AI & Geopolitics

Dear Readers,

Chip-on-Wafer-on-Substrate

Hardly any other country combines so much power in such a small space as Taiwan. In inconspicuous factory buildings, chips are produced that determine the clout of superpowers, the course of the global economy, and the speed of AI progress. Without these tiny structures, there would be no Blackwell era at Nvidia, no next generation of language models, no data streams for the digital industry. But this very backbone is also the geopolitical bone of contention of our time – between a US that wants to reduce dependencies and a China that wants to free itself from them.

In this issue, we take an in-depth look at the global chip war: from Nvidia's daring dependence on TSMC to the US's billion-dollar investments, from Huawei's DUV detours to Europe's secret key player ASML. We show why packaging is becoming perhaps the most underestimated weapon in this conflict – and how closely military maneuvers, trade tariffs, and technological races are intertwined. If you want to understand why the future of AI is being written not only in models but also in silicon, you should definitely read on.

All the best,

Taiwan is on everyone's lips and always at the center of global conflicts. A handful of factories in Hsinchu, Tainan, and Taichung decide whether data centers roll out new AI models, whether defense electronics work, and whether the global hunger for data can still be satisfied. In these clean rooms, the world's most advanced logic chips are manufactured, packaged, and installed in servers—a nervous system that thrives on designs from American companies, depends on Dutch tools, and ultimately converges in Taiwan. Without this chain, there would be no “Blackwell” era at Nvidia, no mass inference, no next wave of generative industrial software. The political-economic punchline: both major powers—the US and China—are simultaneously dependent on Taiwan and eager to reduce that dependence. The question that arises is: will the battle for semiconductors, packaging capacity, and standards escalate into a heated conflict, or will it remain—for now—a war of tariffs, embargoes, and naval exercises?

Nvidia's most advanced artificial intelligence (AI) chip, Blackwell, consists of multiple chips glued together using a complex chip on wafer on substrate (CoWoS) advanced packaging technology offered by Taiwan Semiconductor Manufacturing Co (TSMC), Nvidia's main contract chipmaker.”

— Reuters

Taiwan is the bottleneck for advanced logic chips (5 nm, 3 nm, soon 2 nm) and high-performance packaging (CoWoS/SoIC), without which today's AI accelerators would be useless. Nvidia has its data center GPUs manufactured on TSMC's custom 4NP process; the modules are then assembled using complex 2.5D/3D encapsulation (HBM stacks, interposers)—a bottleneck that has repeatedly impacted supply chains in 2024/25. TSMC has therefore doubled its CoWoS capacity and plans further expansions by 2025/26; nevertheless, packaging remains a structural bottleneck. For Blackwell (GB200), Nvidia switched to CoWoS-L and is dynamically shifting orders between Hopper legacy and Blackwell ramp-up. Crucially, virtually all CoWoS production remains in Taiwan.

This geographical concentration explains why Washington and Beijing are looking at the same facilities: whoever has access or leverage here has a direct influence on the global diffusion of AI. The fact that TSMC is expecting double-digit sales growth in 2025 and plans to start production of 2 nm (N2) at the end of 2025 further increases dependence in the high-performance segment.

Since 2022, the US has been pursuing a dual strategy: negative incentives (export controls, tariffs) and positive incentives (CHIPS funding). Specific examples of tariffs explicitly aimed at reshoring: Tariffs on semiconductor imports from China will rise from 25% to 50% in 2025; at the same time, tariffs on EV batteries, solar cells, polysilicon, and wafers—all key inputs in the electronics and solar supply chains—have been significantly increased. The White House and the USTR officially confirmed these steps in 2024. The logic behind them is to cushion foreign cost advantages, make investment in the US relatively more attractive, and at the same time slow down China's technological rise.

On the positive side, billions are flowing in from the CHIPS Act: Intel received up to $8.5 billion in direct grants (plus potential loans), GlobalFoundries around $1.5 billion for expansion in New York/Vermont, Micron up to $6.16 billion for new memory fabs in Idaho and New York; further allocations followed in 2024/25. The signal effect: Part of the future capacity growth—especially in memory and “mature nodes”—is to take place in the US. For the “cutting edge” (AI GPUs), TSMC will ramp up several plants in Arizona; however, the crucial CoWoS packaging will remain in Taiwan (for the time being), which is why US fabs will only reduce vulnerability, not eliminate it.

TLDR covers the best tech, startup, and coding stories in a quick email that takes 5 minutes to read.

No politics, sports, or weather (we promise). And it's read by over 1,250,000 people!

Subscribe for free now and you'll get our next newsletter tomorrow morning.

China is responding to export controls with massive subsidies, import substitution, and political pressure along the supply chain. Since 2023/24, Huawei has been producing SMIC 7 nm chips without EUV (via DUV multipatterning) in quantities and pushing ahead with its Ascend AI accelerator line. In 2025, Huawei announced the 910C (essentially two 910Bs), with the aim of replacing or supplementing Nvidia's models regulated in China (H20 and similar). At the same time, US agencies report that Huawei will probably not be able to produce more than ~200,000 AI chips in 2025 due to yield and tool limitations—too few to meet nationwide demand. This illustrates the dilemma: progress is real, but without ASML EUV, high-end logic remains at a disadvantage in terms of cost and yield; even 5 nm approximations via DUV are possible, but they are expensive and difficult to scale. YMTC is making rapid progress in memory (high-layer 3D NAND), but here too, sanctions lists and tool bottlenecks are slowing things down.

Lithography machines are the “printing presses” of the chip industry. ASML's EUV systems work with 13.5 nm light generated by tin droplets that are struck by a CO₂ laser ten thousand times per second. Because EUV is absorbed by glass, the scanners use only high-precision mirror optics from ZEISS. The advantage lies less in a single component than in the extreme system integration: vacuum, optics, mechatronics, metrology, and software all work together perfectly. The next step is high-NA EUV, which further increases resolution and is already available at Intel. China is investing heavily, making progress with DUV and reporting laboratory successes with EUV light sources; however, a stable, production-ready EUV scanner remains difficult to achieve in the foreseeable future – partly due to export controls.

At the same time, Beijing is trying to build up its domestic tool base. According to industry data, the share of domestic manufacturing equipment in tools purchased in China grew from ~5% (2020) to a good 11% (2024). Companies such as Naura and AMEC are growing strongly; lithography remains the Achilles' heel. SMEE is still at mature nodes in the commercial sector, with EUV remaining out of reach. The result: China can ramp up volume in mature nodes (28 nm and older) and in parts of memory production—but the gap remains in the high-end segment, which is why Huawei & Co. are creating alternatives but do not (yet) represent a complete substitute for TSMC/ASML.

ASML is the world's sole supplier of EUV scanners and also dominates the market for high-performance DUV systems. Under pressure from the US, the Netherlands has extended licensing requirements, blocked exports of certain DUV models to China, and made it clear that licenses will also be required for some services and updates from 2024.

For China, this means that not only is it more difficult to purchase the latest tools, but the availability of spare parts and software can also be regulated. For Europe, it means that ASML's business in China remains politically risky, but demand from the US/EU/Asia “AI bloc” has compensated for this so far.

“BEIJING/TAIPEI, May 23 (Reuters) - China launched "punishment" drills around Taiwan on Thursday in what it said was a response to "separatist acts", sending up heavily armed warplanes and staging mock attacks as state media denounced newly inaugurated President Lai Ching-te.

The exercises in the Taiwan Strait and around groups of Taiwan-controlled islands beside the Chinese coast come just three days after Lai took office.”

— Reuters

In military terms, Beijing has repeatedly conducted large-scale exercises around Taiwan in 2024/25—openly as “punishment” or deterrence against “separatism.” These maneuvers effectively simulate elements of a blockade (control of sea/airspace, port access, anti-intervention scenarios). Think tanks such as CSIS have played out blockade scenarios for 2025: The largest naval battles since World War II would be conceivable, with both sides suffering heavy losses and the economic effects sending shock waves around the globe. RAND analyses also outline how a tit-for-tat exchange could quickly turn into a “protracted war” with risks of escalation (including a nuclear dimension)—a strong deterrent against a reckless first strike.

The military cost-benefit analysis argues against an amphibious invasion in the short term; in terms of realpolitik, a phased approach of coercive diplomacy is more likely—pressure in gray areas, “standardized” exercises, selective trade and data blockades, cyberattacks, and, if necessary, a temporary “quasi-embargo” at sea. The goal would be to influence Taiwan's political decisions while buying time to ramp up the domestic chip base. Washington's priority therefore remains to reduce vulnerabilities (packaging!), maintain credible deterrence, but define escalation guardrails.

Taiwan is indispensable to the US because it manufactures and packages the cutting-edge AI hardware that American technology companies (Nvidia, AMD, Apple, hyperscalers) value—and because the export control architecture vis-à-vis China relies on this leverage. For China, Taiwan is central because the island—politically a “reunification goal”—also symbolizes the bottleneck of its own technological vulnerability: as long as EUV, advanced DUV generations, and advanced packaging are not replaced domestically, the rise in AI scaling will remain relatively slow. In this structure, the incentive for coercive measures is high—but the leap into a major war remains rationally unattractive due to the costs involved.

In the short term (1–3 years), open war—understood as a large-scale kinetic operation with direct US/Japanese intervention—is unlikely based on current knowledge. Reasons: (1) devastating military and economic costs on both sides (CSIS/RAND), (2) uncertain prospects of success for an amphibious invasion, (3) the option for Beijing to achieve more through gradual pressure than with an all-risk strike. Further cycles of military exercises, airspace/territorial violations, selective blockade measures, cyber operations, and trade leverage are more likely. In the medium term (3–7 years), the risk increases if China advances its chip self-sufficiency to such an extent that threats of sanctions lose their deterrent effect and domestic political or symbolic milestones (e.g., 2027 plans for PLA modernization) create pressure. This is precisely why the US is investing in tariffs, CHIPS subsidies and—critically—packaging diversification: the less the world depends on Taiwanese CoWoS capacity, the less vulnerable it is to blackmail.

In the short term (1–3 years), open war—understood as a large-scale kinetic operation with direct US/Japanese intervention—is unlikely based on current knowledge. Reasons: (1) devastating military and economic costs on both sides (CSIS/RAND), (2) uncertain prospects of success for an amphibious invasion, (3) the option for Beijing to achieve more through gradual pressure than with an all-risk strike. Further cycles of military exercises, airspace/territorial violations, selective blockade measures, cyber operations, and trade leverage are more likely. In the medium term (3–7 years), the risk increases if China advances its chip self-sufficiency to such an extent that threats of sanctions lose their deterrent effect and domestic political or symbolic milestones (e.g., 2027 plans for PLA modernization) create pressure. This is precisely why the US is investing in tariffs, CHIPS subsidies and—critically—packaging diversification: the less the world depends on Taiwanese CoWoS capacity, the less vulnerable it is to blackmail.

Sources:

🔗 https://www.reuters.com/technology/nvidia-ceo-says-its-advanced-packaging-technology-needs-are-changing-2025-01-16/

🔗 https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing

🔗 https://bidenwhitehouse.archives.gov/briefing-room/statements-releases/2024/05/14/fact-sheet-president-biden-takes-action-to-protect-american-workers-and-businesses-from-chinas-unfair-trade-practices/

🔗 https://www.reuters.com/world/asia-pacific/china-starts-military-drills-around-taiwan-days-after-new-president-takes-office-2024-05-23/ 🔗https://www.reuters.com/technology/asml-expects-us-dutch-export-rules-hit-china-sales-by-10-15-2024-01-24/ 🔗https://www.reuters.com/world/china/huawei-readies-new-ai-chip-mass-shipment-china-seeks-nvidia-alternatives-sources-2025-04-21/

ChatGPT is revolutionizing how we work, but most people barely scratch the surface. Subscribe to Mindstream for free and unlock 5 essential resources including templates, workflows, and expert strategies for 2025. Whether you're writing emails, analyzing data, or streamlining tasks, this bundle shows you exactly how to save hours every week.

Subscribe for Your Free Bundle

While the whole world often focuses exclusively on the significant benchmarks of the latest LLMs, the essential bottlenecks are often overlooked: chip production, with its potential for conflict between China and the US, and the equally necessary energy production to supply the chips installed in data centers with electricity.

The battle for the best chips is likely to continue for some time, with sanctions and counter-sanctions between the two superpowers, China and the US. Two things are questionable: 1. Will the US be able to bring a large part of chip production back to its own country? 2. Will China find its own way to produce similarly powerful chips in its own country? The very near future will depend on these two questions. The end is open.