China’s AI ascent leaves Trump a stark choice: escalate or relax chip controls?
Beijing and Washington are locked in an era-defining contest – and China’s rapid technological progress raises questions about limits of containment
US President Donald Trump’s landmark visit to China comes as the US-Iran war disrupts global energy supplies, fuels economic uncertainty and adds fresh strain to Washington-Beijing ties. In this story, part of a series examining how rivalry, interdependence and geopolitical crises are reshaping the relationship between the two powers, we examine how artificial intelligence (AI), chip controls and competing technology ecosystems are redefining US-China rivalry.
Barely managing a few bites of his dinner, he breezed through a marathon of media interviews and accepted a stream of selfie requests from guests. The rock star treatment even spilled into the hotel lobby when he left, where Huang patiently signed autographs for star-struck fans.
The developments underscore the rapid rise of China’s AI industry and its decreasing reliance on Nvidia’s technology, even as chips remain at the heart of the US-China rivalry.
Huang will not be accompanying US President Donald Trump on his trip to Beijing to meet Chinese President Xi Jinping this week, but the same question will loom over the talks: will the head of Nvidia finally manage to get what he wants?
A deal on semiconductors?
That view was widely shared by analysts, with some arguing that export controls had become embedded in the broader strategic rivalry, making them increasingly difficult to roll back.
Ker Gibbs, former president of the American Chamber of Commerce in Shanghai, said technology controls, semiconductors and AI were structural issues that were “too hard to negotiate and too deeply embedded in each country’s strategic priorities”.
Laila Khawaja, research director at Gavekal Technologies, said there was little political appetite in Washington to roll back its export controls. If implemented in full or in part, the Match Act could deter China’s progress in advanced semiconductor manufacturing and substantially disrupt current capabilities, she warned, adding that Beijing was likely to use the summit to remind Washington that escalation carried costs.
Against this backdrop, some analysts believe Beijing’s goal may be less about rolling back existing restrictions than preventing further curbs. Wang Dan, China director at Eurasia Group, said the latter would be a great deliverable for Beijing. “What’s even more important is that the existing export controls aren’t escalated,” Wang said.
The H200 divide
“China is about 40 per cent of the world’s technology industry. To concede that market for the US technology industry is a disservice to our country,” Huang also warned last month, after repeatedly urging the US government to approve sales of the chip to China last year.
Analysts said the drawn-out regulatory process surrounding the H200 illustrates how the chip has become a bargaining chip for broader concessions.
As the semiconductor, AI and robotics industries expand rapidly, there is a massive need to boost computing power
Khawaja of Gavekal said Beijing was forced to balance these two priorities and would eventually allow H200 imports, given rising demand and limited advanced chip capacity at home, although “how it frames the decision will depend on the summit”.
“A positive outcome could see it presented as part of an open, cooperative stance; otherwise, approvals may be granted quietly to meet practical demand without signalling concession,” she added.
Tech war: has China cracked the code?
A year ago, many questioned whether Chinese AI chips could become more than emergency stopgaps for restricted US tech. Now, Beijing is building a commercially viable alternative AI stack that is less dependent on American goodwill.
For the first time, one of DeepSeek’s top-tier AI models was no longer solely dependent on Nvidia hardware and, in some areas, ran more efficiently on Chinese chips.
“Jensen’s concern has become a reality,” said Lian Jye Su, chief analyst at research firm Omdia, citing DeepSeek V4’s adaptation to both Huawei chips and Huawei’s flagship Compute Architecture for Neural Networks (CANN) – widely viewed as China’s closest equivalent to Nvidia’s dominant CUDA software ecosystem.
Morgan Stanley analysts underscored this trend in a recent report, which argued China’s AI graphics processing unit (GPU) market had shifted “from whether domestic chips can participate” to “which vendors will win meaningful share as inference demand scales”.
“China is narrowing the US lead in AI compute not simply at the chip level but also through system-level innovation, supply-chain localisation and increasingly attractive inference economics,” wrote the analysts led by Charlie Chan, a technology researcher leading the investment bank’s Greater China semiconductor coverage.
Zhang Yunquan, a professor specialising in high-performance computing at the Chinese Academy of Sciences, said rising demand for computing power had made inference-friendly chips an area where China could compete internationally. But he cautioned that China remained behind in high-end training chips – with Huawei leading domestic efforts to close the gap.
“Over three to five years, with continued iteration on chipset yields and CANN maturity at cluster scale, domestic chips could handle most of the domestic training workloads,” said Su of Omdia.
This [the Match Act] is an area of bipartisan consensus... And that’s going to limit Trump’s ability to influence that piece of legislation going forward
“As the AI chipset ecosystem in China continues to mature, we will see … more Chinese model providers build AI models optimised for Chinese AI chipsets from day zero,” Su said.
A new equilibrium?
According to the report, Huawei’s Ascend 950 and Cambricon’s Siyuan 690 can outperform Nvidia’s H20 by 50 to 150 per cent in different scenarios when measured by tokens per second – a crucial metric in China’s inference-driven market that reflects both hardware performance and software optimisation.
Most of the leading domestic chips also carry lower costs, with market prices for Huawei’s Ascend 950PR, MetaX’s C600 and Moore Threads’ MTT S5000 cheaper than both the H20 and H200, according to the report.
The authors also identified a fundamental shift in China’s AI chip procurement – from “policy-driven” to “profit-driven” – as major AI companies rush to monetise amid rising token consumption. That means firms able to deliver ready-to-use accelerators quickly could be best positioned to see surging revenue growth.
But as China’s domestic AI ecosystem advances, uncertainty is growing over the future trajectory of US export controls – and the wider technological rivalry between Washington and Beijing.
That balancing act is expected to feature at the coming summit, though the outcome of the talks could be complicated by a stark divide within Washington over the scope and scale of tech restrictions.
According to Marro at the EIU, the executive branch had shown some willingness to keep China dependent on US technology, as seen in the H200 policy, while a hardline Congress was pushing aggressive legislation such as the Match Act.
“This is an area of bipartisan consensus,” he said. “And that’s going to limit Trump’s ability to influence that piece of legislation going forward.”
On chips, Eurasia’s Wang argued that Trump was still unlikely to ease controls on the most advanced chips below the 7-nanometre threshold, though export volumes of less advanced chips were likely to rise – suggesting Washington would continue to manage, rather than sever, China’s access to US technology.
Additional reporting by Kandy Wong