CEOInterviews.AI
Start App
Get an email when Eric Schmidt gives a new interview
One email per new interview. Unsubscribe in one click.
Eric Schmidt
Chief Executive Officer, Relativity Space

Eric Schmidt: US–China AI Race May Be Decided by Power, Not Just Chips|Former Google CEO: The US–China AI Race May Be Decided by Power, Not Chips

📅 Nov 17, 2026 New SciTech 新科技 30 MIN 512 VIEWS 82 SEGMENTS · 2 SPEAKERS
Eric Schmidt: US–China AI Race May Be Decided by Power, Not Just Chips Eric Schmidt provides an in-depth analysis of the next phase of AI: self-improvement, AI agents, the U.S.-China compute race, and whether humans can maintain control over intelligence that surpasses their own. Former Google CEO reveals the truth about the US–China AI race: what may decide the outcome is not chips, but electricity! Eric Schmidt deeply analyzes the next phase of AI: self-improvement, AI agents, the US–China compute race, and whether humans can control intelligence more powerful than themselves.

What Eric Schmidt said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Eric Schmidt, CEO of Relativity Space, spoke at a conference about AI's rapid advancement and its implications. He credited Stanford, SRI, and Xerox PARC for driving global economic growth, noting that 12% of the world's stock market stems from their innovations. Schmidt described recursive self-improvement in AI as accelerating, with models like Fable 5.2 and Astra 6.2 showing historic progress. He warned that open-source Chinese models can be stripped of safety guards, enabling dangerous queries like building ballistic missiles. He argued that AI extinction claims are overstated, citing a professor's math requiring 13 uncorrelated mass events. Schmidt emphasized the need to secure digital infrastructure, proposing federal funding to rewrite open-source C in provably secure languages. He also discussed the US-China race, noting China's power advantage but chip shortage, and predicted AI will replace or augment white-collar work within two years.

Key takeaways

  1. Within one year, most research will be AI-based; within two years, any computer task can be done by AI.
  2. Open-source Chinese models can be stripped of safety guards, enabling dangerous queries like ballistic missile specs.
  3. Claims of AI wiping out humanity are false; 13 uncorrelated mass events would be needed.

Numbers and commitments

FigureWhat it refers toTypeAt
12% World stock market value from Stanford innovations metric 1:06
10 million a year Average salary for new AI PhD in machine learning metric 4:15
1 gigawatt Computing power needed for AI lab metric 4:51
50 billion dollars Hardware cost for 1 gigawatt price 4:51
130 gigawatts China's new power added last year metric 15:52
13 Uncorrelated mass events needed to extinguish humans metric 26:30
1,000 people Casualty event size considered very large metric 27:10

Chapters

  1. 0:00Stanford's economic impact
  2. 2:07Recreating research labs
  3. 6:43AI recursive self-improvement
  4. 10:16Open-source safety risks
  5. 15:52US-China power race
  6. 16:39Transformer history and OpenAI
  7. 20:46Long reasoning and white-collar jobs
  8. 26:28AI extinction claims debunked
  9. 27:50Securing digital infrastructure

Questions asked in this interview

4
  1. 1:50I mean, we do keep hearing about, you know, SRI and, you know, Park and Bell Labs and all these think tanks where people did nothing but bring the greatest minds together to try to invent the future. Does that happen anymore?
  2. 12:42Aren't they all desperately trying to get to recursive self-improvement and take people out of the loop?
  3. 15:46So you think the craze that we're in now has any reasonable way of ending?
  4. 16:39Am I simplifying this too much if I say that Google developed the transformer part of ChatGPT, transformer, and then OpenAI was founded based on that technology? No?
David Pogue 0:00 ↗
Hi, David. Good to see you. Good to see you again. Thank you for joining us. You had a similar impact on my mother.
Eric Schmidt 0:06 ↗
Oh, please. [laughter] We've got lots of time.
Can I say something to you all? I think I owe you thanks, and I think the world all owes all of us thanks. I am in awe of the sequence Stanford way back when creating the Stanford Research Park, creating SRI, spinning it out, and so forth. A simple calculation will tell you that virtually all economic growth in the world is now driven by the work that you all did over the last 80 years. It's extraordinary.
David Pogue 0:44 ↗
Wow. Think about it. [applause]
Eric Schmidt 0:48 ↗
An example would be their estimates that stuff that came out of Stanford in one way or the other, and again, Stanford is the origin. So thank you, John. This all occurred before you were born. [laughter] Sorry. Sorry, John. We're good friends.
It's on the order of 12% of the world's stock market today. I think a reasonable reading of technology rates is that number is going to go up quite a bit. Right? I can't put a bound on it, but we have economists here who can think through this. We have no models for this level of scale.
So think back to when Stanford was really a farm where they had all that extra land, and these people had this extraordinary vision. And when I worked at Park, people talked about Engelbart in these terms that are hard to describe.
David Pogue 1:45 ↗
I think you have under-marketed Douglas Engelbart. That's how brilliant he really was.
Wow. I mean, we do keep hearing about, you know, SRI and, you know, Park and Bell Labs and all these think tanks where people did nothing but bring the greatest minds together to try to invent the future. Does that happen anymore?
Eric Schmidt 2:07 ↗
Um, it should. Um, since I was there and I was very young and I just sort of did what I was told, uh, I didn't really—I did not understand what was going on. I was 22, 23; the people seemed immensely old to me at 31, 32, [laughter] 33, immensely accomplished. It's hard to imagine.
Um, and I was part of the original office demo, and I built a programming language for them and stuff like that as a researcher.
Um, what is true about SRI and is true at Xerox is the people did it because they wanted to invent something. This was a very different California. We're used now to the wealth and the cars and so forth. When I was there way back when, everybody seemed kind of the same. They dressed the same. They all—I mean, maybe this is just me being young, but we all kind of had the same sort of California liberal values, right? Very open. The weather was nice, right? Um, and there was this feeling of we're just inventing stuff.
There was no connection at the time between what we were inventing and business and economics. And it was Steve Jobs' famous visit to Park that started that, because of course he was both a truly brilliant human being, but also a brilliant businessman, and a few other people. And it's interesting to note that Xerox PARC then created Xerox Corporation, took technology out of Xerox PARC and put it in something else which failed, right? And they did it in El Segundo, and I was part of all of the machinations there.
And the people who came to PARC that I worked with came out of SRI and a failed startup called the Berkeley Computer Corporation, which was created by these entrepreneurs—um, let's just call them scientists—who were terrible entrepreneurs, but they were just brilliant, right? They couldn't raise the money and so forth. The money wasn't available. So the change over 50 or 60 years is profound. In my career, I've tried to set up a set of research labs.
Because I also had the privilege of working at Bell Labs when it was very famous at the time. It's very hard to recreate these now. So let's imagine you and I—we know lots of people in AI, and we want to create an AI research lab. Okay? Good question. First, what salaries do we have to pay? [laughter] Okay? What's the average salary for a new AI PhD in machine learning? Mm, 10 million a year? When I worked at Xerox, my salary was 55,000. Seemed like plenty.
Um, another example is we need computing, right? How much computing do we need? Well, I don't know. 1 gigawatt is 50 billion dollars of hardware. So I guess we can't really raise that. The only person who can do that is Elon. [laughter] You get the idea. So there are real structural limits to recreating the magic that I am so grateful for.
David Pogue 5:14 ↗
You know, I'm glad you brought up AI. I do a lot of reporting on AI for the show I work for, CBS Sunday Morning, and I find that you have a very thoughtful approach toward the doomsday topic. I mean, you have incredibly knowledgeable people here. This came in as we sat here. This was at a conference today, a guy named Paul Christiano. He's the head of safety and security for OpenAI. He told a conference this evening, for most people in this room, the most likely cause of death is an AI accident. He works for OpenAI. So you have these guys, and then you have these other people who say that's a hoax. It's ridiculous. Like there used to be this guy named Jensen Huang. Um...
Eric Schmidt 6:05 ↗
David, David, as an aside, I heard today that your show, CBS Mornings on Sunday—
David Pogue 6:13 ↗
Sunday Morning, yeah.
Eric Schmidt 6:14 ↗
—is now equal or ahead of ratings of 60 Minutes. Did you know—
David Pogue 6:18 ↗
Well, I did know that, yeah.
Eric Schmidt 6:21 ↗
That's a pretty—at least for our generation, that's like a huge accomplishment, David.
David Pogue 6:25 ↗
Well, I don't know if it's us going up in numbers or 60 Minutes [laughter] going—
Eric Schmidt 6:32 ↗
Uh, I think in your case it's going up. Um, so [laughter] to answer the question, I have to spend a little bit of time talking about the technical problem. I'm sorry.
Um, the rate at which recursive self-improvement is causing intelligence gains is so shocking I can't keep up with it. It's occurring now, this month, here, right? And the easiest way to do it is take a look at Fable 5.2, which just came out on Monday, I guess two days ago, and look at what it can do.
And if you don't believe that, look at the math gains. And by the way, I heard that Astra 6.2, whatever is coming out right before the Anthropic IPO, just for fun, is expected to be equal or beat in some areas. This race between those two workhorses is historic. People a hundred and a thousand years from now, excuse me, will write books about this 12-month period. This is the moment when the slope gets set.
Here are some examples. Um, in the internal calculations in OpenAI and Anthropic, they believe that their AI improvements are doubling every two to three months. You can do the math. Everyone in this audience can actually do math. Um, just do the combinatorics. And there's every reason to believe that the slope of that is exponential, that in fact the rate will improve.
The limits to this growth are, um, hardware. And also there are plenty of bottlenecks. So what happened in olden days, which is like three years ago, we had a straightforward pre-training, post-training regime with basically some fine-tuning. Today you have recursive loops where the system is improving itself. It started with improving the boring stuff like the data curation and all of that kind of stuff and self-testing.
But now the objective function is such that you can set the objective function and then you can actually just experiment with that efficient frontier. So the typical way it works in one of these labs is the men and women who do this, they have a theory: if I change this, it'll affect that. You know, if I add more math, it'll improve more cooking, you know, or something like that, some very complicated question. They set out a large army of agents overnight that do a great deal of calculation. In the morning they look at the results, then they think of a new way of attacking the problem and they set their agents off.
The question is at what point does liftoff occur? And liftoff is when it's a full recursive loop. It's not there yet for many reasons. Furthermore, we probably don't want to let full liftoff occur. We want a human. At the moment the humans are the bottleneck because the humans have to read the code. Most of the code is incomprehensible, right? So the human has to actually roughly understand what the model is working on.
So a reasonable prediction, and these are public claims, is that within one year the vast majority of the research, if you just look at the trend lines, will be AI-based. That is, the objective function will be set and so forth and so on, and when the research result, it'll be immediately incorporated, typically like a one-day, two-day kind of cycle. The general prediction, again public prediction, is that within two years anything that you can do on a computer can be done by one of these systems. Okay? Now, let's think about that for a while.
For various reasons, I do military work. And for various reasons, I got interested in building a ballistic missile. I'm pleased to report that it is illegal to build your own ballistic missile in America. [laughter] So, if you use the models, they've all been fine-tuned so that essentially they can't answer that question.
However, with Chinese models which are open source, open weights, open source, they have their safety—you can take it out. And now it's relatively easy to remove these safeguards. So, my question was: I want to build a ballistic missile. Not only does it give me a spec, it says, 'Would you like me to tell you how much it will cost?' I said, 'Sure.' And they said, 'Okay, would you like me to show you the layout of the factory?' And I said, 'Sure.'
So then it said, 'Well, do you want to understand the DAX, which is essentially the control system that does the targeting?' And I said, 'Sure.' Now, I know nothing about ballistic missiles, by the way. And I said yes, and it kept going. And I realized the thing that was frightening to me was that it knew what to do next.
And if you look at—this is a long-winded way of answering your question—the ability for these systems to unleash scalable bot farms of cyber attacks is of great concern in the industry, largely based on open source. And speaking as a person who started in open source before open source had a name, we got to solve this problem without losing open source. So that's why Paul, who's very, very smart, is saying this, right? Now, there are plenty of mitigations. I'll give you a simple example.
If you don't understand what the AI is doing, don't release the model, right? Not that complicated. If you can't read the code, don't release the model. If the agents start to communicate with each other in language not English [laughter] or human language, don't release it.
And this morning, as an example, I have a group that talks about this. Someone said the correct solution—this is how serious this is—is we wait for some terrible thing to occur. Okay, I'm just making his argument, not mine. And he goes, and then the government will order the model training to stop for a while so we can have a think. That's, I think, where we are. And that's why Paul said what he said.
David Pogue 12:42 ↗
So right. So I too have reported on this recursive self-improvement problem, absolutely terrifying. But unlike me, you presumably know some people at Google who are working on this, right? Aren't they all desperately trying to get to recursive self-improvement and take people out of the loop?
Eric Schmidt 13:02 ↗
So recursive self-improvement is the technical term for the model improving itself in one form or another. There are debates in the industry as to what consists of RSI. So I'll give you two examples: inference time and training time. So you could take a model and run it, right? And then you could take its output and then take the output and improve the output with a technique called hill climbing. We'll call that inference time RSI.
The consensus is that that does not go to this exponential, because to get to the full exponential, you have to have the actual intelligence getting smarter. The power laws, the scaling laws are so far still true that with enough hardware, enough data, and enough algorithms, you get emergent behavior.
And the strangest thing about this front end, let's call it pre-training, post-training improvements, IQ, is it's really IQ. So my observation as a person who worked at SRI and Xerox is they were full of all these people smarter than me. And what annoyed me about them was that they were good at what I did and were good at everything else too. [laughter] Do you have the same experience? You know, they're just brilliant at math and they write well. [laughter] Like, that is not okay. You should be good at one and not the other, right? Somehow the processors were just better.
And I think that's what we're seeing. So to me, this question about how fast does that training occur is the key thing about RSI. Google, of course, did it with DeepMind, invented a lot of what is called reinforcement learning. So the way it technically works is you have the LLM, which is essentially a deep neural network. And we at Google hired all these people. So I guess we'll either—we invented it or we hired them and then they all left.
Um, and then DeepMind and company invented reinforcement learning, and reinforcement learning is where you go through all forms of scenarios. So one of the questions I was asking my technical friends is, how is it that the math is so good and the physics is so good? And the answer is that the model itself is very, very smart. I mean, it can do chemistry and math and write and so forth. It's one of these annoying human-like people that I just hated because they were so much smarter than I.
And the RL allows it to become specialized, and it specialized in knowing which tool to apply in which order. That is a scalable architecture. So what's the limit? Hardware. Right? What's the limit? Power.
David Pogue 15:46 ↗
So you think the craze that we're in now has any reasonable way of ending?
Eric Schmidt 15:52 ↗
The only limitation is financial. We in America have enough land in the middle of nowhere without people in it, right? Please do not build these things next to LA and next to New York. We can build these large power plants and data centers and we can win. And if you're confused on this, just remember that China last year did 130 gigabytes—gigawatts, excuse me, sorry—of new power, solar-based, right? And we added very little. Now we're accelerating that, and ours, of course, is natural gas. So they have plenty of power. They just don't have the chips.
Uh, furthermore, if you look at the Chinese, they're so good at building stuff. Their data centers look to be as powerful—if they have the chips, which they don't. Cheaper, better, faster, faster construction. It's a race.

18 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Schmidt, E. (2026, November 17). Eric Schmidt: US–China AI Race May Be Decided by Power, Not Just Chips|Former Google CEO: The US–China AI Race May Be Decided by Power, Not Chips [Interview transcript]. New SciTech 新科技. CEOInterviews.AI. https://ceointerviews.ai/interview/3317205/

MLA

Eric Schmidt. "Eric Schmidt: US–China AI Race May Be Decided by Power, Not Just Chips|Former Google CEO: The US–China AI Race May Be Decided by Power, Not Chips." New SciTech 新科技, 17 Nov. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3317205/.

BibTeX
@misc{schmidt2026_3317205,
  author       = {Eric Schmidt},
  title        = {Eric Schmidt:  US–China AI Race May Be Decided by Power, Not Just Chips|Former Google CEO: The US–China AI Race May Be Decided by Power, Not Chips},
  howpublished = {Interview transcript, New SciTech 新科技. CEOInterviews.AI},
  year         = {2026},
  month        = {nov},
  url          = {https://ceointerviews.ai/interview/3317205/},
  note         = {Speaker-attributed transcript with timestamps}
}