The AI Chip War in 2026: A Marathon With No Finish Line

By May 2026, the narrative around the AI chip war has shifted away from “who builds the most powerful GPU” and toward something far more structural: where do chips flow, who controls that flow, and at what cost?

In our previous piece (May 10), we analyzed how the three “AI godfathers” at Davos each revealed their geopolitical calculations through their public statements. Just one day later, new variables are entering the game.

Tightening Controls: The “New Normal” of Chip Warfare

Multiple sources confirm that the US Commerce Department is considering a third major wave of export restrictions on AI chips to China — this time expanding the scope from advanced GPUs to complete datacenter server systems. This follows the A100/H100 cutoff in 2022 and the H800 restrictions in 2024.

Why is the tightening accelerating? The core reason is that China’s AI development speed has exceeded Washington’s expectations. DeepSeek R1’s emergence technically proved that even with restricted chip access, China can train models approaching GPT-4 performance. Rather than prompting reflection, this triggered a classic Security Dilemma response: more restrictions, because each side interprets the other’s capability gains as threats requiring counter-measures.

Jensen Huang’s Dilemma: What Does Losing China Actually Cost?

Nvidia’s FY2025 earnings show China has dropped from 25% of total revenue in 2022 to under 5%. For any company, that’s a massive structural shift.

Yet Huang’s posture is subtly evolving:

Publicly, he continues using the “Sovereign AI” narrative to reassure investors — emphasizing that computing power demand from Europe, the Middle East, and Southeast Asia is filling the gap left by China. In the short term, this works: orders for H200/GB200 series from non-China markets are reportedly booked through 2027.

Privately, Nvidia is navigating a challenge it has never faced before: its ecosystem moat — CUDA — is being eroded from two directions simultaneously:

1. China: Huawei’s Ascend 910B/C series, paired with the MindSpore ecosystem, is building alternative AI infrastructure
2. Domestic US: AMD’s MI300X/400 series is matching or approaching H100 performance in certain workloads

A leaked internal memo quoted Huang as saying: “We’re not selling chips — we’re selling computing democratization.” The subtext: if computing power supply gets “sovereignized” across nations, Nvidia’s position as the global compute supplier actually strengthens. But this requires two preconditions: (1) the restrictions are not permanent, and (2) alternative markets are large enough. Both are currently uncertain.

Lisa Su’s Cold Calculus: Why AMD Isn’t Chasing Nvidia’s China Losses

Unlike Nvidia’s narrative packaging, AMD’s approach is more austere: directly acknowledge that China’s market contribution is limited, then pivot elsewhere.

Lisa Su’s exact words from the latest earnings call deserve closer attention: “Our China AI revenue is not going to be material to our overall AI story.”

Several signals emerge from this:

First, AMD’s AI chip market share is simply not large enough to make China a strategic bet. MI300X series, while competitive on certain HBM configurations, still lags on software ecosystem (ROCm), making customer migration costs high.

Second, AMD’s core profits still come from traditional CPU business (EPYC server chips), which depends heavily on US government trust. If AMD is seen as “getting too close” to China, government contracts are at risk.

Third, Su’s career background (Texas Instruments, Freescale — classic American manufacturing elite) gives her a higher sensitivity to “supply chain security” than Huang. AMD still depends on TSMC for chip manufacturing — that dependency itself forces political caution.

Three Fronts of the Computing Power Contest

Placed in the broader geopolitical framework, the 2026 AI chip war is clearly fighting on three fronts:

Front One: US-China Tech Decoupling

This has been ongoing for four years. Current state: the US controls chip design and manufacturing equipment (ASML lithography, Applied Materials), while China accelerates self-sufficiency in mature nodes (28nm+) and partial advanced nodes. Both sides know full decoupling is unrealistic, but both are vying for advantage in “partial decoupling.”

Front Two: Computing Power Distribution Among Allies

The most overlooked battleground. Japan, the Netherlands, South Korea, and Germany are all building domestic AI computing infrastructure. How Washington allocates chip supply licenses among allies directly affects the strategic cohesion of the allied system. A recent example: Washington is now debating whether to allow Japan to purchase H200 chips for the Self-Defense Forces’ AI projects — unthinkable two years ago.

Front Three: A New Round of Computing Capital Restructuring

Middle Eastern sovereign wealth funds are pouring into AI computing infrastructure. Saudi Arabia, UAE, and Qatar are acquiring large quantities of Nvidia and AMD chips under the banner of “Sovereign AI.” This isn’t merely commercial — these nations are buying computing power to acquire geopolitical voice in the AI era.

China’s Cards: Mature Nodes + Systems Innovation

Western media often overlooks a key fact: China’s real advantage in the AI chip war isn’t advanced process nodes — it’s systems-level innovation.

DeepSeek already proved this. Using limited H100 compute, they achieved GPT-4-comparable performance through MLA (Multi-head Latent Attention) architecture innovation and Mixture-of-Experts optimization. This demonstrates: chip performance matters, but algorithmic efficiency and system integration matter equally — and China’s reserves in these areas run deeper than outsiders expect.

Additionally, China’s mature-node production capacity is exploding. In 2026, Chinese domestic AI inference chips cost roughly one-tenth of an Nvidia H100. While performance gaps persist, they are already commercially competitive in the inference market.

A Marathon Without a Finish Line

To return to the question posed at the beginning: when does the AI chip war end?

Answer: it doesn’t.

The chip war is fundamentally a proxy for great power competition, and great power competition has no end. As long as US-China strategic trust cannot be rebuilt, technology and capital restrictions will persist. This is not pessimism — it’s structural analysis.

For Jensen Huang, this means learning to sustain Nvidia’s valuation narrative while “losing” the Chinese market.

For Lisa Su, this means finding overtaking opportunities within the constraints of “not betting on China.”

For China, this means the era of “buy rather than build, rent rather than buy” is truly over — independent innovation is the only path.

And for those of us observing — the key is understanding the structure of this contest, rather than being carried away by any side’s narrative.

This is an article in the “AI and Geopolitics” series. Previous: “The Geopolitical Choices of AI Godfathers — From Chip Wars to Capital Games”