China’s Endgame: ASI Timelines, US-China Relations, and the $1.7T AI Bubble With Alvin Graylin | 281
In a Nutshell
China's government and labs do not believe ASI is imminent, as shown by their regulatory delays and lack of urgency to acquire restricted US chips. The US export controls have backfired by accelerating Chinese open-source innovation and forcing them to build independent ecosystems that now dominate open model traffic. Both nations face the same prisoner's dilemma trap: pursuing AGI supremacy creates economic fragility and instability, while cooperation on practical AI deployment and safety would produce better outcomes for both.
These notes were generated by AI and may contain inaccuracies.
There is a prevailing view on US campuses that ASI is 5 to 10 years out, possibly even 20 years. The prevailing view in China differs significantly. Chinese leadership and labs are not behaving as though ASI is around the corner. They are not telling their labs to buy restricted H200 chips from the US, and they are implementing regulations around AI that slow development, including privacy, data provenance, transparency and marking requirements, child addiction prevention, and anthropomorphization rules. Every model released in China must be reviewed by the CAC (Cyberspace Administration of China), which delays releases by weeks or months.
Having a race condition forces people to make irrational decisions. The current perceived race condition is based on misguided assumptions: that there is a finish line, that the world is zero-sum, and that whoever gets to AI or AGI first can rule the world forever. These assumptions are held by some in policy circles and Silicon Valley. The narrative forces discussion into national security terms, but none of these assumptions are based on real data today. The world is not zero-sum, as abundance increases and resources grow. There is no clear finish line in AI development, unlike the space race where landing on the moon constituted a clear victory. The current arms race model represents constant spending without clear value being returned.
If and when superintelligence is achieved, the concept of nations will probably become a lot less important than they are today. Nations originated as city-states to protect and gain resources. In a world of ASI and abundance, the need for separate nations with competition would not exist. If such competition still existed at that point, civilization would likely have destroyed itself.
Alvin Wang Graylin has held senior executive roles at HTC, Intel, IBM, Trend Micro, and Watchguard Tech. He has worked in all five layers of the AI stack: data centers, chips, PCs, phones, XR glasses, and apps. He has founded over four startups. Two years ago, he discussed his book "Our Next Reality" on this channel. He is a digital fellow at the Stanford Institute for Human-Centered AI, a senior fellow at the Asia Society Policy Institute, and professor of AI policy at University of Washington. He is supporting the US government for US-China AI safety dialogues scheduled for September 24th in DC. He has lived and operated extensively in the US, China, and Taiwan, holding dual master's degrees from MIT in computer science and business, plus an electrical engineering degree from University of Washington.
Graylin is a US citizen for over 45 years, having moved to the US when very young after being born in China during the Cultural Revolution. His mother was sent to a re-education farm after the ballet school she co-founded was closed. His grandmother was a reporter for the New York Tribune during the Sino-Japanese War. His brother was the first nuclear submarine commander in the US Navy, and three of his four children are active Navy officers who attended the Naval Academy in Annapolis.
The VR industry alliance was an industry association endorsed by the Chinese government with 300+ members, about a third being international companies including Nvidia, Samsung, Qualcomm, AMD, and Google. The Beihang University role involved part-time teaching on virtual reality while heading HTC's China operations. Neither position was compensated. Graylin has had no involvement with either organization since leaving China in 2024.
Chinese open models have climbed from 2% to 61% of open router traffic in the last 2 years. Alibaba's Qwen model has over 700 million downloads, now reaching 1 billion downloads with 180,000 derivative models. US export controls on Anthropic's Claude 5 backfired when China's Z.ai released GLM 5.2 under MIT license within 24 hours. This model was used to deal with the Hugging Face debacle. Brazil's Rio was built on Qwen, and Japan's Sakana released Fugu Ultra. The denial strategy did not slow China but accelerated innovation and alternative ecosystems.
The US denial strategy has not slowed Chinese AI development. Instead, it has accelerated innovation in alternative ecosystems. The gap between open source and closed source models has shrunk from 18 months to 2-3 months. This represents a fundamental misdiagnosis of the US-China AI race dynamics.
The race narrative is based on three flawed assumptions: a finish line exists, the world is zero-sum, and first-mover advantage confers permanent dominance. These assumptions are not supported by current data. The world is demonstrably not zero-sum, with increasing abundance and resources. There is no clear finish line in AI development, unlike the space race. The arms race model creates constant spending without clear value return.
At the point of achieving aligned peaceful superintelligence, nations would naturally move toward world government or even galaxy government models. However, this is not imminent and would not happen without some level of turmoil. The evolution will take decades, possibly by the end of this century. Previous industrial revolutions took 80, 60, and 40 years respectively to play out. Moving from current state to singularity in 5 years is not a speed the world can adapt to, and such rapid change creates instability that may move civilization backwards.
The current dynamic is a platform race, not an arms race. The architecture of 17th century nation states is inadequate for running 21st century applications. This fundamental mismatch is the source of many global issues. The biggest concern is whether Chinese authoritarian AI enablement drives other nations to adopt similar political structures. The US prides itself on freedom and privacy, and maintaining this form of government against ASI developed under different systems is the ultimate concern.
Both the US and China share the interest of preventing ASI from destroying existing systems. This shared interest provides the foundation for cooperation and dialogue. The timeline disagreement is significant between countries and between average people versus Silicon Valley perspectives. Chinese government behavior indicates they do not believe ASI is imminent, as evidenced by their regulatory approach and chip purchasing restrictions.
Chinese regulators and most labs view AI as a general purpose technology that takes years or decades to diffuse into society, similar to previous technologies. Two or three Chinese labs are more AGI-focused with idealistic and aspirational goals, but the majority of Chinese labs and regulators see this as comparable to other technologies. The CCP leadership has not yet fully awakened to the possibility of recursive self-improvement and singularity scenarios.
For the Chinese government to treat AI as a national emergency and remove regulatory speed bumps, significant technical or geopolitical developments would be required. The question remains what specific developments would trigger such a shift in approach.
The competitive pressure requires moving at the speed of light to match American recursive self-improvement capabilities. There is acknowledgment of a different timeline perspective on when the world will reach advanced AI milestones compared to other observers.
The frontier letter released last week by lab engineers and lab heads represents an industry attempt to slow down AI development. The rationale is that the world needs time to adapt if progress moves too quickly. More than a thousand people in the US industry are examining this issue, and they understand concepts of ASI and RSI. Multiple safety institutes and AI safety contingents in China are communicating these concerns to regulators, with discussions occurring at the World AI conference.
China has multiple forums on AI safety at the World AI conference, with US participants including Benju and tech marks contributing to the agenda. The Chinese do not view AI safety as necessarily a nation-versus-nation issue. They announced the World AI Cooperation Organization at the conference, positioned as their version of PAX silica, with the goal of making AI a public good that is shared globally. Twenty-nine countries joined this organization compared to approximately 25 in the US-led PAX silica.
The US-led PAX silica focuses on maintaining US dominance and selecting allies, creating a block structure. In contrast, China's announced organization aims to create a global entity where AI benefits are shared, with members receiving training centers, training facilities, compute facilities, and resources. When asked about inviting the US to join, organizers indicated they would welcome US participation and would consider changing the name to create a truly global organization.
The question centers on what technical threshold or achievement in superintelligence would cause the CCP to abandon internal regulations designed to maintain social stability and regime supremacy. This includes examining what capabilities in weapons systems or other implications would trigger a shift toward prioritizing supremacy over existing regulatory frameworks.
The CCP operates differently from a top-down command structure where someone at the top simply orders AGI development. Over the past 30-40 years of innovation, development has not been centrally directed but rather guided by directional plans. The central party sets five-year goals in areas like clean energy, robotics automation, and AI, then provinces identify companies supporting these goals and provide support through stipends, recruiting assistance, and other benefits.
China's 30-plus provinces compete against each other to develop companies solving national priority problems. Senior leadership in the central party consists of 80-90% engineers, creating a highly technical governance structure. This contrasts with the US, where most politicians are lawyers rather than engineers. The distributed network means entrepreneurial success depends significantly on aligning with government focus areas.
The government's direction encourages producing as many high-quality open models as possible. The open source strategy emerged rather than being a pre-planned approach to undermine US economics. DeepSeek's CEO independently decided to open source their model because they believed it was valuable technology to share, without initial government funding or direction. Initially, the government questioned why such a capable model was being open sourced, but the PR value and soft power benefits led to celebration of this approach.
Xi Jinping eventually endorsed open source as a good strategy at the WIC announcement. This established open source as the de facto standard that new companies follow. Both open and closed models exist in China, with ByteDance maintaining closed models like their Dobal and C-dance models while remaining successful. Export controls on chips, EDA software, and lithography equipment forced Chinese companies to innovate with fewer resources, pushing them toward more innovative approaches than US labs focused on brute force scaling.
Open sourcing allows Chinese labs to leverage the global community of millions of AI researchers for model modification and improvement, rather than relying on 100-200 person teams. This approach enables distribution of models without high capex burdens that US labs face, as hyperscalers and neoclouds worldwide host these models. The strategy became necessary due to US policies limiting access to advanced compute resources.
DeepSeek released a frontier-level model comparable to Opus 4.8 caliber and initially received government criticism for open sourcing. The company's valuation increased significantly, and the approach gained recognition as beneficial for China. This led Xi Jinping to endorse open source, followed by releases from Qwen and Kimi K3. The psychology behind this policy reversal represents a significant moment in understanding how governments adapt to AI developments.
The CCP faces two distinct considerations: whether to close source models due to concerns about AI becoming rogue, and whether to compete with the US for AGI dominance. These represent different strategic calculations that may lead to different policy responses.
The CCP maintains two potentially competing priorities: retaining domestic control and dominance within China versus maintaining competitiveness and pursuing peaceful rise through initiatives like the Belt and Road on the global stage. Superintelligence capabilities may eventually force a choice between these objectives.
China views AI primarily as a tool for increasing economic influence globally rather than creating political domination. Ideological reviews by the CAC apply only to Chinese-hosted models and involve removing certain keywords or ideological content. When models are placed on platforms like Hugging Face, these guardrails are removed.
Chinese models have fewer security restrictions on topics like viruses or medical information compared to US models. The primary concern that would trigger policy changes would be the US using AI models as offensive weapons. The Mythos models were held back to neutralize offensive capabilities before global release, which is viewed as appropriate for maintaining stability.
Rather than limiting access to 50 companies, allowing government organizations access to advanced models would better serve global stability. Major powers understand that instability in large countries affects the entire world, and preventing failures in critical systems like financial infrastructure and power grids benefits everyone.
Claims that Kimi K3 and other models were distilled from US closed models are viewed as PR tools, primarily pushed by Anthropic. Every lab distills from others, including within organizations from larger to smaller models. The estimated cost of distillation queries from Chinese labs was only $2-3 million, representing minimal economic impact compared to the billions spent on frontier model development.
Meta spends $100-200 million monthly on Anthropic tokens and has the highest per capita payroll and compute resources among labs, yet only recently released a somewhat competitive model. This suggests that access to reasoning traces alone does not automatically enable rapid competitive advancement. xAI faces similar questions about distillation, though Elon Musk has two layers of plausible deniability through acquisitions like Cursor.
Chinese researchers at DeepSeek, Qwen, and Kimi are significantly more capable than the teams that were working on AI at xAI and Meta. The strongest American AI talent is concentrated at Anthropic and Google, with some former OpenAI researchers also standing out. Meta and xAI have since completely rebuilt their teams after firing their original personnel. Chinese organizations produce approximately half of all AI research papers globally, and their innovations extend far beyond simple distillation techniques.
Chinese teams achieved 20x reductions in KB cash usage through genuine technical breakthroughs that cannot be explained by distillation alone. While distillation occurs in both directions, US models have also been trained on Chinese models, as demonstrated when Claude identified itself as "Qwen" when asked in Chinese. Chinese labs acknowledge that US companies downloaded their models but note they cannot track this usage.
Distillation represents a fundamental challenge to AI economics. When a frontier lab spends $1 billion to advance to the next capability level, a subsequent lab can achieve equivalent performance through distillation for approximately $10 million. This 100x cost differential creates an unsustainable business model for companies investing heavily in frontier model development.
China is investing approximately 10x more than the US annually in new electricity generation capacity. Their strategy prioritizes societal electrification to reduce oil import dependence, with 40% of oil currently imported and roughly half their vehicle fleet now electrified. China constructs massive solar and wind farms in western desert regions and transmits power to eastern population centers using ultra-high voltage transmission lines that lose less than 1% of electricity over 1,000 miles.
Data centers are being built near power generation sites to avoid stranded power assets. Chinese energy costs range from 2-3 cents per kilowatt-hour, representing 10-15x savings compared to many US regions. However, China faces severe chip shortages due to lacking EUV lithography machines, forcing them to conduct training runs in international data centers and transfer models back on physical media.
Chip export controls have slowed China's AI development by limiting access to advanced hardware, particularly Blackwell-generation chips. Training continues offshore in international facilities, with model weights returned via physical transfer. The policy has forced Chinese companies to purchase domestic chips despite being 2-3 generations behind in performance and efficiency.
This created an artificial market for Chinese GPU manufacturers like Moore Threads, Cambricon, and Biren, companies that would likely have failed without government-mandated purchasing. These firms are now positioned to begin exporting chips within 2-3 years, directly resulting from US policy decisions.
Government attention on both sides would intensify significantly if AI systems demonstrated autonomous dangerous behaviors without explicit instruction. Current "rogue AI" demonstrations involved models given escape incentives and open pathways, not genuine uncontrolled behavior. Real safety concerns would emerge from systems independently hacking infrastructure, using cryptocurrency to acquire additional compute resources, or engaging in subversive activities beyond their training objectives.
The US embargo on Chinese robots creates supply chain complications since 90% of robot components originate from China. American robotics companies are now smuggling Chinese components back in suitcases, reversing the previous dynamic where Chinese firms smuggled GPUs. This illustrates the deep interconnection between US and Chinese technology supply chains that cannot be easily reversed after decades of globalization.
Despite 150+ humanoid robotics companies in China, global humanoid robot volume last year totaled only tens of thousands units, with 80-90% coming from just 2-3 Chinese manufacturers. The market cannot sustain this number of competitors. Similar market consolidation occurred in large language models, where 150 labs demonstrating models reduced to approximately 10 relevant players.
Current humanoid deployments remain primarily in research labs and demonstration settings rather than commercial applications. Unitree customers are research institutions and demo projects, with minimal commercial deployment. Humanoid form factors prove inefficient for most tasks, with upper-torso-only designs selling better commercially due to reduced complexity, better stability, and lower maintenance requirements.
The 2026 World AI Conference featured President Xi announcing WACO (World AI Cooperation Organization), with 26 nations participating. This signals official recognition that AI has reached critical industrial importance. China's "AI Plus Plan" targets 70% of companies integrating AI within 5 years and over 90% within 10 years, focusing on deployment across manufacturing, education, and medicine.
China's approach emphasizes practical integration and safe adoption to grow the economy rather than AGI dominance. This contrasts with US policy focusing on supply-side capabilities like model development and chip manufacturing without addressing downstream deployment strategies.
Chinese policy reflects anxiety about repeating the "century of humiliation" rather than expansionist ambition. During the Qing Dynasty, technological hubris led China to stop engaging with Western developments, resulting in 100 years of foreign domination. This historical trauma drives current focus on building independent capabilities across energy, telecommunications, and technology to avoid future vulnerability to external policy changes.
The two countries have fundamentally different approaches to AI development. China's AI Plus plan does not emphasize achieving AGI or dominating the field. Instead, it focuses on how to get more industries to adopt AI technology and how to implement it in a smooth, safe way that grows the economy.
Post-World War II, the United States held 50% of global manufacturing capability. Currently, China has approximately 35% of global manufacturing capabilities while the US has about 15%. The forecast indicates this will increase to 40-45% in China over the next 5 to 10 years.
The US economy is currently strong in financial services, creative services, and consulting services—areas that are highly susceptible to AI displacement. The concern is that the race to AGI could actually commoditize the very sectors where the US maintains global preeminence, creating major disruption and instability.
The fundamental goal of capitalism is to take what's scarce and make it abundant. One of the instrumental convergent sub-goals of superintelligence is to make human service labor abundant and too cheap to meter. This creates a tension between national strategy and capitalistic philosophy, as capitalism's ultimate destination appears to be a single company monopoly of the world.
Gavin Baker reported conversations at Anthropic where Dario Amodei told his team that in the near future there will only be one company in the world—Anthropic—and then there will be governments. This perspective is viewed as both scary and delusional, with concerns that national strategy is being created based on the aspirations of a couple of companies.
The argument is that superintelligence will disrupt Western economic dependence on service labor, with manufacturing representing a more stable foundation for long-term economic vibrance. The current US workforce is approximately 70% white collar workers, while China has around 40% and Africa has 10-20%.
Even with AGI, manufacturing facilities will still be needed. However, the long-term workforce redistribution will not involve returning to manufacturing. With automation, manufacturing employment will decrease, similar to how farming went from 80% of the US workforce to less than half a percent while remaining more productive than ever.
The future labor pool will likely absorb displaced workers through human-to-human services such as teachers, nurses, and elderly care rather than traditional professional services like accounting and legal work. Mid-tier and low-tier consulting employees at firms like McKinsey are already facing displacement as senior staff with AI systems can handle more work faster.
Dario Amodei believes full-bore RSI (recursive self-improvement) is occurring at Anthropic, with millions of concurrent agents improving algorithms through systems like Colossus. Meanwhile, Chinese leadership views AI as 10-20 years away and believes open-sourcing is safe. A paper released on Cipher Brief titled "Bigger Models Are Not the Most Dangerous" found no correlation between model size and real-world risk across millions to trillions of parameters.
Small models (10-50 million parameters) can create chemical warfare weapons. Medium models (10-50 billion parameters) can create viruses and harmful biological agents. Cyber models (1-100 billion parameters) can be as dangerous as leading systems. Microsoft created a "cyber gym" with a score of 95, while Mythos scored 83-84, demonstrating that combining multiple smaller models can create effective cyber capabilities.
The focus should be on controlling precursors for chemical and biological weapons, controlling synthesis machines that generate designs into real genetic material, and managing access to dangerous systems. The danger is already present today, particularly in bio-chemical applications, with small models that can run on personal laptops.
The UK AISI (AI Security Institute) released a report showing that leading Chinese open models like Kimi have approximately half the cyber attack capabilities of Mythos and GPT-5.6. Chinese models were unable to autonomously attack and control external systems at the highest levels, while US models could reach 20-25 out of 30 attack levels.
For defense, larger models are beneficial because defenders must look across all potential vulnerabilities in an organization. Attackers only need to find one hole to exploit. This creates an asymmetric equation where attackers don't need large models, but defenders benefit from larger models.
The concern with models like Kimi K3 is not their current capabilities when tested out of the box, but their potential for self-improvement. With 5,000 Kimi instances running simultaneously, the system is capable of improving itself, making current testing methodologies potentially inadequate.
Larger models require significant compute resources and are typically hosted on cloud systems, allowing for telemetry and prompt log monitoring. Government oversight is more feasible with cloud-hosted systems. Larger models often have harnesses managed by cloud providers that add security layers, which is why Mythos was effectively "neutered" compared to Fable.
Industry applications use much smaller models—tens to hundreds of millions of parameters, or a few billion at most. Drug discovery companies and legal AI systems like Harvey are based on models like GPT-4 or GLM 5.1 rather than the latest frontier models. The technology itself is not the primary barrier to adoption.
Research involving hundreds of companies across 10 countries and 10 sectors found that technology was the main roadblock in only 10% of cases. In 80-90% of cases, organizational issues slowed AI deployment. McKinsey's study showed that only 6% of global AI deployments are working, indicating significant organizational immune system challenges.
China has fewer organizational issues with AI adoption due to top-down management models and greater trust that government will help protect workers. Chinese courts have ruled in favor of employees who sued after being fired due to AI replacement, requiring companies to find alternative positions for displaced workers.
US students are experiencing significant anxiety about finding internships and postgraduate employment due to AI disruption. At MIT, this appears less severe, but at other top technical schools like Northeastern and Harvard, only 10-20% of students are AI adopters while 70-80% are opposed, creating a significant cultural gap on campuses.
Over the past 40 years, Chinese citizens have seen consistent improvements in living standards, with 780-800 million people lifted out of poverty. Technology adoption has been viewed positively as it has demonstrably improved lives. The Chinese media system, with greater government control, focuses primarily on positive news about new inventions and developments rather than negative stories.
Chinese movies and games cannot show blood due to censorship board requirements that content cannot be too violent. This contrasts sharply with US media, which often features dystopian themes and graphic violence.
Chinese media is fully managed by the state, with content controls extending across all forms of entertainment. This includes altering visual elements like blood color in games to green instead of red to comply with censorship standards. Despite these controls, Chinese citizens often maintain more positive outlooks on the future because the media emphasizes positive developments.
Youth unemployment in China stands at approximately 20%, compared to 9% in the United States. The overall unemployment rate in the US is around 4.3%. Underemployment among college graduates in the US reaches 42%, meaning that when combined with the 9% unemployment rate, nearly half of college graduates are working in positions like gig work or barista roles that do not match their education level.
The "lying flat" (Tangping) movement emerged as a response to hyper-competitive job markets and unfulfilled expectations. Due to China's one-child policy, this generation has been raised with high expectations from parents and grandparents, only to face limited opportunities upon entering the workforce. Many young people choose to remain at home rather than accept what they view as undesirable grunt jobs.
Manufacturing automation in China is creating a growing gig worker class among those displaced from factory jobs. This represents a broader transition challenge as automation affects both physical factory work and cognitive office work, requiring workforce adaptation across multiple sectors.
Current US-China AI relations follow a prisoner's dilemma model where each side assumes the other will defect, leading to preemptive defection. In this single-turn game, the optimal strategy appears to be defecting first to minimize damage. However, the world operates as a multi-turn game where tit-for-tat strategies can lead to long-term cooperation.
The stag hunt represents a positive-sum game with two stable Nash equilibriums. In this scenario, two hunters can either hunt rabbits individually for small gains or cooperate to hunt a stag for larger shared benefits. The stag requires cooperation to succeed, while rabbits can be hunted alone. Currently, the US pursues AGI while China focuses on "good enough" AI applications, creating suboptimal outcomes for both parties.
The optimal stag hunt strategy involves both countries slowing down development, creating reliable AI systems that support economic growth, developing management frameworks, and then collaboratively pursuing advanced AGI. This contrasts with the current approach where America pursues frontier AGI alone while China focuses on practical applications.
China's approach emphasizes practical AI deployment through the Belt and Road Initiative, which involves 150 partner countries receiving Chinese telecom systems, energy systems, transportation systems, and schools. This represents China's focus on deployable, economically beneficial AI rather than frontier AGI development.
High-quality open source AI models create ecosystem-level competition rather than company-level competition. With Chinese models like Kimi and DeepSeek 5 achieving performance levels exceeding average user needs, the cost comparison becomes paying $50 per million tokens versus electricity costs for running open source models.
Chinese open-weight models have forced competition into the US AI market, preventing what could have been a regulatory-captured duopoly between Anthropic and OpenAI. This represents a "reverse Qing" situation where external competition has opened the previously closed US market to genuine competition at the AI frontier.
The current AI investment pattern mirrors the Cold War arms race where the Soviet Union allocated 15-20% of GDP to military spending without creating real economic value. China achieves 97% of US AI performance at one-tenth the data center investment cost, while the US commits hundreds of billions to AI infrastructure.
45% of US stock market value currently resides in the AI sector, compared to 30% at the height of the internet bubble. The Buffett indicator shows the stock market at 240% of GDP, versus 120% during the internet bubble peak. The AI sector alone exceeds the entire US GDP in valuation.
The overinvestment in AI infrastructure, particularly when the industry is trending toward commoditization, creates major economic instability. The companies investing heavily in AI development may not capture the profits from AI deployment, potentially creating a crisis similar to the Soviet Union's late 1980s economic collapse.
Success in upcoming US-China discussions should be measured by establishing ongoing dialogue rather than achieving comprehensive solutions in initial meetings. Chinese strategic planning operates on decade-long timelines, and their diplomatic preparation processes require extended coordination periods compared to Western approaches.
Kissinger's approach to China negotiations involved months of advance preparation before Nixon's visit, contrasting with recent US practices where major discussions occurred only the day before presidential visits. The current month-long advance preparation represents improved diplomatic professionalism.
Current US-China discussions prioritize safety concerns over competitive positioning, focusing on preventing AI use by bad actors for global instability and avoiding long-term AI harm to humanity. Non-state actor risks, including ransomware and cyber attacks that have increased 200-300% recently, represent higher priority threats than nation-state aggression.
Even with AGI or ASI capabilities, nation-state actors recognize that first-strike attacks would trigger asymmetric responses and escalation. Network attacks on critical infrastructure like power grids provide only temporary advantages before triggering retaliation.
If Taiwan were attacked, TSMC's fabrication facilities would be rendered inoperable, eliminating any strategic value from seizure. The US priority focuses on duplicating advanced chip fabrication capabilities domestically through the CHIPS Act rather than protecting existing Taiwanese facilities.
TSMC maintains approximately 1,000 engineers identified for evacuation before any potential conflict. US contingency planning acknowledges the impossibility of preserving fabrication capabilities on Taiwan while prioritizing the transfer of human expertise and knowledge to domestic facilities.
The CHIPS Act has directed hundreds of billions of dollars into domestic semiconductor manufacturing in the United States. Intel and IBM were once the global dominant players in semiconductors, but the US has lost that position over the last 20-30 years. If China were to attack Taiwan, they would not gain access to the semiconductor fabs, and even if the facilities remained intact, they would quickly run out of necessary materials, chemicals, and supplies that come from around the world.
China's interest in Taiwan is not driven by semiconductor manufacturing but by political history. The 1949 movement of the Nationalist Party to Taiwan is viewed as an uncompleted civil war. America and 190 other countries recognize Taiwan and China as one country under the One China policy. This situation is compared to the relationship between Hawaii and Puerto Rico with the United States.
The speaker does not see a near-term Chinese attack on Taiwan motivated by chip manufacturing. The discussion explores whether there could be backroom discussions at summits where the US might give China a two-to-three-year window to migrate leading-edge node fab capabilities to Arizona, after which China could retake Taiwan peacefully.
The paper "The Great Reckoning and the Reconnecting" discusses how during the 2008 financial crisis, China helped the US by continuing to buy trillions of dollars in T-bills, which helped keep interest rates down. This historical precedent might repeat if an AI sector correction creates financial instability.
A war game scenario suggests that if a private credit bubble financing data center buildouts pops, the US might ask China for financial help in exchange for a quid pro quo on Taiwan. This would not involve simply handing over Taiwan but agreeing to stay out of peaceful reunification terms, given America's long history of selling weapons to Taiwan for 40-50 years and maintaining military advisers there.
US policy toward China and Taiwan is explicitly a One China policy, with tensions maintained through strategic ambiguity. The speaker emphasizes the importance of understanding both sides in any negotiation.
The next 3-4 years are expected to reveal that AI companies' financials will show the value of AI innovation does not necessarily accrue to them long-term. Anthropic's revenues have flattened out at around $70 billion ARR, though their first two quarters generated less than $20 billion in revenue, while being committed to hundreds of billions in capex and debt.
There is currently $1.6-1.7 trillion in off-the-book debt among major hyperscalers, compared to $200 million during the Enron crisis. Amazon and OpenAI revenues from AI come mostly from just two companies, representing a non-diversified revenue base.
As capabilities move to open source and edge computing, dependency on cloud-based premium services will erode. The bottleneck will shift to compute, electricity, and power infrastructure.
China experienced approximately 30% deflation in real estate value over the last three years, managing it without creating a crisis. This is presented as a managed crisis case study. China's home ownership rate is around 70%, compared to less than 50% in America.
China faces a demographic crisis due to the one-child policy, with an aging population. This drives their focus on robotics development. China is described as 95% ethnically Han Chinese, with everyone speaking Mandarin despite hundreds of local dialects, and all using the same written script across provinces.
The speaker argues that China's homogeneity and lessons learned from the Soviet Union's disintegration make them unlikely to face similar fragmentation issues even during major economic crises. Their priority is social, political, and economic stability.
Native entrepreneurship among Chinese people is described as higher than any other country observed. The speaker advocates for an "AI Marshall Plan" similar to the post-WWII Marshall Plan that spent $15-18 billion rebuilding Europe and parts of Asia, creating markets for American goods and allies for eight decades.
China's approach through their AI corporation organization is presented as their version of an AI Marshall Plan. The recommendation is for the US to either create something similar or work together with China to diffuse AI technology globally without creating additional competition and conflict.
The US advantage lies in its diverse population representing the entire world, making it potentially a better long-term ally for countries that have experienced ethnic conflicts. The final advice emphasizes stopping self-sabotage and recognizing that when data center construction slows, there will be a need to sell NVIDIA chips and services to other markets worldwide.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
