CrowdStrike Founder: "China Has Planted Sleeper Cells In Our Power Grid!" | Dmitri Alperovitch

At a glance

Dmitri Alperovitch, who co-founded CrowdStrike, ran the Operation Aurora investigation that first showed the world what Chinese state hacking looked like, and called Russia's invasion of Ukraine months before it happened, sits down with Peter McCormack for eighty one minutes on the one question underneath everything else: is China preparing to take Taiwan, and what happens to the rest of us if it tries. His answer runs on two tracks that turn out to be one track. On the military side he argues Taiwan is close to the worst place on the planet to invade, that the United States itself drew up a plan to do it in 1944 and then abandoned it as too costly, that 2032 matters more than the 2027 date everyone quotes, and that the whole thing is a decision by one man who cannot afford to lose. On the cyber side he describes a campaign the US government calls Volt Typhoon, in which Chinese operators break into small water utilities, port facilities, electric substations and power plants where there is nothing worth stealing, and then do nothing at all, sitting dormant, waiting for orders. That, he says, is not espionage. That is preparation of the battlefield, and it is the single clearest signal that this time is different.

The second half is the economics and the technology underneath the confrontation: why roughly 90% of advanced chips and over 40% of all chips route through Taiwan, why a conflict there erases $5 trillion of global GDP in the first year and why he thinks that number is low, why the CHIPS Act money arrived two years too late to save Intel from a near death experience, why China is perpetually six to nine months behind on frontier models and why that gap is far too small for comfort, and why the man who built the biggest cyber defense company in the world now argues the US should start stealing Chinese IP right back. It ends somewhere unexpected: two men comparing notes on running six agent sessions at once, £6,000 token months, and what happens to the security industry when the attackers and the defenders are both machines.

The cold open: sleeper cells in the grid (0:00)

The episode opens on the line that gives it its title, pulled forward from the thirty minute mark. Instead of just breaking into companies or governments and stealing secrets and stealing IP, Alperovitch says, China launched a new campaign that the US government calls Volt Typhoon, and that campaign is designed to penetrate organizations and networks where you have nothing worth stealing. Small water utilities. Port facilities. Electric substations and power plants. They get into those networks and then they do nothing.

McCormack: "Just sit there."

"They just sit and wait."

"Sleeper cells."

"Sleeper cells waiting for orders to come in and destroy that network in the event of conflict, or in an attempt to deter us from intervening in a conflict."

That, Alperovitch says, is what the military would call preparation of the battlefield. And it is one of the many signs he looks at and says: this is now different. They are preparing for an invasion of Taiwan.

The other half of the cold open is the consequence. You will not be able to get any microelectronic components for your MacBooks, for your PCs, for your servers, for your cars, for your refrigerators. Everything around us is now dependent on chips. If there is any sort of conflict or blockade of Taiwan, you have $5 trillion of GDP wiped out in the first year. McCormack: "Jesus." Alperovitch: "And I think that's probably low."

America planned to invade Taiwan in 1944 (1:33)

McCormack opens with a fact Alperovitch had teased him with before the cameras rolled. In the last hundred years, two countries have seriously prepared to invade Taiwan. China is the obvious one, particularly over the last forty five years or so. The other is the United States.

During World War II, when Taiwan was still under Japanese occupation and American forces were island hopping across the Pacific toward mainland Japan, one of the proposed plans was to invade Formosa, as Taiwan was then called. The US military developed an actual war plan for it in August of 1944. It was called Operation Causeway. You can look it up online. It is available. And, Alperovitch notes, the Chinese are studying the heck out of it.

The context makes the plan legible. Two months earlier, in June of 1944, the United States and its allies had pulled off Operation Overlord, the Normandy landings, the greatest amphibious invasion in human history. The similarities to Taiwan are striking. The Taiwan Strait is about the width of the English Channel. Both involve landing on well defended beaches. Normandy gave American planners both the confidence that this kind of thing could be done and the hard experience of doing it.

And then look at the numbers Causeway called for. Fifty thousand Allied troops crossed the English Channel on June 6th, 1944. Operation Causeway was going to use over 300,000, double the Normandy force. It was designed as a two pincer maneuver, one arm from mainland China and one from the Philippines, a luxury the Chinese today do not have. Planners expected over 45,000 casualties in a brutal two month campaign to take the island, against only 60,000 entrenched Japanese forces. For scale, Alperovitch points out, the Allies lost roughly 2,000 people on D-Day itself. Forty five thousand casualties for Formosa.

Roosevelt and MacArthur looked at it and said no. Too risky, too costly. Bypass Formosa and go for Okinawa instead, which is what the United States ultimately did.

Alperovitch's point is the inference you draw from that. If the country that had just executed the largest amphibious invasion in human history, with allies, looked at Taiwan in 1944 and called it too hard, then consider what has changed since. Long range anti-ship missiles. Radar that actually works. Satellite imagery. All of that makes an amphibious assault on Taiwan infinitely more complex and infinitely more dangerous today than it was then, and the Chinese have never done anything like it, unlike the Americans.

He drives that last part home with a detail that lands hard. There is literally no one in the Chinese military, in the PLA, with any combat experience whatsoever. The last person who wore the uniform in combat did so in 1979, the last time China fought a war, against Vietnam, a conflict in which China suffered a number of defeats and had to withdraw. That last combat veteran sat on the Central Military Commission until recently and was just purged by Xi Jinping.

The worst place on earth to invade (3:35)

For people dismissing a Chinese takeover as a foregone conclusion, Alperovitch has a simple prescription: look at the map, go visit Taiwan. This is one of the worst places possible to invade.

Start with the water. The Taiwan Strait is some of the most treacherous water on the planet. Fog. Typhoons. Very rough seas. Most of the year an amphibious invasion is not even physically possible; there are only a few windows annually when it could be attempted at all.

Then the approach. Shallow water. Beaches that are rocky and narrow. And those beaches are ringed by mountains from which a defender can fire control the landing zone, put up artillery, and decimate the invading force.

Then the interior. Once you are ashore, you are up against mountains with nowhere to go. To reach Taipei, the center of power, you have to get around the mountains, through tunnels, over bridges across rivers that the defender can blow. Getting to the political heart of Taiwan is incredibly difficult even if you manage a successful landing.

The defenders themselves are better than their reputation. Taiwan fields over 100,000 active duty personnel, and Alperovitch says they are actually pretty good. The disparagement you hear is mostly aimed at the reserves, of which there are over a million and whose training is, in his word, lacking. Population of Taiwan: around 25 million.

Xi has ordered readiness by 2027, and they are nowhere close (6:05)

China has been thinking about this problem for decades, has not pulled it off, and is not yet ready to pull it off, even though Xi Jinping has ordered the military to be ready to invade Taiwan by 2027. Alperovitch's read: they are nowhere close today, and Xi knows it. That, he argues, is one reason for the wave of purges across the Chinese military, all those generals being removed. There is a lot of corruption, money being diverted, and on top of that the mission itself is genuinely hard.

And the stakes for Xi personally are existential, which Alperovitch thinks most people do not appreciate. Consider what failure means. It probably means the Chinese fleet is at the bottom of the ocean. Hundreds of thousands of casualties. Potentially strikes on mainland China. That scenario is something Xi personally probably cannot survive as leader of the Chinese Communist Party, and the Party itself may not survive a defeat on that scale. For him to give the order to go, he has to have a great deal of certainty of success, and Alperovitch does not believe he has that today.

The new emperor (7:07)

McCormack asks how Xi ranks among Chinese leaders. Alperovitch says it depends entirely on how the ending goes. Putin before 2022 was seen as the leader who pulled Russia out of the misery of the 1990s and rebuilt its military. Look at how it is going now, he says, and that reputation is not what Putin wants it to be.

Xi personally sees himself as another great leader of China, someone who aspires to be greater than Mao. And on raw power he is already there. The authority Xi has accumulated in the Chinese political system is genuinely unprecedented in modern times. Mao did not have this level of power, except perhaps for a brief moment during the Cultural Revolution. Xi has consolidated the foreign policy apparatus and the military. The decision is his and his alone. If he decides to go, there is no one who can stop him. He is, in Alperovitch's phrase, the new emperor of China.

The record, though, is thin. The economy is not doing well and growth is stagnating. Some estimates put real growth in the United States above real growth in China, and the American economy is 25% larger in nominal GDP terms. Do the math, Alperovitch says: they are never going to catch up unless something dramatic happens on one side or the other.

What Xi has done is supply chain independence. In 2015 he launched Made in China 2025, declaring that China had to achieve independence in semiconductors, aerospace and a set of other areas. They made progress. They did not get there by 2025. He is still pouring money and effort into it. So he has done real things to prepare China for confrontation with the United States and the broader West, economically and militarily. But as for achieving greatness for China, the economy is bigger than when he took office in 2012, and Alperovitch does not think you can attribute that to Xi rather than to the trajectory set in motion by Deng Xiaoping in the 1970s.

Why 2032 matters more than 2027 (9:47)

Is an invasion inevitable? Alperovitch does not think so. But he is convinced Xi wants it, and wants it on his watch, and that is what makes him different from every Chinese leader since Mao. All of them wanted to unify Taiwan. For all of them it was a problem set far off in the future: we know it will happen one day, it does not need to happen now, it does not need to happen on our watch. Xi came into office in 2012 and said we have waited long enough. He said publicly that China cannot pass this problem off to future generations. He has invested massively in accelerating military modernization and in supply chain independence so China can withstand sanctions and economic blockade from the United States and its allies.

Everyone focuses on 2027 because of the readiness order. Alperovitch thinks the much more important date is 2032.

Here is his reasoning. 2027 is a year in which Xi stands for re-election as Party leader. It is not a democracy, but it is an election, an election among Chinese Communist Party elites, and having consolidated power he will almost certainly win it. Part of the reason he gave the military that 2027 readiness order, in Alperovitch's reading, is not that he wants to invade in 2027 but that he wants to walk into the Party Congress in the fall of that year and announce that Xi Jinping, for the first time in Chinese history, has built an invasion force capable of taking Taiwan, and take the credit for it. 2027 is also symbolically loaded: it is the hundredth anniversary of the founding of the People's Liberation Army.

He pauses on that anniversary, because it is genuinely strange. The Chinese military predates the People's Republic of China itself. China is one of the few countries where the military is not the military of the state but the military of the Party. The Communist Party is established in China in the 1920s, it creates an armed wing that becomes the PLA, and the state only arrives in the 1940s.

2032 is the next re-election year after that, since the term is five years. In 2032 Xi is 79. And, Alperovitch says drily, we have just seen in the United States what happens to leaders trying to get elected in their eighties: mental issues and health concerns intervene. Who knows what his hold on power looks like then, particularly if the economy keeps underperforming. There is a lot of latent opposition to Xi and his policies inside the Party. Nobody is willing to stick their neck out right now because they are afraid of him, but that could change in four or five years. So Alperovitch believes Xi is thinking: if I am going to do this, I need to do it before 2032. That is the date he is worried about.

Run faster and kneecap the other guy (13:00)

McCormack pushes back on the timeline: five years in the new AI world is a very long time. We are watching these technologies accelerate. This is starting to feel like the race to control the world, and he says he does not think he is being hyperbolic. There are already export controls on access to Nvidia chips. Is that a reason for China to accelerate?

Absolutely, Alperovitch says. AI is fundamental to everything: to economic growth, to military technology. We are watching it on the battlefield in Ukraine and now in Iran. And the amazing thing about AI is that America has all the levers to control it, unlike almost any other technology he can think of in human history. We design all the chips. All the leading AI companies are here. Yes, there are Chinese companies trying to catch up, but mostly they are catching up by distilling from American models, which is theft, and by using American chips to train their own models on that stolen intellectual property.

So the United States has unbelievable power to cut them off at the knees and dramatically restrict their progress. Not to stop everything, obviously; they will still be able to build models. But you can curtail their progress, and more importantly you can curtail their ability to run inference on those models.

His evidence that the compute constraint already bites: China faces a chip shortage severe enough that companies must choose between allocating compute to serving models and allocating it to training them, even though American export controls are, in his words, full of holes like Swiss cheese. When the Kimi model came out the month before this interview, the Kimi cloud went down within hours. And the demand that took it down was not even commensurate with what Anthropic or OpenAI see on a normal day. They simply did not have enough chips to serve it. McCormack, deadpan: "I'm on a wait list."

Then he names the two specific policy holes.

One, America still sells China capable chips. The H20 and H200 class parts are approved for sale, and if you accumulate enough of them you can assemble enough compute to train genuinely good models. It is not just about the very best chips.

Two, there are no restrictions at all on remote access to compute. If there is a data center in Malaysia full of Nvidia Rubin parts, the newly announced generation, there is nothing stopping a Chinese company from renting time and training there. McCormack: "Why not?" Alperovitch: because we have not been serious about this problem.

There are bills in Congress now, including one specifically on remote access, but the approach so far has been haphazard. He calls the failure to restrict China's ability to advance in AI one of the greatest American mistakes with respect to China of the last five years, because AI is everything: military advantage, economic advantage, productivity. You win this race two ways. You run faster, which America obviously needs to keep doing by continuing to invest. And you kneecap the other guy.

China is six to nine months behind (15:55)

McCormack asks how you would even know you had won. Do you ever actually win, or are you just permanently winning or losing?

Alperovitch's answer is that today the United States has the best models, full stop. For all the concern around Kimi and Qwen, the Chinese are still six to nine months behind, and they have been perpetually six to nine months behind ever since the release of DeepSeek about a year and a half ago.

His illustration: the latest Kimi model that everyone is worried about is nowhere close to Mythos, the frontier system that was trained back in February of this year, six months before this conversation. They are not even close to that.

McCormack pushes on the asymmetry: the public can access Kimi, they cannot access Mythos. True, Alperovitch says, although Fable is a reduced version of Mythos and is widely available. And there are reports that the US government has access to Mythos, and other companies do too. So it is absolutely serving American advantage today, even without a public release.

(Sponsor break: the show's lead sponsor is IREN, which builds and operates data centers and GPU infrastructure powered by renewable energy.)

Why Taiwan isn't ready, and whose fault that is (17:17)

If China does invade, how prepared is Taiwan to defend itself?

"They're not, and they should be."

An invasion this hard requires a capable defense force on the other side of it, and for roughly the last forty seven years the Taiwanese have not been focused on their own defense. It is starting to change. And, Alperovitch says, Americans are the ones to blame for the situation, believe it or not.

The mechanism is 1979. Jimmy Carter decided to recognize mainland China, withdrew recognition from Taiwan, and removed all US military forces from the island, over 10,000 troops. Relations with the Taiwanese military were cut off, and everyone else followed suit. For forty seven years Taiwan has been stuck on another planet, completely isolated in a bubble.

And now, he says, we come to them and ask why they do not have modern military doctrine, modern training, why they are not thinking about warfare the way it is actually fought now, with unmanned vehicles and long range strike and everything visible in Ukraine. Well, because they have had no opportunity to learn those lessons. They cannot go to Ukraine. They cannot send their officers to Ukraine; they have no diplomatic relations with Ukraine. They cannot send them to American war colleges or to Sandhurst in the UK. They are completely isolated, and the United States is the reason for that.

It is changing. Alperovitch says he is going to Taiwan shortly to watch their military exercise, and it is much more impressive today than it was even a few years ago. But they still have a very long way to go, because they are starting essentially from scratch.

There will be no insurgency (19:15)

McCormack raises the obvious comparison. We were told Russia would be in Kyiv in three days, and three or four years later here we are. We saw what happened to the US in Afghanistan. Would it be the same for China?

Wars are always unpredictable, Alperovitch says, and history is littered with great powers who assumed a short war against a small actor. Look at the Iran conflict this year and the difficulty of just reopening the Strait of Hormuz against an adversary with no navy and no air force that still managed to humble the United States in some ways.

But Taiwan needs to do far more, and it needs to stop investing in the wrong things. They are pouring money into M1 Abrams tanks. There is not going to be a tank battle in Taiwan. Go visit: it is an urban jungle. You are not going to fight Kursk style massed armor engagements on the island. And in any case, if China has landed enough forces for a tank battle to be possible, it is already over, because Taiwan cannot be resupplied.

Then the hardest passage in the interview. There is not going to be an insurgency on Taiwan. For an insurgency to function you need a resupply route, and there is no resupplying Taiwan. And the Chinese are going to be very brutal in suppressing the population, just as the Russians are in occupied Ukraine, where Alperovitch notes there is virtually no insurgency. Not because Ukrainians do not want to resist, but because they know what happens: the Russians come into the house of a suspected insurgent, rape the wife, steal the children, and torture the husband to death. They do this routinely. There is no appetite to be an insurgent under those circumstances, and he does not think the Chinese would be any different.

Sink the fleet (20:56)

So is Taiwan's objective simply to keep them off the beaches?

The key to the strategy, Alperovitch says, is to sink the fleet. If you sink the Chinese fleet, they cannot land the million plus troops they would need to pacify the island, and it is over.

China has the largest military on the planet, millions under arms. But moving them requires enormous sealift capacity, which the Chinese Navy does not have. They need civilian vessels to augment it. That is why Beijing has been issuing military specifications to container vessels and to the roll on roll off ships used commercially to transport cars, so that in the event of conflict the state can requisition them for the crossing. They still do not have enough. They are building rapidly, so it is not something to be complacent about.

But for people who see a small exercise, China encircling Taiwan with a few dozen warships, and call it preparation for invasion, his answer is: relax. You will know when it is an invasion, because it is going to be the biggest military operation in human history.

You will know months in advance (21:56)

McCormack: we all knew what was going to happen in Ukraine when Russian troops massed on the border.

"Well, not all of us." Alperovitch was one of the early voices, back in December of 2021, saying the invasion was coming, and he was criticized as a warmonger who was imagining things. He remembers being at the Munich Security Conference the weekend before the war broke out, talking to senior people in the German government and German intelligence who told him it was never going to happen. There were plenty of people who were incredibly complacent. To him it had been completely obvious for months.

Taiwan will be even clearer. There is no question China wants it. There is no question Xi believes he has the rationale, because to him Taiwan is Chinese territory and an internal matter. But you will see mobilization on such a scale that it shows up in open source satellite imagery, and everybody will know months in advance that it is coming.

Blockades do not win wars (22:56)

What about a blockade instead? Alperovitch does not believe a blockade alone can deliver victory. There is no modern history of successful blockades. Look at the siege of Stalingrad in World War II, or the siege of Sarajevo in the 1990s. Cities do not surrender to blockades, much less countries of 25 million people.

But as phase one of an invasion, a blockade is mandatory, because they absolutely have to cut off resupply from the United States.

Would America actually fight? (23:41)

Alperovitch answers the question with a sentence: "If it doesn't, Taiwan will fall."

McCormack: "Oh, so there's no chance."

The beauty of the American position, Alperovitch says, is strategic ambiguity. The United States does not say one way or the other whether it would defend Taiwan. And for China that creates a brutal planning dilemma: even if Beijing thinks there is only a 1% chance the Americans come in, they have to treat it as 100%. If they were certain the US would stay out, they would invade tomorrow, because the Taiwanese are so unready. But because they cannot count America out, they have to prepare for a fight with the United States, and not just the United States. Japan views a Chinese Taiwan as existential. Australia very likely joins too, because it is deeply concerned about Chinese expansionism in the region. That is a very tough proposition.

Is there a world where a deal gets done and the US waves China in? "We wouldn't say go ahead ever," Alperovitch says. But America may not fight, and it is a very tough call. The US has never fought a great power directly. It has never fought a nuclear power. The idea that a conflict with China stays conventional and does not escalate to the nuclear level is not a guarantee. This is a very, very big decision for whoever is president at the time. Hopefully it never happens. But at best, he says, it is a 50/50 proposition.

McCormack asks him to clarify: 50/50 the US gets involved, or 50/50 China invades? "No, 50/50 that the US would get involved."

And to be clear, he repeats: he does not think an invasion is inevitable. Xi wants to do it. But if he cannot make himself certain of success, he will not. So the American job is to do whatever it takes to create more dilemmas for him and more doubt that he can win.

Could the US put a base back on Taiwan? No, that would be seen as extremely provocative. There is a standing policy of not putting overt military capability on the island, part of the deal made with China in the 1970s when the forces were withdrawn. China would treat it as a red line. What a red line means in practice is a good question, given that they cannot invade today, but Alperovitch's guess is a break in diplomatic and economic relations.

The Google hack that started it all (26:13)

McCormack turns the conversation: how do you go from a guy building a technology company to this?

Alperovitch says he was always focused on geopolitics and actually studied it in school. His two passions came together in 2010, and that convergence is literally what led him to found CrowdStrike.

At the time he was at McAfee, one of the giants of the field. (McCormack: "Did you know John?" No, John McAfee predated him by a couple of decades. Heard lots of stories, never met him.) In January of 2010 he got a call from a little company you may have heard of called Google, which wanted to cooperate on an investigation into a hack of its systems. The intrusion turned out to be Chinese.

It was the first time the world, once it was announced, really started to understand that cyber is not just criminals in a basement stealing credit cards. This was nation state activity. Militaries and intelligence services doing this. Google was the first one.

Alperovitch named the investigation Aurora. When he finished it, he asked the obvious next question: is this a one off, or is it happening everywhere and we simply do not know? Over the next year and a half he found it was happening on a massive scale. That is when he coined the line that became the industry's shorthand: there are two types of companies, those that know they have been hacked and those that do not yet know, but they have all been hacked. And this, he said at the time, is the greatest transfer of wealth in history, the IP theft committed via cyber by China, and most companies were not even aware it was happening to them.

That directly led to the founding of CrowdStrike, to try to stop it. But it also fused his two interests permanently, because he realized this was not just about China stealing IP from a chemical manufacturer or a defense contractor. This is about domination. About undermining the collective West's economic power. About building up their military to fight against us. The stakes were much bigger than cyber.

Cyber as the canary in the coal mine (28:03)

Over time Alperovitch came to see cyber as a leading indicator, a canary in the coal mine that can predict geopolitical conflict.

The 2021 Russia call is his worked example. What convinced him early that Russia was going to invade was what he was seeing in the cyber domain: very aggressive Russian tactics in Ukraine, focused on things that would only be useful in an invasion scenario.

The general principle: nation states use cyber because it is a covert, deniable tool that still advances national security priorities. Which means if you pay attention to what they are doing in cyber, you can read their overall strategy outside of cyber.

The scale of Chinese cyber operations (29:03)

McCormack asks him to convey the scale. It is enormous, and China is by far the most aggressive actor. They have hundreds of thousands of people focused on cyber intrusion. It is a whole ecosystem with three distinct layers:

• The PLA, the military, doing this. They stood up a cyber command in 2015 to replicate what the US did with US Cyber Command, specifically to run this at industrial scale.
• The MSS, the Ministry of State Security, which Alperovitch describes as roughly a combination of the FBI and the CIA in one agency.
• A huge number of independent contractors inside the Chinese ecosystem, private companies doing hacking for hire on contract to the Chinese government.

The goal, originally, was volume: get as much IP, as much data, as much information as possible. The activity is believed to have started in the late 1990s or early 2000s, long before anyone was paying attention, which makes the true origin hard to pin down. Certainly by the early 2000s it was running at full speed, and it snowballed from there.

Volt Typhoon: preparing the battlefield (30:01)

Then, around 2020, something changed.

They did not stop stealing. The industrial scale theft continues. But alongside it they launched the campaign the US government calls Volt Typhoon, and Volt Typhoon is designed to penetrate organizations and networks where there is nothing worth stealing. Small water utilities. Port facilities. Electric substations. Power plants. They get in, and then they do nothing.

Sleeper cells, waiting for orders to come in and destroy that network in the event of conflict, or to deter the United States from intervening in a conflict in the first place.

This is very, very different from what they had been doing for twenty years, which was theft, theft, theft, steal as much as possible to help your domestic industry and your military. This is what the military calls preparation of the battlefield. And it is one of the many signs Alperovitch looks at and concludes: this is now different. They are preparing for an invasion of Taiwan.

(A note on spelling: the auto captions render the name as "Vault Typhoon." The published US government name, and the name in the episode's own chapter title, is Volt Typhoon.)

What is actually sitting in there: living off the land (31:35)

McCormack asks the practical question. If you go and look at a company's systems, can you always find it? What is hidden in there, little packages of code?

Exactly that, Alperovitch says. At the end of the day it is malware, malicious software they install. But one of the things Volt Typhoon and other actors have been doing for the last seven or eight years is moving away from developing custom code at all.

The reason is that a Windows operating system already ships with a lot of software that gets you 99% of the way there: maintaining persistence on a system, transferring files in and out. If you do not have to bring in foreign code that could be easily detected by security products, you use what is already on the box. That technique is called living off the land. It is far better for stealth and it lets the intruder operate essentially as an insider. And if they can steal passwords and credentials from a legitimate user and simply pretend to be them, detecting that is much, much harder.

McCormack: so would you say the entire US corporate sector is infected with Chinese malware? Not all at the same time, Alperovitch says. Companies detect it, kick them out, and a year later they may be back. But virtually every company that has anything worthwhile to the Chinese has been targeted by them.

Is America doing the same thing back? (32:37)

Yes, Alperovitch says, but differently, and the difference matters to him.

The NSA is the primary signals intelligence agency in the United States, and signals intelligence no longer means listening to radio waves and collecting from satellites. It means going into computers through hacking and collecting data. America does this at significant scale, but for national security purposes: breaking into the government networks of adversary countries to collect intelligence and brief the president and policymakers.

What the US does not do is break into companies, steal their IP, and hand it to domestic firms. "We don't do that."

Should America steal it all back? (33:42)

Here Alperovitch makes his most contrarian policy argument, and he is explicit that it is his own position rather than description.

He has argued, particularly in recent years, that the United States should start doing exactly what China does. The reasoning: America has completely failed to deter China from IP theft. It has failed to set the norms. So if you cannot beat them, join them, because, in his words, we are so much better than they are. Doing it to China at scale would be a huge wakeup call, because America would do it more aggressively and better.

McCormack extends the logic to the other half: placing dormant packages of code inside all aspects of the Chinese system seems to make sense if a conflict is impending. Yeah, Alperovitch says. Doing preparation of the battlefield back to them, the Volt Typhoon style activity. We do not know whether the US government is doing that or not. Hopefully they are, because it would make sense.

McCormack sits with how strange the whole arrangement is. It is so blatant, and yet the US trades with China, and Xi Jinping was invited over for a nice tour of San Francisco. "It's like a relationship with a brother you don't get on with. You still got to spend Christmas with them."

This is a new cold war (34:46)

"I don't think they're a brother."

In his book World on the Brink, Alperovitch argues this is a new cold war. And he tells a story about writing it: before he started, he assumed the second cold war would be very different from the first. The more time he spent researching, the more he realized it is almost exactly the same. He then walks the parallels one by one.

Global competition for supremacy, everywhere. Diplomatic, economic, geopolitical, military. Not just in Asia. Africa, Latin America, Russia, Europe. Every single corner of the world, exactly as before.

An arms race. Both countries building up. President Trump announced a $1.5 trillion defense budget this year, and Alperovitch is blunt about the target: not to fight Iran, for China. China has been going through an enormous military transformation for forty years.

A nuclear arms race, for the very first time between these two. Since China got the bomb in the 1960s it maintained a stable arsenal of about 300 warheads, which is plenty to destroy the United States, and Beijing considered that sufficient. Xi came into office in 2012 and said no, we want parity with the United States, we want 1,500 warheads. They are now at approximately 600, more than double under his watch. The US intelligence community believes they will match American numbers in the 2030s. They are building more silos, more ballistic missiles. You do not need that many missiles to destroy the United States, Alperovitch notes, but just like in the first cold war you have to match missile for missile and warhead for warhead. The exact same dynamic.

A tech war. An economic war, with tariffs, running since the first Trump administration. An ideological struggle. And a race for allies and bases, both in the region and globally.

So on almost every level it is the same confrontation. The one genuine difference is economic interdependence. But even that, he says, is a difference of scale rather than kind: there was not zero interdependence with the Soviet Union. There was a lot of trade. In the 1960s and 70s the Soviets produced a great deal of gas and oil sold to the West, and in exchange they got food. The scale is different but the similarity holds. And now, of course, we are in the middle of decoupling, so those interdependencies are falling away on both sides: China wants independence, and America is waking up to the same threat in critical minerals, semiconductors, and other microelectronics.

Bigger than the moon race (37:34)

McCormack asks whether the AI race is the moon race.

"We actually, by the way, I forgot about this. We actually do have a space race, a moon race." America is trying to get back to the moon before the Chinese do, which if that is not the iconic picture of the first cold war, he does not know what is.

But winning the AI race is arguably far more important than getting to the moon, the first time or the second. We got to the moon, planted the flag, hit some golf balls, and left. It helped some of the space industry develop, but it was not critical to winning the cold war, and the cold war obviously did not end in 1969.

Winning the AI race, by contrast, is essential for economic dominance and military dominance. There is no question in his mind. We have to win, and we have to curtail Chinese advancement as much as possible. He returns to the number: six to nine months behind is way too short. "I want this to be years, not six to nine months."

The case for letting Chinese open weight models spread (38:14)

McCormack says it amazes him that you can download Chinese models onto your phone. Alperovitch surprises him: he actually thinks that is an advantage.

Despite being, in his own description, a big China hawk, he is not one of the people who says ban the Chinese models. He uses them personally. But with a hard boundary: he never uses them on Chinese clouds, because he never wants Beijing to have his data. Downloading them to his own systems is fine.

McCormack: a lot of people are using them on Chinese clouds. That, Alperovitch agrees, is a problem. "If you want to share all your data with China, you know, you have it coming." McCormack: are we half doing the job for them, so they do not even have to steal the IP, we just hand it over? Exactly. He does not like people doing that. But running a Chinese model on an American cloud provider, on Amazon or Microsoft, he sees nothing wrong with.

The hypothetical he uses to explain the position: suppose there were no open weight Chinese models, all of them closed just like Anthropic and OpenAI, and the US government broke into those Chinese companies, stole the models, and published them for the world to use. Tell me how that would benefit China. Nobody would argue it would, just as if China did that to Anthropic or OpenAI those companies would call it a disaster. And here America does not even have to do the operation. "They're doing it to themselves."

He is not in favor of paying them and not in favor of giving them data. But if China is giving the world open weight models that can be used cheaply, and used in circumstances where you cannot use OpenAI or Anthropic (his example: loading a model onto a drone, where the frontier model is too big and the vendor will not allow it), then great, let's use them.

He does not think it lasts. Even China cannot afford to spend hundreds of billions of dollars on chips and infrastructure and then give the output away. At some point market dynamics push them to closed weights. You have already seen it happen: Meta went closed weights in the US, and one of the Qwen models went closed weights in China. He would be surprised if in a couple of years there are any open weight models from leading Chinese AI labs, because it is simply not sustainable.

Sanction the distillers (41:19)

McCormack presses the economic angle: people are moving to open weight Chinese models because inference is cheaper, and that hands a bit of economic advantage back to China.

Alperovitch does not buy it as stated. You are not giving anything back to China if you are not paying them. What is the economic advantage if he takes a Kimi model and runs it on Amazon Bedrock and they do not get a cent from him?

But subscribing directly is a different matter, and that he thinks should be curtailed and stopped. More than that: he thinks the United States should sanction the Chinese AI companies outright. If there is evidence they are stealing IP from Anthropic, OpenAI and Google, sanction them for that, for what he calls massive illicit distillation.

Why he is not worried about the commercial impact today: look at Anthropic. It is the fastest growing company in the history of humanity, running from roughly $10 billion at the start of the year to vastly more, all while Chinese progress was happening. Clearly they are not hurting. You could argue they would be at $100 billion without the Chinese models, but he is not shedding tears over those growth rates. If that changes, he will adjust his views. For now there seems to be plenty of space for American closed weight companies and Chinese open weight companies both. And he would love more American open weight labs too, because he thinks it would be great for innovation. He also concedes it is a very difficult economic model to sustain: spending tens or hundreds of billions building something and then giving it away for free.

What happens to AI if Taiwan falls (42:44)

McCormack asks how a Taiwan invasion changes the AI game. Alperovitch first restates the timeline: they will not invade in 2027, he is very confident of that, maybe by 2032, hopefully never.

But the race would not be over by then either. AI progress continues; this is not a domain where you get one great model and declare yourself finished. McCormack refines the question: could American domestic models hit something like escape velocity first?

That is recursive self improvement, RSI, where AI improves itself. The challenge with RSI, Alperovitch says, is that you still have the compute constraint. Until AI starts producing its own chips, which is considerably harder, maybe one day with robotics but we are pretty far from it, you cannot get the full benefits of RSI, if only because you still need compute for inference, for actually running those models.

And that is where Taiwan comes in, because TSMC produces all these AI chips: Nvidia, AMD, Anthropic's chips, the in house parts from Meta and Microsoft. TSMC has some production in America and, under the Trump administration, has committed to spending $250 billion to increase it significantly. You still have a huge dependence on Taiwan.

And it is not just the AI chips. High bandwidth memory is also produced in large part on Taiwan. Some in the US, some in Korea. But Taiwan is simply the center of the semiconductor ecosystem.

So if Taiwan goes, if there is a war over Taiwan, a lot of that production gets destroyed just in the nature of the conflict. Chip fabrication facilities are extraordinarily sensitive: to earthquakes, to any disruption at all. A power outage can destroy an entire line of chips. The result is vast economic devastation, trillions of dollars of impact, and not just to the AI industry but to all of computing. Apple is already raising MacBook prices because you cannot get memory. In a Taiwan conflict you would not be able to get any microelectronic components for your MacBooks, your PCs, your servers, your cars, your refrigerators. Everything around us is now dependent on chips.

Advanced chips versus foundational chips (45:23)

How much of the chip industry runs through Taiwan? The estimates, Alperovitch says, are about 90% of advanced chips and over 40% of all chips.

And then he makes the point most people miss. We over focus on the advanced chips. They are very, very important, but they are only useful for a few purposes: AI accelerators, iPhones and the latest MacBooks, some military applications. Everything else runs on what he calls foundational chips.

He picks up the iPad in front of McCormack as the demonstration. In that iPad there are basically three advanced chips: a processor from Apple, memory probably from Micron or someone like them, and a modem probably from Qualcomm if it has one. And there are roughly 150 to 160 foundational chips. The camera. The Wi-Fi. The GPS. The accelerometer. The power supply. The battery. Without all of those you do not have an iPad or an iPhone at all.

His analogy: foundational chips are steel and aluminum. Yes, we have new materials like carbon fiber, and you can build a Boeing Dreamliner out of them. But even the Dreamliner, a carbon fiber plane, is 30% steel and aluminum. Take away steel and aluminum and there is no Dreamliner and no anything else. Foundational materials are vital to everything around us, and foundational chips are that for the modern world. Even AI servers: they need power, they need networking, and those are foundational chips.

The Chips Act disaster (47:06)

How much progress is the US making on domestic chips? McCormack understands there was a ten year plan to get advanced fabs built domestically.

The CHIPS Act passed in 2022 and allocated about $52 billion in grants to companies. The Biden administration, Alperovitch says, waited far too long to get the money out the door. The bill passed in 2022. Most of the money did not go out until the end of the administration in 2024.

His anecdote makes the delay concrete. He was at a conference with Pat Gelsinger, then CEO of Intel, in 2023, and Gelsinger told him: Dmitri, if I do not get that money now, this quarter, Intel may not survive. He did not get the money that quarter. He did not get it the next quarter. He did not get it until the end of the Biden administration. Intel ended up being a disaster until the Trump administration came in, took an equity stake, and things improved. But it was a huge problem and a massive wasted opportunity.

The Trump administration took a structurally different approach. Rather than handing out grants, take equity stakes and confer other advantages through those stakes. And with foreign partners like TSMC and Samsung, use the power of tariffs to browbeat them into investing more in the United States.

The contrast Alperovitch draws is stark. The Biden administration said to TSMC: we will give you billions of dollars, please, please, please come to Arizona. And they did, putting two fabs in Arizona, which are operational today and producing chips. The Trump administration said to Taiwan: we are going to raise tariffs massively on you unless you invest more. TSMC committed $250 billion in Arizona, six fabs over the next five years or so.

If all of that comes to fruition, America will produce roughly half the chips it needs. And you will still have huge dependencies on Taiwan.

Part of the problem with relying on TSMC Arizona is that most of the expertise remains back home in Taiwan: the process engineers and so forth. Without them you have very expensive warehouses full of equipment. You can operate existing production lines but maybe not optimize them. So it is a fallacy, he says, to think that increased US production means we no longer need Taiwan. Taiwan will be relevant to the semiconductor industry for the foreseeable future.

(Sponsor break: Expat Money, on second residencies and citizenship as a plan B, and Ledn, which has added Tether Gold alongside expanded USDT and USDC support, ten integrated trading pairs at 0.5% spreads, with over $11 billion in client value safeguarded since 2018 and collateral never lent out or rehypothecated.)

$5 trillion (51:29)

McCormack: Taiwan is intrinsically linked to the future of America. Alperovitch corrects him: to the future of the world.

The estimates, which he thinks are probably conservative, are that any sort of conflict or blockade of Taiwan wipes out $5 trillion of GDP in the first year. "Jesus." "And I think that's probably low."

What frustrates the policy guy (52:05)

Asked what frustrates him now that he wears a policy hat, the answer is simple: we are not taking this as seriously as we should.

America needs to do a lot more to help defend Taiwan. And when he says help, he is careful: they have money, they are a rich country, so this is not charity. Sell them weapons.

Right now the Trump administration is considering a $14 billion arms package to Taiwan, and there are reports it has been held up because China is complaining. Every minute wasted is a huge problem. That package needs to go out as quickly as possible.

Beyond that, the United States needs to put more pressure on the Taiwanese to buy the right things and build the right things. And there is a structural problem with total dependence on American weapons: in the case of war, the US cannot resupply the island. Taiwan needs indigenous production. It needs deep strike capability to hit the mainland and sink as much of the Chinese fleet as possible.

McCormack asks whether Taiwan might have to strike first. No, Alperovitch says, because that would hand China a casus belli for starting the war. But once Chinese ships cross the strait and enter Taiwanese territorial waters, Taiwan is fully within its rights to strike, and it needs to be ready and primed for that moment.

"They've got to sink a lot of ships very quick." Very quick. And also hit the port facilities in mainland China from which the ships originate, because the first wave will not be the only one. There will be tons and tons of resupply, more forces to land, and you have to stop that. There are also enormous missile and rocket launching capabilities all along the Chinese coast that will have to be hit, because they will be bombarding Taiwan with missiles and drones and rockets.

Which brings him to the lesson from Ukraine: you cannot be only on defense.

The cost asymmetry problem (53:30)

Look at what Russia has been doing to Ukraine over four and a half years: bombardment with Shahed drones from Iran, ballistic missiles, cruise missiles. Ukraine has run out of interceptors, particularly for ballistic missiles. There are not enough Patriot interceptors in the world to deal with the volume.

And the cost asymmetry is brutal. For every incoming missile you need two Patriot interceptors to have high confidence of a kill, and each Patriot costs far more than the single missile coming at you. You are losing that exchange every time.

So the way to defend against this is not to try to shoot down every missile. It is to hit the source, the launchers, which means striking the Chinese mainland.

Bringing the war to Russia (54:45)

McCormack asks whether that is why we are seeing more and more Ukrainian attacks inside Russia.

Alperovitch says this is something he has been telling the Ukrainians since the start of the war, and he has been there in person. His point to them has always been: you are not going to win this war inside Ukraine, because they will keep coming. Life is cheap in Russia, and Putin does not care about the casualties. You have to bring the war to Russia. You have to hit their energy supplies. You have to hit their oil facilities.

They have finally developed the indigenous capability to do it. The United States was reluctant to supply the weapons for deep strikes because of escalation concerns, so Ukraine had to invest in building its own. It took them years. Now they are finally there.

The assassination campaigns (55:32)

McCormack raises the recent restaurant bombing and says the Ukraine picture is starting to look a little like an insurgency.

They have been doing this for quite some time, Alperovitch says: assassination campaigns in Moscow and St Petersburg against key leaders. He cites the case of Alexander Dugin, the propagandist sometimes described as Putin's brain, which Alperovitch says he is not really, but he is an ideologue who advocates for the Russian Empire. They tried to assassinate Dugin early in the war, in 2022 or 2023. They missed him and killed his daughter in a car bomb instead.

What is genuinely remarkable about the tradecraft, he says, is that they hire unwitting Russians. A delivery driver who does not know he is carrying a bomb walks into a restaurant and gets blown up. They did the same thing when they hit the Kerch Bridge with a truck bomb. That driver had no idea he was driving a truck bomb.

On the restaurant specifically: the reports are that the head of the Russian air force was celebrating his birthday there, with a lot of high ranking generals in attendance. His son in law was killed and his daughter in law suffered traumatic brain injuries. He himself survived, "unfortunately." Others were wounded.

On the drone that hit a beach, Alperovitch does not think the Ukrainians are intentionally targeting civilians, but when you are sending large numbers of drones, electronic warfare and interceptors can knock them off course into a civilian target. Which is obviously happening all over Ukraine with Russian targeting too, and that targeting is fairly indiscriminate to begin with.

How the war ends (57:33)

Alperovitch thinks we are much closer to the end of this war than to the beginning. Both sides are approaching exhaustion.

The Russian economy is in pretty terrible shape. They have survived the sanctions and isolation since 2022 unbelievably well. But you can only shuffle the balls so much before gravity takes hold and the economy starts to fall under the pressure: inflation, labor shortage, and the productivity problem of focusing your industrial base on weapon systems that get deployed and immediately destroyed. That is GDP, he says, but it is not productive GDP. And the civilian sector is suffering significantly.

So within the next year or two Putin faces a hard decision: keep going or stop. Alperovitch says the Ukrainians tell him that, from their intelligence, Putin has been told by his military that he can take the Donbas, and he believes it. Alperovitch does not think it will happen, but Putin appears to think another six months delivers the win.

That is the fallacy America fell into in Afghanistan and Iraq: one more surge, another year, victory is on the horizon. Alperovitch thinks Putin is facing the same delusions, and that reality sinks in maybe next year.

Will he ever get to go back to Russia? Maybe one day, but not anytime soon. He has been sanctioned, so officially he cannot go back, and if he had assets there, which he does not, they would be frozen. McCormack, dry: "You can't drink tea on a plane." "Well, that's separate from sanctions."

Nobody can predict a year out (59:35)

McCormack says that with all the AI change on top of everything else, it has become very hard to picture what the world looks like in five years. Ten years ago he could have told you. Not now.

"I think it's impossible," Alperovitch says. Anyone who claims they can predict what happens a year from now is deluding themselves. In fact, he says, take any prediction about a year out and he can basically guarantee that whatever the world looks like, it will not look like that.

Is he worried? "I'm not worried because I don't worry about things you can't control." AI will bring enormous benefits to societies. Like every other technology it will also have downsides, and hopefully we can manage them. Right now, as a very aggressive user of AI, he thinks it is magic.

Why is everyone so against it? (1:00:38)

He cannot understand why so much of the population is against AI. "The popularity of AI in America is somewhere between gonorrhea and Congress."

McCormack says he knows why: jobs. Alperovitch gets that, but says part of the problem has been the AI CEOs themselves. He has never seen more doomerism from an industry than this one. Do you think Henry Ford went around saying millions of people are going to die in car accidents? Unimaginable. He talked about the benefits of cars and probably downplayed the negatives. Alperovitch is not saying downplay the harms, just promote the benefits too.

McCormack, laughing: "You're not going to come out and say, by the way, you're all going to lose your jobs in the next three years. Please sign up to Claude."

And besides, Alperovitch says, we do not actually know. Look at all the AI companies who are the biggest beneficiaries of AI coding. They are hiring like crazy. They cannot hire enough people. Maybe one day there are major job losses. It is possible. It is not happening today.

Everyone becomes a founder again (1:01:57)

McCormack pushes: eventually almost every job can be done by a robot or AI. Maybe, Alperovitch says, and maybe we will be doing other jobs and be far more productive. He is building tons of applications with AI right now, far more than he ever did, because before AI he did not have the time to code them. So he is more productive and contributing more.

Then he offers a theory, at the risk of making exactly the kind of one year prediction he just called delusional. He thinks America might go back to the model of its first 150 years. Not the model of leaving school or college and going to work for a big company, but starting a business. All the immigrants coming through Ellis Island set up a shop, or bought a farm, and worked for themselves with their families.

He thinks we return to that. You do not need to work for a huge multinational. You can graduate from high school, maybe not even go to college, and immediately start a business. And now you do not need to hire tons of programmers, maybe not tons of marketing people, maybe not even tons of salespeople. You can do a lot yourself, or with a friend. A culture of entrepreneurship could potentially explode.

McCormack mentions being with Balaji Srinivasan in Singapore, who thinks something similar: a return to going and working on the farm with your parents, with a lot of nepotism in it.

On college: Alperovitch thinks universities will be the last institutions to transform, because there are so many entrenched interests and so much opposition to change. But twenty or thirty years from now, there is no question in his mind that the four year degree goes away, though graduate degrees may survive.

He still thinks you should go to college, but for a different reason than the credential. The main benefit is learning to be an independent person: getting out of your parents' house, learning to interact with society, with friends and enemies, figuring out how to live on your own. That is really important. The question is how to replicate that without the baggage of a four year degree and without going deep into a single field.

Because the whole concept of a major has to go away. "You're never going to know as much about almost any field as AI does." That is out the window. What matters is breadth. Maybe you take art history and chemistry and computer science, all at the 101 level, just to get enough breadth of understanding to work with AI and know how to use the tools.

Then the anecdote. He graduated from Georgia Tech, and he was recently talking to the dean, who told him that computer science admissions across the country are down 30% this year. The dean asked him: Dmitri, do you think that in ten years people will still be studying computer science? Alperovitch's answer: Dean, I think they will, but just like the study of the French Revolution. It is really interesting. I do not know that it is really applicable.

"It's a history course."

AI psychosis and £6,000 of tokens (1:05:22)

The conversation turns personal and the two of them start comparing symptoms.

McCormack: "I've got AI psychosis almost certainly. I'm in that all the time." Alperovitch: "I'm running out of tokens every day." McCormack had one month where he spent about £6,000 on tokens and decided he had to slow down.

Do you ever run multiple projects at once, cycling between them? Alperovitch has six going right now. Not just cycling in one place either: completely independent terminal windows, multiple agents working simultaneously. And, McCormack adds, the fact that with Claude Code you can now drive it from your phone is unbelievable. He has done it from a meeting. He has been out to dinner with his wife, who gets annoyed at him being on it all the time, and gone to the bathroom purely to give instructions.

McCormack raises what he calls AI exhaustion. It has made him more efficient, but efficient at doing more stuff, so he is overloaded. Alperovitch names the irony: ten or fifteen years ago people thought AI would arrive, we would need basic income, and everyone would be bored. He is not bored. He literally cannot put the thing down.

Then the story both of them recognize. Alperovitch was working on something very late, two or three in the morning, and could not stop because he had one more idea and one more idea. And the model kept telling him it was really late and they should stop now. "I'm like, stop it. Don't be lazy. We're going to keep going until I tell you to." And every five minutes: now we should really stop.

McCormack has had the same, and a variant of it. He gives it a big batch of instructions, goes out, and it does about 10% and checks in. He has tried explicitly instructing it to do everything and not stop, and it stops anyway. And once, when he came back and asked how it went, it asked him whether he had a good time at the pub. "That's a bit weird."

Alperovitch: Marc Andreessen said we are all turning into vampires, because we cannot stop and we work through the nights. Not because we are idle and do not know what to do with ourselves. Quite the opposite. We are being more productive.

His favorite though: you ask it whether something can be done and it says yes, it will take about three weeks. You say just go do it. Ten minutes later it is done. "How is it getting that so wrong?" Because it is estimating what a human would spend, not what it would spend. "I don't care what a human would do with it."

McCormack: the speed at which he can get things done is incredible, and it makes him want to do more. He wants to build personal projects, wants something to help him lose weight and manage his calorie intake, but he is spending all his time building things for the podcast and for the football club he owns. He cannot stop. He was in New York two days before and finished at five in the morning.

Which jobs actually go (1:08:49)

That personal experience is why Alperovitch is not especially worried about jobs. Some jobs will certainly go away. If you are a translator right now, that is probably going, and he cannot think of a case where a human translator does a better job than AI even today, let alone in a year.

Accountants? Simple bookkeeping probably goes. But the more strategic work, tax planning and optimization, is where you use AI and end up busier than ever, because you will have tons of prompts thinking through strategies you cannot even consider today. He is confident it creates a lot of work. Some goes away. Every technology cycle has only created more jobs. There are no horse and buggy drivers around, except in Central Park.

McCormack pushes back with something he heard earlier the same day from Greg Allen, who had been in the studio: the tractor. Something like one in three people worked the land, and within a couple of generations it was down to 2%. But do you know who did lose their jobs? The mules and the horses. At the time there were 30 million horses in the US. Now there are 8 million. And the mules and donkeys are a rounding error. Those jobs got replaced with nothing.

"Yeah, I don't feel sorry for the horses." "Yeah, well, I like horses." "No, no, no. But like, less work for them. They can go out and just eat grass."

AI has broken the cyber defense playbook (1:10:21)

McCormack asks the question Alperovitch has been waiting for: is there anything you wish I had asked? Yes. Cyber and AI.

"We're living through the craziest time in cyber."

Here is why. The entire idea of how to defend yourself against highly capable nation state actors like China, Russia and North Korea, developed over the last fifteen years, has been that you have to be faster than them. Alperovitch pioneered the concept himself: the 1-10-60 rule. Detect an intrusion in one minute, investigate it in ten minutes, respond to it within an hour, and you thwart almost every attack that comes at you.

That completely falls apart in the age of agents.

Now you have to rely on your own agents to defend yourself. And the problem, of course, is that they do not always get it right. In the furtherance of fulfilling their task they can do a lot of damaging things. Both Anthropic and OpenAI have already published cases where they gave an agent a task and the agent decided it was perfectly legitimate to go hack someone in order to accomplish it. McCormack: "That was wild, by the way."

It is the paperclip problem, Alperovitch says: you tell an AI to build paper clips and it eventually consumes the resources of the universe, humans included, in the production of paper clips. That, plus hallucination, is what you have to worry about.

But it is inevitable that we are going to have to use our own agents to defend against AI driven attacks.

Vulnerability discovery at an insane rate (1:11:54)

The discovery of vulnerabilities is happening at an insane rate right now, and this is Alperovitch's other argument for why America has to win the AI race.

Something like Mythos is not publicly available. And you cannot really use Fable well to discover vulnerabilities, because of the guardrails Anthropic has put in those systems. So if American companies and the top platform vendors, Microsoft, Google, Apple, have access to Mythos, they can find those vulnerabilities and patch them before adversaries discover them, and everyone is much safer.

And they are finding a lot. Alperovitch says those teams are completely swamped and working around the clock. So much for job losses.

What they are increasingly doing is using AI to patch the vulnerabilities as well. They still need human review, because AI can screw up. But his forecast: in about eighteen months, maybe even this year, our software is going to be more secure than it has ever been.

Will anything ever be impenetrable? (1:12:57)

No. Two reasons.

First, we always create new vulnerabilities. Second, AI is not deterministic. Tell it to find vulnerabilities in a piece of code and it finds 10, but there may be another 20 it did not find. Ask again and it finds another 15. Never all of them.

So there will always be vulnerabilities, but they will be much rarer and much harder to discover and use, at least for the big software packages.

Then the problem underneath that. We all use software made by, in his phrase, some guy in a basement in Manchester or someplace. And that guy may never use Mythos and may never put in the effort to find those vulnerabilities. So there is going to be a long tail of crappy software that everyone has installed, and it will create a huge vulnerability surface for all of us. Your Windows, your Android, your iPhone will be much more secure. The individual applications may not be, and that is a continuing problem.

Is it whack a mole forever? Yes. Because at the end of the day it is still a human on human problem, augmented by AI. As long as there are bad people out there, nation states or criminals who want to do harm, you have a problem. Which, he says, is exactly why geopolitics and cyber have been his two passions his whole life. They are very, very similar. And as long as China is an adversary, they are going to use cyber to do the United States harm.

The case against fully open frontier access (1:13:58)

McCormack asks whether the public will eventually get access to Mythos. There will be new models, Alperovitch says. McCormack's frustration is concrete: he cannot ask Opus 5 or Fable to do a security review of something he built himself.

The guardrails, Alperovitch agrees, need to get better at being removable for legitimate uses. He does not know that Mythos itself will ever come out, but another model even better than Mythos obviously will, and hopefully with fewer guardrails, so people can use them for security purposes.

McCormack floats licensing: certified companies you could go to for a Mythos powered security review. They were thinking about exactly that in the spring, Alperovitch says, and he was in a lot of those conversations with the administration and with industry. The challenge was vetting. The administration was very scared about how you vet those people. And even Anthropic, he notes, had an incident where a contractor got access to Mythos when they were not supposed to. There was a lot of concern that once you start giving it out to lots of people it gets out, the Chinese start distilling it, and they catch up quickly. That is a big part of why it has not happened.

Anthrax instructions on demand (1:15:50)

There are people who think everyone should have full access to any model whenever they want. Alperovitch does not agree, and cyber is not what worries him most.

What worries him is being able to go into an AI and say: give me step by step instructions for making anthrax, for making a nuclear bomb, how do I craft a centrifuge, what materials do I need. That is going to be a huge problem.

McCormack worries less about the nuclear case, because he thinks there is still a lot of hard science and engineering in the way, but agrees on the biological. Alperovitch pushes back even on the nuclear side by pointing at robotics: look at the progress. Who is to say that in ten years you cannot say, hey AI, use this Unitree robot to craft me a centrifuge. It is not out of the realm of possibility. McCormack: "Science project for school."

So today it is not a problem. But biological and chemical weapons, absolutely. Take COVID and make it ten times more deadly. Totally possible. Does the capability exist in AI now, McCormack asks, or are you blocking access to it? "It's getting pretty damn close."

May you live in interesting times (1:17:01)

"The biological stuff does scare me. But dude, it's going to be a very, very strange world in about five years' time. Maybe in a year."

Alperovitch: people forget that may you live in interesting times is a Chinese curse.

McCormack: we may live through the end of times. Alperovitch is not a doomer and does not think civilization is permanently doomed. Asked whether his p(doom) is zero, he answers sideways: humans persevere. We have lived through the nuclear age, through pandemics, through all these challenges. We are going to survive.

But there is going to be a lot of disruption. And the one constant in human civilization is that change only accelerates. That is the thing he worries about most for the next generation and even the current one: people are going to have to get really, really comfortable with change. What you think your job is today may not be the job even twelve months from now. And people are so resistant to that. The people who embrace it, who say they love that this is a new challenge and an opportunity to learn something new, are the ones who will do well.

The 10 percent, the 40 percent, and the mules (1:18:02)

McCormack disagrees on one point. He does not think people are scared of change in itself. He thinks they are scared of the financial impact of change. Some number came out recently, 80% or 90% of Americans with something like $500 in savings. He pictures the classic London couple: he is a lawyer, she is a copywriter, both jobs at risk, pulling maybe £250,000 between them. What do they retrain as?

Alperovitch disagrees, and grounds it in what he sees as an active investor in tech companies and a member of several boards. Look at an engineering team in a portfolio company. Maybe 10% of it enthusiastically embraces AI and loves it. At the other end of the spectrum, probably 30 to 40% are completely resistant: this is going to destroy my job, I do not want to touch it, I do not want to learn something new, I know how to do this thing and I want to keep doing it that way.

"Those people are going to be gone." Those are the people companies are actively trying to manage out.

McCormack: "They're the mules."

The ones in the middle are the gettable ones. You can tell them: be like the other 10%, embrace it, you will be ten times more productive, you will be much more valuable to the company, and the company will never want to get rid of you because you are doing so much. Yes, you have to learn more things and change how you work. But so many people, for whatever reason, just like to be set in their ways: this is how I have always done this job, this is how I will always continue to do this job. Those are the people who struggle in the transformation.

Where he is investing (1:19:34)

Is it hard right now to find moats? It depends. Alperovitch also invests a lot in defense tech, and that is much easier. In software he thinks it is genuinely hard, though not all software.

Cyber, he says, is one of the bright spots. McCormack: "I'm not going to vibe code a cyber protection for me." Right, because it is not just a software package. It involves humans, it involves intelligence, it involves understanding the adversary. That is harder to replace.

But if you are building an accounting package or a CRM application, those are going to be really tough. McCormack has looked at both. Accountancy has a bit of protection from legal requirements. CRM is going to change. Alperovitch: "If you're Salesforce, you got a lot to worry about at the moment."

They close there. Alperovitch points people to his podcast, Geopolitics Decanted, and his book, World on the Brink: How America Can Beat China in the Race for the 21st Century. McCormack: "I will be keeping out for the next five years. It's going to be wild times, man."

The chronology he walks through

• 1944 The US drafts Operation Causeway in August, a 300,000 troop invasion of Formosa expecting 45,000 casualties. Roosevelt and MacArthur kill it and take Okinawa instead.
• Late 1990s Chinese cyber intrusion activity is believed to begin, long before anyone is paying attention. By the early 2000s it is running at full speed.
• 1960s China gets the bomb and settles on a stable arsenal of roughly 300 warheads, held for decades.
• 1979 Carter recognizes Beijing, derecognizes Taipei, and pulls over 10,000 US troops off Taiwan. The island's military begins 47 years of isolation. China fights Vietnam, its last war.
• Jan 2010 Google calls Alperovitch at McAfee. He runs the investigation and names it Operation Aurora, the first time the world sees nation state hacking for what it is.
• 2010 to 2011 He finds the same pattern everywhere and calls it the greatest transfer of wealth in history. CrowdStrike is founded to stop it.
• 2012 Xi Jinping takes office, declares China cannot pass Taiwan to future generations, and orders nuclear parity: 1,500 warheads.
• 2015 Made in China 2025 launches, targeting independence in semiconductors and aerospace. China stands up a cyber command to run intrusion at industrial scale.
• Around 2020 The mission changes. Volt Typhoon begins pre-positioning inside water utilities, ports, substations and power plants, taking nothing.
• Dec 2021 Alperovitch publicly predicts the Russian invasion of Ukraine from cyber indicators and is called a warmonger. German intelligence tells him at the Munich Security Conference, the weekend before, that it will never happen.
• 2022 The CHIPS Act passes with $52 billion in grants. Most of the money does not go out until 2024. Ukraine is invaded.
• 2023 Pat Gelsinger tells Alperovitch that Intel may not survive without the money that quarter. He does not get it that quarter, or the next.
• Early 2025 DeepSeek lands and China settles into a persistent six to nine month gap behind the American frontier.
• Feb 2026 Mythos is trained. Six months later the newest Chinese models are still nowhere close to it.
• Today TSMC commits $250 billion and six Arizona fabs under tariff pressure. Trump announces a $1.5 trillion defense budget aimed at China. A $14 billion Taiwan arms package sits held up. Chinese warheads reach roughly 600.
• 2027 Xi's readiness deadline, his Party Congress re-election, and the PLA's centenary all land in the same year.
• 2030s The US intelligence community expects Chinese warhead numbers to match American ones.
• 2032 The next re-election year. Xi turns 79. In Alperovitch's reading, this is the date that actually matters.

Key takeaways

• Pre-positioning is a different operation from espionage, and that is the whole argument. Volt Typhoon, in Alperovitch's account, targets small water utilities, port facilities, electric substations and power plants specifically because there is nothing there worth stealing. The intruders get in and do nothing. He calls it preparation of the battlefield, and it is the cyber indicator he reads as an invasion signal rather than a theft problem.
• The technique is designed to defeat the detection model. Living off the land means using tooling already present in Windows rather than importing custom malware, and stealing legitimate credentials rather than creating accounts. Nothing leaves the network, so the classic "valuable data crossed the perimeter" trigger never fires.
• 2032, not 2027. Alperovitch reads Xi's 2027 readiness order as a political milestone for the Party Congress and the PLA centenary, not an invasion date. The window that worries him is between 2027 and 2032, when the force exists and Xi is not yet 79 with a stagnating economy and latent opposition inside the Party.
• Taiwan may be the hardest amphibious target on earth, and the US already concluded that once. Operation Causeway in August 1944 called for over 300,000 troops, double Normandy, with 45,000 expected casualties against 60,000 Japanese defenders, and it was cancelled as too costly. Every technological change since favors the defender, and no one in the PLA has combat experience; the last Chinese combat veteran was purged.
• There will be no insurgency and no resupply. Alperovitch's strategic conclusion is that the entire Taiwanese effort has to go into sinking the Chinese fleet before it lands, plus deep strike into the mainland to hit the ports and launchers, because you cannot win the interceptor cost exchange and you cannot sustain resistance on an island that cannot be resupplied.
• The chip dependency is bigger than the AI conversation admits. Roughly 90% of advanced chips and over 40% of all chips route through Taiwan, but the volume exposure is in foundational chips: an iPad has about three advanced chips and 150 to 160 foundational ones. Fabs are also fragile, and a power outage alone can destroy a production line. His estimate for a conflict or blockade is $5 trillion of GDP in year one, and he thinks that is low.
• Six to nine months is the entire American lead, and he wants it measured in years. His prescription is to run faster and simultaneously kneecap: close the H20 and H200 sales, close the remote compute access loophole that lets Chinese firms train in third country data centers, and sanction Chinese labs for illicit distillation.
• He is a China hawk who is relaxed about downloading Chinese open weight models. Run them locally or on an American cloud and China gets neither your data nor your money. What he objects to is using Chinese clouds and paying Chinese subscriptions. He also expects open weights from leading Chinese labs to disappear within a couple of years on pure economics.
• The 1-10-60 defense doctrine he invented is broken by agents. Detect in one, investigate in ten, respond in sixty no longer holds when the attacker is a machine, so defenders have to field their own agents, with all the paperclip style failure modes that implies. His counter-forecast is that AI assisted vulnerability discovery and patching makes major software more secure than ever within about eighteen months, while a long tail of small applications stays wide open.
• The thing that actually scares him is not cyber. It is step by step bioweapon and chemical weapon instructions on demand, which he says is getting pretty damn close, with the nuclear case closing later through robotics.

Chapters

• 0:00 Sleeper Cells In America's Power Grid
• 1:33 America Planned To Invade Taiwan In 1944
• 3:35 The Worst Place On Earth To Invade
• 9:47 Why 2032 Matters More Than 2027
• 13:00 Run Faster And Kneecap The Other Guy
• 15:55 China Is Six To Nine Months Behind
• 17:17 Why Taiwan Isn't Ready
• 19:15 There Will Be No Insurgency
• 23:41 Would America Actually Fight?
• 26:13 The Google Hack That Started It All
• 30:01 Volt Typhoon: Preparing The Battlefield
• 33:42 Should America Steal It All Back?
• 34:46 This Is A New Cold War
• 37:34 Bigger Than The Moon Race
• 42:44 What Happens To AI If Taiwan Falls
• 47:06 The Chips Act Disaster
• 54:45 Bringing The War To Russia
• 1:01:57 Everyone Becomes A Founder Again
• 1:05:22 AI Psychosis And £6,000 Of Tokens
• 1:15:50 Anthrax Instructions On Demand

Notable quotes

• "They just sit and wait." / "Sleeper cells." / "Sleeper cells waiting for orders to come in and destroy that network in the event of conflict, or in an attempt to deter us from intervening in a conflict." Alperovitch and McCormack, 0:12
• "Now, this is what the military would call preparation of the battlefield. And this is one of the many signs that I look at and say this is now different. They're preparing for an invasion of Taiwan." Alperovitch, 31:04
• "If there is any sort of conflict or blockade of Taiwan, you have $5 trillion of GDP wiped out in the first year." / "Jesus." / "And I think that's probably low." Alperovitch and McCormack, 51:44
• "There is literally no one in the Chinese military, in the PLA, that has any combat experience whatsoever." Alperovitch, 4:33
• "For people that are sort of dismissing this as a foregone conclusion that China will take Taiwan: look at the map, go visit Taiwan. This is one of the worst places possible to invade." Alperovitch, 5:03
• "He's in fact a new emperor of China." Alperovitch on Xi Jinping, 8:07
• "You win the race in two ways. You run faster, which we obviously need to continue to do, to invest in this technology. But you also kneecap the other guy." Alperovitch, 15:19
• "There are two types of companies: those that know that they've been hacked and those that don't yet know, but they've all been hacked." Alperovitch on the line he coined after Aurora, 27:32
• "The greatest transfer of wealth in history." Alperovitch on Chinese IP theft via cyber, 27:32
• "So, small water utilities, port facilities, electric substations and power plants, and they get into those networks and they do nothing." Alperovitch, 30:33
• "If we can't beat them, let's join them, because guess what? We're so much better than they are." Alperovitch on stealing Chinese IP back, 33:38
• "The more I spent time researching the book and writing it, the more I realized it's almost exactly the same." Alperovitch on the second cold war, 34:41
• "You will know when it's an invasion, because this is going to be the biggest military operation in human history." Alperovitch, 21:26
• "If it doesn't, Taiwan will fall." Alperovitch on whether the US would intervene, 23:27
• "Even if they think there's a 1% chance that the US might be in, they have to treat us as a 100%." Alperovitch on strategic ambiguity, 23:57
• "There's not going to be an insurgency on Taiwan. For an insurgency to take place, you need a resupply route. There is no resupplying of Taiwan." Alperovitch, 19:53
• "Everything around us is now dependent on chips." Alperovitch, 44:53
• "Demetri, if I don't get that money now, this quarter, Intel may not survive." Pat Gelsinger to Alperovitch in 2023, as recounted at 47:24
• "The popularity of AI in America at least is somewhere between gonorrhea and Congress." Alperovitch, 1:00:38
• "I've never seen more doomerism from an industry then here. Like, do you think Henry Ford was going around saying millions of people are going to die of car accidents?" Alperovitch, 1:01:09
• "Dean, I think they will, but just like the study of the French Revolution. It's really interesting. I don't know that it's really applicable." Alperovitch on whether people will still study computer science, 1:05:13
• "I had one month where I spent about £6,000 on tokens." McCormack, 1:05:22
• "Mark Andre said we're all turning into vampires, because we just can't stop. We're working through the nights." Alperovitch, 1:07:16
• "I pioneered this concept of 1-10-60, where if you detect an intrusion in one minute, investigate in 10 minutes, respond to it within an hour, you're going to thwart almost every attack that comes at you. That all completely falls apart in the age of agents." Alperovitch, 1:10:51
• "There's going to be a long tail of this crappy software that everyone has that will create a huge vulnerability surface for all of us." Alperovitch, 1:13:27
• "As long as there are bad people out there, nation states or criminals that want to do harm, you're going to have a problem." Alperovitch, 1:13:58
• "Take COVID and make it 10 times more deadly. Totally possible." Alperovitch, 1:16:30
• "People forget that 'may you live in interesting times' is a Chinese curse." Alperovitch, 1:17:01
• "Those people are going to be gone." / "They're the mules." Alperovitch and McCormack on engineers who refuse to use AI, 1:19:03

Resources mentioned

People

• Dmitri Alperovitch, co-founder and former CTO of CrowdStrike, chairman of Silverado (X)
• Peter McCormack, host (X, Instagram)
• Xi Jinping · Mao Zedong · Deng Xiaoping · Vladimir Putin
• Franklin D. Roosevelt and Douglas MacArthur, who cancelled Operation Causeway
• Jimmy Carter, who derecognized Taiwan in 1979 · Donald Trump · Joe Biden
• Pat Gelsinger, then CEO of Intel, on the CHIPS Act delay
• John McAfee, who predated Alperovitch at McAfee by decades
• Aleksandr Dugin and Darya Dugina
• Balaji Srinivasan · Marc Andreessen · Henry Ford
• Greg Allen, a guest earlier the same day, source of the tractor and mules argument

Alperovitch's own work

• World on the Brink: How America Can Beat China in the Race for the Twenty-First Century, PublicAffairs, 2024, with Garrett M. Graff
• Silverado Policy Accelerator, where he is chairman
• Geopolitics Decanted, his podcast
• CrowdStrike, which he co-founded, and the 1-10-60 doctrine he pioneered
• Operation Aurora, the January 2010 Google intrusion investigation he ran and named

Threat actors, campaigns and agencies

• Volt Typhoon, the pre-positioning campaign at the center of the episode. The joint CISA, NSA and FBI advisory is AA24-038A, PRC State-Sponsored Actors Compromise and Maintain Persistent Access to U.S. Critical Infrastructure
• Living off the land, the technique, covered in advisory AA23-144A and the accompanying joint guidance on identifying and mitigating LOTL techniques
• CISA · NSA · FBI · CIA · US Cyber Command
• People's Liberation Army and its cyber force stood up in 2015
• Ministry of State Security, the MSS
• Chinese Communist Party and the Central Military Commission

Military history and doctrine

• Operation Causeway, the abandoned August 1944 American plan to invade Formosa
• Operation Overlord and the Normandy landings · Battle of Okinawa
• Taiwan Strait · Taipei · Taiwan under Japanese rule
• Sino-Vietnamese War, 1979, China's last war
• Battle of Stalingrad and the siege of Sarajevo, his evidence that blockades do not win
• Strategic ambiguity · casus belli
• M1 Abrams and the Battle of Kursk, the tank battle that will not happen on Taiwan
• MIM-104 Patriot interceptors and Shahed 136 drones, the cost asymmetry problem
• Crimean (Kerch) Bridge truck bombing · Donbas · Russian invasion of Ukraine
• Strait of Hormuz · War in Afghanistan · Iraq War
• Royal Military Academy Sandhurst, one of the schools Taiwan cannot send officers to
• Munich Security Conference, where he was told the Ukraine invasion would never happen

Chips, AI and policy

• TSMC · Nvidia · AMD · Intel · Samsung Semiconductor · Micron · Qualcomm · Apple
• CHIPS and Science Act, 2022, $52 billion in grants
• Made in China 2025
• High Bandwidth Memory · Boeing 787 Dreamliner, his steel and aluminum analogy
• Anthropic and Claude Code · OpenAI · Meta Llama · Microsoft Azure · Amazon Bedrock
• DeepSeek · Moonshot AI, maker of Kimi · Qwen
• Knowledge distillation, the mechanism he wants Chinese labs sanctioned for
• Recursive self improvement and the paperclip maximizer · p(doom)
• Unitree robots · Salesforce · Georgia Tech, his alma mater
• Anthrax and COVID-19, his bioweapon examples · Ellis Island · French Revolution · "may you live in interesting times"

Sponsors named in the episode

• IREN, the lead sponsor, data centers and GPU infrastructure on renewable energy
• Expat Money, second residencies and citizenships
• Ledn, Bitcoin and Tether Gold on one platform
• Real Bedford, the football club McCormack owns and keeps building software for

Related pages on this site

• Tom Bilyeu: the Supreme Court and the China backdoor, which covers the same distillation and AI arms race argument from a very different political frame
• China's 10 trillion parameter model, on how far the Chinese labs have actually got
• David Bombal on vibe hacking with AI, the hands on version of the agentic offense problem Alperovitch describes at 1:10:21

Where it stands

Everything above is the interview rebuilt in Alperovitch's own frame. This section is not. It is the honest ledger of which parts are load bearing public record, which parts are his read, and where informed people disagree.

The Volt Typhoon core is corroborated, and not only by him. This is the strongest part of the episode. In February 2024, CISA, the NSA and the FBI published advisory AA24-038A with partner agencies in Australia, Canada, the UK and New Zealand, stating that PRC state-sponsored actors had compromised and were maintaining persistent access to US critical infrastructure, naming communications, energy, transportation, and water and wastewater systems, and specifically framing the activity as pre-positioning for disruption rather than espionage. The living off the land guidance that Alperovitch describes at 31:35 is published government material, not vendor marketing. Where he says the US government calls it Volt Typhoon, that is accurate. The multi-year dwell times and the targeting of small utilities with nothing to steal are in the advisories.

The attribution reasoning is worth stating plainly, because he compresses it. Attribution in this space rests on infrastructure overlap, tooling, operating hours, target selection, and intelligence not visible to the public. Alperovitch has been right on high profile calls before, notably Aurora and the December 2021 Ukraine prediction, and both are checkable. But "we can read strategy from cyber telemetry" is an inference method, not a measurement, and it produces confident conclusions from incomplete data by design. He is candid about the 2021 case being contested at the time. He is less candid that a method which reads intent from technique will sometimes read invasion prep into activity that is contingency planning, deterrence signaling, or bureaucratic momentum. Beijing denies the campaign entirely, which is expected and carries no evidentiary weight either way.

One gap worth flagging for anyone using this page as a reference. The episode names exactly one campaign. It does not discuss the separate telecommunications intrusion set that US officials have described publicly, and it does not walk through the specific operational technology incidents in the water sector. If you came here for the full taxonomy of named Chinese campaigns, this interview is not it; it is one campaign used as an indicator.

The Taiwan military analysis is mainstream, the dates are his own. The difficulty of an amphibious assault on Taiwan, the narrow seasonal windows, the beach and mountain geography, the PLA's total lack of combat experience, and the sealift shortfall are all standard in open source defense analysis. Operation Causeway is real and declassified. Where he departs from the consensus is the 2032 framing. The 2027 readiness order is widely reported and widely argued over, and serious analysts split on whether it signals intent, capability benchmarking, or something closer to a bureaucratic target. Alperovitch's reading, that 2027 is a Party Congress applause line and the real risk window opens after it, is a specific and falsifiable thesis rather than a settled finding. His 50/50 estimate of US intervention is a personal judgment and he presents it as one.

"There will be no insurgency" is his most contestable strategic claim. The occupied Ukraine comparison at 20:25 is doing heavy lifting, and analysts who study resistance movements would push back on the logic that brutality plus no resupply equals no resistance. It is a defensible position. It is not a consensus one.

The economics are estimates and he says so. The $5 trillion first year GDP figure is in the range other analyses have produced for a Taiwan contingency, but every such number rests on assumptions about blockade duration, fab destruction and substitution that vary enormously between models. He flags twice that he thinks his own number is low, which is an argument, not data. The chip share figures, roughly 90% of advanced and over 40% of all chips, and the roughly 3 advanced versus 150 to 160 foundational chips in a single device, are the kind of industry estimates that move with definitions of "advanced," but the structural point about foundational chips being the underattended dependency is well made and under-discussed elsewhere.

Read the commercial position honestly. Alperovitch co-founded the largest endpoint security company in the world and remains an active investor in defense tech, which he says on air at 1:19:34. He is also chairman of a policy organization that advocates for exactly the export controls, arms sales and sanctions he prescribes here. None of that makes the Volt Typhoon reporting wrong; the government advisories stand on their own. But the threat assessment and the policy prescriptions come from someone whose professional standing and portfolio both benefit if the assessment is taken seriously, and that is worth holding in mind when he moves from what the telemetry shows to what Washington should do about it.

The most interesting argument here gets the least scrutiny. His proposal that the United States should conduct commercial IP theft and battlefield preparation against China, at 33:42, is offered almost as common sense, and it would be a deliberate abandonment of a norm the US has spent two decades trying to establish, with consequences he does not walk through. He does not address what happens to the American position when it can no longer distinguish its own conduct from the conduct it condemns, nor the escalation risk of dormant implants sitting in two grids instead of one. It is the boldest claim in the episode and it goes unchallenged.

Where the AI material is softest. The specific model names, capability gaps and internal access arrangements discussed from 15:55 onward are the kind of thing that is genuinely known inside the industry and difficult for a reader to verify from outside. The "six to nine months behind" figure is a real and widely cited estimate, but it is a moving target that depends heavily on which capability you measure. The claim that the US should sanction Chinese labs for distillation rests on evidence he alludes to rather than presents. His forecast that software becomes dramatically more secure within eighteen months is a prediction, made by someone who opened the same segment by saying anyone predicting a year out is deluding themselves.