from PCB interconnect to on package which looks more similar to chip scaling.
It's a finger in the wind approximation with lots of confounders, but historically it seems to work out that way for most circuit things.
At some point the agents will be good enough that you can tell them "here's $100, go make me money" and they will, maybe not a lot and not all the time, but the EV given the cost of inference will be positive.
Then why would any AI company exist when they could just use their money to buy tokens from another AI company and make money for zero effort? There would be no incentive to be a provider.
Not to mention inflation would grow to match or outpace the rate you could earn on these guaranteed AI gains.
This seems somewhat obviously true, to some degree, but it all depends on the prompt, I think?
In the equilibrium, profits fall to zero in capitalism, but there's never equilibrium, everything is changing. Profit comes from insights that others haven't seen, from access to opportunities that others don't have, or through monopolistic control generating economic rents.
For that AI to make money, it has to have a harness that gives it one of the above things, which does seem possible.
But strictly time-limited, because as soon as someone else figures out what you are doing, there's no moat to prevent them from telling their AI to do it too.
Markets reach equilibrium very fast when everyone has access to the exact same tools and information.
Those rewards seem like they would be captured by the providers and AI companies though who would use them first to make themselves money. You would be left with whatever they didn’t pursue with their first movers advantage.
I guess the question boils down to, what remote work could a human do to earn money over the internet?
So how many companies/individuals pay a random spam bot contacting them 100€ to upgrade their website? I guess start with targeting gullible and trying to scam them will be the more successful experiment for them. (If run unrestricted)
Hmm I might need to rephrase. Initial challenge was to use GLM 5.3 Flash and I was on the non-Flash version for the whole vibe coded build. Just wasn’t paying attention and I didn’t realize that one session was a quarter of the month’s spend and 150% of the budget (big price difference between models)
It was a bit of a silly challenge, wasn’t sure how workable, learned a lot in the process about what actually drives usage / costs, and how to keep both under control
Not OP, but I've used both quite a bit and I think GLM 5.3 (non-flash) is a vastly better model. The flash variant is good for general workhorse agents, but it doesn't seem to reason holistically about code and over-engineers solutions to each specific problem it solves. But if you use GLM 5.3 to write a detailed plan with little to no ambiguity, GLM 5.3 Flash executes it just fine for a fraction of the price.
I have a hard time justifying GLM 5.3 these days. It’s slightly better than Flash but rarely enough to justify the much steeper price. We chose to use usage-based billing only so are very sensitive to model price.
When text wuality or for pure but adwansed coding is concerned i always pick glm 5.3. The flash is awesome for everything that dosent really matter though.
Ill tell you from my personal use glm 5.3 flash was better then deepseek 4.1 flash, also deepseek liked to yap in his reasoning traces soo fucking much, the yapping was fast but the task was so slow to complete...
Interesting, I have roughly the opposite experience regarding speed: DeepSeek V4.1 Flash gets stuff done way more quickly for me than GLM 5.3 Flash. (I was getting 300+ tokens/sec with DS vs ~100 with GLM in my testing.) I agree that GLM 5.3 Flash is a slightly better model, but for me at least, it's not a huge difference, and I'd rather have DeepSeek's speed.
I've only compared it in a few tool assisted design tasks, and deepseek performed significantly better (and faster). I haven't had them do enough coding to get a feel for that, though.
