The US-China AI Gap Has Narrowed to Six Months, Reveals Recent Conversation with Former Google CEO Eric Schmidt

The AI Gap Between China and the US Has Shrunk to Six Months! Former Google CEO Eric Schmidt Reveals in Latest Conversation..

Behind this change of tone, several things left a deep impression on him. China doesn’t have the most advanced chips, so it finds ways to make models cheaper; frontier performance is still being pursued, but deployment speed is already fast; open‑weight models cost less, and in markets that cannot afford to pay high token prices long‑term, they are actually easier to roll out.

He had just come to China a few months ago, and after watching, gave a very straightforward assessment: the United States is better at inventing cutting‑edge things, while China excels at actually putting them into products, machines, and real‑world scenarios. This led him to a judgment: the strongest models in the future may still come from the United States, but the ones that ordinary people around the world end up using may not be the strongest. Below, enjoy:

Source | IN咖

Layout | 唐唐

Editor | 艾暄

“I tried to use DeepSeek, but US law does not allow it.”

10Oct22th, in San FranciscoMasters of Scaleat the summit, in a conversation, formerCEOEric·Schmidt(Eric Schmidt)said.

He said that when the restrictions on China’s access to advanced AI chips, they thought these measures would be effective for about a decade, but now the gap between China and the United States in the AI field may have narrowed to just six months.

Schmidt’s assessment is worth a look.

2001, he joined the three‑year‑old Google asCEO, over the next ten years, together with founders Larry·Page, Sergey·Brin together, grew Google into a global technology giant.

After leaving Google’s frontline management, his influence gradually extended to the US technology and national security policy arena.2016 he was appointed the inaugural chair of the US Department of Defense Innovation Board, later serving as chair of the US Artificial Intelligence National Security Commission, participating in advancing USAI, advanced semiconductor and national security strategy formulation. Now isRelativity Space‘s CEO.

2021, the committee he led released a final report of 756 pages, systematically discussing how the United States can maintain its global leadership in artificial intelligence and how to respond to China’s technological catch‑up.

Schmidt is also one of the earliest figures in the US tech community to take China’s AI competitiveness seriously.

As early as2017, he warned his American counterparts that China’sAI development pace should not be underestimated.

2025YearDeepSeekAfter DeepSeek’s rise in 2025, it further reassessed the actual impact of the U.S. advanced chip restriction policy and published an article discussing China’s shift from a technology follower to an innovator.

In this latest conversation, he not only affirmed China’s progress in algorithm efficiency and open-weight models, but also emphasized that China willAItechnology into practical applications.

“We are the side that invents cutting‑edge technology, while they are the ones who actually deploy it.”Schmidt, who visited China just a few months ago, suggested that American hardware entrepreneurs should go to China and see it for themselves.

He also predicts that over the next decade, thanks to lower costs and broader applications, ChineseAImodels could achieve a larger global user base.

“<span class.

Oops error. Need correct HTML.

“In the end, most people in the world will use Chinese models.”

However, Schmidt has not changed his stance of hoping the United States maintains its technological lead. But he also believes that the US‑China AI competition is not a zero‑sum game.

This roughly 27‑minute conversation also shows how a long‑time key figure in shaping U.S. technology strategy is reassessing China’s AI competitiveness. The host is the well‑known U.S. business mediaFast Companyformer editor‑in‑chief Bob·Safian(Bob Safian).

The following isSmart Investor(ID: Capital-nature)the latest compiled conversation, shared with everyone.

01

From scienceAI gap between China and the United Stateshas narrowed to about6months

 Safian  Focusing on China’s AI development, we often hear heated rhetoric that makes it hard to discern the truth. Therefore, I’m glad to invite you today to help us clarify these issues.

The United States originally had a clear first‑mover advantage in artificial intelligence, but now people are beginning to doubt whether that advantage can be sustained. You once said that the gap between China and the United States in AIhas narrowed to about six months. What significant changes have occurred recently?

 Schmidt  The biggest change that occurred in the past year, and something everyone here should notice today, is that the cost of software has dropped from being very expensive to almost zero. When the price of something falls dramatically, its quantity increases dramatically.

So, when I think about the coming years, what comes to mind is: we will have an enormous number of software applications. What will these software do? Who will develop them? How will they collaborate with each other?AIHow will intelligent agents operate? And so on.

China is also pondering the same question.

In my view, the US‑China competition is the most decisive competition for your generation, our generation, and our era. Because both sides are sparing no effort to develop reasoning and intelligence systems that surpass human capabilities.

This does not mean that these systems will become like humans; rather, it means they will possess abilities that surpass human capabilities, i.e., so‑called superintelligence. In fact, in some respects, such abilities have already emerged.

Take a look at those outsideWaymoautonomous vehicles. Their visual perception is better than that of human drivers, especially because they don’t drive under the influence. You get what I mean, right?

Back to US‑China AI competition.

A few years ago, I was also among those pushing to restrict China’s access to advanced chips. At the time, I thought these restrictions would be effective for about ten years.

A year or two ago, China seemed to be about one to two years behind the United States. 6 months or so, and many studies support this assessment.

How did they do it?

After all, didn’t we restrict China from accessing advanced chips? That’s true. But the Chinese are very clever; they have found ways to use the algorithms and knowledge we developed to achieve comparable technological capabilities with roughly one‑tenth the energy consumption, even though there are still gaps in some performance metrics.

Now, the entire industry is debating whether this gap will continue to shrink in the future or widen again.

But at least for now, it can be fairly said that this is already a competition between the two superpowers of China and the United States. In my view, this is the most decisive rivalry of our era.

Competing for leadership in the intelligence arena is, without exaggeration, a competition of millennial magnitude. Once such intelligence truly emerges, our children, grandchildren, great‑grandchildren, and countless future generations will be fundamentally shaped by the changes we see today.

 Safian  Some people think that constantly emphasizing China’s progress inAIfield as merely an excuse to delay regulation. Do you think this claim has any merit? Are you concerned about this issue?

 Schmidt  People always come up with all sorts of explanations.

I was deeply involved in discussions between China and the United States. Before Dr. Kissinger passed away, I worked with him to promote communication between the two countries. When he made his final trip to China to meet national leaders, I accompanied him.

Chinese people are well aware of the nature of this competition and understandAImay bring, but they will not stop.

So, one way to answer your question is to ask the opposite: if we want to slow down the development speed of AI, what exactly should we do?

How many cutting‑edgeAIlabs? Five, six, seven? Roughly that many. China? Also five, six, seven. As a result, you are dealing with a problem involving about12participants.

In China, in theory the government can order these labs to act according to government requirements. But in the United States, you have to get the relevant companies to reach some kind of voluntary agreement.

Thus, this issue involving 12 parties, was simplified to a problem between China and the United States.

The next question is, since China already believes it is falling behind, how could we possibly persuade them to slow down? That is simply impossible.

We must find other ways, while continuing to advance technology development and maintaining China’s leading advantage, to ensure the safety of this technology.

02

I once tried to invest in DeepSeek, but U.S. law does not allow it

 Safian  In some respects, China has structural advantages. The government can designate AI as a national strategy, guiding all parties to move in the same direction. Whereas you previously thought that the United States’ advantage lies precisely in its diversity and decentralization. These are two fundamentally different development models.

So, the Chinese AI industry’s actions—are there any that American leaders should learn from?

 Schmidt  I’ll give you a simple analysis framework.

China’s energy supply capacity is very strong, and it has a large amount of renewable energy.Ten years ago, the United States also said it would achieve this, but we failed.

China does not require companies to wait up to seven years to connect to the grid, and there is no comparable debate here. They are building data centers on a massive scale, and the scale is enormous.

Moreover, even though China’s financial markets face more restrictions, China can still provide financing for startups.

It is illegal for Americans to investDeepSeekis illegal, I know this because I have tried it myself. Trust me, this is indeed prohibited under US law.

（The U.S. Treasury’s foreign investment security rules took effect on January 2 2025, imposing prohibitions or filing requirements for U.S. persons investing in certain Chinese artificial‑intelligence, semiconductor, and quantum‑information‑technology companies.）

DeepSeek is currently seeking financing. From what I understand, they recently completed a large financing round, and the funds appear to be secured. My Chinese friend told me the money comes from the private sector, and they believe the business is economically viable.

There is also a very complex open‑source issue involved here, which we can discuss later.

But fundamentally, China has been moving forward on its own path. What they lack most now is chips.

Here in the United States, we have everything we need. We have brilliant talent, a wide range of debates, and an astonishingly deep financing market.

Is there any other place in the world where you can raise $50 billion just by sending a text message? Apart from the United States, probably no other place can do it.

Thank heavens, the US financial industry is providing funding for the hobbies of people like us. As for the result, we’ll wait and see.

But the United States lacks energy, and also lacks data centers. What is the right solution? Build your own power supply system.

There is a saying reportedly from Tocqueville:“America is great because of its people’s optimism, its vast lands, and the absence of a king.”

At least we can still measure ourselves from these three aspects.

The United States indeed has a lot of land, and many areas are vacant. So we can build data centers in those places. This benefits the local economy and also allows us to build supporting power generation facilities ourselves.

That’s what we need to do.

According to current estimates, by 2030 or 2031 we will need about 60 to 80 gigawatts of power supply capacity. Whether sufficient funding can be raised for these projects remains to be seen.

Based on my estimate, for each 1 GW of power demand, the related infrastructure would require an investment of roughly $50 billion. By the way, 1 GW is roughly equivalent to the capacity of a large nuclear power unit. And as far as I know, the United States does not build nuclear units that quickly.

03

China chooses open source, the US chooses closed source, which path will win?

 Saffian  You just mentioned open source. China’s leading position in open‑source models—or rather, open‑weight models—has propelled Chinese AI’s global dissemination. For the business leaders present today, is developing on top of China’s open‑source models a wise choice?AIin the world. For the corporate leaders here today, is it a wise choice to develop on top of China’s open‑source models?

 Schmidt  Let’s first clarify the concepts a bit.

First, open source(open source). When I was at Berkeley, I took part in the early development of the open‑source movement and have always been very proud of it. That was a long time ago, but the code I wrote back then is still open source today. If you want to see those very old pieces of code, you can still find them.

In addition, there are open weights(open weights). Open weights refer to the practice of investing substantial funds to train a model and then releasing the resulting parameters publicly. These parameters are extremely large, potentially reaching the scale of trillions. (Open weights do not necessarily imply fully open source.)

China’s strategy is largely open‑source and open weights, while the United States’ strategy is mainly closed‑source and closed weights.

From the perspective of a U.S. company, that of course makes perfect sense. Training a model costs1hundred million to5billion dollars, right?

 Saffian  Yeah, since we spent so much money, why give it away for free?

 Schmidt  That’s what lawyers would say.

I’ve argued with lawyers for 50 years. I’ve always believed that you should offer some things for free, because that can build a platform.

The basic idea of the open‑source movement is to first build a platform, allowing others to develop their own things on it.

Closed-source movement is different. In the early years Bill·Gates and Steve·Ballmer‑led Microsoft is a typical example. They focus mainly on end users; whether you make the product well or not, ultimately the users will judge it.

In the United States, open source has largely won and has become the foundation of much of our technical work today. However, most largeAImodels remain largely closed source. There are good reasons for this, and I am not criticizing them.

Of course, there are also some different examples.

For exampleReflection AI just released a model with 500 billion parameters. NVIDIA’s Nemotron has a model of comparable scale, and both are very powerful. Google also has a model called Gemma 4, a very powerful smaller-scale model. All of these models are available for download.

So, this is not simply a straightforward confrontation between China and the United States.

China chose open source and open weights because it is the most sensible strategy for them. They were originally at a disadvantage, needed to acquire users quickly, and were willing to bear the risk of their technology being taken by others.

So far, this approach seems to be working, at least they have secured funding.

Therefore, the competition between open‑source/open weights and closed‑source/closed weights will become a crucial factor in shaping the future landscape of the AI industry.

If you ask the open‑source companies or the closed‑source model companies in San Francisco, they’ll all tell you that their strategy is better because they believe they can stay ahead.

If you ask Chinese people, they will also tell you that their strategy is better because they believe it can keep them ahead.

So, what we are facing is a genuine competition.

However, open‑source models and open‑weight models also face a problem that is difficult to solve. We have devoted a lot of effort to this, but so far, I think we have not found an effective solution.

Just think about it: closed‑source models are in the hands of companies, and those companies are bound by the laws of the countries where they operate,the CEO must also obey the law; whereas in China, the government has strong control over AI labs.

But open-weight models are different.

Once model weights are made public, even if developers set safety restrictions, it is difficult to control how others use these models.

There is now a technology called“Safety Mechanism Removal”technology, which can essentially strip away the original safety constraints of a model, yielding an unrestricted model. These models may be used for network attacks, and developers can hardly control where they end up or who uses them.

What is even more concerning is that software costs are dropping rapidly.

Do you remember what I said at the beginning? Software costs are approaching zero. In the coming months, the expansion speed of software capabilities will be astonishing. That’s why tech workers in San Francisco and elsewhere are very concerned, and some are starting to ask whether we can letAIthe development speed slow down?

But as we discussed earlier, AI labs in both China and the US slow down simultaneously, it is almost impossible.

This is a very tricky problem. Many people think that in the future12 months, some kind of cybersecurity incident may occur, making everyone truly realize the seriousness of this issue.

 Safian  So, if I am a corporate leader who decides to use Chinese models because they are cheaper for me, does that count as being unpatriotic? Is it lacking long‑term vision?

 Schmidt  the CEO of an American company is, to some extent, patriotic, but they also care a lot about money.

There are already many such examples: people use small servers to run open-weight models to solve specific problems, becausethe tokencost is lower.

If you deploy the model yourself, the token cost can be reduced to about one‑third to one‑fifth of the original. Of course, these models are not as powerful.

So, I’m concerned that in the future we might see a scenario where the most cutting‑edge models dominate the fields we care about most—such as mathematics, chemistry, and other incredibly significant breakthroughs.

For mathematicians, this is nothing short of an apocalyptic disaster! The fundamental landscape of their entire world, in the past50years had hardly changed, even dating back to when Gödel proposed the relevant theory. But suddenly, this world was overturned within a month. They will face a crisis concerning the meaning of their own existence.

This is the change that cutting‑edge models will bring.

But for most of us, solving financial accounting problems doesn’t require such advanced mathematical skills. Trust me, financial accounting isn’t that complicated.

Therefore, in the future many users may adopt these cheaper open-weight models to handle a variety of everyday tasks.

Is this okay? In the United States, this is likely legal.

But is doing this a good idea? That depends on how you view risks such as geopolitics.

Of course in China, people use open-weight models.

04

US‑China AIScoring the competition item by item: China may lead in deployment speed and global influence

 萨菲安  We have prepared a scoreboard. I will present several metrics, and you decide who has the advantage: China or the United States. Let’s start with model performance.

 Schmidt  USA. According to current benchmark tests, the United States is leading.

One way to understand this is that each AI lab has already released models, and also has next‑generation models under development.

From the information disclosed by these labs, the performance of the next‑generation models will be outstanding.(scaling law)still holds, and even more astonishing breakthroughs are forthcoming.

So, for the model performance item, I choose the United States.

 Safian  Okay, the next item is model cost.

 Schmidt  China.

As I just said, Chinese models’Tokencost is at least three times cheaper.

In many ways, this is actually because they were forced to cut costs. We didn’t provide them with high‑performance chips, so they turned to algorithmic improvements, supervised fine‑tuning, and some other very impressive technical approaches.

Interestingly, once Chinese people invent something in the open-source field, Americans immediately adopt it. Thank you, China!

 Safian  Okay, next is energy. I guess this item should be China, since you mentioned it earlier.

 Schmidt  Alright, you guessed right.

I recall that China’s relevant energy construction figure last year was about 560 gigawatts. That’s an astonishing number.

This is the benefit brought by national industrial policy. If they need energy, they will build on a large scale, and the vast majority of it will be renewable.

 Safian  Okay, let’s look at another metric about leadership and talent.

 Schmidt  I will give this item to the United States.

But the United States is doing some very foolish things. I suppose none of you here today would be surprised.

Arrest the brightest people, then expel them from the country or refuse to grant them the crucial visas that would allow them to come to the United States—definitely a terrible strategy.

We give them education, let them earn PhDs, and then we kick them out. Real genius, right?

Really stupid.

 Safian  The next item is security control.

 Schmidt  The United States has a better security control system.

A series of meetings have been held in the past, and I participated in some of them. At these meetings, the various labs agreed to share security practice experiences, etc.

I think, overall, we should expect the government will eventually regulate AI.

It’s not that I welcome regulation; it’s becauseAIwill permeate every aspect of society, inevitably leading to some adverse consequences, and those consequences will drive the introduction of regulation.

Security is the most obvious area among them, such as AI alignment with human goals and related issues.

SafiAn  The next item is deployment speed.

 Schmidt  China. We are the side that invents cutting‑edge technology, and they are the side that truly deploys it.

I just visited China a few months ago, and I’ll go again next month.

AIThe speed at which AI is being applied to everything in China—from digital systems to mechanical equipment—is simply astonishing. Moreover, they love showing you various demo products.

There is simply no equivalent product in the United States.

So, if anyone here is developing hardware today, I encourage you to visit China and see for yourself, because in this field, China is outpacing us.

This item clearly belongs to China.

Safian The last item is global influence.

 Schmidt  If measured by impact over the next ten years, this item belongs to China.

What is the biggest difference between free products and expensive products? Free products will have more users, right?

The vast majority of countries in the world lack sufficient economic strength and wealth to bear the kind of costs we require, includingTokencosts and other various expenses.

I understand why American companies price it this way, and I’m not trying to criticize that.

But for me, one thing has become very clear:In the end, most people in the world will use Chinese models.

This has huge geopolitical significance. Most of the geopolitical experts I interact with haven’t truly realized this yet.

But once you think about it a little, the answer becomes obvious.

On one side are free, cheap, and sufficiently usable models; on the other side are expensive, powerful models, but they mainly serve the elite, wealthy individuals, major enterprises, and top research institutions.

05

AI is not a zero‑sum competition; the US and China can cooperate and advance together

 Safian  We—did we miss something? Is there a possibility for both the United States and China to be winners? Or is this destined to be a zero‑sum game?

 Schmidt  I don’t think AI is a zero‑sum game; China is not our enemy.

I have experienced such a market before; although the two companies competed with each other, they ultimately created an entirely new market together.

Therefore, the ideal scenario is that American elites and Chinese elites can engage in some form of rational dialogue. Of course, under the current U.S. administration, this is clearly impossible.

But if possible, we should sit down and discuss:“How exactly should we manage all of this? Can we reach consensus on certain matters?”

For example, create a “No Surprise Principle”(no-surprise rule).

I learned one thing while working at the Pentagon: when a country launches a missile, it notifies other countries. Those countries then point their satellites at the missile launch.

There are two reasons for doing this. First, to confirm that the launch party indeed carried out the launch as previously described; second, they can also use it to train their own missile launch monitoring system.

This is a coordinated arrangement reached by all parties, aimed at ensuring everyone’s safety. You certainly wouldn’t want a missile to be launched suddenly without anyone knowing where it will go.

This is completely reasonable, and such a system arrangement 50 years ago was already established.

So I think,the United States and China can list a series of perfectly reasonable cooperation items, and the ultimate goal of these collaborations is to achieve human prosperity and development.

Finally, I want to say, AIis the most important thing that will happen in the rest of our lives.

AIis the foundation of scientific progress, and scientific progress is exactly what I am currently working on. Scientific progress creates wealth, and wealth creates opportunities, well‑being, and the various benefits that society can enjoy.

Because of what we are doing today in the field of artificial intelligence, the future lives of our children and grandchildren will be very, very bright.

Yes, we face many problems. But I think the important thing is that we must continue to give this our all.

The speed of technological innovation has become so fast that even I can’t keep up, and I’ve been in this industry for a long time.

06

AI has changed the way I work, invest, and run a business

 Safiian  To the business leaders present today, is there any particular message you would like to convey? If you hope that the United States will ultimately win, what different actions should they take now?

 Schmidt  First, start usingAI.

I especially love this scenario: someone walks up to me and goes on about artificial intelligence. I ask them:“Have you actually used it yourself?”

He said:“No, there isn’t.”

Alright, maybe he’s smarter than me.

But I think you have to use these things yourself.

The experience that truly blew me away was when I started having AI perform long-chain thinking and deep reasoning.

My reaction at the time was:“Oh my god, it really came!”

At that time, I wasn’t sure if I fully understood what this thing in front of me actually was, but I knew it had arrived.

It changed the way I think about life, the way I work, invest, and run a business.

Here’s a very obvious example, forAIsystems, the finance function is actually relatively easy to handle. Why not letAIgenerate financial reports? Then you and theCFOcan discuss directly around that report. That would be really interesting.

So, what about inventory management in manufacturing? Safety management? Cash‑flow forecasting? You might be sitting there thinking:“Oh, these things are so boring.”No, these things are transformational.

Because once your way of thinking changes like this, new ideas will keep popping up. One idea after another, and then another.

This is why being able to live in an era like today is such a precious thing.

You might also want to see:

Is AI killing “traditional companies”?

What is it that 64% of people never manage to overcome in their lifetime?

A 6 billion yuan acquisition that helped a Suzhou father and son rake in 94 billion yuan

AI content enters the experiential era—how Loopit turns every idea into a playable world?