One Stack Is Closing. The Other Is Just Adding Layers.

Jensen Huang framed the stakes with unusual bluntness this spring. The Nvidia CEO warned that US export controls are not merely costing him sales in China — they are pushing Chinese firms to build an alternative to the entire American AI platform, from chips to the software that runs on them ^[1]. What he described is now visible in the numbers, and it reveals something counterintuitive about how tech decoupling actually works. Both countries want the same thing: control over their own AI stack — the chips, the models, the data centers. Both are pouring billions into domestic alternatives. But they are arriving at opposite outcomes, and the reason is the constraint itself. China cannot buy Nvidia's best chips or the leading Western AI models. American companies can buy from anyone, including each other and their competitors. The side that is blocked is building a genuinely self-contained ecosystem. The side that is free is layering in-house alternatives on top of dependencies it never sheds. The Chinese replacement is measurable and accelerating. Nvidia's share of the Chinese AI chip market has fallen from 95% to 40% this year and is projected to drop to 8% in 2026 ^[2][3]. Huawei's Ascend chips are taking the lost ground, and the model companies are following. Zhipu AI trained its GLM-5.2 frontier model entirely on 100,000 Huawei Ascend processors, competing with OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 at less than a tenth of the cost ^[4]. DeepSeek optimized its V4 model specifically for Huawei Ascend ^[5], and once that integration proved out, ByteDance, Tencent, and Alibaba all ordered Huawei's next-generation Ascend 950 ^[6]. China's Ministry of Commerce cites US export controls directly as the trigger for this push ^[7]. On the American side, the same de-risking instinct produces something structurally different. Amazon's in-house Trainium chips have reached a $20 billion annual run rate, and Amazon is in talks to sell them to outside customers ^[8][9] — while simultaneously committing to buy roughly one million Nvidia GPUs through 2027 ^[9]. Microsoft is routing production traffic for Excel, Word, Teams, and GitHub Copilot through its own MAI models, with the stated goal of eliminating payments to outside labs ^[10].

We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost — Mustafa Suleyman

Yet Microsoft is also exploring hosting DeepSeek's Chinese open-source model on Azure for a cheaper Copilot tier ^[11] — reaching across the very divide it is supposedly walling off.

How each side builds its own AI stack — and what it still buys from outside

China: replacing: Nvidia China share: 95% → projected 8% ^[2][3]. Zhipu GLM-5.2 trained on 100,000 Huawei Ascend 910B, frontier-level at <1/10th the cost ^[4]. DeepSeek V4 optimized for Huawei Ascend ^[5]. ByteDance, Tencent, Alibaba all ordered Huawei Ascend 950 ^[6]. Huawei's Tau Scaling Law targets 1.4nm without EUV by 2031 ^[12]. Beijing: 2 trillion yuan for domestic data centers, 80% domestic-tech requirement ^[13]. Model exports restricted as national security assets ^[14].

US: layering: Amazon: Trainium at $20B run rate, talks for third-party sales, still buying 1M Nvidia GPUs through 2027 ^[8][9]. Microsoft: MAI models replacing external labs, also exploring hosting DeepSeek on Azure ^[10][11]. Apple: exploring US chip production, buying Chinese memory chips ^[15][16]. Alphabet: most complete US integration (TPUs + Gemini), still buys Nvidia ^[17][8]. US custom silicon co-designed with Broadcom and Marvell, not built from scratch ^[18][19].

There is a deeper structural reason the American layering differs from China's replacing. US "in-house" chips are not truly built in-house. Amazon's Trainium is architected by Marvell. Google's TPUs are co-designed with Broadcom, which also designs chips for Meta, OpenAI, and Anthropic ^[18][19]. American hyperscalers are inserting a specialist intermediary between themselves and Nvidia. Chinese firms are building on Huawei's domestically designed Ascend chips — a cleaner break, even if those chips are not yet as efficient as Nvidia's.

2.7% Anthropic's remaining performance lead over Chinese models, per Stanford's 2026 AI Index — The gap the export controls were supposed to preserve ^[20]

If the Chinese stack were demonstrably inferior, the strategy of denial would be working — keeping China a generation behind. But the gap is measured in single digits. China leads the world in AI patents and publications, and researcher migration to the US has dropped 80% ^[20]. The controls forced the construction of a fully domestic stack that turned out to be near-peer, while the US stack still runs on Taiwanese-foundry chips, Nvidia GPUs, and a web of specialist intermediaries. US officials frame the export controls as a tool to maintain America's lead over China ^[21]. The irony is that the controls have done more to build China's self-containment than America's. The side with less freedom to trade now has a more self-contained ecosystem than the side with every option open. TSMC still manufactures 90% of the world's advanced processors, and the Western supply chain is expanding, not closing ^[22]. But growth through trade is not the same as independence through construction. The question now is whether Washington adjusts. Having driven China to build a domestic stack that now nearly matches its own, the export-control strategy has run out of room to deny technology China has learned to make itself. The next move — tighten further, or accept the bifurcation and compete on terms — will define whether the two stacks stay parallel or begin to pull apart in directions no policy designed them to go.