The AI Race: U.S. vs. China || Sept 22, 2026
The AI race is in full swing. China is chasing digital sovereignty, but there's a glaring hole in that plan...it doesn't have the advanced manufacturing capabilities to achieve it. Here's the real kicker: nobody does.
China might be a bit more flexible with its ethics and data manipulation, but that means nothing without the advanced chipmaking equipment. China can produce less efficient chips, but has incredibly low yields on leading-edge chips. To produce the highest-level chips, there are roughly 65 countries involved in a complex supply chain.
Which means that not even the U.S. could go out and do this on its own (unless it underwent a 20+ year tech ecosystem rebuild). America's real advantage in the AI race comes from the broad coalition of allied countries all working together to produce high-end semiconductors.
Looking for something else?
The links below will take you to the Zeihan on Geopolitics YouTube page, each of our social media channels, info on how to purchase Peter Zeihan's books, our website, and more. Click the box to navigate to your desired page.
If there is still something else that you need, the helpful links tab below will direct you to anything you might need. This includes the Patreon support page, how to submit a question for Peter, and more.
Comments (25)
Bro, get a microphone with a wind guard! Appreciate the content :)
The 21st century arms race.
AMERICA🇺🇸 still the best,,
So in a nutshell one group is a communist/fascist amoral money grabbing hater of human rights and freedoms and the other is the PRC.
Is any of this going to matter once globalization collapse is in full swing and the ability to manufacture chips goes away. It's a race to see if we get Skynet or globalization collapse first.
I would love to hear Peter's take on Elon Musk's Terra Fab and the Lithography ring to replace (or at least supplement) the reliance on ASML. At some point the ASML choke point for chips will be duplicated or leap frogged since it is technology. Same goes for TSMC. memory, etc. Nothing on the horizon to validate true discussion yet, other than Terra Fab, but it will inevitably happen at some point because technology progresses. I still like the US as long term AI chip leader because the choke point for chip Fab we own is a natural resource. Check out Maxinomics episode on Spruce Pine. Although, I guess the purest quartz in the world could probably be manufactured with enough resources pointed at it. Bottom line, nobody knows where this is gonna go and who will be the “winner” in 20 years. Shit is moving fast. We might even be able to grow chips through biology with the stuff AI is uncovering. Ok… maybe I’m going too deep for a comments section : )
Very relevant, not too deep!
The new IBM Northpole chips are 12nm. Combined with leaner distilled LLMs, things could get interesting.
Would be great if we could use AI to get rid of the Demorat party
Would be great if AI could help MAGA grow brains. Oh, wait... that's your plan, isn't it? You're going to replace shit for brains with AI slop. Wow! Big step up. What a triumph! /s
If you ask me, I think China has the foundations for achieving AI supremacy in at least a year or two. Unlike the United States, China has been investing rapidly in a bigger high voltage transmission lines and power generation infrastructure in general… Well, also not increasing their carbon output. I think a lot of investors are going to be shocked when say the people in the greater Memphis metropolitan area finally have enough in dealing with the fumes from Elon Musk data center there which is spewing CO2 across the city, causing health issues… the public backlash to the asthma causing machine and children will be severe The United States is in a sprint drinking energy drinks while China is jogging drinking water and Gatorade when it comes to the fundamentals of the Hardware necessary for AI
Carbon dioxide (CO2) does not cause health issues, or asthma. It does cause global heating, though. Data centers are the marginal new power user / builder. They're responsible for most of the new demand and most of the new power generation being built, but they're still a small fraction of the overall power use in the U.S. The new power generation they're paying for is almost always cleaner and more efficient than legacy plants. Getting upset by data center pollution is mostly for nutters. People who care about pollution should be focused on shutting down the legacy plants, taking advantage of how the data centers are lowering the costs for new solar/wind/battery/natgas/nuclear.
Long story short, AI not gonna stick around for long once globalization fully dies. Remember these chips burn out every 3-5 years and needs to be replaced.
Only if you run them at max power & bad cooling... Multiple studies now have shown the mean time to failure for advanced GPUs when setting power caps and undervolting the chips is 10 years provided you can keep the vram memory and the hotspot on the GPU cores under 70c...the AMD r9700 Pro AI GPUs I run on my local workstation have a power cap at 210w each (factory stock is 330w) and an undervolt of -75mv (enables high processing speeds at lower power, provided it's stable) and my total performance decrease compared to stock is only 6.7% while at the same time keeping the hotspot and vram max at 68c. Whether or not data centers are doing this is unknown, but its definitely is possible to extend the shelf life 2.5-3x of reported figures.
There is something to be said about doing more with less...because of the constraints imposed on them with the export bans, through a combination of learning how to reverse engineer the logic of US models (distillation) and genuinely inventive novel solutions they are making their models FAR more efficient with similar capabilities to US labs with a rapidly narrowing gap. An open source Chinese model I'm currently using which is a developer technology preview called Qwen 3.8 Flash Next, does BETTER in an advanced software coding benchmark (Terminal Bench 4.0) than Claude Opus 4.8 Max. The reason this is important is threefold: 1. The benchmark was released AFTER the models were released, meaning both Opus and QFN did not cheat: neither model trained on the test while it was being developed. 2. Anthropic does not publicly release the size of their models in regards to parameters (more parameters allow the model to store more facts, understand more complex reasoning), but it's estimated within silicon valley at anywhere from 3 to 5 TRILLION parameters for Opus 4.8...Qwen Flash Next is 125 BILLION parameters, and beat it in one of the hardest coding tests out there, with 24-40x less parameters which is crazy but not actually the craziest part. I run that model on a desktop computer that cost me about $4500 all-in. To run Opus 4.8 on Max effort, you need roughly 150,000 of the best Nvidia GPUS, running in a 300 megawatt data center, that cost some UNGODLY amount of money (anywhere from $10 to $15 billion PLUS operating costs). 3. Opus 4.8 was released on May 26th of this year as the state-of-the-art US model. Qwen 3.8 Flash Next (again a free open source developer preview) was released August 28th, just 3 months after. The thing Peter and many others underestimate, is that once the intelligence is good enough in these Chinese models, they will be able to use AI to design their own chips AND how to manufacture the hardware to build them. Meanwhile they are creating artificial backlogs for power transformer and plant equipment for US buyers to slow our data center builds down as a middle finger for chip bans we imposed on them. The world we are about to enter is going to be simultaneously incredible and chaotic at the same time. Buckle up.
You are absolutely right. I find Peter's 's reflections on technology at best behind the times but mostly woefully inadequate. How do you keep up with the rapid technology developments if you spend your days hiking in the wilderness? I still like watching his videos and especially reading the comments. He is still superb on geostrategic insight.
My takeaways from the video (all with a grain of salt): it might take china 9 datacenters to have the same output as 1 datacenter in the us - but from that they gain - the hardware knowledge how to make the chips themselves - something no other country is doing (everybody else needs the cooperative work) - the adjusted software knowledge - and as we saw with the initial deepseek stuff huge gains can be achived when you are ressources deprived and **have** to optimize - from both the above comes independence - they can spit in the soup of openai, google, anthropic etc. - by spitting in the soup they can hit the US economy Since they invested HEAVILY into wind- and solarfarms (contrary to the US right now - they hinder/pay of new wind farms in the sea) they have ample power - which is quite a problem in the US (from what we hear) - so they **can** build 9 datacenters for every one in the US. It costs them dearly - but the above described gains might outweigh it for them...
Exactly, they have insane amounts of spare power reserves to pull from while at the same time making their models FAR more efficient in the amount of vram required for the intelligence capabilities. And now we're in a spot in the US where China is creating artificial backlogs for the power plant and transformer equipment we need...sucks to be on the other end of export controls!
Annoying when wind noise wipes out chunks of audio, forcing us to join the dots.
American's can't get these chips either. At least not you or me without significantly overinflated costs. It's literally become a billionaire tech-bros only economy. You're spot on about the ethics of these people too. At this point, I'm almost hoping that AI is a bubble of malinvestment ready to pop.
I'm not sure the average Joe could ever afford any advanced technology in its early stages. We have (paid for) access, though, which is cool, and the things are affordable enough that companies and universities can buy them. Most folks aren't going to want to run their own cluster anyway; that's. a dedicated hobbyist sort of thing.
You do not need advanced hardware if your software performs on par with the competitors while consuming less power. So they are focusing also on software component of of AI while American techno-kings can allow themselves to relax as hundreds of billions are simply given them with no restraint-why bother then.
Fun fact. Little over 3400 Chick-fil-As and "The U.S. has roughly 4,000 existing data centers, while some 3,000 more have been announced or are under construction, according to Apollo Global Management"
Yes
Who's the little visitor at the 1 min mark?