IBM might not have a data strategy or AI plan but he isn’t wrong on the inability to generate a profit.
A bit of napkin math:
NVIDIA claims 0.4J per token for their latest generation
1GW plant with 80% utilisation can therefore produce 6.29 10^16 tokens a year.
There are ~10^14 tokens on the internet. ~10^19 tokens have been spoken by humans… so far.
Don't know what the source is, but it feels missing a few orders of magnitude. Surely it only counts text? I can't imagine there are only so few data on the internet if you count images and videos.
Tokens are, roughly speaking, how you pay for AI. So you can approximate revenue by multiplying tokens per year by the revenue for a token.
(6.29 10^16 tokens a year) * ($10 per 10^6 tokens)
= $6.29 10^11
= $629,000,000,000 per year in revenue
Per the article
> "It's my view that there's no way you're going to get a return on that, because $8 trillion of capex means you need roughly $800 billion of profit just to pay for the interest," he said.
$629 billion is less than $800 billion. And we are talking raw revenue (not profit). So we are already in the red.
But it gets worse, that $10 per million tokens costs is for GPT-5.1, which is one of the most expensive models. And the costs don't account for input tokens, which are usually a tenth of the costs of output tokens. And using bulk API instead of the regular one halves costs again.
Realistic revenue projections for a data center are closer to sub $1 per million tokens, $70-150 billion per year. And this is revenue only.
To make profits at current prices, the chips need to increase in performance by some factor, and power costs need to fall by another factor. The combination of these factors need to be, at minimum, like 5x, but realistically need to be 50x.
The math here is mixing categories. The token calculation for a single 1-GW datacenter is fine, but then it gets compared to the entire industry’s projected $8T capex, which makes the conclusion meaningless. It’s like taking the annual revenue of one factory and using it to argue that an entire global build-out can’t be profitable. On top of that, the revenue estimate uses retail GPT-5.1 pricing, which is the absolute highest-priced model on the market, not what a hyperscaler actually charges for bulk workloads. IBM’s number refers to many datacenters built over many years, each with different models, utilization patterns, and economics. So this particular comparison doesn’t show that AI can’t be profitable—it’s just comparing one plant’s token output to everyone’s debt at once. The real challenges (throughput per watt, falling token prices, capital efficiency) are valid, but this napkin math isn’t proving what it claims to prove.
> but then it gets compared to the entire industry’s projected $8T capex, which makes the conclusion meaningless.
Aren't they comparing annual revenue to the annual interest you might have to pay on $8T? Which the original article estimates at $800B. That seems consistent.
I just misread the article, as it seems to bounce around between $nX capex for nY gigawatt hours in every paragraph.
But it looks like the investments are $80MMM for 1GW. Which, if true, would have the potential to be profitable, depending on depreciation and electricity costs.
Broad estimates I'm seeing on the cost of a 1GW AI datacenter are $30-60B. So by your own revenue projection, you could see why people are thinking it looks like a pretty good investment.
Note that if we're including GPU prices in the top-line capex, the margin on that $70-150B is very healthy. From above, at 0.4J/T, I'm getting 9MT/kWh, or about $0.01/MT in electricity cost at $0.1/kWh. So if you can sell those MT for $1-5, you're printing money.
> So if you can sell those MT for $1-5, you're printing money.
The IF is doing a lot of heavy lifting there.
I understood the OP in the context of "human history has not produced sufficiently many tokens to be sent into the machines to make the return of investment possible mathematically".
Maybe the "token production" accelerates, and the need for so much compute realizes, who knows.
That signifies that your company is not appealing to impressive candidates for some reason or another. Companies that offer good pay, some other great benefits in the place of good pay, or kind of okay pay but very interesting work have no trouble getting people, especially in today's market.
No. Impressive candidates are applying to jobs that pay somewhat reasonably, even if it's below what they expect. If candidates who are desperate are still completely skipping over a company, that says something about that company.
I am also hiring, in Europe with very good work/life balance but modest salaries and like the parent I'm also not that impressed with candidates, so to me the other explanation is that candidates have a wildly incorrect estimation of what is a somewhat reasonable pay in 2025.
The position with FAANG like salaries have reduced drastically. Companies paying 6 figures just to have the privilege to have an entry level developer with this then seen as magical skill of being able to type code was a dream that is over. Look at salaries of engineers in other industries, breaking 6 figures needs a lot of seniority, $150k is rarely heard of for ICs.
I don't know what the market is like in the EU. I can just tell you North America is really bad with a deluge of talent. And sadly, many can't really live off of an EU salary.
breaking six figures in California isn't that impressive unless you are literally single and out of college. But that quickly gets eaten up when rent is 2k+/month and you have school loans to pay off. When you're not paying nearly six figures for college and have your taxes built into most of your day to day life, you don't need six figures.
I understand things are different there, but I thought even in California 6 figures at entry or mid level was a thing only in software engineering, so my point was that premium over other professions was evaporating.
We pay well. Very well in fact. We’re a small company though.
I have a harder time hiring here than at my previous position with a much larger company, even though my current employer is superior to my old one in every way except for brand recognition.
And from others in this space (who have at times tried to recruit me) this is not a problem unique to my company.
What's your area, what qualities are you looking for, and what's your filtering process?
If you're not in a major hub and ATS is filtering out all the good candidates not gaming the system, the results are inevitable. If you're looking for a senior for junior or less pay... Well, it's easier to keep searching than take a job that literally can't pay rent in some high COL areas.
If in the end we can just have .spy on some files that have performance critical functions in them and the rest is just normal python, this could be down right amazing.
We recently swapped out mypyc optimised module for a rust implementation to get a 2-6x speed up, and not having to do that would be great.
Also, when I read about the language features which make Python intrinsically slow, generally I think "I never use that." e.g. operator overloading meaning you need to do all these pointer dereferences just to add two numbers together. Yes, I get that pytorch and numpy rely critically on these. But it makes me wonder...
Could you disable these language features on a module-by-module basis? That is, declare "in this sub-tree of the source code, there shalt be no monkey-patching, and no operator overloading" and therefore the compiler can do a better job. If anybody tries to do the disallowed stuff, then it's a RuntimeError. Would that work?
It probably would but it'd likely not be trivial. Because how do you recognize monkey-patching in a dynamically typed language? I can think of many edge cases here.
See my other comment in the thread. I would argue that anything that uses arcane dynamic stuff in python should be renamed to .dpy and the vast majority of the commonly used constructs retain .py
The issue in HN threads like this is that everyone is out to promote their own favorite language or their favorite python framework that uses dynamic stuff. The majoritarian and hacker-ethos of python don't always line up.
Like Chris Lattner was saying on a recent podcast, he wrote much of Swift at home on nights/weekends over 18 months. We need someone like that do this for spy.
I’m pretty sure it’s smartphones and aging society.
I’ve been gathering data on the natural experiment that occurred due to differences in proliferation of smartphones across countries. The sex surveys aren’t consistent but that is a very strong factor. Look at the hockey stick curve in the paper here.
The rest of the decline is (in my evaluation) best explained with the increase in average age of the 18-64 year old demographic.
> The rest of the decline is (in my evaluation) best explained with the increase in average age of the 18-64 year old demographic.
I wondered about this before reading the article, but there's a chart in the article comparing narrow ~10 year age ranges to their past levels, and the decline is persistent in each group, so I don't think it can be explained by the overall population being older.
The thing that is often missed is that with high taxes and resulting high labour cost, you get high productivity. Simply because low productivity is not profitable.
This results in bad service, high quality goods and strong utilisation of capital goods.
But as others have noted the wealth disparity is increasing thanks to new policies and low interest rates leading to asset inflation.
As someone who sees the outcome of people losing everything to sophisticated scammers/ fraudsters and thieves and how little authorities are able to do, nah, the overreach is not in sight.
There are more criminals than abusive IRS agents. And usually when people tell me stories like that, there is more to it..
This. I maintain an ecommerce platform written in Python. Even with Python being slow, less than 30% of our request time is spent executing code, the rest is talking to stuff over the network.
I think articles like this cast too wide a net when they say "performance" or "<language> is fast/slow".
A bunch of SREs discussing which languages/servers/runtimes are fast/slow/efficient in comparable production setups would give more practical guidance.
If you're building an http daemon in a traditional three-tiered app (like a large % of people on HN), IME, Python has quietly become a great language in that space, compared to its peers, over the last 8 years.
SRE here, that horizontal scaling with Python has impacts as it’s more connections to database and so forth so you are impacting things even if you don’t see it.
A bit of napkin math: NVIDIA claims 0.4J per token for their latest generation 1GW plant with 80% utilisation can therefore produce 6.29 10^16 tokens a year.
There are ~10^14 tokens on the internet. ~10^19 tokens have been spoken by humans… so far.
reply