Falling behind China on deep tech and chips

The Short Version

In three very different conversations, the same worry keeps coming up: the contest between the United States and China is increasingly a contest over technology, and the U.S. can't assume it will stay ahead. Semiconductor analyst Dylan Patel and AI researcher Nathan Lambert explain why chips are now the front line: America's AI lead depends on Nvidia's designs and on TSMC's factories in Taiwan, export controls are an attempt to limit how much AI China can run, and cutting China off may push it toward Taiwan [2]. Elon Musk puts the rivalry in historical terms, with China's economy possibly growing to two or three times America's, a Taiwan flashpoint he calls certain from China's standpoint, and a near-term bottleneck in chips, transformers and electricity [3]. Former Navy pilot Ryan Graves makes a more unusual version of the argument: China's state money can fund long-shot science that American venture capital avoids, and secrecy that made sense when the U.S. had no peer now risks leaving it behind [1]. They differ on details and remedies, but all treat technological leadership as a national security question, not just an economic one [1][2][3].

Table of contents

What They Said

Dylan Patel and Nathan Lambert: chips, controls and Taiwan

Lex Fridman's long conversation with Patel, founder of the semiconductor research firm SemiAnalysis, and Lambert, a research scientist at the Allen Institute for AI, was prompted by the DeepSeek moment, when a Chinese lab's cheap, open-weight reasoning model shook the AI world and Nvidia's stock [2]. Their account is the most detailed in these sources on where the U.S. is ahead, where it is exposed, and what policy is trying to do.

What DeepSeek showed, and what it didn't. DeepSeek achieved genuine engineering breakthroughs [2]. Its model has over 600 billion parameters but activates only about 37 billion per token, using an unusually sparse mixture-of-experts design, and a memory-saving technique called multi-head latent attention [2]. Because it trained on H800s, Nvidia chips allowed in China with reduced interconnect bandwidth, it went below Nvidia's standard communications library and programmed GPU cores close to the assembly level [2]. As Patel puts it, "necessity is the mother of innovation" [2]. Export controls, in other words, pushed Chinese engineers to get more out of weaker hardware.

But the famous low cost figure was misunderstood. It covered only the final pre-training run of V3 on about 2,000 H800s, excluding research and experiments that typically take two to four times that compute, and excluding R1's reinforcement learning [2]. SemiAnalysis believes DeepSeek has closer to 50,000 GPUs [2]. Its parent, the hedge fund High-Flyer, claimed in 2021 to have the largest A100 cluster in China, 10,000 GPUs, before export controls [2]. Patel says there has been "a 1200 x reduction in cost to achieve the same level of intelligence as GPT-3" in about three years, and DeepSeek sits on that trend line rather than below it [2]. And "no one spent more than a billion dollars on a model that's released publicly" [2]. The lesson is not that China leapfrogged the U.S. overnight, but that a capable Chinese lab can stay close to the frontier with fewer and weaker chips.

How export controls work, and leak. Patel traces how U.S. controls evolved: first limiting both compute and interconnect (which permitted the H800), then compute only, and now the China-specific H20, which has cut compute but equal or better memory bandwidth and capacity than the H100 [2]. Nvidia shipped about a million H20s to China last year and recently canceled orders for about two million more, which Patel reads as a sign further restrictions are coming [2]. He argues the H20 is actually well suited to reasoning [2]. On smuggling, he estimates 200,000 to 300,000 GPUs were routed to China last year through places like Singapore and Malaysia, including a man flying first class to China with a Supermicro server, and says ByteDance legally rented huge numbers of GPUs from Western clouds until new AI diffusion rules [2].

What controls can and can't do. Lambert reframes the goal: China will always be able to train models, but controls limit how much AI it can run, and reasoning models make running AI far more compute-intensive [2]. OpenAI's o3 solved ARC-AGI puzzles using about 1,000 samples at an estimated five to twenty dollars per question [2]. Fridman summarizes Anthropic CEO Dario Amodei's case: if super-powerful AI arrives around 2026 and confers military advantage, it is better for a democracy to lead [2].

The timing bet. Here the two guests add important nuance. Lambert's own AGI estimate is later than the AI CEOs': "For me, it's probably after 2030," discounting their two-year forecasts because of their incentives [2]. Patel expects capabilities around 2027 or 2028 but says the cost of running them at scale will slow deployment, and he worries that China could point such systems at military uses faster, as it did with drones [2]. His sharpest point is that export controls are a bet on timing: if AI is not transformative within five to ten years, controls effectively help China win long term, because China can outbuild the U.S. in power and industry and will subsidize its own chips [2]. Controls only make sense if AI changes society soon, which he believes [2]. He adds that China has only just started to take AGI seriously: DeepSeek's CEO recently met a top Chinese official, and a trillion-yuan (about 160 billion dollar) AI subsidy followed [2].

Taiwan. The biggest risk, Patel says, is that cutting China off from critical technology pushes it toward a blockade or takeover of Taiwan [2]. He notes that the world has historically been most peaceful under a single dominant power, and that transitions between powers rarely go smoothly [2]. The stakes in Taiwan come from TSMC. As the cost of each new fab doubled every few years, now 30 to 40 billion dollars or more for a leading-edge fab, companies that designed and built their own chips dropped out, and TSMC aggregated everyone's demand [2]. Its edge is also cultural: top graduates of Taiwan's best university go there for modest salaries, and after earthquakes engineers simply show up to fix the fabs [2]. Only three places do leading-edge semiconductor R&D: Hsinchu in Taiwan, Hillsboro in Oregon (Intel) and Pyeongtaek in South Korea (Samsung) [2]. TSMC's Arizona fabs, roughly 20% of its 5-nanometer capacity, would stop working within a year or two without Taiwan, and almost every product, from cars to fridges, depends on TSMC [2].

Can the U.S. bring chipmaking home? Patel estimates that fully doing so "would take a decade and a trillion dollars," compared with a CHIPS Act of about 50 billion dollars versus China's roughly 200 billion a year in subsidies [2]. Intel, once the manufacturing leader for 20 years, lost its way through management mistakes, including saying no to the iPhone, and Patel calls its dire state a national security concern [2]. Meanwhile, China is racing to dominate the long tail of older chips, like power and analog chips, and Patel says export controls have accelerated that [2]. Fridman argues that importing the world's best talent should not be controversial [2].

Power and scale. The race is also physical. Patel says data centers could rise from two or three percent to 10% or more of U.S. electricity by the end of the decade [2]. GPT-4 was trained on about 20,000 A100s drawing 15 to 20 megawatts; by 2024 Meta was at about 128,000 GPUs and 150 megawatts, and xAI built a 200,000-GPU cluster in Memphis [2]. Companies are turning to natural gas because building power plants and transmission is slow [2]. He sees Trump-era deregulation, such as easy permitting on federal land, as something that could accelerate the race [2].

Influence through models. Patel and Lambert also raise a softer form of competition: whose values are built into widely used AI. Patel notes that American-trained models are quietly making British spelling obsolete, and asks, "What happens when the models are backdoors not just to computer systems, but to our minds?" [2]. He cites Sam Altman's line that "superhuman persuasion will happen before superhuman intelligence," while saying he doesn't believe DeepSeek has done anything like this [2]. Lambert adds that Anthropic research shows planted phrases in pre-training data can trigger hidden behaviors [2]. Fridman reads Mark Zuckerberg's remark that there will be a global open-source standard and it matters that it be American [2]. Lambert calls R1's permissive license a major reset, the first truly frontier model that is open and commercially usable without restrictions [2].

Elon Musk: economics, Taiwan and the electricity crunch

Musk's conversation with Fridman, recorded in late 2023, frames the rivalry through history [3]. A self-described lover of war history, he points to Athens and Sparta, which both saw the Peloponnesian War coming and still could not avoid it [3]. "And the foundation of war is economics," he says [3]. In the conversation, he describes the situation plainly: "the economy is likely to be two, perhaps three times larger than that of the US" [3].

China's strengths. Musk is candid about what China does well. "The sheer number of really smart, hardworking people in China is incredible," he says, and "the architecture in China that's in recent years is far more impressive than the US," pointing to train stations, buildings and high-speed rail [3]. He has spent a lot of time in China and met its leadership many times [3].

Taiwan. Musk sees Taiwan as the clear flashpoint. From China's perspective, he says, it is like a U.S. state that isn't part of the country, "more significant than Hawaii" [3]. He adds that in China's view "the only reason it hasn't been is because the US Pacific fleet" [3]. His bottom line: "China has been very clear that they will incorporate Taiwan peacefully or militarily, but that they will incorporate it from this standpoint is 100% likely" [3]. That matters for chips because, as Patel explains, Taiwan is where TSMC's leading-edge manufacturing and R&D sit [2].

Hopeful signs. Musk is not purely alarmed. He notes that China has historically been inward-focused rather than acquisitive, having fought brutal civil wars instead, and says "people in China think about China 10 times more than they think about anything outside of China" [3]. He argues the U.S. is also unusual in not being acquisitive, having helped rebuild Europe and Japan after World War II [3]. His general view is that ignorance, not other people, is the real enemy [3].

The technology bottlenecks. Musk's other contribution is a physical view of the AI race. He predicts a sequence of constraints: a silicon shortage today, then a shortage of voltage step-down transformers in about a year ("You need transformers to run transformers"), then general electricity shortages in about two years [3]. "The initial constraint being literally voltage step down transformers. Because you've got power coming in at 300,000 volts. And it's got to step all the way down eventually to around 0.7 volts," he explains, and "the industry is not used to rapid growth" [3]. He told a gathering of utility companies to prepare for electricity demand to roughly triple, driven in the short term more by electrified transport and heating than by AI [3]. He argues for far more batteries, because the grid is sized for the worst second of the worst day of the year with almost no buffering [3]. Patel's numbers on data center power and xAI's gas-turbine-powered Memphis cluster fit the same picture [2].

Regulation and openness. Musk says "I've been pushing for some kind of regulatory oversight for a long time" on AI, calling himself a Cassandra on the subject [3]. He wants an objective third-party referee that can see what leading players are doing [3]. At the same time he describes regulation's absurd side, such as SpaceX having to play sonic booms to a seal through headphones, and says Tesla answers to over 100 regulatory agencies [3]. On open source, he generally favors it, perhaps with a delay of around six months [3]. That tension, between oversight and speed, runs under the whole competition debate.

Ryan Graves: deep tech, state money and secrecy

Graves, a former F/A-18 pilot who now leads the nonprofit Americans for Safe Aerospace, comes at the question from an unusual direction: unidentified anomalous phenomena (UAP) [1]. His argument about China is less about chips and more about who is willing to fund long-shot science.

The funding gap. Graves believes China is one of the biggest reasons the U.S. government is opening up about UAPs [1]. The American model relies on venture capital, which avoids research that takes a decade or more to pay off; China can pour state money into long-shot science [1]. If China is investigating the same phenomena and recovered objects with more resources, he argues, keeping the topic buried in classified programs risks being outcompeted [1].

What alarmed scientists. He says he's met scientists who were alarmed at an international conference when Chinese researchers requested collaboration on topics once mocked in the West: gravity manipulation (now called extended electrodynamics) and cold fusion (now "low energy nuclear reactions") [1]. Rogan brings up a researcher who claimed a gravity breakthrough and later reportedly went to China [1]. These are Graves's and Rogan's accounts, and the sources give no further detail about them.

Secrecy's cost. Graves agrees that secrecy made sense when the U.S. was the lone superpower, but "when we're operating in a world where we have near-peer or peer adversaries," the only answer is to activate America's scientists and engineers [1]. He wants a public, unclassified national effort led from the White House, drawing on weather radar and satellite data, Department of Energy computing, the National Science Foundation and the Office of Strategic Capital [1]. He proposes aligning UAP research with things the Pentagon already wants to buy, such as total airspace awareness [1].

Quantum and data. Graves and Rogan also touch the computing race. Rogan relays Marc Andreessen's description of quantum chips solving problems that would take longer than the universe's lifetime, and worries about encryption, financial markets and infrastructure becoming vulnerable [1]. Graves adds that China has reportedly been storing encrypted data so it can decrypt it later [1].

A domestic example. The drone incursions over the East Coast, which prompted the conversation, illustrate his broader point about capacity. Authorities couldn't lock onto the objects with infrared, electronic warfare attempts failed, and legal knots, such as needing approval up to the deputy attorney general to intercept signals, slowed any response [1]. His summary: "It's hard, but it's very solvable," and "just a resource problem it's not a technology problem" [1].

Where They Agree, and Where They Don't

Agreement: China is a peer competitor in technology. Graves describes a world of "near-peer or peer adversaries" [1]; Patel and Lambert describe a Chinese lab staying close to the frontier despite export controls [2]; Musk describes an economy that may grow to two or three times America's and a workforce and infrastructure he openly admires [3].

Agreement: state money versus market money. Graves contrasts American venture capital, which avoids decade-long bets, with China's state funding [1]. Patel makes the same contrast in chips: a CHIPS Act of about 50 billion dollars versus roughly 200 billion a year in Chinese subsidies, and a trillion-yuan AI subsidy that followed DeepSeek's CEO meeting a top official [2].

Agreement: Taiwan is the flashpoint. Musk calls China's incorporation of Taiwan certain from its standpoint [3]. Patel warns that cutting China off from chips may push it toward a blockade or takeover, and explains why that would matter so much: TSMC's leading-edge production and R&D sit there, and even its Arizona fabs depend on Taiwan [2].

Agreement: physical capacity matters. Musk's transformer and electricity bottlenecks [3] and Patel's projection of data centers reaching 10% or more of U.S. electricity [2] both point to power and manufacturing as the real constraints, not just algorithms. Graves's call to marshal national labs, satellite data and computing is a version of the same idea for science [1].

Where they differ: how to respond. Patel supports export controls only on the premise that AI will be transformative soon, and warns they can backfire by accelerating China's own chip industry and raising the risk over Taiwan [2]. Lambert emphasizes that controls limit how much AI China can run, not whether it can train models [2]. Graves's remedy is openness: less secrecy, more public science [1]. Musk's emphasis is on building capacity (chips, transformers, power, batteries) and on some regulatory oversight of AI [3]. Fridman adds talent: importing the world's best people should not be controversial [2].

Where they differ: how worried to be. Musk offers reassurance alongside alarm, noting China's history of being inward-focused rather than acquisitive [3]. Patel is more worried, observing that power transitions rarely go smoothly and that China could turn AI toward military uses faster [2]. Lambert's later AGI timeline (after 2030) implies less urgency than Patel's 2027 or 2028 [2]. Graves's concern centers on exotic research areas that most scientists don't take seriously, so his claims are the most speculative of the group [1].

Where they differ on openness. Patel and Lambert see open models as a battleground for influence, with Zuckerberg arguing the global standard should be American and R1's license resetting the field [2]. Musk favors open source with a delay [3]. Graves wants openness in government science [1]. None of them treats secrecy as a sufficient strategy on its own.

Why It Works

The sources explain why technology leadership has become a security issue through several mechanisms.

Compute is power. If super-powerful AI confers military advantage, whoever can run the most of it matters [2]. Reasoning models generate tens of thousands of output tokens, so memory usage balloons and far fewer users can be served per GPU, which is why limiting chips limits AI deployment [2]. Patel worries that China could apply AI to military uses faster, as it did with drones [2].

Chokepoints create leverage and risk. Leading-edge chipmaking is concentrated in three places, and the most important is Taiwan [2]. That gives the U.S. and its partners leverage through export controls [2], but it also creates a single point of failure: TSMC's Arizona fabs would stop working within a year or two without Taiwan [2], and China views Taiwan as a core interest it will incorporate one way or another [3].

Scale economics favor whoever keeps investing. Fab costs doubling every few years is what concentrated manufacturing in TSMC [2]. Subsidy gaps (50 billion versus roughly 200 billion a year) and China's push into older chips [2] suggest that over time, sustained state investment can wear down a lead, which is exactly Patel's argument that controls help China if AI's payoff is slow [2]. Graves applies the same logic to basic science: state money can fund what venture capital won't [1].

Constraints drive innovation on both sides. Export controls pushed DeepSeek to program close to the hardware and squeeze more out of H800s [2]. Musk's view that first you make a technology work and then make it efficient [3] fits the same pattern. Restrictions can buy time, but they also teach the restricted side to do more with less.

Physical infrastructure is slow. Power plants and transmission take time to build, which is why companies turn to natural gas [2], and transformers are about to become a bottleneck [3]. A country's ability to build fast is itself a technology advantage.

Information and influence. Models can carry hidden values or planted behaviors [2], and stored encrypted data could someday be decrypted by quantum computers [1]. Technology leadership also shapes who controls information.

Where It Applies, and Where It Doesn't

Where the argument is strongest. AI chips and chipmaking, where Patel and Lambert give detailed evidence on export controls, smuggling, TSMC's dominance and the cost of catching up [2]; and power and grid capacity, where Musk and Patel agree on looming shortages [2][3].

Where it depends on timing. Export controls are only worth their costs if AI becomes transformative within about five to ten years, according to Patel [2]. If Lambert's later timeline is right, the balance of costs and benefits could look different [2].

Where claims are less settled. Graves's examples of China pursuing gravity manipulation and low-energy nuclear reactions come from conversations with scientists and a conference request, and the connection to UAPs is his interpretation [1]. Rogan's quantum computing fears are relayed secondhand [1]. The report that China stores encrypted data for later decryption is presented as a reported claim [1]. These points support the general theme of a deep-tech race but are much weaker evidence than the chip-industry details.

Where the picture is more hopeful. Musk argues China has historically been inward-looking rather than expansionist [3], and Patel and Lambert say neither of them thinks the Chinese government subsidized DeepSeek or timed its launch to short Nvidia; DeepSeek simply ships fast [2]. Competition does not have to mean conflict, though Taiwan remains the danger point [2][3].

Where "falling behind" overstates it. Patel and Lambert say the DeepSeek moment was real but widely misunderstood: its cost figure covered one training run, its prices sit on a long-running cost curve, and U.S. firms still lead in total compute, with Google holding the biggest total compute and xAI the largest single cluster [2]. Nvidia remains the company clearly making money [2]. The risk is less that the U.S. has already lost and more that its lead rests on fragile foundations.

What To Do With It

1. Read AI news through the hardware. When a headline claims a cheap breakthrough, ask what the cost figure includes; DeepSeek's covered only a final training run on about 2,000 H800s, not research or reinforcement learning [2].
2. Watch export controls and their loopholes. Follow which chips are allowed (H800, then H20) and how many move through third countries or cloud rentals; Patel estimates 200,000 to 300,000 smuggled GPUs in a year [2].
3. Track Taiwan as a technology issue, not only a military one. The leading edge of chipmaking and R&D sits in Hsinchu, and TSMC's U.S. fabs still depend on it [2]; Musk calls China's intentions toward Taiwan certain from its standpoint [3].
4. Compare investment, not just announcements. Weigh U.S. programs like the roughly 50 billion dollar CHIPS Act against China's roughly 200 billion a year in chip subsidies and its new AI subsidy [2].
5. Treat electricity as part of the race. Expect transformer and power shortages to limit AI buildouts [3], and watch whether permitting and grid expansion keep up with data centers heading toward 10% or more of U.S. power [2].
6. Support talent and open science. Fridman argues importing the world's best talent should not be controversial [2], and Graves argues that with peer adversaries, activating America's scientists and engineers matters more than secrecy [1].
7. Pay attention to whose models the world uses. Open models spread values and possible hidden behaviors [2]; the fight over an American versus Chinese open standard is part of the competition [2].
8. Prepare for data security threats. If adversaries are storing encrypted data to decrypt later, as Graves reports, sensitive data's long-term protection deserves attention now [1].
9. Keep both risks in view. Patel warns that falling behind and provoking a crisis over Taiwan are both dangers [2]; Musk argues for peace and mutual understanding even while describing Thucydides-style tension [3].

Sources

1. 2244 - Ryan Graves · The Joe Rogan Experience
2. DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters · Lex Fridman Podcast
3. Elon Musk: War, AI, Aliens, Politics, Physics, Video Games, and Humanity · Lex Fridman Podcast