China's triumph despite sanctions: why the East is so successful with AI - a DeepDive

While Washington banned exports of Nvidia's H20 accelerators in spring 2025, Chinese laboratories simultaneously presented language models that ranked on par with GPT-4o and o1 in benchmarks. Just three months later, the US government lifted the H20 ban – a concession to economic pressure and the threat of supply shortages for strategic minerals such as gallium and germanium. This raises the question once again: 

How is China managing to continue developing cutting-edge AI despite tough computing sanctions?

Beijing is pursuing a multi-pronged approach. On the one hand, it is focusing on algorithmic efficiency: modern mixture-of-experts architectures such as DeepSeek-V2 activate only about ten percent of the total parameters per token, thereby drastically reducing the training effort. On the other hand, the domestic hardware industry is growing stronger: According to benchmarks close to the company, Huawei's Ascend 910B achieves up to 80 percent of the A100's performance. Where both are insufficient, gray markets fill the gap – estimates suggest that Nvidia GPUs worth over a billion US dollars entered the country through unofficial channels in the second quarter of 2025 alone.

This combination of software elegance, confident chip strategy, and passionate improvisation is forcing the US to keep coming up with new countermeasures – and at the same time revealing the limits of traditional export controls. Whether the latest H20 deal marks a respite or the start of the next stage of escalation will determine the global AI race in the coming months.

“Chinese AI developers have faced U.S. restrictions on access to the world’s most advanced AI chips, including those from chip leader Nvidia, since 2022. The Biden administration in December again tightened export control rules. But the developers have found workarounds.”

— Wallstreet Journal

Shortly before midnight, hundreds of thousands of green status LEDs flicker in a data center on the outskirts of Shenzhen. While the US has been blocking the export of high-performance GPUs since 2022, language models are being developed here that compete with GPT-4 in benchmarks. The scene highlights a paradox: How are Chinese developers managing to train world-leading models despite political restrictions on the necessary computing power? Washington's increasingly tight restrictions – from the A100 and H100 bans in 2022 to the blockade of the specially slimmed-down A800/H800 and the H20 ban in April 2025 – are well documented.

Nevertheless, China's range of models grew rapidly, from Baidu ERNIE 4.5 to DeepSeek V3 to Tencent's Hunyuan. This discrepancy leads to the guiding question that structures this article: Why does Beijing apparently need significantly fewer transistors and watt hours to deliver world-class AI?

In 2 years you will be working for AI

Or an AI will be working for you

Here's how you can future-proof yourself:

Join 1,000,000+ pros at companies like Google, Meta, and Amazon that are using AI to get ahead.

Sign up and start learning AI

The trend toward sparsity provides a first clue. Large language models can be built in a modular fashion so that only a fraction of all parameters are active per token. DeepSeek-V2, for example, has a nominal 236 billion parameters, but only 21 billion are calculated simultaneously – reducing training costs by up to 90%. Engineers take limited GPU budgets into account right from the design stage: Federated learning distributes the computing load across clusters in different provinces, which flattens power peaks and allows existing cards to be used for longer.

A quote from an internal DeepSeek paper sums up the attitude: “Scaling is not an arms race for parameters, but an optimization problem for global resources.” Such theses contradict the US paradigm of “bigger is better” that will dominate until 2023 and shift the game toward software elegance.

“According to a report from South China Morning Post, Wang Tao, Executive Director, Chairman of the ICT Infrastructure Managing Board at Huawei, recently stated at the World Semiconductor Conference and Nanjing International Semiconductor Expo that Huawei’s advanced Ascend 910B AI chip achieves an efficiency of up to 80% compared to NVIDIA’s A100 when training large-scale language models. In terms of specific test performance, it surpasses NVIDIA’s A100 AI GPU by 20%.“

— trendforce.com

At the same time, China is aggressively investing in its own accelerators. According to benchmarks close to the company, Huawei's Ascend 910B achieves 80% of the A100's performance, and even exceeds it in some workloads – without any US IP in the silicon. Since April 2025, Baidu has been equipping its new supercluster with 30,000 Kunlun P800 chips; it can train models beyond the 100 billion parameter class or fine-tune a thousand customers in parallel.

Even if the yield of the 910 series remains low and software porting is laborious, the balance of power is shifting. Developers are getting used to native toolchains, and each new generation reduces dependence on Nvidia.

“A US probe into Chinese AI firm DeepSeek’s possible circumvention of US export controls has cast Singapore in the spotlight. Nvidia’s financial filings revealed that 22% of its third-quarter revenue was billed to Singapore, making it the second-largest market after the US. However, Nvidia insists that these are “bill to” locations, not “ship to” addresses.”

— sanctionsassociation.com

Nevertheless, Nvidia cards remain in high demand – and are finding their way into the country in large quantities via back channels. Between April and July 2025, GPUs worth over a billion US dollars were imported past export controls, often in small shipments via Hong Kong or Singapore. Repair shops in Guangdong dismantle defective A100 modules, replace memory chips, and put them back into service: up to 500 cards per month.

Even training data travels unconventional routes: terabytes are flown to Bangkok in backpacks full of high-performance SSDs, pre-trained on rented servers there, and then played back. Nvidia is now publicly warning against “unofficial chips without warranty.”

While Chinese companies were exploring unofficial avenues, Beijing put targeted pressure on Washington. The ban on gallium, germanium, and later antimony in December 2024 hit US and European semiconductor supply chains hard. In the spring of 2025, China further restricted exports, and prices for high-purity germanium quintupled within twelve months. The message was clear: those who withhold chips will not receive raw materials.

“Gallium and germanium are used in semiconductors, while germanium is also used in infrared technology, fibre optic cables and solar cells. Antimony is used in bullets and other weaponry, while graphite is the largest component by volume of electric vehicle batteries.”

— Reuters

At the same time, Nvidia CEO Jensen Huang lobbied in Washington and warned of sales losses of US$5–6 billion. After a meeting with President Trump in mid-July 2025, there was a change of course: H20 deliveries to China can resume, and an “RTX Pro” series is in the works.

Economic pressure and diplomatic pinpricks thus had a pincer effect on the US administration.

The analysis points to a multidimensional recipe for success: algorithmic efficiency reduces the need for high-performance cards, in-house chips fill gaps, gray markets smooth out peaks—and political countermeasures provide leverage in Washington. Taken together, this creates an ecosystem that is not dependent on a single technology or a single supplier country.

China is demonstrating that technological leadership is not solely linked to the number of newly delivered GPUs. What matters are clever architectures, diversified hardware sources, and a foreign trade policy that uses raw material and sales markets as bargaining chips.

This resolves the discrepancy outlined at the beginning: limited computing resources are compensated for by organizational and strategic expertise. Whether the reapproval of the H20 marks a respite or the beginning of a new race remains to be seen.

What is certain, however, is that any further barriers are more likely to fuel recent creative impulses than slow them down – a lesson that export policymakers and chip designers alike should take to heart. Perhaps the most pressing question for both sides is no longer whether restrictions work, but how we can coexist in a lasting, innovation-driven equilibrium.placeholder text

Sources:

🔗 Wall Street Journal, „China’s AI Is Catching Up Fast“, Dez 2024
🔗 CSIS, „Limits of Chip Export Controls“, Apr 2025
🔗 DeepSeek‑V2 Paper, arXiv 2405.04434
🔗 Efficient Training through FL+MoE, arXiv 2507.05685
🔗 TrendForce Bericht zu Huawei Ascend 910B, Juni 2024
🔗 Reuters, „Baidu Cluster mit 30 000 Kunlun‑Chips“, Apr 2025
🔗 Tom’s Hardware, „$1 bn Nvidia GPUs Smuggled“, Jul 2025
🔗 ACSS Singapore, „Great Chip Shuffle“, März 2025
🔗 Tom’s Hardware, „Underground Repair Shops“, Jul 2025
🔗 Times of India, „Suitcases voller SSDs“, Jul 2025
🔗 Reuters, „China banns Gallium, Germanium“, Dez 2024
🔗 Reuters, „Export‑Einbruch Germanium“, Jul 2025
🔗 New York Post, „Huang trifft Trump“, Jul 2025
🔗 Reuters, „H20‑Lieferung genehmigt“, Jul 2025

Looking at the current dynamics, an increasingly asymmetrical but highly strategic race between the US and China is emerging in the field of generative AI. The US backtracking on H20 exports was not a sign of détente, but rather an expression of economic and geopolitical pressure – especially from Nvidia, which generates over 20% of its revenue in China.

In the short term, a more restrictive but differentiated US strategy is to be expected. Instead of a total ban, there will be small-scale approvals, dynamic performance limits, and possibly even firmware-based control of GPUs to prevent export goods from entering the country via third countries or being used for military purposes. At the same time, there is discussion in the US about a licensing system that would require AI models above a certain compute footprint to be registered or approved – similar to an “OpenAI/Anthropic Only” licensing policy.

China is responding on three fronts:

China's leverage in the form of rare earths remains particularly explosive. If Washington goes too far again, countermeasures at the raw materials level are likely – most recently, prices for high-purity germanium rose by over 400%. The US is more vulnerable than it admits. Its dependence on China for raw materials for semiconductor production is structural.

The US can currently delay China's AI advance, but it cannot stop it. Every new export restriction provokes innovation and strategic relocation in China. The pressure on Western companies such as Nvidia to choose between economic freedom and geopolitical loyalty is increasing. What is emerging is a global dualism: two AI worlds, connected via open-source code but separated by chips, data, and political narratives. In the long run, China seems to have the advantage—not because of greater resources, but because of its compulsion for efficiency. And perhaps, paradoxically, that is the real superpower.