The Geopolitical Chokepoints of Artificial Intelligence

A guest post by Julian Alexander Brown. Helium, DRAM, HBM and the finite AI race.

Good Morning,

The year 2026 is in many ways the year of AI datacenter and semiconductor bottlenecks. As Nathan Warren over at Exponential View has noted:

⭐ The Helium Shortage

“The Iran War has created a choke point in the supply of helium, a byproduct of natural gas processing and LNG production that’s used in more than 20 steps of semiconductor fabrication….”

Helium, of all things? Helium’s unique physical properties make it irreplaceable for modern chip fabrication. Qatar is the source of 34% of global helium supply.

Exponential View, read the note.

• The Strait of Hormuz closure has blocked exports, leaving a significant portion of global supply stranded. South Korea is the most exposed, sourcing 64.7% of its helium from Qatar.

This could exasperate the HBM Bottlenecks that was already in place with rising prices. Samsung and SK Hynix produce over 60% of global memory and supply most HBM used in Nvidia’s AI GPUs.

• HBM is a major bottleneck for AI GPUs. HBM is essentially a 3D skyscraper of DRAM dies. This creates a "manufacturing bottleneck" that slows down the entire AI supply chain.

• When military strikes in the Iran War hit the Ras Laffan Industrial City in Qatar. This single facility accounts for roughly 30% to 33% of the world's helium supply. How geopolitics and the semiconductor industry intersect is obviously way more complicated than many of us realize.

⭐ The Datacenter Slowdown

New data center announcements fell by half in Q4 2025.

In Ed Zitron’s latest piece: The AI Industry Is Lying To You, he goes over some interesting issues here.

The source of Paul Kedrosky and Zitron appears to be data from Wood Mackenzie. This accelerating slowdown of datacenter announcements have many factors, including significant local community opposition in many U.S. States and projects.

“Of the 241GW of disclosed data center capacity, only 33% of it is actually under active development.” According to the sources, planned capex remains skewed towards large, speculative projects.

⭐ DRAM Shortage - RAMageddon - related to our HBM Problem

So as of March 2026, the industry is experiencing what analysts are calling "RAMageddon"— a severe, structural shortage of both High Bandwidth Memory (HBM) and conventional DRAM.

What appears to be happening, and I’m not an expert here, is that unlike previous "boom-bust" cycles caused by temporary oversupply, this shortage is driven by a fundamental reallocation of global silicon wafer capacity toward AI. The real bottleneck for GPUs is TSMC, but let’s not even go there today.

ARM Announces AI Chip

In somewhat unrelated AI chip news: ARM finally announced its much anticipated AI chip. It’s called Arm AGI CPU. This marks a historic departure from the company's 35-year-old business model of purely licensing intellectual property (IP) to partners like Apple, Qualcomm, and Nvidia. Instead of just selling blueprints, Arm is now selling physical chips to the likes of Meta, OpenAI, SAP, Cloudflare and others.

They announced this in an ARM Everywhere keynote on March 24th. I found it all fairly interesting. The company claims its flagship AI chip is specifically designed to handle the orchestration required for Agentic AI—autonomous systems that reason and act with minimal oversight.

ARM AGI CPU, ready by end of 2026 for Meta, OpenAI, etc…

I asked the talented Julian Alexander Brown before either of these trends manifested to look into the bottlenecks around AI. He writes:

Julian Alexander Brown

Julian is an exciting writer for me to connect the dots. He explores the nexus and frontiers of governance, geopolitics, economics, energy, and technology. MSc Global Governance & Diplomacy at Oxford (2025–26); experience at the White House and U.S. State Department. His research explores how emerging technologies, energy transitions, and great-power competition are reshaping transatlantic diplomacy and the global order. He’s a Second Lieutenant, U.S. Army Reserve.

Recent Works

1. Memory Is the New Bottleneck: Why Micron Is Still Mispriced

2. The AI Power Boom Is Accelerating the Energy Transition

3. A Guide to the Exponentials

4. The Doom Cycle: Why the U.S.–China AI Race Is Breaking the World Order

Increasingly for me writers who explore the macro of the frontiers of emerging tech and AI are truly exciting and valuable. Suffice to say that Julian produced for us a 4,000 word exploration on the topic.

Synopsis

Of today’s piece:

The guest contributor offered us the following summary and introduction commentary:

“For much of the past decade, AI progress appeared to be driven by ideas that diffused easily across borders. That model no longer holds. Today, frontier artificial intelligence is constrained by geopolitical chokepoints: access to advanced chips, the ability to deliver large amounts of electricity quickly, and the capital and institutions required to build and operate massive data centers. Software efficiency continues to improve, but it accelerates competition rather than leveling it. As a result, frontier AI capability is becoming geographically concentrated in places where power, silicon, finance, and political capacity align, reshaping the global distribution of technological advantage.”

Immersed in the academics of tech geopolitics and the strategic implications of emerging technologies, Julien’s practical experience in the Army and U.S. State Department look very promising to me for his future coverage.

⭐ The Geopolitical Chokepoints of Artificial Intelligence

By Julian Alexander Brown, early 2026.

Why Chips, Power, and Geopolitics Now Decide Who Wins the AI Race

National AI capability index by country, decomposed across key inputs including computing power, data, algorithms, human capital, economic resources, regulatory environment, and model performance, illustrating the uneven distribution of AI capacity across the global system. — Source Harvard University, Belfer Center

For much of the past decade, the rise of artificial intelligence appeared to be governed by inputs that diffuse quickly. Advances in architectures, algorithms, training techniques, and data practices spread rapidly through academic papers, open-source code, and global talent markets. Because these inputs were highly replicable and only weakly tied to geography, it was natural to assume that AI capability would spread broadly across countries and firms.

Comparison of AI diffusion to other earlier technologies. — Source: Microsoft AI Diffusion Report 2025.

We Are Here

That assumption no longer matches reality.

Frontier AI today is increasingly constrained by chokepoints. The binding constraints have shifted away from knowledge and technique toward physical and institutional capacity. At the center of this shift sit frontier AI data centers, where the largest models are trained and where the most demanding inference workloads are anchored. These facilities are where electricity is converted into computation at scale, and their feasibility depends on assembling inputs that are slow to expand, expensive to replicate, and unevenly distributed across the world.

In practice, frontier AI depends on the simultaneous availability of two limiting inputs: access to the most capable AI chips and access to large, reliable supplies of electricity. Software efficiency remains essential, but it now operates within these constraints rather than replacing them. Algorithms determine how efficiently compute is used, but the scale, cost, and location of frontier AI capability are set by silicon, power, capital, and the institutions that govern them. Empirical scaling laws, energy economics, and geopolitics together explain why the highest performance deployment of AI is becoming geographically concentrated even as algorithmic innovation continues to diffuse globally.

⭐ Energy as the dominant scaling variable

Once access to frontier AI chips exists, electricity becomes the principal variable that determines how large AI systems can grow. In 2024, data centers consumed roughly 415 terawatt hours of electricity, about 1.5 percent of global electricity demand. Multiple projections suggest this could more than double by 2030 as AI workloads move from intermittent bursts to sustained high utilization on specialized hardware.

According to the International Energy Agency (IEA), global data center electricity demand is projected to reach between 700 TWh and over 1,700 TWh by 2035. Driven by AI, this doubling or tripling of consumption establishes data centers as a primary driver of future global energy demand. — Credit to IEA

At the frontier, this demand concentrates in a small number of extremely large, continuously operating facilities. Frontier AI training centers routinely draw between 100 and 500 megawatts of power around the clock, placing them in the same category as heavy industrial plants. A single 300-megawatt facility consumes roughly 2.6 terawatt hours per year, comparable to the annual electricity use of a mid-size city.

Projected power growth of leading AI data centers, rising from near zero to gigawatt scale by 2026 and 2027 and approaching the electricity demand of major cities. — Credit to Epoch AI.

Many AI data centers announced or under construction in 2025 and early 2026 are even larger, designed as multi-building campuses that begin in the hundreds of megawatts and scale toward gigawatt-level power demand over time. Instead of expanding incrementally, these projects are planned from the outset to host multiple frontier training clusters. This raises the baseline for what qualifies as large-scale AI infrastructure and pushes electricity from an operating expense into a system-level constraint.

Geographic footprint of major planned frontier AI data centers, illustrating the land area and scale of xAI Colossus, OpenAI Stargate, and Meta Hyperion projects relative to dense urban environments. — Source Epoch AI.

Power delivery as a timing constraint

Falling energy prices alone do not remove this constraint. Solar power and battery storage have become much cheaper over the past decade, but lower prices do not make large amounts of electricity available faster. Frontier AI data centers require hundreds of megawatts of reliable power, delivered to specific sites on timelines set by competitive model development rather than utility planning cycles. Meeting that demand depends on permitting speed, grid interconnection, transmission capacity, and the ability to bring new generation online quickly. Only a small number of locations can do this at scale.

As I have written about previously, under these near-term constraints, AI demand is pushing investment toward modular generation and storage technologies such as utility-scale solar and batteries. These technologies dominate marginal capacity additions not because they fully solve reliability challenges, but because they can be permitted and built far faster than large thermal plants or long-lead grid expansions. At frontier scale, delays matter more than electricity prices. A year without power means a year without training runs, missed model iterations, and lost compounding advantage. Energy abundance sets the long-run ceiling of AI capability, but the speed and location of power delivery determine who reaches that ceiling first.

Software efficiency, scaling laws, and why algorithmic efficiency is not a chokepoint

Algorithmic efficiency advancements have reduced the amount of compute required to reach a given level of performance by roughly an order of magnitude. OpenAI has estimated that algorithmic progress alone reduced the compute needed to achieve fixed language-model performance by about tenfold between 2012 and 2022, independent of hardware improvements. A 2024 study by Epoch AI found that for LLMs, the compute required to reach a performance threshold has actually halved roughly every 8 months, significantly outpacing Moore’s Law. Comparable efficiency gains have appeared across vision, speech, and multimodal systems, where breakthroughs are mirroring the rapid progress seen in text-based models.

Training compute of notable machine learning models over time, illustrating the exponential growth in compute used at the AI frontier. — Source Epoch AI

Crucially, these gains have occurred across the entire frontier rather than within any single firm or country. Nearly all leading AI labs have adopted similar techniques over the same period. As these methods spread rapidly through academic publications, open-source frameworks, shared tooling, and global talent markets, software efficiency diffuses faster than any other input in the AI stack. It cannot be stockpiled, export controlled or reliably confined to a particular geography.

For this reason, algorithmic efficiency is not a geopolitical chokepoint. It does not determine who can build frontier AI systems. Instead, it determines how quickly the frontier advances for everyone.

⭐ Scaling laws and the limits of efficiency

Empirical scaling laws also explain why these efficiency gains do not level outcomes. Improvements in training methods reduce the compute required to reach a given level of performance, but as documented in the Epoch AI Capabilities Index, additional compute continues to produce meaningful gains at the frontier. As a result, efficiency triggers a Jevons paradox: the phenomenon where increasing the efficiency of a resource leads to an increase, rather than a decrease, in its total consumption. Recent industry analysis and the IMF notes that lowering the price of intelligence through better algorithms makes it economically viable to scale even further, effectively driving a “resource race” that far outpaces traditional efficiency gains.

Epoch Capabilities Index (ECI) scores of major AI models by release date, illustrating the pace and distribution of frontier capability growth across leading organizations. — Source Epoch AI

DeepSeek as a “stress test”

China’s DeepSeek provides a concrete stress test of these dynamics. DeepSeek achieved significant advances in training efficiency, delivering strong benchmark performance at far lower compute cost than many observers expected. These gains narrowed the apparent distance to leading U.S. closed models relative to open models. The initial reaction to DeepSeek reflected a genuine surprise at how much performance could be extracted from a given amount of compute through better algorithms and training regimes.

Cost of the cheapest large language model per million tokens by performance level (MMLU), illustrating the rapid decline in inference costs across leading model providers. — Source Bain

But narrowing the gap is not the same as overtaking the frontier. At no point did DeepSeek clearly surpass the leading models from OpenAI, Google, or Anthropic across the broader capabilities that define frontier performance. More importantly, the frontier did not stand still. U.S.-aligned labs quickly absorbed similar efficiency techniques and applied them to much larger compute budgets, continuing to advance overall capability. What initially appeared as a disruption increasingly resembled a compression of timelines rather than a reordering of leadership.

The conditions that enabled this compression are unlikely to persist. DeepSeek’s models were trained on export-compliant Nvidia accelerators that, while constrained relative to the most advanced U.S. systems, belonged to the same general hardware generation and supported similar software stacks. At the time, many U.S. frontier labs were still training at scale on A100- and early H100-class hardware, limiting the effective system-level gap. That proximity mattered. Algorithmic efficiency is most powerful when underlying hardware generations and system characteristics remain broadly comparable, allowing software optimizations to bridge marginal performance deficits.

That parity is now breaking down. As Chris McGuire documents in his Council on Foreign Relations analysis, U.S. export controls are designed to prevent China from accessing successive generations of frontier AI hardware. Even where recent policy changes have allowed limited access to H200-class accelerators, these permissions do not extend to Blackwell and subsequent platforms. These upcoming platforms represent step changes in system-level performance, efficiency, and memory.

Comparison of projected best-in-class AI chip performance from Nvidia and Huawei, illustrating the growing divergence in frontier compute capability. — Source CFR

As frontier AI advances on roughly annual hardware cycles, exclusion from successive generations produces a compounding disadvantage. Algorithmic efficiency can reduce the compute required to reach a given capability, but it cannot substitute for missing system-level throughput or the aggregate compute enabled by very large clusters running continuously on the newest hardware. Consequently, as hardware asymmetries widen, the scope for software-only leapfrogs narrows rather than expands, solidifying the lead of those who control the physical stack.

Relative contributions of compute scaling and algorithmic progress to AI performance, alongside the rapid growth of installed NVIDIA GPU compute capacity by generation. — Source Epoch AI

DeepSeek thus illustrates both the power and the limits of algorithmic efficiency. Efficiency can compress timelines, but it cannot overcome persistent asymmetries in access to frontier chips, energy, and capital. The initial shock faded because efficiency diffused and the frontier advanced again. Going forward, widening hardware gaps make a repeat of that moment less likely, not more.

⭐ Frontier chips as the gating constraint

At the frontier of artificial intelligence, access to advanced chips is shaped less by markets than by geopolitics. U.S. export controls on China, first tightened under the Trump administration and significantly expanded under the Biden administration in October 2022, were not designed simply to block access to Nvidia’s most advanced AI chips. They were designed to constrain China’s participation in the semiconductor ecosystem that makes frontier AI possible.

At the frontier of artificial intelligence, access to advanced chips is shaped less by markets than by geopolitics. U.S. export controls on China, first tightened under the Trump administration and significantly expanded under the Biden administration in October 2022, were not designed simply to block access to Nvidia’s most advanced AI chips. According to recent CFR analysis, these controls are part of a broader strategy to constrain China’s participation in the entire semiconductor ecosystem that makes frontier AI possible.

Comparison of projected best-in-class AI chip performance from Nvidia and Huawei, illustrating the growing divergence in frontier compute capability. — Source CFR

As Chris Miller explains in Chip War, semiconductor leadership depends on a tightly integrated global supply chain spanning chip design, fabrication, manufacturing equipment, software, and specialized labor. U.S. export controls target this entire system. They impose licensing requirements, often with a presumption of denial, not only on advanced chips but also on manufacturing tools, electronic design automation (EDA) software, technical data, and even the ability of U.S. persons to support advanced Chinese semiconductor projects.

As this supply chain is global, the controls extend well beyond U.S. firms. Their effectiveness depends on allied cooperation, secured through a mix of legal leverage, diplomatic pressure, and shared strategic interests:

• The Netherlands has fully adopted the U.S. “systems” approach in its export licensing. This has effectively halted ASML from shipping its most advanced immersion DUV lithography tools to China. Consequently, once existing orders are fulfilled, ASML forecasts its revenue share from China will fall from approximately 33% in 2025 to roughly 20% in 2026.

• Japan continues to enforce strict semiconductor export controls through its Ministry of Economy, Trade, and Industry (METI) on 23 categories of advanced manufacturing equipment. This includes restrictions on critical deposition and etching tools from firms such as Tokyo Electron and Nikon, aimed at preventing their use in sub-14 nm logic chip manufacturing.

• South Korean chipmakers Samsung and SK Hynix now must seek explicit annual approval to ship equipment to their factories in China. This new annual license framework limits their ability to upgrade those fabs and gives the United States decisive leverage over the production of advanced memory, including High-Bandwidth Memory (HBM) essential for AI chips.

• Taiwan has effectively blocked Chinese firms from the world’s most advanced fabrication. TSMC no longer manufactures 7 nm or more advanced chips for Chinese customers, following U.S. Department of Commerce directives to ensure that chips made for China remain at least two generations behind the global cutting edge.

Global competition across the AI stack, showing how leadership is distributed across firms and countries from data and software to chips, manufacturing tools, and raw materials. — Source Brookings Institution.

Technical chokepoints in lithography, fabrication, and design

At the center of this constraint sit extreme ultraviolet (EUV) lithography machines produced by ASML. These tools are essential for manufacturing chips at five nanometers and below. As of early 2026, each standard EUV system costs over $200 million, while next-generation High-NA EUV tools, required for sub-two-nanometer production, now approach $400 million per unit.

Timeline of the expansion of U.S. and European semiconductor export controls, illustrating the progressive extension from lithography equipment to advanced manufacturing tools, software, AI models, and compliance obligations. Source: Deloitte Insights.

China has never received commercially viable EUV tools under the current export regime. Replicating them would require breakthroughs across optics, plasma physics, materials science, and software that took leading firms’ decades and tens of billions of dollars to achieve. Even advanced deep ultraviolet (DUV) tools depend on restricted components, servicing, and software updates, further limiting China’s ability to maintain its existing machines and operate near the frontier.

Fabrication, the hyper-precise process of printing billions of microscopic transistors, presents another challenging barrier for China. Frontier AI chips are manufactured almost exclusively at TSMC using advanced logic processes that pack billions of transistors onto a single chip while maintaining extremely low defect rates. Recently TSMC has successfully initiated mass production of its 2nm (N2) node, utilizing a record-breaking $56 billion capital expenditure plan to meet the insatiable demand of AI hyperscalers’. Export controls prevent Chinese firms from taping out leading-edge designs at TSMC and restrict access to the equipment upgrades and process knowledge required to sustain high yields.

Despite heavy state investment, Chinese domestic fabs remain several generations behind. While China’s SMIC’s is nearing 5nm capabilities, its yields for advanced AI chips like the Ascend 910C are reportedly hovering between 60% and 70%. This is significantly lower than the 80–90% yield rates typical of TSMC’s mature advanced nodes. Consequently, Chinese firms continue to struggle with yield, reliability, and cost, even when nominal node sizes appear close on paper.

Fabrication, the hyper-precise process of printing billions of microscopic transistors, presents an almost insurmountable barrier for China. Frontier AI chips are manufactured almost exclusively at TSMC, which currently produces over 95% of the world’s most advanced semiconductors. This near monopoly is built on an unparalleled technology lead and massive scale. Export controls prevent Chinese firms from taping out leading-edge designs at TSMC and restrict access to the equipment upgrades and process knowledge required to sustain high yields. Despite China’s heavy state investment, including the $49 billion “Big Fund” Phase 3, Chinese domestic fabs remain at least three to five years behind.

Design software and human capital reinforce these constraints. Electronic design automation (EDA) tools from U.S. firms such as Synopsys and Cadence, which are used to design, simulate, and verify complex chips before fabrication, are indispensable for modern AI hardware. While some restrictions on EDA tools were temporarily eased in July 2025 following trade negotiations involving rare earth minerals, they remain a de facto monopoly and a persistent lever of geopolitical pressure. According to recent 2026 legal alerts, these software chokepoints are increasingly codified alongside broader outbound investment curbs. Furthermore, limits on U.S. persons working on advanced Chinese semiconductor projects restrict access to tacit knowledge, that cannot be replaced by capital investment alone.

Semiconductor sales by country in 2021, showing the United States’ dominant share of global semiconductor revenue relative to other major producers. —Source CSIS

⭐ System-level advantage and capital lock-in

The effects of these controls are visible in market outcomes. Nvidia’s chief executive has stated that the company’s share of the Chinese AI accelerator market fell from roughly 95% to effectively zero following export restrictions. Before the controls, China accounted for an estimated around 25% of Nvidia’s data center revenue.

The Council on Foreign Relations analysis explains why this matters for AI specifically. Frontier advantage today is determined less by individual chip specifications than by system-level capability at scale. Nvidia’s edge lies in a tightly integrated hardware, software, and networking stack that converts electricity into usable compute more efficiently than any alternative.

Crucially, Nvidia’s main competitors are either American, such as AMD, or based in allied countries like South Korea, where firms such as Samsung and SK Hynix operate within the same U.S.-aligned ecosystem. Competition exists, but it does not erode U.S. strategic advantage; it merely shifts market share within a closed network.

Crucially, Nvidia’s main competitors are either American, such as AMD, or based in allied countries like South Korea, where firms such as Samsung and SK Hynix operate within the same U.S.-aligned ecosystem. Competition exists, but it does not erode U.S. strategic advantage; it merely shifts market share within a closed network.

These hardware advantages are then locked in by capital intensity. Frontier AI data centers are among the most expensive assets in the global economy. In 2025, foreign direct investment in data centers exceeded $270 billion, with individual campuses routinely requiring $5 billion to $10 billion in upfront investment.

These hardware advantages are then locked in by capital intensity. Frontier AI data centers are among the most expensive assets in the global economy. In 2025, announced (FDI) in data centers exceeded an estimated $270 billion, accounting for more than one-fifth of all global greenfield investment. Individual campuses are scaling to unprecedented sizes, such as the $25 billion “Frontier” mega-campus in Texas, which features 1.4GW of capacity. These individual campuses routinely require $5 billion to $10 billion in upfront investment, creating a financial barrier to entry that only the best-capitalized nations and firms can overcome.

Amortized hardware and energy costs of training frontier AI models over time, showing the rapid escalation in training expenses as model scale increases. — Source Epoch AI

Firms and countries that can deploy capital earlier build larger clusters sooner, train larger models first, and justify still larger investments. When access to frontier accelerators diverges, aggregate compute diverges with it. Capital does not merely fund frontier AI; it compounds and stabilizes leadership once chip and energy advantages exist.

This is why frontier chips function as a gating constraint. Export controls do not block a single technology that can be reverse engineered; they constrain an entire, evolving ecosystem. Each new hardware generation raises the bar again, forcing China to chase a frontier that continues to advance inside a network of allied firms and nations.

⭐ The centrality of geopolitics

These constraints reassert geopolitics as a decisive force in frontier AI. Participation at the frontier now depends on control over a small set of inputs that are unevenly distributed and politically mediated: access to successive generations of advanced chips, access to the upstream ecosystems that produce them, access to large volumes of reliable energy, and access to capital markets capable of financing infrastructure at scale. How countries combine these inputs increasingly determines how they participate in the global AI system.

This does not imply that innovation elsewhere ceases. Algorithmic breakthroughs, architectural changes, and new training techniques will continue to emerge globally, including in Europe, China, Israel, India, and across parts of the Global South. But innovation heavily interacts with physical and institutional constraints:

• Sovereign AI Initiatives: Nations like the UAE and the UK are launching sovereign compute reserves to ensure national security and operational continuity.

• The Infrastructure Ceiling: While open-source models continue to improve, they often lag behind the frontier because they cannot be paired with the massive scale of hardware available to U.S.-aligned firms.

• Energy as a Chokepoint: The AI-energy nexus has become so critical that data center expenditures are forecast to reach $4 trillion by 2030, straining power grids and making energy stability a core metric of national power.

A system of differentiated participation

A small number of countries relate to the AI stack end to end. The United States sits at the center of this group, combining leadership in advanced chip design, AI software ecosystems, electronic design automation (EDA) tools, and capital markets with decisive influence over export controls. Close allies control complementary layers of the stack: Taiwan anchors leading-edge fabrication through TSMC; the Netherlands controls lithography through ASML; Japan supplies critical manufacturing equipment; and South Korea dominates advanced memory (HBM). Together, these countries control the full production chain that defines the frontier and advances with each hardware generation.

A larger group of countries does not control the full stack but possesses enough enabling inputs to exercise meaningful leverage. These countries lack semiconductor chokepoints, yet compensate through different combinations of energy, capital, institutional capacity, and geopolitical alignment. Within this group, influence takes different forms depending on national context.

Some countries, such as Saudi Arabia and the United Arab Emirates, use surplus energy and capital to secure access to U.S.-aligned chips and attract frontier-scale AI infrastructure, including large data center projects like the 5GW G42 Stargate campus. Canada, Australia, and parts of Northern Europe occupy a similar position, pairing political alignment and strong institutions with energy resources. While these countries do not shape chip design, they can train competitive national models, host large-scale training and inference workloads, and embed AI deeply into their domestic economies.

Other countries within the same group, including much of the European Union, Japan, Israel, and India, exercise leverage through different channels. Despite more limited energy or infrastructure scale, they can still build strong national AI models, lead in high-value applications such as cybersecurity and medical devices, and shape global rules through regulation and standards, including frameworks like the EU AI Act. Their influence flows less from hosting the largest compute clusters and more from adoption, governance, and integration of AI across their economies, often relying on external providers for frontier-scale workloads.

China, as discussed already, occupies a distinct and complex position. It has scale, talent, capital, and energy, and it will remain a major AI power. But it lacks secure access to successive generations of frontier chips and key upstream inputs such as EUV lithography and advanced fabrication services. Despite a recent shift in U.S. policy allowing case-by-case exports of chips like the Nvidia H200, on its own that will not be enough to fundamentally change China’s predicament and lack of leverage in the AI supply chain. In this context, innovation can narrow gaps, but China is unlikely to replace a world-leading international supply chain all within its own borders in the foreseeable future..

Most other countries relate to the AI stack primarily as adopters. This includes much of Latin America, Africa, Central Asia, Southeast Asia, and parts of Eastern Europe. These countries may have talent and demand but often lack the energy surplus, capital depth, or geopolitical alignment required to host frontier infrastructure. Their access to advanced AI is mediated through foreign platforms, cloud providers, and external capital, shaping where value, learning, and influence ultimately accumulate. Without the ability to build or host these models, they face a growing digital divide, as the strategic control and capital surpluses of AI companies are likely to remain highly concentrated in the countries that are already high-income.

Global AI diffusion by economy, measured by AI user share, illustrating uneven adoption across regions and income levels. — Source: Microsoft AI Diffusion Report 2025.

The result is not a simple hierarchy but a differentiated system of participation in the AI stack. Some countries control the full stack, while others dominate critical inputs. Many exert influence by hosting large-scale infrastructure, shaping adoption, or setting rules and standards, and a broad set participates primarily downstream.

At this stage of AI development, power flows less from isolated technical breakthroughs and more from how effectively countries position themselves within the physical and institutional layers of the stack. Durable advantage depends on the ability to convert energy, capital, and geopolitical alignment into sustained access to compute, infrastructure, and markets over time.

About the Author

Julian Alexander Brown is a master’s student in Global Governance and Diplomacy at the University of Oxford, where he studies tech geopolitics and the strategic implications of emerging technologies. He serves as a Second Lieutenant in the U.S. Army Reserve and has worked across government and the private sector, including at the White House, the U.S. Department of State, and Red Cell Partners, as well as teaching in the D.C. Jail. He earned his bachelor’s degree in International Relations and Political Science from The George Washington University, with additional study at Sciences Po Paris.

Follow his Notes. Visit his Archives in the first months of his writing. It was first launched in September 2025.

Addendum

Some of my field notes are as follows:

As of early 2026, HBM has become a literal bottleneck for the global economy.

HBM (High Bandwidth Memory) is the primary "AI memory" used in accelerators like Nvidia’s Blackwell and AMD’s Instinct series.

With the Semiconductor boom it’s not just BigTech hyperscalers spending serious money on Capex.

⭐ Capital Expenditures (Capex) are a Major AI Bottleneck

The above group are spending roughly $200 Bn. on Capex in 2026. This means by 2027, hyperscalers and major Semiconductor players will be spending over $1 Trillion a year on capex. (Sure not all of them are U.S. based, but the same supply-chain that is global)

• Total Capex 2025: approximately $160.2$ billion

• Total Capex 2026 (Forecast): $204.95$ billion

• Total Increase: $44.75$ billion

• Percentage Increase: 27.9%

For countries like China, Capex is a major bottleneck to be able to build out the AI Infrastructure required even though they have more total available energy. The Venture Capital system of the United States is very dominant. The CCP still subsidizes innovation and their own priorities significantly. China has strategic advantages in things like biotech, quantum computing, space-technologies, advanced materials, some national defense areas - that are not easy to verify and how AI will accelerate them is not fully clear yet.

Funding Advantages

While Moonshot AI recently raised $1 Billion (the bones of the new Cursor model) which is a lot for a Chinese LLM lab recently, OpenAI said it expanded its record funding round to $120 Bn. That’s a 120x difference. While ByteDance and Alibaba are significantly raising Capex in AI and datacenters, it’s still not very much.

Yet somehow China still leads in Open-source AI, some aspects of Physical AI and how its own emerging technologies compound together.

High Capacity

Watch the hearing.

⭐ China’s Growing Lead in AI Talent Density

China’s talent density in engineering and AI research is also a kind of bottleneck. They will have more graduates in Machine learning than ever. Let’s be clear here, China has significantly overtaken the U.S. in domestic talent cultivation. I won’t dwell on this point, but it bears mentioning.

• According to the MacroPolo Global AI Talent Tracker (updated in late 2025), China now produces nearly half (47%) of the world’s top-tier AI researchers, compared to just 18% produced by the United States.

• The numbers don’t lie. China currently produces nearly twice as many Science and Engineering PhD graduates as the U.S., a gap that is projected to widen through 2026.

• Now to put this in perspective, remember - Chinese institutions now account for more AI research publications than the U.S., UK, and EU combined.

Switching gears a little. The U.S.-China Economic and Security Review Commission (USCC) in its March 2026 report.

Read the following carefully:

Key Findings

• China has opted to go all in on an open-source approach to AI. Most Chinese labs publish model source code and weights. They also charge far less to use high-end products than their global competitors. This has resulted in the acceleration of global uptake of Chinese AI and created a feedback loop where widespread adoption drives iteration, then further adoption. As of publication, Alibaba’s Qwen models accounted for the largest model ecosystem on Hugging Face, with over 100,000 derivatives.

• This open ecosystem enables China to innovate close to the frontier despite significant compute constraints. Chinese labs have narrowed performance gaps with top Western large language models. They have also developed key architectural and training advances that are now industry standards.

• Open model proliferation creates alternative pathways to AI leadership. China’s strategy prioritizes data curation and refinement through the deployment of embodied AI in manufacturing, robotics, and research where specialized, real-world data from widespread use may compound into advantages that proprietary U.S. models cannot easily replicate, even if they maintain technical superiority on benchmarks.

• China’s open AI model strategy and its manufacturing dominance are mutually reinforcing. As the Commission’s 2025 Annual Report documented, China’s industrial base generates “interlocking innovation flywheels” across adjacent sectors. Open models accelerate this dynamic by enabling low-cost AI deployment across factories, factories, logistics networks, and robotics—generating real world data that feeds back into model improvement. Beijing has built the institutional infrastructure to exploit this advantage, designating data as a formal factor of production and permitting enterprises to carry data assets on their balance sheets.

• U.S. export controls primarily target the digital loop—restricting access to advanced chips used for frontier model training—but are not well suited to addressing the physical loop of deployment-driven data creation and accumulation across China’s manufacturing base. As open models reduce the compute required for effective deployment, China’s ability to generate proprietary industrial data at pace and scale becomes increasingly independent of access to cutting-edge hardware. 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.

A Subcommittee on Cybersecurity and Infrastructure Protection hearing entitled, “DeepSeek and Unitree Robotics: Examining the National Security Risks of PRC Artificial Intelligence, Robotics, and Autonomous Technologies and Building a Secure U.S. Technology Base.” - WATCH.

Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance

Read.

• Dual-Loop Strategy: China’s approach relies on two reinforcing “loops”: a digital loop, where open-source AI models (like Alibaba’s Qwen) drive rapid global adoption and community iteration, and a physical loop, where these models are deployed across China’s massive manufacturing base to generate proprietary industrial data.

• Neutralizing U.S. Export Controls: By focusing on high-efficiency open-source models that require less computing power, China is narrowing the performance gap with Western AI. This strategy allows them to innovate close to the technological frontier despite U.S. restrictions on advanced AI chips.

• Industrial Data Advantage: China is prioritizing “Embodied AI”—integrating artificial intelligence into robotics and factories. This generates a specialized “feedback loop” of real-world operational data that proprietary U.S. models cannot easily replicate, potentially giving China a long-term lead in the physical application of AI.

• Global Ecosystem Dominance: Chinese open-source models are becoming the global standard for startups and developers (accounting for an estimated 80% of U.S. AI startup usage). This widespread adoption allows Beijing to set technical norms and standards, challenging the market dominance of “closed-loop” American companies like OpenAI and Google.

I started AI Supremacy to think about the U.S. vs. China in emerging technologies. This is the meaning behind the name, in case you were ever curious. The United States and China are locked in a contest over artificial intelligence that will shape the global balance of power for decades. 🌍 It’s my belief that this strategic rivalry hastens the advent of technological progress and AI.

Further reading: China’s Tech long game

China's technology long game

Thanks for Reading!

Appendix China AI:

Newsletters relevant to covering China AI and tech. Shortlist (in no particular order):

1. Hello China Tech by Poe Zhao

2. High Capacity, by Kyle Chan

3. Recode China AI, by Tony Peng

4. China Talk, by Jordan Schneider

5. Tech Buzz Insider, by Rui Ma

6. AI Proem, by Grace Shao

7. The Dunne Insights Newsletter, by Michael Dunne

8. Concurrent, by afra

9. Interconnected, by Kevin Xu

10. Sinocism, by Bill Bishop

11. Semianalysis, by Dylan Patel and his team.

12. The CNBC articles of Evelyn Cheng

13. Sinocities, by Andrew Stokols

14. DigiChina Update, by DigiChina Project - A project of the Stanford Program for Geopolitics, Technology, and Governance

15. The Special Competitive Studies Project (SCSP) - Eric Schmidt related, SCSP is supported by a team of technologists, national security professionals, and subject-matter experts.

16. ASPI’s Cyber and Tech Digest, by Australian Strategic Policy Institute.

17. China in Space, by Jack C.

18. The China Space Monitor, by Blaine Curcio

As well as so many others. Missing to me are China Newsletters that focus on Space technologies, China’s robotics companies, semiconductor industry, national defense, China’s stocks, and specifically China’s AI startups and open-source AI labs.

China trying to reverse engineer ASML’s and EUV Machines

ASML (Netherlands) relationship with China and China’s attempt to reverse engineer their core technology is also fascinating to watch. ASML is a major bottlneck for China and its semiconductor sustainability goal.

China has frantically tried though: With a supposed "Manhattan Project" style effort in Shenzhen reportedly produced a prototype Extreme Ultraviolet (EUV) light source. While this isn't a full "machine" yet, it’s a critical component. Sources suggest the project was accelerated by hiring former ASML-trained Chinese engineers to replicate specific subsystems.

ASML does a lot of its sales in China.

The gap between China and ASML still appears immense:

ASML and China also appears to be decoupling.

• In 2024, China was ASML’s largest market, accounting for 41% of revenue due to a massive "pre-ban" stockpiling of DUV (Deep Ultraviolet) machines.

• By 2026, this share is estimated to drop to 20%. This "air pocket" is being offset by a massive surge in Taiwan (estimated 35%) and South Korea (estimated 25%), driven by the high-volume rollout of 2nm AI chips and High-NA EUV systems.

If you think about it, global leaders like TSMC, Nvidia or ASML represent significant chokepoints and bottlenecks for the future of AI. Each one is worthy of greater study and easily their own coverage.

⭐ AI Choke Points and Bottlenecks

Don’t be fooled however the Semiconductor boom, Datacenter rollout and so-called “AI revolution” and full of chokepoints and bottlenecks in the 2020s:

Further considerations:

• Advanced Packaging (CoWoS): The industry’s “true bottleneck.” Even if silicon wafers are abundant, the specialized process of stacking memory and logic (Chip-on-Wafer-on-Substrate) is controlled almost entirely by TSMC, with lead times for AI GPUs stretching up to 52 weeks.

• The HBM “Supercycle” Crunch: High-Bandwidth Memory (HBM3E/HBM4) is in perpetual shortage. SK hynix and Micron have reportedly locked up 40% of global DRAM supply for AI hyperscalers, creating a massive price-hike for consumer electronics.

• Lithography Monopolies: The reliance on ASML’s High-NA EUV machines remains an absolute bottleneck; without these $350M+ tools, training “Frontier” models (GPT-6 equivalents) is physically impossible.

• ASIC Fragmentation: Major players (Google, Amazon, Meta) are moving away from Nvidia toward custom silicon (TPUs/Trainium). This is creating a “fragmented stack” where software optimized for one chip cannot easily run on another.

• The “Pax Silica” Blocs: New U.S.-led alliances are forcing countries to choose between Western hardware stacks or Chinese open-source ecosystems, creating a “bipolar” AI world.

• Rare Earth Dominance (Venezuela & Africa): China has secured new agreements for rare earth minerals (crucial for AI hardware magnets and sensors) in regions like Venezuela, sparking potential U.S. trade retaliation.

• Sovereign AI Infrastructure: More nations (Saudi Arabia, UAE, France) are demanding “on-soil” data centers and proprietary models to avoid reliance on Silicon Valley, leading to a “protectionist” data environment.

• The “Talent Drain” Trap: While the U.S. remains the top destination for top-tier AI researchers, tightening immigration and security clearances have created a “talent bottleneck” that slows down the deployment of new labs.

Datacenter Specific

• Grid Capacity: 4–10 year wait times for high-voltage utility interconnections in Tier 1 markets.

• Power Hardware: 2.5-year+ lead times for large transformers and high-capacity switchgear.

• Liquid Cooling: Monopoly-level shortages of Quick Disconnect (QDC) couplings and Coolant Distribution Units (CDUs).

• HBM Supply: Near-total allocation of High Bandwidth Memory through 2026, stalling new GPU deployments.

• Advanced Packaging: Throughput limits at TSMC’s CoWoS facilities delaying finished AI accelerator delivery.

• Specialized Labor: Critical shortage of MEP (Mechanical, Electrical, Plumbing) contractors for 100kW+ rack densities.

• Zoning & Permitting: Increasing local moratoriums on “Gigascale” (500MW+) projects due to water and noise concerns.

• On-Site Generation: Bottlenecks in the supply of Small Modular Reactors (SMRs) and industrial-scale natural gas turbines for “off-grid” centers.

Finally of course if you believe the AI bull market is an AI bubble, debt and corporate bonds could become a bottleneck for future development as the demand for compute is likely to outpace actual real-world capacity.

I hope this gave you some food for thought.