The AI Infrastructure War: Chips, GPUs, and the Race for Computing Power

En bref : The world’s biggest tech companies are spending close to $700 billion building AI infrastructure in 2026, but the real battle is geopolitical. US export controls, TSMC’s manufacturing monopoly, and a global scramble for sovereign compute are turning AI chips into a strategic resource as contested as oil. This article maps the three fronts of the AI infrastructure war and what they mean for every country that depends on imported computing power.

$700 Billion and Counting

In 2026, the world’s largest technology companies will spend close to $700 billion building the physical machinery of artificial intelligence. Microsoft is tracking toward $120 billion or more. Meta plans $115-135 billion. Amazon’s cloud division has budgeted $200 billion. These figures — collectively greater than the GDP of Sweden — cover data centers, networking, cooling systems, and above all, the specialized chips that make AI possible.

But behind these staggering capital expenditures lies a quieter, more consequential struggle. The AI infrastructure race is not just a commercial competition between tech giants. It is a geopolitical contest over who controls the supply of computing power — and who gets cut off from it. Export bans, chip embargoes, and semiconductor nationalism are turning silicon into a strategic resource as contested as oil was in the twentieth century.

This is the AI infrastructure war, and understanding how the AI revolution depends on physical supply chains is the first step to navigating it. The fight unfolds on three fronts: trade policy, manufacturing concentration, and the scramble for sovereign compute.

The Export Control Regime

In October 2022, the US Bureau of Industry and Security issued export controls that reshaped the global semiconductor landscape. The rules restricted the sale of advanced AI chips — including NVIDIA’s A100 and H100 GPUs — to China, along with the equipment needed to manufacture them domestically. Unlike previous restrictions that targeted specific end-users, the 2022 rules blocked entire categories of chips based on performance thresholds, drawing a line across the global computing landscape.

NVIDIA responded by designing downgraded chips — the A800, H800, and later the H20 — that slipped under the performance limits. Chinese AI labs snapped up roughly one million H20 chips in 2024 before the Trump administration banned even those in April 2025, reversed course in mid-2025 by approving the H20 alongside AMD’s MI308 in exchange for a 15% revenue-share arrangement, then announced plans in December 2025 to allow NVIDIA’s H200 — the most powerful AI chip ever greenlit for Chinese export — with a 25% surcharge.

The whiplash carries real costs. For NVIDIA, China represented roughly 20-25% of data center revenue before the controls. For Chinese AI companies, each policy reversal makes the case for domestic alternatives more urgent.

The TSMC Dependency Problem

If export controls represent the political dimension of the AI infrastructure war, TSMC’s manufacturing dominance represents the structural one. Taiwan Semiconductor Manufacturing Company fabricates approximately 90% of the world’s most advanced chips — the sub-5nm processors powering NVIDIA’s GPUs, Apple’s phones, and virtually every custom AI accelerator built by Google, Amazon, and Microsoft.

This concentration in a single company, on a single island, in one of the world’s most geopolitically sensitive regions, is the AI industry’s greatest vulnerability. A disruption to TSMC’s operations would halt the global supply of advanced AI chips within weeks.

TSMC’s 2026 capital expenditure of $52-56 billion reflects both confidence and hedging. The company is expanding its Arizona mega-campus to a $165 billion, six-fab complex, but even at full buildout, Arizona will produce only a fraction of Taiwan’s output. The lithography supply chain compounds the risk: ASML, a Dutch company, is the sole manufacturer of the extreme ultraviolet (EUV) machines required for advanced chip production. Standard EUV machines cost over $200 million each, while the next-generation High-NA systems cost approximately $380 million apiece. ASML plans to produce only about 20 High-NA systems per year by 2028.

The Sovereign Compute Scramble

The combination of export controls and manufacturing concentration has triggered a global scramble for sovereign AI compute — national infrastructure that does not depend on another country’s goodwill.

The EU Chips Act aims to mobilize 43 billion euros in public and private semiconductor investment, with additional state aid decisions on first-of-a-kind facilities representing over 31.5 billion euros. Intel’s Fab 52 in Arizona runs on the company’s 18A process with $7.86 billion in CHIPS Act funding. The CHIPS and Science Act overall has catalyzed over $630 billion in private semiconductor investment across 140 projects in 28 US states.

China’s approach has been more aggressive by necessity. Huawei’s Ascend 910B, roughly comparable to NVIDIA’s A100, is widely adopted by Chinese AI labs. The newer Ascend 910C approaches but does not match H100 performance, delivering approximately 800 TFLOPS FP16 — roughly 60-80% of the H100’s capability depending on the workload, according to independent benchmarks by DeepSeek researchers. SMIC manufactures these chips at 7nm using older DUV lithography to work around ASML’s EUV export restrictions — functional but at higher cost and lower yield.

Singapore has positioned itself as a neutral hub for chip design and advanced packaging, leveraging political stability and engineering talent to attract companies navigating the US-China divide. Saudi Arabia and the UAE are building massive AI data center campuses funded by sovereign wealth. India’s Semiconductor Mission, backed by an incentive framework of 76,000 crore rupees (approximately $9 billion), targets chip packaging, testing, and domestic manufacturing as a first step.

Supply Chain Chokepoints

Beyond TSMC and ASML, less visible chokepoints create systemic risk throughout the AI infrastructure war.

High-bandwidth memory (HBM) is manufactured by only three companies: SK Hynix, Samsung, and Micron. SK Hynix holds an estimated 62% market share in HBM shipments. Rare earth elements required for manufacturing — particularly gallium and germanium — are dominated by China, which imposed export controls on these materials in 2023 as retaliation. Advanced packaging has emerged as its own bottleneck: NVIDIA has reportedly secured over 60% of TSMC’s 2026 CoWoS allocation, meaning that even unlimited fabrication capacity would not solve the assembly constraint.

The GPU cloud market reflects these supply dynamics. CoreWeave, which projects $12-13 billion in 2026 revenue, built its business by securing GPU allocation early. Access to compute has become a competitive moat in itself.

Who Has the Advantage

The United States retains the strongest overall position — American companies dominate chip design (NVIDIA, AMD), cloud infrastructure (AWS, Azure, Google Cloud), and the software ecosystems tying them together. But the US does not control manufacturing. That power sits with TSMC in Taiwan and ASML in the Netherlands.

China is closing the gap faster than many expected. DeepSeek demonstrated in early 2025 that frontier-class model performance could be achieved with fewer, less advanced chips through algorithmic efficiency — a finding that undermined the premise that export controls alone could contain Chinese AI progress.

Europe has invested heavily in subsidies but lacks a domestic chip design champion. The EU’s strength lies in its control of critical equipment (ASML) and regulatory influence.

What Comes Next

Three dynamics will shape the next phase of the AI infrastructure war.

First, the export control regime will remain volatile. The tension between national security objectives and commercial interests ensures ongoing policy unpredictability. Companies building supply chains around regulatory assumptions are building on sand.

Second, manufacturing diversification will accelerate but remain incomplete. TSMC’s Arizona fabs, Intel’s domestic expansion, and Samsung’s Texas operations will collectively reduce — but not eliminate — the concentration of advanced manufacturing in Taiwan.

Third, the definition of “advanced” AI infrastructure will shift. As compute scaling techniques improve and inference-optimized architectures mature, efficient inference on mid-tier hardware could prove as consequential as frontier training on cutting-edge GPUs. The NVIDIA GPU economy may evolve from a hardware monopoly into a platform ecosystem where software lock-in matters more than silicon specifications.

The AI infrastructure war is not a contest any single nation wins outright. It is a long-running negotiation — conducted through trade policy, capital investment, and engineering innovation — over who builds the computing foundation of the twenty-first century.

Frequently Asked Questions

Sources & Further Reading

• US Export Controls on Advanced Computing and Semiconductors — Bureau of Industry and Security
• TSMC 2026 Capex Spend of $56 Billion — Data Center Dynamics
• ASML Annual Report 2025 — ASML
• Huawei Ascend 910C vs Nvidia H100 — IEEE Spectrum
• CHIPS and Science Act Investment Tracker — Semiconductor Industry Association
• AI Capex 2026: The $690B Infrastructure Sprint — Futurum Group
• European Chips Act Overview — European Commission
• TSMC Intends to Expand Arizona Investment to $165 Billion — TSMC Newsroom