Of course they will. Tokens are valuable, you can always spend a finite budget on specialized tokens or fewer and higher quality tokens, size of user base and engagement gives you a flywheel moat that is difficult for newcomers to compete with. The market is complex and easy to oversimplify.
My new startup tokencoin will blah blah blah exchange rate, (something AI writes here), 3. profit (more AI), benefiting all human kind and helping our users scale up their productive intelligence!
It's hard and complex to enter any mature market. The vast majority of firms that attempt to enter a new market fail. LLM's have no more than this normal moat.
Isn’t capital and momentum a moat? Sure Chinese models use distillation but I don’t see them training models from scratch. At least not today. But maybe as chips get cheaper and they have Chinese made ones?
Apparently not much of one. There are, what, 5 or more companies with frontier models? And open weights models like MiniMax are snapping at their heels
There are many markets where open source has been nipping at heels for a long time.
Obviously product areas differ for reasons structural and happenstance. But there is definitely a pattern that occurs, where open source fast follows commercial advances, benefiting from having a clear target to develop for.
Which is of course, a great service. Even if it never unseats the commercial version, it forces the owners to reinvest more in improvements, by undermining their moats. As well as providing a much better value alternative version for many people.
I’m not technically familiar but I remember someone saying that models like MiniMax basically skip the cost of training by using distillation to basically “steal” the models from OpenAI or Anthropic, and that these companies now have various defenses against this. What happens when MiniMax has to do the full work themselves?
> The report estimates that carbon emissions from models with the least efficient inference are over 10 times as high as those with the most efficient inference. DeepSeek’s V3 models were estimated to consume around 23 watts when responding to a “medium-length” prompt, while Claude 4 Opus was estimated to consume about 5 watts.
This makes absolutely no sense. I suppose they meant watt hours, and that's a weird way to explain carbon emissions...
"On the other hand, Perrault noted that 'Epoch AI independently estimates Grok 4’s emissions to be significantly higher at approximately 140,000 tons of CO₂.'"
I realize these are still estimates, but when the other independent analysis nearly doubles the outcome I'm not left feeling optimistic. One could argue some numbers from others are underestimates... which of course just bums me out all the more!
I agree 100% if those are the only two options. I guess my point is that it's fair to assume that Elon's crew is doing the bare minimum in terms of efficiency and pollutant mitigation-- at least when compared to other data centers who do legally compliant business with real power companies.
The "China leads in robotics" seems to be unaffected by AI. The China line is basically on the same trajectory since 2012. The chart does no belong in the article.
They also lead the world in EV production on paper, but in practice a large portion of those numbers might be driven by government pressure, not actual demand [1].
I’d personally take this data with a big grain of Goodhart’s law.
The graph says "new industrial robots installed", which is a bit misleading. For example the newest BYD factories are still stuffed with German/Japanese robots.
What's worse is that this the predictable result of a choice that America made decades ago and continues to make.
Outsourcing manufacturing capacity to China and letting domestic manufacturing skills atrophy and institutional knowledge die out was a choice that many people opposed but were ultimately helpless to stop because the people making the decisions ignored them and did it anyways for personal gain is how we got here.
You'd think that the supply chain shocks that we saw during COVID would be a wake up call that would have jolted people into action.
You'd think that Ukraine-Russia war would have been a wake up call that would have jolted people into action.
You'd think that the recent failures by the US military in Iran and the depletion of years of missile stockpiles would have been a wake up call that would have jolted people into action.
I'm at a loss to explain it. It's like the American oligarchs want to weaken America, or at least are willing to do so if it means that they have greater control over it. Maybe they don't care about manufacturing capacity because they know that America is ultimately a nuclear protected island and that even if things continue to decline they'll be safe to rule it like a king?
Worth calling out AI sentiment among young people is not nearly so rosy: https://news.gallup.com/poll/708224/gen-adoption-steady-skep...
Also nobody will ever have a moat, so the graph of investor stupidity is going through the roof.
Of course they will. Tokens are valuable, you can always spend a finite budget on specialized tokens or fewer and higher quality tokens, size of user base and engagement gives you a flywheel moat that is difficult for newcomers to compete with. The market is complex and easy to oversimplify.
My new startup tokencoin will blah blah blah exchange rate, (something AI writes here), 3. profit (more AI), benefiting all human kind and helping our users scale up their productive intelligence!
It's hard and complex to enter any mature market. The vast majority of firms that attempt to enter a new market fail. LLM's have no more than this normal moat.
Well yes that’s my point: AI does not suddenly do away with the market.
Isn’t capital and momentum a moat? Sure Chinese models use distillation but I don’t see them training models from scratch. At least not today. But maybe as chips get cheaper and they have Chinese made ones?
> Isn’t capital and momentum a moat?
Apparently not much of one. There are, what, 5 or more companies with frontier models? And open weights models like MiniMax are snapping at their heels
There are many markets where open source has been nipping at heels for a long time.
Obviously product areas differ for reasons structural and happenstance. But there is definitely a pattern that occurs, where open source fast follows commercial advances, benefiting from having a clear target to develop for.
Which is of course, a great service. Even if it never unseats the commercial version, it forces the owners to reinvest more in improvements, by undermining their moats. As well as providing a much better value alternative version for many people.
I’m not technically familiar but I remember someone saying that models like MiniMax basically skip the cost of training by using distillation to basically “steal” the models from OpenAI or Anthropic, and that these companies now have various defenses against this. What happens when MiniMax has to do the full work themselves?
>Chinese models use distillation but I don’t see them training models from scratch
Maybe because they don't have to. If someone is doing the heavy work and they can take output of that, it's a win for them.
[dupe]
https://news.ycombinator.com/item?id=47758028
Source: https://hai.stanford.edu/ai-index/2026-ai-index-report
> The report estimates that carbon emissions from models with the least efficient inference are over 10 times as high as those with the most efficient inference. DeepSeek’s V3 models were estimated to consume around 23 watts when responding to a “medium-length” prompt, while Claude 4 Opus was estimated to consume about 5 watts.
This makes absolutely no sense. I suppose they meant watt hours, and that's a weird way to explain carbon emissions...
Besides the lead in robotics for China, those Grok emissions charts are the thing that most leap off the page.
"These estimates should be interpreted with caution. In the case of Grok, they rely heavily on inferred inputs drawn from public reporting"
That chart doesn't really pass the sniff test.
The rest of the quote you began continues:
"On the other hand, Perrault noted that 'Epoch AI independently estimates Grok 4’s emissions to be significantly higher at approximately 140,000 tons of CO₂.'"
I realize these are still estimates, but when the other independent analysis nearly doubles the outcome I'm not left feeling optimistic. One could argue some numbers from others are underestimates... which of course just bums me out all the more!
I don't know if I would want to do too much sniffing around the Methane power they are using over at xAI.
https://www.theguardian.com/us-news/2025/jul/03/elon-musk-xa...
That's definitely a very visible use of carbon generating fuel, but I'd choose methane over coal power plants all day.
I agree 100% if those are the only two options. I guess my point is that it's fair to assume that Elon's crew is doing the bare minimum in terms of efficiency and pollutant mitigation-- at least when compared to other data centers who do legally compliant business with real power companies.
> Training AI models can generate enormous carbon emissions
Sure, but what I'd really like to see is a graph for how much carbon is generated serving these models globally.
I still don't understand the State of AI in 2026.
The "China leads in robotics" seems to be unaffected by AI. The China line is basically on the same trajectory since 2012. The chart does no belong in the article.
China’s robotics lead holy cow.
They also lead the world in EV production on paper, but in practice a large portion of those numbers might be driven by government pressure, not actual demand [1].
I’d personally take this data with a big grain of Goodhart’s law.
[1]: https://www.bloomberg.com/features/2023-china-ev-graveyards/
The graph says "new industrial robots installed", which is a bit misleading. For example the newest BYD factories are still stuffed with German/Japanese robots.
China’s manufacturing lead in a graph
It striking, but says nothing about AI.
Don't they have ten times more people than the next highest country (Japan) though?
What's worse is that this the predictable result of a choice that America made decades ago and continues to make.
Outsourcing manufacturing capacity to China and letting domestic manufacturing skills atrophy and institutional knowledge die out was a choice that many people opposed but were ultimately helpless to stop because the people making the decisions ignored them and did it anyways for personal gain is how we got here.
You'd think that the supply chain shocks that we saw during COVID would be a wake up call that would have jolted people into action.
You'd think that Ukraine-Russia war would have been a wake up call that would have jolted people into action.
You'd think that the recent failures by the US military in Iran and the depletion of years of missile stockpiles would have been a wake up call that would have jolted people into action.
I'm at a loss to explain it. It's like the American oligarchs want to weaken America, or at least are willing to do so if it means that they have greater control over it. Maybe they don't care about manufacturing capacity because they know that America is ultimately a nuclear protected island and that even if things continue to decline they'll be safe to rule it like a king?
Profits generated by AI: <not graphed>
The absence speaks volumes.