Wagering our lives, on what and why?
June 2, 2026
The View from the Other Side of the Pacific
By Scott Ortkiese
Imagine you are a senior strategist at the Korea Development Bank in Seoul, or a deputy director at Japan’s Ministry of Economy, Trade and Industry, or a senior analyst inside China’s National Data Administration. Your job is to watch the great American AI infrastructure blitz and make a recommendation about what your country should do next.
You watch $500 billion committed to the Stargate Project, with OpenAI, SoftBank, Oracle, and MGX pledging to build the largest AI computing complex in history, starting in Abilene and Shackelford County, Texas. You watch xAI break ground on a $20 billion data center in Southaven, Mississippi, pushing Elon Musk’s training compute toward two gigawatts in the greater Memphis corridor alone. You watch Microsoft, Alphabet, Amazon, and Meta collectively plan more than $300 billion in capital expenditures for 2025, with McKinsey projecting $5.2 trillion of total data center investment required by 2030.[1][2][3][4]
And then you lean back, and you think: Let them build it.
Because what the American AI infrastructure bet gets fundamentally wrong is the premise that the most powerful AI system is automatically the most valuable one. The real battle is not over who controls the largest cluster of Nvidia GPUs. It is over who controls the data those GPUs need to do anything genuinely useful. And on that question, the United States has a dependency it is not talking about honestly.
Compute Is the Factory. Data Is the Ore.
To understand the strategic error, you need to grasp a distinction that dominates boardroom AI conversations but rarely makes it into policy debates: the difference between a model’s general language capability and its domain-specific operational value.
General large language models, trained on the publicly available internet, are extraordinary at tasks that require broad pattern recognition across a diversity of topics. They can write software, summarize legal documents, explain thermodynamics, or draft a marketing brief. They are horizontal tools. But when an enterprise tries to use one of these models to optimize a steel rolling process, predict a yield failure in a semiconductor fab, or refine the casting parameters for a turbine blade, the model immediately confronts a wall. It does not know what it does not know. It lacks the proprietary, process-specific data that defines the actual problem.[5][6]
This is not a theoretical limitation. JPMorgan reportedly holds approximately 500 petabytes of proprietary financial data, while ChatGPT was trained on roughly one petabyte of publicly available text. The model that wins a benchmark tournament and the model that can actually run your business are not the same model. The former requires compute. The latter requires data that was never posted on the internet.[^7]
McKinsey captured the industrial stakes in May 2026: “When AI models are widely available, companies can build strategic moats through proprietary data, embedded workflows, scale, and customer lock-in.” The domain-specific LLM platforms market already stood at $6.8 billion in 2025 and is projected to reach $52.4 billion by 2034, precisely because enterprises are learning this lesson. The model is the commodity. The data is the moat.[8][9]
What Asian Industrialists Actually Control
Here is what is sitting on the other side of the Pacific, largely undigitized, largely uncontested, and almost entirely irreplaceable.
Korea’s Samsung Electronics and SK Hynix together account for roughly two-thirds of the world’s DRAM memory and a dominant share of NAND flash storage. Their process engineering knowledge, refined across decades at the atomic level of material science and lithography, is not available for training on any public dataset. TSMC, the Taiwanese foundry that fabricates the chips inside every device that runs modern AI, went so far as to fire staff and launch legal proceedings in 2025 over suspected leaks of its 2-nanometer process details, described as trade secrets of extraordinary sensitivity. The 2-nanometer node is among the most complex manufacturing processes ever developed by humanity. It does not exist on the internet. It exists in the accumulated proprietary know-how of a small number of Asian fabs.[^10]
POSCO in South Korea runs blast furnace operations refined over sixty years with such precision that their process data is treated as a national industrial asset. Hyundai Steel spent years building an AI-powered Smart Enterprise system that merges sensor data, process control modules, and production management across its entire steel manufacturing value chain, explicitly designed to keep that intelligence on-premise. In Japan, the ARUM software platform trained on more than five thousand tools and ten billion machining conditions, encoding decades of precision manufacturing expertise into an AI that can generate complex CNC machining programs in fifteen minutes versus the sixteen hours required by a skilled technician. That training corpus is privately held, deeply proprietary, and was built from lived physical experience in Japanese machine shops, not scraped from a webpage.[11][12]
This is the concept of Monozukuri, Japan’s philosophy of manufacturing as a craft discipline, now being systematically digitized and weaponized. The tacit knowledge embedded in a master machinist’s hands, a veteran furnace operator’s instincts, a process engineer’s intuition about a yield anomaly, is what Japan’s Keidanren calls the foundational layer of Society 5.0, the fusion of cyberspace and physical space that only becomes strategically meaningful when the physical side contains something no one else can replicate.[13][14][^15]
The South Korean government formally recognized the stakes in April 2026, committing 48 billion Korean won through a supplementary budget specifically to convert master craftsmen’s tacit knowledge into AI training data before the aging workforce that holds it retires. The urgency behind that investment communicates exactly what is at stake: that knowledge, once gone, cannot be reconstructed from the internet, from synthetic data generation, or from the most powerful language model in Memphis.[^16]
Korea and Japan’s Strategic Calculation
The Korea Institute for Future Strategy articulated the problem with unusual directness in a January 2026 policy brief: despite Korea’s global strength in semiconductors, shipbuilding, and defense, its manufacturing sector faces a structural vulnerability because “during the process of AI adoption in manufacturing, there is a heightened risk that tacit knowledge from production sites may be converted into data and transmitted to, accumulated in, and controlled by overseas servers.”[^17]
This is the exact mirror image of America’s framing. Washington talks about AI dominance in terms of compute, model parameters, and benchmark scores. Seoul is talking about the risk that Korean factory intelligence will be silently extracted into foreign AI systems and held hostage. The brief argues that the “core of manufacturing AI competitiveness lies in securing and exercising control over high-quality, domain-specific data, that is, in achieving data sovereignty,” and explicitly criticizes existing “sovereign AI” discussions for their disproportionate focus on model sovereignty rather than data sovereignty.[^17]
In practical terms, this means that Nvidia’s flagship collaboration with Samsung, SK Group, and Hyundai Motor Group, deploying a combined 260,000 Blackwell GPUs across Korean enterprises, is being architecturally structured to keep the resulting industrial AI data within Korean infrastructure. Samsung’s new AI megafactory integrates its semiconductor manufacturing process into an AI-powered platform to optimize development and production in real time, but the knowledge generated by that platform, the yield correlations, the defect signatures, the process window optimizations, is Korean.[18][19]
Japan’s approach is embedded in the Society 5.0 vision, which since 2016 has explicitly proposed building a “digital twin” of every element of Japanese society in cyberspace before reflecting it back into the physical world. What sounds like techno-utopian language is, at the strategic level, a program to ensure that the intelligence amplified by AI is grounded in irreplaceable Japanese physical-economy data rather than on compute capacity owned by foreign hyperscalers. Japan differs from other countries in terms of language and cultural background, and can only get so far by relying on AI developed elsewhere, as Keidanren noted in 2023, before it must develop indigenous capability.[14][13]
Korea and Japan are not blind to the Stargate buildout. They are watching the experiment and building their own plays from a position of data strength. They do not need to win the compute race. They need to ensure that when AI becomes as common as electricity, the most valuable applications require their data to work.
China’s Two-Loop Strategy: The Most Sophisticated Play
If Korea and Japan are playing careful defense, China is playing aggressive offense, and its strategy makes the American infrastructure bet look even more one-dimensional.
The U.S.-China Economic and Security Review Commission published a working paper in March 2026 titled “Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance” that lays out the mechanism with remarkable clarity. China has built two reinforcing feedback loops. The first is digital: by open-sourcing its AI models and pricing them far below American competitors, China is accelerating global adoption of Chinese AI infrastructure. Alibaba’s Qwen models already account for the largest model ecosystem on Hugging Face, with over 100,000 derivatives. Roughly 80 percent of U.S. AI startups reportedly use Chinese base models for their derivative products. When you give your tools away and everyone builds on them, you control the standards.[20][21]
The second loop is physical and far more consequential in the long run. By deploying open AI models at low cost across China’s manufacturing base of factories, logistics networks, and robotics systems, China is generating real-world industrial data at a scale that no amount of internet scraping or synthetic generation can replicate. As the USCC paper notes, “Epoch AI estimates that leading U.S. AI companies may exhaust high-quality publicly available training data for AI language models between 2026 and 2032, reshaping the competitive landscape as the next frontier shifts toward companies’ proprietary data for specific use cases.”[^21]
China recognized this structural shift a decade ahead of the current conversation. Xi Jinping declared data “a new production factor” in 2017. The CCP Central Committee formally designated data as the fifth factor of production alongside land, labor, capital, and technology in 2020. In 2023, China’s Ministry of Finance established national data asset accounting standards, becoming the first country to allow enterprises to carry data on their balance sheets as intangible assets. China is not just building data centers. China is building a legal, institutional, and physical infrastructure to extract, own, and compound the AI value latent in its industrial base.[^21]
China’s “AI+ Manufacturing” initiative, translated and published by Georgetown’s Center for Security and Emerging Technology in March 2026, is sweepingly specific. The policy calls for large AI models embedded across aerospace, pharmaceuticals, semiconductors, new materials, and software, with 100 high-quality industrial datasets, 1,000 AI agents, and 500 benchmark application scenarios by 2027. The emphasis is not on frontier model benchmarks. It is on agentic AI that generates proprietary operational knowledge within specific industries. China is building the world’s largest industrial data collection operation, and it is doing so while spending a fraction of what American hyperscalers spend on compute. For 2025, U.S. AI capital expenditure by Microsoft, Amazon, Meta, and Google alone was at least $350 billion, against less than $40 billion for China’s major cloud providers. The asymmetry is not a sign of American strength. It is a sign that America and China are competing in different games.[22][21]
The Structural Dependency America Won’t Discuss
Here is the uncomfortable geometry. The United States is building the world’s most powerful AI compute infrastructure. But the highest-value enterprise applications of that infrastructure require domain-specific training data. The most irreplaceable domain-specific data in the world originates inside the manufacturing and industrial operations of East Asia. That data is currently being actively protected from foreign extraction. The AI systems running on American compute will, in their most valuable enterprise configurations, depend on data flows that Asian strategists are working systematically to keep under domestic control.
This is not a hypothetical risk. The Korea Institute for Future Strategy identified it in January 2026 as an active and urgent vulnerability, warning that AI adoption in Korean manufacturing without data sovereignty protections creates a direct pathway for Korean industrial intelligence to accumulate on foreign servers. The policy response, on-premise AI infrastructure, domestic manufacturing AI consortia, standardized data ownership contracts, is explicitly designed to prevent American and Chinese hyperscalers from owning the derivative intelligence generated by Korean factories.[^17]
The U.S. AI trade policy blind spot extends into the supply chain itself. As the Coalition for a Prosperous America noted in May 2026, the AI data center buildout has exposed structural failures in U.S. industrial policy, including the basic fact that the hardware powering American AI, the TSMC-fabricated chips, the Samsung memory, the SK Hynix DRAM, originates in Asia under Asian control. U.S. export controls target the digital loop, restricting Chinese access to advanced chips used for frontier model training. But as the USCC paper concludes, “this gap in the U.S. policy framework means that even successful controls on training compute may not prevent China from building AI advantages rooted in its physical economy.”[23][21]
You can build the world’s largest data center in Memphis. You still depend on Asian data flows.
The Foundational Mistake: AI Systems as Knowledge Creators
The deepest error embedded in the American AI infrastructure narrative is a category confusion between AI systems as tools and AI systems as originators of knowledge.
The dominant framing, implicit in trillion-dollar capital allocation decisions, is that sufficiently powerful AI systems will eventually synthesize new, high-value knowledge from first principles, making historical domain expertise progressively less relevant. The scaling laws hypothesis, the empirical finding that model performance improves predictably with increases in parameters, training compute, and dataset size, underwrote this bet. If you build big enough, the model generates its own understanding.[^21]
But scaling is hitting diminishing returns. OpenAI itself described its February 2025 GPT-4.5 as its last model to rely primarily on scaling pre-training rather than reasoning. Novel architectures, improved post-training techniques, and chain-of-thought reasoning have delivered performance gains without relying on increased model size. DeepSeek-V3 was trained at a reported cost of approximately $5.58 million, dramatically cheaper than GPT-4’s development cost of over $100 million, while performing comparably on many benchmarks. The compute moat is dissolving faster than the capital it consumed.[24][21]
What does not dissolve is the physical knowledge that underpins precision manufacturing. The yield improvement curve in semiconductor fabrication is not generated by an LLM reasoning from first principles. It is generated by thousands of engineers running controlled experiments on production equipment over decades, producing a proprietary dataset of correlations between process variables and output quality that cannot be reconstructed by any model trained only on publicly available text. Weinan E, a Chinese Academy of Sciences academician and machine learning scholar, argued in September 2025 that future AI progress will depend less on building larger models and more on securing high-quality, diverse data.[^21]
This is what the competitive debate actually comes down to. The value in enterprise AI is not concentrated in the model. It is concentrated in the alignment between the model and the proprietary dataset that makes the model relevant to a specific problem. A domain-specific language model platform trained on POSCO’s sixty years of blast furnace telemetry data will outperform any general frontier model on POSCO’s problems, regardless of which company’s gigawatts of training compute produced the base model.[6][9]
The Strategic Patience of the Observer
Asian sovereign strategists have a structural advantage that American capital markets cannot replicate: patience.
The Stargate buildout is committed capital. OpenAI, SoftBank, Oracle, and MGX have locked in a $500 billion, four-year program. Microsoft, Alphabet, Amazon, and Meta are spending more than $400 billion in 2026 alone on AI infrastructure. Goldman Sachs analysts observed in April 2026 that the scale of AI infrastructure investment is most determined by assumptions around silicon useful life, data center cost and complexity, factors which remain highly uncertain. This capital is deployed under assumptions about where AI value concentrates. If those assumptions are wrong, the write-downs will be historic.[25][1][^21]
Asian sovereign investors, meanwhile, are watching the experiment. Deloitte noted in February 2026 that APAC sovereign investors are “moving from caution to action on AI,” but the action is deliberate: expanding into cloud, semiconductors, and governance with careful attention to the data sovereignty implications of each investment. Asia Pacific was already the fastest-growing AI data center market in the world in 2025, attracting over $150 billion in infrastructure investment. But the character of that investment is different from the American build: it is oriented around on-premise sovereign infrastructure, domestic AI consortia, and government-backed data collection platforms rather than hyperscale compute sold as a service.[26][27]
The Korean business community’s call for a Korea-Japan data-driven manufacturing cooperation framework, articulated at the Korea Chamber of Commerce and Industry Summer Forum in 2025, reflects exactly this logic. Korea excels in AI speed and throughput; Japan in precision data and process depth. Together, as the Korea Chamber chairman warned, failure to revitalize manufacturing with AI could force a significant portion of Korean industry out of existence within a decade. But revitalizing manufacturing with AI means owning the AI’s knowledge base, not subscribing to it from a Texan data center.[^28]
Why This Matters for Every Enterprise AI Decision
The implications of this analysis run through every serious enterprise AI investment being contemplated right now.
General LLMs are rapidly commoditizing. Chinese labs offer frontier-comparable performance at a fraction of U.S. model prices: Kimi K2.5 costs four times less than GPT-5.2 while matching it on composite benchmark scores. The base model is not where the value lives. The question every enterprise leader must now answer is whether the proprietary data they control, or have access to, is sufficient to make AI operationally transformative rather than merely competent. For manufacturing enterprises, that question leads directly to the tacit knowledge problem: the expertise embedded in your most experienced workers, encoded in decades of process data, embedded in the institutional memory of how your specific equipment behaves under your specific conditions.[^21]
McKinsey is blunt about the implication: when AI models are widely available, competitive moats are built through proprietary data and embedded workflows, not model access. The enterprise that does not own its industrial data, that allows it to flow to a platform provider’s servers, training derivative models that can eventually be sold to competitors, is surrendering its most durable competitive asset.[^8]
This is not a hypothetical concern. It is precisely the scenario the Seoul National University Institute for Future Strategy warned about in January 2026 when it identified the “heightened risk that tacit knowledge from production sites may be converted into data and transmitted to, accumulated in, and controlled by overseas servers.” The companies most exposed to this risk are those in industries with high proportions of tacit knowledge: precision manufacturing, specialty chemicals, process engineering, pharmaceutical production, energy operations.[^17]
Conclusion: The Origin and Control of Real-World Knowledge Is Decisive
The United States has made a bet that the decisive variable in AI competition is compute, and has committed sums of capital that dwarf the GDP of most nations to prove it. The bet may produce remarkable general-purpose AI capabilities. It will not, by itself, produce the domain-specific knowledge that makes those capabilities operationally indispensable in the industries that actually run the world.
China recognized this asymmetry first, institutionally designating data as a factor of production while the United States was still debating chip export controls. Korea is now racing to digitize its aging workforce’s tacit manufacturing knowledge before it retires. Japan is weaving the physical intelligence of its industrial economy into a cyber-physical architecture designed to keep that intelligence domestically controlled. Nvidia can sell 260,000 GPUs to Korean conglomerates, but the knowledge generated by those GPUs running in Samsung’s fabs and Hyundai’s factories will be Korean.[19][16][14][21]
The origin and control of real-world knowledge will be decisive because AI systems cannot create it. They can only organize, amplify, and apply it. A general frontier model is an extraordinary amplifier. But an amplifier with no signal produces only very sophisticated noise. The signal, the hard-won, embodied, process-specific knowledge of how the physical world actually works, was accumulated before any of these data centers existed, in the factories, fabs, shipyards, and refineries of Asia. It will not be surrendered because the largest data center on earth sits in Memphis.
For sovereign strategists in Tokyo, Seoul, and Beijing, this is not a threat. It is leverage. And they can watch the American experiment play out with perfect composure, adjusting their own plays before committing capital, knowing that the experiment’s success depends on data flows they control.
References
1. About the Stargate Project – This project represents the largest investment ever made in the field of AI, with an immediate commi…
2. Announcing The Stargate Project – OpenAI – The Stargate Project is a new company which intends to invest $500 billion over the next four years…
3. Tech leader xAI investing more than $20 billion in Southaven – The project will represent a corporate investment exceeding $20 billion and will create hundreds of…
4. Inside the relentless race for AI capacity – Visual and data journalism – In 2017, Apple pledged to invest $2bn in the construction of a data centre spanning 83 acres. Tech c…
5. Applying LLMs to Enterprise Data: Concepts, Concerns, and Hot-Takes – Ask GPT-4 to prove there are infinite prime numbers, while rhyming, and it delivers. But ask it ho…
6. Domain-Specific Language Models vs General LLMs – ITTech Pulse – DSLMs vs general LLMs: learn how enterprises compare language models to balance AI flexibility, doma…
7. AI data analyst won’t work because proprietary data is locked inside… – Enterprises have two options to do that – share the data with a remote LLM or host their own. Compan…
8. Building competitive strategic moats with AI – McKinsey – When AI models are widely available, companies can build strategic moats through proprietary data, e…
9. Domain-Specific LLM Platforms Market Research Report 2034 – The domain-specific LLM platforms market was valued at $6.8 billion in 2025 and is projected to reac…
11. Hyundai Steel applies AI, big data for advanced production… – The new system features AI technology and big data to advance the entire process from manufacturing…
12. AI in Manufacturing | The Government of Japan – AI-based software capable of writing complex machining programs in place of skilled technicians is t…
13. Using AI to Realize Society 5.0 for SDGs (2023-06-09) – As we make increasing use of AI to bring about Society 5.0 for SDGs, we should avoid making utilizat…
14. Society 5.0 – Society 5.0 was first proposed as a human-centered society in which economic development and the res…
15. Monozukuri: Japanese Factory Automation Philosophy – We call this tacit knowledge. Monozukuri in factory automation uses sensors to turn that intuition i…
16. South Korea launches AI project to turn factory know-how into data… – The Ministry of Trade, Industry and Energy said on 28 it will 추진 the “AI model development project b…
17. AI Transformation of Korea’s Manufacturing Sector and Data… – Institute for Future Strategy Issue Brief Vol. 47“AI Transformation of Korea’s Manufacturing Sector…
18. Nvidia boosts manufacturing with Hyundai, Samsung, and SK – The two firms are shifting from strategic adoption of advanced software platforms and infrastructure…
19. Nvidia to supply 260000 chips to Korea for AI factories with… – Nvidia will deploy 50,000 GPUs to Samsung to build a new AI megafactory that integrates the entire s…
20. [PDF] How China’s Open AI Strategy Reinforces Its Industrial Dominance – China’s strategy prioritizes data curation and refinement through the deployment of embodied AI in m…
21. How China’s Open AI Strategy Reinforces Its Industrial Dominance – As the Commission’s 2025 Annual Report documented, China’s industrial base generates “interlocking i…
22. Implementation Opinions on the “AI + Manufacturing” Special Initiative – Read our translation of China’s industrial policy for the application of AI technology in the manufa…
23. America’s AI Boom Has a Trade Policy Blind Spot – The AI data center buildout has exposed three distinct failures in U.S. trade and industrial policy.
24. DeepSeek: how a small Chinese AI company is shaking up US tech… – Chinese artificial intelligence (AI) company DeepSeek has sent shockwaves through the tech community…
25. Tracking Trillions: The Assumptions Shaping the Scale of the AI… – The scale of AI infrastructure investment is most determined by assumptions around silicon useful li…
26. $150B+ AI Data Center Infrastructure Enters the Industrial Phase – How power, capital structure, and sovereign policy turned APAC into the global center of gravity for…
27. APAC sovereign investors move from caution to action on AI – Deloitte – APAC sovereign investors embrace AI, expanding into cloud, semiconductors, and governance. Discover…
28. “Manufacturing Survival Requires Korea-Japan Data Partnership… – There is a growing perception that, as in the case of OCI’s Korea-Japan joint venture, cooperation b…
