Skip to main content
Silicon Epoch
Chapter 10 · Geopolitics & The Chip Wars · Updated July 2026

Geopolitics: Sovereign silicon & supply.

The software of AI is global and open, but the physical supply chain is hyper-concentrated. Control of leading-edge semiconductor lithography and advanced packaging forms the core axis of modern superpower competition.

01

TSMC Global Expansion & Taiwan Node Ramp

TSMC fabricates the physical substrate of frontier AI. N2 (2nm) volume production commenced in Q4 2025 at Kaohsiung Fab 22 and Hsinchu Fab 20[2]. Backside power delivery nodes (N2P, A16) are scheduled for volume production in H2 2026. Global fab expansions are ongoing to de-risk geographic concentration:
  • •Arizona (Fab 21): Projected to increase output by 80% in 2026 compared to 2025; 2nm production targeted by 2028[1].
  • •Kumamoto (Japan): First fab projected to increase output by 130% year-on-year in 2026[1].
  • •Dresden (Germany): ESMC joint venture construction continues, establishing sovereign European capacity[1].
  • •CoWoS Packaging Constraint: Advanced packaging (Chip-on-Wafer-on-Substrate) remains the binding bottleneck. TSMC is ramping capacity from 35K in 2024 toward 120K–130K wafers/month by late 2026, with NVIDIA securing ~60% of total output[?].
02

EUV Lithography Scanners & ASML Monopolies

ASML (Veldhoven, Netherlands) retains a monopoly on Extreme Ultraviolet (EUV) lithography tools. The next-generation High-NA EUV scanners (0.55 NA) print chip layers at 8nm resolution. The EXE:5000 pre-production system was delivered to Samsung's Hwaseong Campus in March 2025[30], and production-scale EXE:5200 tools are shipping now:
  • •Intel Lead: Intel Oregon received the first R&D tools; commercial EXE:5200 shipments are scheduled for late 2026/early 2027 to fabricate 14A nodes[30].
  • •TSMC R&D: TSMC purchased High-NA systems (costing up to $400M each) for R&D at Hsinchu, delaying full production deployment until economics improve[30].
  • •Samsung Deployment: Samsung received its first system in 2024 and is installing two systems in 2025–2026 for 2nm logic and high-density DRAM[30].
03

U.S. BIS Export Controls & Loophole Crackdown

The U.S. Bureau of Industry and Security (BIS) enforces strict limits on advanced computing items. Under the January 15, 2026 final rule, exports of NVIDIA H200 and AMD MI325X chips to China/Macau shifted from presumption of denial to a case-by-case review, subject to strict conditions[32]:
  • •Performance Thresholds: Review applies to chips with Total Processing Performance (TPP) < 21,000 and DRAM bandwidth < 6,500 GB/s[32].
  • •Export Tariffs & Caps: Licenses impose a 25% revenue tariff, a 50% volume cap compared to U.S. shipments, and strict third-party testing[32].
  • •Cloud Loophole: The House passed the Remote Access Security Act (369-22) to block Chinese entities from renting GPU compute via foreign cloud nodes[31].
04

China's Sovereign Semiconductor Stack

In response to Western bans, China has injected USD 28.3 billion into its National Semiconductor Fund, designating SMIC and Huawei as national champions to build an independent domestic silicon ecosystem[33]:
  • •SMIC Fabrication: SMIC is producing 7nm (N+2) using older DUV multi-patterning at 20-40% yields. 5nm pilot runs are underway targeting 2026 mass production[33].
  • •Huawei Ascend 910C: Two-chiplet design delivering 780 TFLOPS BF16. Huawei shipped 805K total accelerators in 2025 (653K being 910C)[34].
  • •Huawei Ascend 950PR & DT: Next-gen architectures. 950PR delivers 1.56 PFLOPS FP4 with 112GB HiBL memory; the training variant (950DT) features 144GB HiZQ memory[34].
  • •Atlas 950 SuperPod: Massive Chinese cluster combining 8,192 Ascend 950DT chips in 160 cabinets, claiming 6.7× NVL144 compute capacity[34].
05

Sovereign AI Clouds & Regional Jurisdiction

Nations are bypassing US hyperscalers by building localized AI stacks to protect national data sovereignty. The UAE's G42 has pioneered portable jurisdiction frameworks to keep model workloads compliant[?]:
  • •Digital Embassies & Greenshield: UAE's G42 and Microsoft partner to integrate agentic AI under local legal umbrellas, ensuring data residency and security[?].
  • •Europe's CADA & Chips Act 2.0: In June 2026, the EU introduced the Cloud and AI Development Act to triple domestic datacenter capacity, contracting SAP and T-Systems for sovereign cloud stacks[?].
06

National Security AI & Defense Mandates

Superpowers have transitioned to an "AI-first" national security posture, streamlining commercial procurement and deploying zero-trust architectures for defense command networks[?]:
  • •NSPM-11 Mandate: Issued in June 2026, National Security Presidential Memorandum-11 directs U.S. defense agencies to accelerate multi-vendor frontier model integration[?].
  • •Zero Trust Defense: Compressing sensor-to-shooter latency using autonomous edge-compute agents, backed by hardened cybersecurity model defenses[?].
Supply Chain Visualized

The fabric chokepoint.

ASML's lithography scanners flow into TSMC's manufacturing foundries in Taiwan. From there, advanced chips route freely to U.S. hyperscalers, while exports to Chinese domestic entities face strict BIS licensing bans and third-party restrictions.

ASML (Netherlands)EUV/High-NA Lithography100% ChokepointTSMC (Taiwan)N2 GAA / CoWoS-S90% Advanced FabU.S. Hyperscalersunrestricted capex scalingChina DomesticAscend 950 / SMIC 5nm
Hyperscaler Demand Signals

Surging Bids for Export-Compliant Chips.

Despite the regulatory tariffs and caps, demand for U.S. hardware inside China remains unprecedented. ByteDance has prepared over $14 billion in purchase orders for NVIDIA H200 processors for 2026, contingent on case-by-case BIS export license approvals[32].

Concurrently, Chinese tech firms are committing billions to Huawei's domestic Atlas 950 platform, with ByteDance signing a $5.6B hardware deployment agreement for the upcoming Ascend 950 series[34].

← Previous: Chapter 09 · Infrastructure
Table of Chapters
Next: Chapter 11 · Open vs Closed →