There’s long been a disconnect between concerns about the massive impact of AI data centres on electricity demand and claims by Sam Altman and others that the impact is really modest. Ed Zitron recently posted a summary of OpenAI’s 2025 accounts which helps to clarify things a bit.
In short, if you look at actual electricity demand needed for current AI use, it’s small. And that doesn’t change if demand grows at high but plausible rates. On the other hand, if you look at what is needed to justify the current valuations of AI and its competitors, the implied growth is staggering.
Starting with demand, the OpenAI accounts break out current revenue (what people are paying to use ChatGPT etc) and the cost of revenue (the costs of the data centres used to answer queries)
The cost of revenue is about $8 billion (rounded for simplicity). A conveniently scaled measure of electricity output is a Terawatt hours (TWh). Roughly equal to the annual output of a typical 1GW coal-fired or nuclear power station operating 24/7 is about 10 TW
If 25 per cent of that is electricity, and electricity costs 10 cents/kWh, the implied production electricity use is:
(0.25× 8 * 10^9) / (0.10* 10^9) =20 TWh
or the output of two large power plants running full time (I’ve previously derived the same estimates from reported cost per token)
That is about 0.5 per cent of U.S. electricity demand. Doubling this to cover the broader U.S. generative-AI industry gives a convenient round estimate of 1 per cent of U.S. electricity demand in 2025.
The US Energy Information Authority estimates demand growth of 15 per cent year, which implies doubling over five years
So the U.S. AI sector would rise from about 1 per cent to about 2 per cent of U.S. electricity demand by 2030.
It’s important to remember that this is growth in electricity demand. If AI becomes more efficient in converting electricity into tokens and tokens into useful answers, say doubling every two years, the number of answers would grow tenfold.
A new load equal to 1 per cent of current US demand would be quite manageable in aggregate, requiring the output of perhaps four of five 1GW power stations. The problems would be of the kind we are already seeing – data centers clustered into a few locations and pricing structures where the costs are shifted onto existing consumers. These are fixable problems.
The problem is that this is nowhere enough to make OpenAI and others sustainably profitable, let alone to justify the market valuations we are seeing. It’s harder to analyse these valuations than to measure current demand, but I’ll give it a go.
Assuming a price earnings ratio of 30, a trillion dollar valuation requires profits of around $30 billion a year. Assuming that profit margins can be maintained at 30 per cent, that requires revenue of $100 billion a year, and costs of $70 billion a year, around 10 times the current level for OpenAI alone. Taking account of Anthropic and the AI-boosted valuations of Meta, Alphabet (Google) and Space X, it would be more like 50 times . And, indeed, these numbers are conservative compared to the projections were seeing.
That gives growth equal to 50 per cent of current US demand, which is utterly impossible. Even with data centers spread across the developed world the growth in demand would be unfeasible and any attempt to deliver it would be catastrophic.
I’m going to call BS on this. There is simply no way that demand can grow enough.
OpenAI already has 900 million weekly active users, so there’s not much room for growth on the extensive margin. As for growth in revenue per users, I’m going to appeal to a combination of introspection and casual observation.
As regards introspection, I’m a heavy user mainly for enhanced search and testing ideas. I’m using OpenAi’s $20/month plan, as well as playing with other options like Mistral and Copilot. If OpenAI doubled their price, I’d switch to one of the others. If they all raised prices to (say) $100/month, I’d buy a high-powered computer with Nvidia chips and switch to local hosting.
Most people I know are not massively online, and use the free version of ChatGPT. They are unlikely to pay for it. A lot will be satisfied with Google’s (awful) Gemini which is “free”, that is, funded by advertising. Google already dominates the advertising market, so there’s no potential for growth there.
That leaves business uses. The short-lived fad for “tokenmaxxing” suggested the possibility of huge growth in demand. However, the disappointing results have brought that back to ground. Both at the firm level and at the aggregate economy level, there’s no indication that AI is delivering the productivity gains required to justify massive new expenditure.
Still, this yawning gap helps to explain the ferocity of the debate over AI and LLMs. I think stoclk market valuations are crazy, and don’t treat them as reflecting reality. I follow the EIA and expect that the hyperscaling boom will go the way of tokenmaxxing. On this view, electricity use for LLMs will remain modest.
Those most hostile to AI assume that the hyperscalers will deliver on their promises at least as far as making massive investments in data centres. That gives you the disaster scenario.
On the other hand, many critics point to the huge volume of useless slop produced by AI. These two criticisms contradict each other. There’s a partial reconciliation with Sturgeon’s Law (90 per cent of science fiction is crud. But 90 per cent of everything is crud).
Both viewpoints lead to the conclusion that we should not have rapid growth in data centres. The disaster scenario suggests a need for urgent action. The sceptical version suggests that building power stations to support data centres will rapidly prove non-viable, leaving a legacy of stranded assets, mostly gas-fired power plants that will be an obstacle to a clean energy transition
For the moment, the hype about AI and data centers has not produced a comparable boom in power plant construction. Indeed some projects have been canceled or delayed. Others like the Trump-branded Fermi project appear to be nothing more than vaporware.
From a range of social, economic and environmental viewpoints, the best thing to do regarding AI would be to abandon massive expenditure on hyperscaling and focus on learning how to use and manage AI to yield the best ratio of benefits to cost.
{ 26 comments }
Thomas Gordon Hewitt 06.24.26 at 4:11 pm
The big use case isn’t chat, it’s what has come to be called ‘agentic’. Ai processes are automatically spawned, and these can span other processes. So very complicated processes -such as “please create run and debug some software that does X, Y, and Z”. These can sometimes run for days before producing a ‘finished’ product. So the demand is essentially unlimited, depending only upon the user’s budget.
Now I think progress in efficiency will be faster than you expect. Less expensive alogorithms might cut the amount of required work per token by an order of magnitude or more. This stacks upon whatever gains can be made on the hardware side. Never the less, damands from our new AI masters could well grow exponentially.
Bob 06.24.26 at 9:21 pm
Thanks for this John. The other concern with AI-related data centres is their consumption of water for cooling. Perhaps that concern is also overstated.
DMSeattle 06.24.26 at 10:33 pm
All I know is that I use Gemini a lot and it does incredibly useful work. And if I was forced to I would even pay for it!
So I’m very bullish on AI. Will there be problems? Duh of course there will be problems. Does that mean that in some investors are over enthusiastic? Duh. It’s capitalism so of course there will be irrational enthusiasm. But something’s happening Mr. Jones.
The first thing we should be doing is not using Public money to invest in private investment. Let the people who will be making the money take the risk. In that case that’s pretty much end of the discussion.
somebody who knows what ai is for 06.24.26 at 11:38 pm
a foolish notion. you must understand that the poor have to die of heatstroke so that elon can make his anime girlfriend have bigger breasts. if you don’t support this outcome enough, the entire world economy will permanently collapse.
D. S. Battistoli 06.25.26 at 4:19 am
John, thank you for kicking off a useful conversation, and grounding it in hard numbers.
While I would like to come back to this in order to consider total effects, I think there are some trends that need to be highlighted:
Your current model keys off American energy consumption. The US, which houses roughly 4% of global population, accounts for roughly 15% of global electricity consumption. (Recall the old saw that we’d need 4.5 earths for energy and resource demand to be sustainable if everyone around the world used energy as profigately as do Americans). So when we talk about any sector driving 1-2% increases in American energy usage, we are talking about (a) energy usage that is already unsustainable given global population distributions and (b) energy demands that would represent a higher proportion of total generated output in almost all other parts of the world, with the differential smallest in the most energy-intensive economies (the Global North) and largest in the most energy-efficient ones (clustered in the Global South).
AI token usage has been demonstrated thusfar to follow the Jevons Paradox: declining (token) unit costs lead to more-than offsetting increasing (token) usage, driving increasing total costs (and, to our point, increasing energy consumption). Now, there is very little data on which to go here (mainly OpenAI reporting) and the length of the reporting window is short, but there is a solid chance that AI, if it is not subject to a major revaluation (a big if) may continue to increase its energy demand regardless of efficiency gains.
Related to, and in part driving the above trend, AI token usage growth is not primarily driven by chat trends now: it is driven by agentic AI, which is fantastically more token-hungry than is chat: 50-500x so on a per-task basis according to a February 2026 estimate.
Now, let’s pivot to the Global South.
We can start by looking at Ghana, which in the last 15 years, managed to nearly eradicate dumsor (brownout) rates due to massive generation- and transmission-capacity improvements that were closely tied to rising levels of government indebtedness that led to government default on loans and massive macroeconomic disturbances. And this is before people started trying to build AI data centers in Africa.
Most of the continent is not even as fortunate as Ghana, a middle-income country. There is a massive debate surrounding such initiatives as the Sustainable Development Goals around electricity generation and distribution and the World Bank’s Power Africa initiative, which really start to bite in lower income countries. Even before taking into account rising energy demand from data centers, there is a massive trade-off issue being hotly debated: Should new electricity capacity be directed to currently unconnected consumers, the macroeconomic benefits of which connection are apt to be modest-to-negative even as the human-rights implications are positive? Or should new power capacity be directed to industry, which is likely to help countries on a macroeconomic level, while also potentially increasing popular unrest (ongoing brownouts faced by poor- and middle-income households despite growing generation capacity were among the factors in the “Gen Z” protests that brought about the fall of the Madagascar government in 2025). And again, this is before taking into account the rising power-generation requirements of AI, which in Africa are of magnitudes far greater than 1-2% of current generational capacity.
Now, a counterargument might be that Africa shouldn’t build data centers, but instead buy tokens from offshore ones. This argument, which at one level might resolve the data-usage problem on a continental scale, would tend to aggravate balance-of-payment crises across the continent. (One Afro-optimist position, articulated by Joe Studwell, has Africa becoming a net creditor to the world for agentic task completion, assuming that tasks be completed by humans, and looking at continental-level population trends globally. If we assume that a significant set of task activities are assigned to AI agents these agents will be cost-domiciled in the Global North, the Studwell payment-rebalancing assumptions are put under stress).
I have not as yet quantified these concerns, though it should be near-trivial to do so in a future step. For now, my goal is to underline the importance of separating out token demand in agentic and chat use-cases and not using current levels of American macroeconomic energy production and consumption as a denominator in an equation that looks to calculate the presentist and future sustainability of AI energy demand, either in a “skeptical” or in a “disaster” model of future cost and energy-consumption trends.
Alex SL 06.25.26 at 5:24 am
i agree with your analysis of the likely demand ~ price curve at the end, but I am a bit confused about the electricity use. You start calculating it as 25% of the 8 billion cost of revenue.
First, what else would that cost of revenue substantially be if not the electricity used for inference plus a bit extra so that the people supplying the compute can keep the compute operation running, pay back their loans, and make a bit of a profit at the end? Therefore, shouldn’t the percentage of cost of revenue that is electricity cost be well over 50%? I just don’t really see what other large items go in there.
Second, is OpenAI honest about its inference costs? I have no way of judging that myself, but based on the bizarrely large ad budget and the observation that some Chinese companies do the same, Zitron has suggested that OpenAI may have hidden a good part of their inference costs for subsidised/gifted/trial queries in the ad budget.
Third, what about the electricity demand for training? To my understanding, that wouldn’t be in cost of revenue but in R&D, although it really will have to continue indefinitely to avoid model drift.
Fourth, just on a purely anecdotal basis, if I run an LLM on my laptop through Ollama or AnythingLLM, its fan achieves levels of rotation that it never does doing any of the analyses I do for work, or even running a Teams meeting or suchlike. So, this surely looks to me as if running an LLM eats a lot more electricity than any normal computer operations.
So, to this point, I am already quite certain that 25% of 8 billion times two is probably a lower floor of the actual electricity demand of generative AI.
But fifth, one to two percent of the electricity demand of a nation seems enormous to me. Leaving aside that every percent matters if we want to avoid increasingly catastrophic levels of climate change over the next few decades, this isn’t only about amounts of electricity in isolation, it is about what amount of electricity for what purpose. If I were told we need another 1-2% of power to supply people with air conditioned home, with infrastructure for commuting, with health care, I would say, sure, that is an investment worth making. But we are only talking about running computers here. And in terms of economic outcomes, what are these 1-2% of power actually used for?
Gork, is that true? Spam. Fraud. Plagiarism washed through a bot, so everybody seems to think that makes it okay. A bunch of very generic-looking, often flawed images used for cheap advertising, to make artists unemployed. Weird videos and shrimp Jesus images for clicks. CSAM generation. Summarise this document so I don’t have to do my job (as an expert hired and trusted for my expertise). Draft a text for me so I don’t have to do my job (ditto). “Do research” (University of Google but now with a chatbot optimised to flatter the user). Save some time coding. De-skill office workers, and keep students from learning critical thinking and how to do research themselves.
To put it kindly, that is not worth the investment. For that outcome, 1-2% of a nation’s electricity is a deeply offensive waste, a bleeding wound to the mere concept of useful economic activity.
John Q 06.25.26 at 6:28 am
TGH @1 Agentic use is the hyperscaling case I describe. The assumption is that the agents will do such marvellous things that users (mostly businesses) will pay incredibly large amounts to run them. That seemed plausible during the tokenmaxxing fad, much less so now. The big problem is that all AIs are fallible, as are people. But with LLMs and humans you can check as they go. With agents, the errors just propagate
Bob @2. Water use grows in line with (steam-based) electricity. Small currently, large if the hyperscalers are right.
DSB @5 Thanks for useful contributions. I will think more about this
Alex:
1.The other big cost of compute is the Nvidia chips used for it, which don’t last that long. Then there’s the physical costs of the data centre.
2. You can add in the marketing budget shown in Ed’s accounts. That turns a small profit into a small loss, but it doesn’t change the electricity calculation
3. That’s the whole point of the post. The training budget is for hyperscaling. Updating existing LLMs with new content just requires a web crawl similar to what Google does continously
4. Even running at maximum, a modern laptop uses energy comparable to an incandescent lightbulb, which also generates plenty of heat. So, this supports the point that you can do current AI with very little energy
5. As you present it, this is an argument against computers of all kinds, not specific to AI. The Internet is a bunch of silly rants and cat videos, funding by meretricious advertising. Calculators do sums for you, instead of you doing your job with pencil and paper – kids don’t learn. And of course pencil and paper take the place of memorisation which is the only true learning. I’ll leave the airconditioning argument alone – a whole different can of worms
Alex SL 06.25.26 at 9:13 am
Okay, that makes sense. It is a different form of resource wastage, but not electricity except to the degree that electricity is needed to produce the GPUs.
Agreed that it isn’t a multiplication, but I was merely arguing that the real number is likely higher than your first estimate.
Not sure what hyperscaling means in this context, I must admit. Hyper kind of means more than usual, and scaling means making things larger, so “even larger models”? But be that is may, that is emphatically not how this works. If a trained model does an ‘agentic’ web crawl, the results only end up in the context window of the query and are gone when a new session is started. The only ways of “updating existing LLMs” are starting over from scratch with the entire training process or doing ‘posttraining’. But the latter isn’t updating information either (e.g., this celebrity is now dead), it is merely to fiddle with the behaviour of the model (e.g., avoid giving this kind of response, that is too woke/that is too offensive/that would be a security risk). That leaves starting over, which is why the AI companies constantly do that. Thus, training costs are cost of revenue, the AI companies just find it more convenient to pretend otherwise.
I was merely making the point that running LLMs is expensive compared to other computational tasks. That’s what that first L means, in one sense. But also yes, I can run an LLM on my work laptop (my private one has insufficient memory), if I am happy to use a compressed model that is effectively useless compared to the ones that the major companies provide on their data centers via API, and if I am happy to wait 5-10 min for that effectively useless response. That is okay for me to play around with just to test if indeed I can do something useful on device (the answer always being “no”), but it will not be acceptable to Joe McAverage who wants this thing to summarise meeting notes or suchlike.
On the one hand, yes. If you gave me a button that deleted all of Facebook, all of Twitter, and the 95% worst videos on Youtube, I wouldn’t hesitate for a second, and nothing of value would be lost to humanity. But on the other hand, the disagreement between us here is that I see a lot more value in emails, web search, social media, and virtual meetings than I do in the unreliable plagiarising chatbot. Whenever I try to use LLMs for anything, they fail me, and a lot of what they work okay-ish for had much more efficient non-LLM solutions already, be it search, translation, or voice to text. Also, and that is the crux of this post, I see a lot more value per electricity used in emails, web search, social media, and virtual meetings than in the chatbot, outside perhaps of assisted coding. (Which worked 1.5 times of the three times I tried it, and I am charitably counting the time I had to go through ca. six iterations until the script finally worked as a half-success instead of five failures and one success).
nonrenormalizable 06.25.26 at 9:25 am
Looking slightly broader than AI, there has been a big jump in data centre construction and energy usage over the past few years.
Via https://www.idc-a.org/insights/Sk2PFZCost98Eqhk2FUA, from May 2026:
For context, 1GW is slightly more than is needed to power one million American homes.
You could argue that the IDCA, which produced this report, has some of the same vested interests in maintaining the hype about growth in this sector.
I think you’re right to question how much further this expansion can continue on to the future given realistic limits on upgrades to electricity grids etc.
But the growth to date, some (not insignificant) fraction of which is down to AI usage, is significant and will have significant effects, e.g. https://www.theguardian.com/technology/2026/apr/24/officials-hugely-underestimated-impact-of-ai-datacentres-on-uk-carbon-emissions
marcel proust 06.25.26 at 12:36 pm
@JohnQ: But with LLMs and humans you can check as they go. With agents, the errors just propagate
Scott Cunningham is an empirical microeconomist who used to blog (substack?) primarily about the intellectual history of the credibility revolution and natural experiments in economics. In the last 8-10 months perhaps the majority of his posts have been about his experience with various versions of Claude. He is a very optimistic bleeding edge sort, but also very much aware and open about the problems he has run into and how he (tries) to resolve them. He has discussed what he uses Claude for, the benefits and to some extent the costs, and much more, his successes and failures with Claude, and what he has tried to minimize both the seriousness and frequency of the failures.
ManySome of his posts, esp. about Claude, are not Substack paywalled. (I am not a paying reader and I know that I have read a number of the posts that are now paywalled. I think he must have gone back and changed the settings for a number of them. Sorry)Two of his most interesting have been:
1) about a presentation at the Kennedy School this spring where he took an article that he was confident his audience was very familiar with, and had Claude analyze, replicate and I believe extend it, from start to finish during the talk. Then, to finish his talk, he presented the slide deck of its own analysis that Claude had created.
2) Having 300 instances of Claude create analyses of the minimum wage, and giving each instance a slightly different prompt (which set of previous analyses to build on). The results were very interesting, almost chaotic in the butterfly wings sense of the word.
None of this addresses the economic issues/costs that the OP raises, but it does address the usefulness of the technology. I am a few years older than JohnQ and retired, so I expect never to (have to) master Claude (or other AI) any more seriously than I have google search or wikipedia over the last quarter century. However my children and their spouses, who are in early middle age, nearly 20 years out of college, have been fully incorporating it into their work lives and at least 2 of them, one in the private sector the other in the public sector, are astounded with how much it has allowed them to scale up their work. It is not as though they give a command and 20 minutes later, they have something useful; rather, 3-4 hours later, with a reasonable amount of back and forth as Claude returns with revised results from the most recent command, they have something useful. And during that 3-4 hours, they are having to review, and tell Claude how to revise, “Claude’s” work for only several minutes every 20 minutes. Thus over 3-4 hours, to the extent they can multi-task, they can do so incredibly productively
Kenny Easwaran 06.25.26 at 4:48 pm
I’m surprised that you’re so credulous about the claims that tokenmaxxing was a “fad” and therefore agentic coding use won’t grow. It’s true that there was a temporary fad of companies (notable Meta) rating workers on the basis of how many tokens they used, which was broken when they realized how many hundreds of millions of dollars they were spending on these tokens. But even as companies have pulled back on that mis-alignment of incentives, Anthropic’s revenue seems to be growing quickly, and the people I know who are using these models are doing more with them (and more people are doing so).
Alex SL 06.25.26 at 6:26 pm
Darn, I keep forgetting what happens with the formatting if I begin paragraphs with numbers. Sorry about that. Those were meant to reflect the five numbers of #7.
marcel proust,
Whenever I read how using Claude or some other LLM revolutionises somebody’s work or massively increases their efficiency, I wish there would be a bit more detail. Because I struggle to understand what it could do for me apart from either committing plagiarism (e.g., having it write part of my manuscripts, reports, or grant proposals and then putting my author name onto them), having it write a Python script once in a month at most, or summarise texts that I would rather read myself because I don’t trust its judgement, because it doesn’t have any. Those are uses but hardly scaling up anything.
What I see among colleagues who have gone in on genAI is mostly that third use case (e.g., summarising a meeting recording instead of simply jotting a few notes down while in the meeting), or generating images for conference talks or events. It does not seem to increase their efficiency, as far as I can tell, but it certainly makes me worry whether I can trust their judgement.
Garrett Wollman 06.25.26 at 8:29 pm
It is inescapable that hyperscale data centers are being built today where the largest subsidies, cheapest (almost always highly regulated) electricity and least burdensome environmental regulations are, which for the US means mostly the South. And even with that, projects are getting delayed or canceled simply due to the inability to deliver power to those sites, because utilities and regional transmission organizations have 5-to-10-year pipelines of projects; they can’t build transmission infrastructure at the speed that a building contractor can throw up a single-story steel-framed building with no interior partition walls. And so we see data-center developers building their own on-site power plants, exclusively fossil-fueled, if the turbines are even available. (GE Vernova has a huge order backlog as well, fueling its own insane stock price.)
My employer is part of a multi-university consortium which operates a data center in western Massachusetts, where power is not cheap (although land in deindustrialized mill towns is). I just looked up the tariff, and the data center — which is connected to the transmission network on the generator side of a major hydro plant — is 11.5 cents per kWh and $4.69 per kW (peak half-hour average demand). The average demand for the whole facility is over 6.5 MW, and over a 24-hour period it used a bit under 160 MWh. (At least the peak-to-average ratio is fairly close!) That’s a $55m monthly power bill, for a facility that’s used solely for research computing, has no outside commercial customers, isn’t fully built out, and isn’t running flat out with ML training tasks 24×7.
(This is still cheaper than renting the hardware from Amazon, Google, Microsoft, or Oracle, even accounting for capital costs.)
Thomas Gordon Hewitt 06.25.26 at 11:46 pm
I see huge potential upsides to the technology. I’ve been trying my (retired) hand at math education. I can find stuff that’s better than what I’ve gotten from a score or more books put together
.
It can also create excellent study plans, which can also measure your progress, diagnose weaknesses, and create exercises to cover them. So for those with a strong auto-didactic streak, it is apparent that the new AI can supersize their actual learning. But I fear the vast bulk of students will use (are using) it to cheat on homework and tests. The fact that we are heavily credential bond in our reward systems really works against genuine learning here.
So I fear we will create a strongly two tier cognitive society. The few percent with the discipline to use the AI for skill boosting will thrive, while the surrounding Hoi-Polloi deskill themselves.
As per output of Engineers and scientists, my guess is an order of magnitude increase in per researcher productivity. There is some early indication that Jevon’s paradox will increase demand for science/engineering, so that employment likely won’t go off a cliff.
Thomas Gordon Hewitt 06.25.26 at 11:54 pm
Now electricty use per se, is unlikely to be a planetary carry capcity issue. What matters is how we generate and distribute that energy. With the exceptions of Nuclear, including potentially Fusion, it must rapidly transition to 100% renewables. The race for data-centers could help here, however too many are choosing natural gas based power as a source for their massive proposed new capacity. Pressure to force them towards renewables plus storage is very important. Otherwise the climate future looks bleak indeed.
Now grid storage will soon move beyond the Lithium battery stage. Sodium Ion batteries are coming on strong -even GM is jumping in big time, with grid storage as a target. Couple that with 25% of better solar efficiency and we relly can cover the deamnd with renewables. But again near term incentives are needed.
Here AI can help with the facilities planning and expansion, and with shortening the development pipeline for the new energy tech.
Satori 06.26.26 at 3:17 am
I think I’ve heard that that even more so than Power, Water usage by data centers presents a fairly immediate problem for people living nearby.
John Q 06.26.26 at 5:31 am
I don’t want to do the “left and right disagree with me, so I must be right” centrist thing, But the comments are evenly divided between people who think I’ve been too generous to AI and those who think I’ve underestimated it.
The main explanation, as with previous big innovations, is that AI has had a huge impact in some areas and very little in others, and a mix of negative and positive. So, depending on where you are you see very different things.
Looking in aggregate, and assuming agentic stuff delivers more of the incremental gains we’ve seen with LLMs, I estimate AI will raise productivity growth by about 1 per cent a year, and double the amount of electricity used in the IT sector. That’s a lot as economic changes go, comparable to previous ITC innovations – mainframe computing, personal computers, the Internet itself. But it leaves most of our lives unchanged.
marcel proust 06.26.26 at 1:43 pm
Alex SL:
I am not the one to provide the detail, since I don’t use it. It would quickly generate into an instantiation (what a wonderful word: makes me sound so learned, so competent, so … pretentious) of whisper down the lane. If you have the time and interest, try to read the links to Scott Cunningham’s work. He has gone into great detail about how he uses it (i.e., step by step): how he works with it to avoid plagiarism and other problems that you and others have mentioned: how it can fail in summarizing articles and papers, and what he does to have it generate reliable summaries that he can trust. Most of what I know about AIs I learned
in kindergartenfrom his accounts.I trust the work habits of my child & child-in-law, and both report the need for careful review of the results, and for several rounds of pushing back on Claude until the results are satisfactory. The one in the public sector is, as a matter of fact, in math education at an inner city school, working with teachers to design better lesson plans and curricula, much like what Thomas Gordon Hewitt describes in his second paragraph.
engels 06.26.26 at 6:22 pm
Come friendly bombs and fall on Slough
https://www.theguardian.com/environment/2026/jun/26/slough-is-like-an-experiment-europes-largest-datacentre-hub-leaves-town-sweltering
Meanwhile, the human brain consumes about 15W. There are 180 million people unemployed worldwide and 270 million children out of school.
Alex SL 06.27.26 at 10:25 am
marcel proust,
Thanks for the response, and fair, you are reporting second-hand information from people you trust.
Extrapolating from my own work and what I see in colleagues, I am just very confused how “3-4 hours later, with a reasonable amount of back and forth as Claude returns with revised results from the most recent command, they have something useful” translates into “scale up their work” as opposed to “I get frustrated and simply do it myself in 20 minutes tops”.
And also, as always, how do I know after the 3-4 hours that the results are correct unless I did not need the AI in the first place? I am really very puzzled by the claims of efficiency gains because whenever I see results where I know the correct answer, they are at best “off” if not logically inconsistent or flat out wrong. Literature reviews cite works to support statements that those works do not support, or the bot gets stuck on a small aspect of the larger space I asked it to report on. AI-generated code is unnecessarily verbose and non-modular, which suggests to me that larger software projects using AI will likely be even more bloated than the already bad software that we have all been trained to accept by Microsoft over the last decade or so.
When I see people use AI for meeting notes, they have to spend more time editing the AI output to make it acceptable than I would have spent taking those notes myself. A few months ago I was part of a breakout session where we made notes in a shared Word doc. When the session ran out, one of the colleagues suggested to have Copilot summarise the (drum roll) one and a quarter A4 pages worth of notes we had made. I regret having to admit that I was unable to suppress a derisive noise, but it was trivial to skim the notes and formulate a three point summary to present to the larger group before she even ran that request.
What are people doing here? Does nobody want to use their brain anymore for an efficiency gain of at most one minute, in this case in a workshop where that minute would at any rate only be spent waiting for the main session to resume?
And by definition, the kind of work generative AI is intended to help with does not lend itself to automation like adding three screws to a gadget on an assembly line or cutting planks to a defined size did, and where it is easy to quality test every 1,000th output to see if the machine still works to specifications. They are generally one-off tasks that need to be checked individually. I do not understand what we gain from using AI except what is now called cognitive surrender.
Thomas Gordon Hewitt,
I am a scientist, and I can guarantee that there will not be an order of magnitude increase in productivity for my profession, unless you measure it in terms of AI-generated slop papers submitted to journals that then have to reject them, decreasing the editorial productivity of science journals. The main bottleneck in science is data generation: specimen collection, lab experiments, testing prototypes, DNA sequencing, glasshouse trials, microscopy, etc. None of this can be sped up by a word guessing model. Some of it can be sped up by robotics, but even that won’t work in some other areas, because a bacterial culture or experimental plant grows at its own speed.
nonrenormalizable 06.28.26 at 5:13 pm
Some nearer-term effects of the AI + data centre boom:
Utility boss warns US faces blackouts due to power supply shortfall
Trader Joe 07.01.26 at 11:03 am
Thanks for this analysis, its quite interesting.
I would note that if you look to the prospectus for Xai and Anthropic, they are assuming long term margins of more like 50% than the 30% you use. I would view this a plausible but far from certain as it assumes adoption and scale effects plus better energy usage LT. Id further note that most major tech businesses have margins in the 50%+ range so its not a “pie in the sky” assumption.
If they do achieve those levels, and that’s what valuations seem to be assuming right or wrong, your calculation changes pretty dramatically as to profitability and may even allow for some price cutting which could in turn increase demand.
To be sure, prospectuses are just that, but its where your calculations differ most sharply from Wall Street’s.
Thomas Gordon Hewitt 07.03.26 at 10:31 pm
The hope of near-monopoly rents accruing to the frontier labs is highly likely to be a major financial dissapointment. The Opensource models are closing the gap, producing results as good as the frontier models did four to eight months ago. My son who is using it to increase his software productivity at work is cutting token charges by using Open source models for all but the most challenging tasks. So instead of monopoly rents, expect near comodity prices.
Alex SL 07.04.26 at 11:05 am
Thomas Gordon Hewitt,
Agreed. There are various claims floating around in the discourse on generative AI that are all, in a sense, unrelated. So, even if the claim that chatbots revolutionise science and create massive efficiencies in all knowledge work were to come true (they won’t), that doesn’t mean there isn’t a speculative bubble going on that may dwarf the dot com bust and 2008. Be it open weights models or competition between corporate models, be it ChatGPT versus Claude or Claude versus DeepSeek, the moment anybody tries to raise prices to where they make a profit, most users will decamp to where prices are still subsidised.
What many fans of the large AI corporations seem to be missing is that SaaS, social media, and cloud service providers are only able to extract monopoly rents because it is very difficult for a large company to switch from MS Excel to LibreOffice Calc or to uproot their entire server and database infrastructure, or for a social media user to get their entire network of friends to switch to Mastodon. There is strong vendor lock-in. None of that applies with chatbots. While some of them may perform better or worse on this or that benchmark, ultimately they all kind of work the same way, and even now we can watch vibe coders switching between models rather easily.
Bringing this back to the original topic, this means that electricity consumption will in the long run, after the over-investment in data centers and the overvaluation of tech stocks have popped and likely sent the world into a years-long depression, settle into an equilibrium where the users have to pay for the electricity cost of their chatbot queries, and then they can and will have to decide how much of their yearly income to invest into summarising work emails or cheating on homework. It will become no more than a ‘normal’ part of our collectively suicidal resource overuse. The problem right now is that not even that demand curve applies, because venture capital is speculatively subsidising chatbot providers to shovel tens of billions of dollars per year into a furnace, so electricity use on training and using them is higher than it would be if it was budgeted even only semi-rationally.
Thomas Gordon Hewitt 07.06.26 at 4:43 am
The recent retreat by Meta, choosing to rent out much of their compute resources rather than shooting for the ceation of a digital god is an indication that overspending by frontier labs is becoming a major concern. It has sent memory prices downward, as demand had outstripped supply, but that may be changing. There is still a lot of potential for dramatic (order of magnitude) reduction in the energy cost per token. It’s always been the case that improvements in algorithmic efficiency have roughly matched improvements in electronics. Also coming 2.5D electronics promises another decade of scaling to electronic cost performance.
Tm 07.06.26 at 7:44 am
nonrenormalizable 9: these data seem implausible, especially the figure for Germany (“5.5GW, consuming 9.5 percent of the nation’s electricity”). It’s hard to believe that data centers have grown to use almost 10% of Germany’s electricity without anybody having noticed. Germany’s public electricity use has been declining markedly especially since 2018. This is generally explained by a decline in energy intensive industry due to high energy prices, efficiency gains, and the increase in solar power (often consumed directly without being fed into the grid). Why doesn’t the data center boom show up in these data?
Of course energy consumption should be quantified in TWh, not in GW (power). It’s hard to make anything with these GW figures.
Comments on this entry are closed.