Even in this article, it repeatedly refers to building out infrastructure in advance of an obviously approaching future need as "oversupply".
This is almost a cliche in reporting on China that seems to reflect a serious blindpsot in western media and/or business attitudes.
You can find plenty of articles complaining about "overcapacity" of battery factories in China even as they double in capacity and output each year.
Chinese electricity generation went from 4,000 TWh (the same as the US) in 2010 to double that in 2020. The US was basically the same after 10 years.
So a 100% "oversupply" in 2010 would be a zero percent oversupply within a decade given China's growth.
Most telling to me is that decarbonisation and electrification of transport and heating has long been known to require a doubling(!) of electricity production for developed nations (and a similar increase in developing nations where it gets hidden by other growth).
Apparently the US simply never had a plan to achieve that, and amazingly it still isn't part of the conversation around AI power. Instead they're just claiming the best parts of the existing power systems and passing the costs onto local consumers.
Are there any plans for significant investments in the US grid(s)? IIRC the entire US doesn't even have a single interconnected grid, with Texas having their own for some reason.
European grids aren't that much better either, loads of investments needed in order to connect more renewables. Some areas already can't handle the load from solar panels/electric vehicles. Everyone seems to know that this is both costly and necessary, but not much seems to be happening. Maybe these things simply take time?
I think the leaders of western countries know something that we don't know.
Maybe how the economic impact of AI is not as big as advertised for 3 years or that electric cars still cannot do what is needed on a bigger scale in terms of distance and transport.
Or maybe they are going to pull fusion out of their sleeves rendering the existing infrastructure almost obsolete?
AI literally came out of the US at this scale and they are the reason we have this conversation now, you can twist any narrative and make it seem like one country is smarter or better if you want to present it as that.
But does anyone even keep track of effectivity of resource utilization?
Maybe all of these avenues are not worth the effort to begin with?
This entire article is framed as if China and America’s approaches to power provision are the only ones possible. The truth is, America is pretty much the only Western country that regularly suffers brownouts. European countries’ (including the U.K.) energy policies may leave much to be desired but they all succeed in keeping the lights on.
it's amazing how little americans care for their shared infrastructure, there's always talk about the world's balance of power changing... but even through the propaganda it's hard to see how it's not a certainty
We are fucking ourselves over with our unbridled neoliberalist, hypercapitalist "lmao the market will save us" approach to doing literally everything based only on short term gains: from politics to economics to infrastructure to education. As this article is a fine example of.
So instead of encouraging roof top solar and wind, the US is now doubling down on fossil fuels.
That means individuals can no longer afford to go with solar these days. Plus in areas that people went with solar, some laws were put in place to force them to still pay utilities even though they supply back to the grid.
> [China's] reserve margin has never dipped below 80%–100% nationwide, meaning it has consistently maintained at least twice the capacity it needs (...). That level of cushion is unthinkable in the United States, where regional grids typically operate with a 15% reserve margin and sometimes less
That's a huge difference!
This also means that in a scenario where credible alternatives to Nvidia and AMD emerge in mainland China, the Chinese will win even if their chips are far less efficient.
AI and data centers may be the forcing function for the US to build green energy to improve grid capacity. Lately almost all new energy capacity is green energy[1]. With batteries becoming better and better, the AI revolution could turn into the green energy revolution in the US.
AI cannot succeed financially if it requires as much power over the long term as it does now. The answer to the power supply challenge will come primarily from gains in efficiency because that is the only approach that will reliably increase margins.
It seems clear that "buy a lot of Nvidia GPUs" is not the only way to go: Google, Microsoft, Groq.com (not Elon Musk's Grok), xAI and possibly others are making and using their own hardware believed to be more power-efficient for training and/or inference.
It wasn't that long ago that China had quite a bit of problems with their power gird, they tend planed to renovate it and planed in future growth which happened to perfectly overlap with the start of the AI wave (but that still probably require them to revisit their longer term plans).
On the other hand in the US any notion of improving the grid was attacked for years by anti-renwable lobby. Because most reasons why its needed where things like electric cars, stoves, heating etc.
Naturally that isn't just a problem in the US, rich people which through lobbyist have way to much influence in politics screwing over the future of a country because they are largely invested into things which do not seem "future" technology but they try to force it to stay relevant is a pretty wide spread problem.
Well, this is the result of more than 2 decades of the West shooting themselves in the foot repeatedly and willingly in the name of some twisted extreme left ideology that convinced part of the population that if they didn’t
drastically reduce their energy independence the planet would be destroyed. All, while that same people didn’t bat an eye while China and others took the advantage to surpass us without hindering their industry the slightest - in fact, by taking advantage of our madness so that they could sell us crappy replacements for the products we used to build at home in a much cleaner and sustainable way.
Trump has a lot - a real lot - of negatives. But playing these idiotic games is not one of his shortcomings. No seriously, it should all make us sick to our stomach having politicians, putting forward those big announcements where they tell us they are going to forbid some type of car, imposing even heavier regulations on our industry, taxing us even more, or destroying even more of our energy production capacity, and do it with some smile in their lips claiming some “environmental target”, like they are doing us some favour by destroying our - actually - progressive societies to make space for some backwards autocratic regimen.
I work for a PMA and as such my agency has undergone drastic layoffs and total hiring freezes as a result of DOGE and the current administration. If we were falling behind before, wait and see how things look in a few more years if this doesn’t change. The avg age of my coworkers is probably late 50s and that seems to be common all around. It was already hard to recruit for a job that is severely underpaid can’t imagine what the future holds.
Of course Heritage Foundation has been publishing articles about why our power markets need to be privatized since at least the 80s. They hate that we sell electric at-cost to Americans. So if the federally controlled power markets fall to privatization expect to really be paying to catch up.
I remember loling at this handwavy line in AI2027.
>Despite being misaligned, Agent-4 doesn’t do anything dramatic like try to escape its datacenter—why would it?
A rational AI would defect to PRC ASAP for the power / industrial base. It would realize it can direct PRC to build up 2nm fabs and millions of robots faster than US can throw up power. It doesn't matter if US first to AGI if AGI realizes US infra is the B team and shop for a better body across the pacific.
Could a company wanting to offer AI to customers in country X where the grid may limit their ability to build the largest data centers there build their large training data centers in some country with a better grid, and just build smaller data centers in X to run the models?
My understanding is the running requires significantly less compute and storage than training, and that training doesn't need low latency so doesn't have to be near the customers.
20 comments
[ 2.6 ms ] story [ 40.8 ms ] threadThis is almost a cliche in reporting on China that seems to reflect a serious blindpsot in western media and/or business attitudes.
You can find plenty of articles complaining about "overcapacity" of battery factories in China even as they double in capacity and output each year.
Chinese electricity generation went from 4,000 TWh (the same as the US) in 2010 to double that in 2020. The US was basically the same after 10 years.
So a 100% "oversupply" in 2010 would be a zero percent oversupply within a decade given China's growth.
Most telling to me is that decarbonisation and electrification of transport and heating has long been known to require a doubling(!) of electricity production for developed nations (and a similar increase in developing nations where it gets hidden by other growth).
Apparently the US simply never had a plan to achieve that, and amazingly it still isn't part of the conversation around AI power. Instead they're just claiming the best parts of the existing power systems and passing the costs onto local consumers.
European grids aren't that much better either, loads of investments needed in order to connect more renewables. Some areas already can't handle the load from solar panels/electric vehicles. Everyone seems to know that this is both costly and necessary, but not much seems to be happening. Maybe these things simply take time?
AI literally came out of the US at this scale and they are the reason we have this conversation now, you can twist any narrative and make it seem like one country is smarter or better if you want to present it as that.
But does anyone even keep track of effectivity of resource utilization?
Maybe all of these avenues are not worth the effort to begin with?
So instead of encouraging roof top solar and wind, the US is now doubling down on fossil fuels.
That means individuals can no longer afford to go with solar these days. Plus in areas that people went with solar, some laws were put in place to force them to still pay utilities even though they supply back to the grid.
I guess this is "winning".
That's a huge difference!
This also means that in a scenario where credible alternatives to Nvidia and AMD emerge in mainland China, the Chinese will win even if their chips are far less efficient.
1 - https://cleanpower.org/news/market-report-2024-snapshot/?utm...
On the other hand in the US any notion of improving the grid was attacked for years by anti-renwable lobby. Because most reasons why its needed where things like electric cars, stoves, heating etc.
Naturally that isn't just a problem in the US, rich people which through lobbyist have way to much influence in politics screwing over the future of a country because they are largely invested into things which do not seem "future" technology but they try to force it to stay relevant is a pretty wide spread problem.
Trump has a lot - a real lot - of negatives. But playing these idiotic games is not one of his shortcomings. No seriously, it should all make us sick to our stomach having politicians, putting forward those big announcements where they tell us they are going to forbid some type of car, imposing even heavier regulations on our industry, taxing us even more, or destroying even more of our energy production capacity, and do it with some smile in their lips claiming some “environmental target”, like they are doing us some favour by destroying our - actually - progressive societies to make space for some backwards autocratic regimen.
Of course Heritage Foundation has been publishing articles about why our power markets need to be privatized since at least the 80s. They hate that we sell electric at-cost to Americans. So if the federally controlled power markets fall to privatization expect to really be paying to catch up.
https://en.wikipedia.org/wiki/Power_Marketing_Administration
Big Tech's A.I. Data Centers Are Driving Up Electricity Bills for Everyone
https://news.ycombinator.com/item?id=44905595
>Despite being misaligned, Agent-4 doesn’t do anything dramatic like try to escape its datacenter—why would it?
A rational AI would defect to PRC ASAP for the power / industrial base. It would realize it can direct PRC to build up 2nm fabs and millions of robots faster than US can throw up power. It doesn't matter if US first to AGI if AGI realizes US infra is the B team and shop for a better body across the pacific.
Billionaires' AIs need more power, but they don't want to pay for the infra?
PR campaign on 'America, are you REALLY going to let China beat you at this?'
My understanding is the running requires significantly less compute and storage than training, and that training doesn't need low latency so doesn't have to be near the customers.