22V Research held a Webinar on US-China competition in the AI technology stack with special guest Paul Triolo of DGA-Albright Stonebridge Group. Paul is a leading expert on the geopolitics of tech and China’s tech sector development.
The link to the replay is HERE.
KEY TIME STAMPS AND DISCUSSION POINTS
02:19 US-China Tech Tensions
- Trump’s December decision to allow sales of the Nvidia H200 chip marks a major inflection point in US tech policy towards China, which has become much more flexible than it was in the first Trump term or under Biden. Under the new influence of Silicon Valley, and in recognition of China’s dominance in rare earths, Trump is in favor of a “sliding scale” that seeks to keep the US ahead of China, but does not attempt to freeze China’s tech sector development.
- The policy stance is still nebulous, and parts of the bureaucracy (and Congress) still advocate for tight controls. But major new restrictions on China’s tech sector are unlikely so long as broader US-China relations remain stable ahead of Trump’s visit to China in April.
15:09 Outlook for US Chips Sales in China
- There is major demand in China for Nvidia’s H200s from companies such as ByteDance, as domestic firms are not yet ready to manufacture advanced chips at scale. Nvidia is likely looking at several million units sold in the initial phase, with the first H200s arriving in China this spring, but the policy dynamics are complicated. The US has yet to clarify the licensing process for Nvidia and AMD.
- Beijing is still working out whether it will impose its own limitations or conditions on imports of Nvidia chips. However, Chinese regulators are unlikely to impose major restrictions on the ability of major domestic tech firms to use Nvidia chips, as long as these firms are also working with domestic chipmakers (such as Huawei) as part of their overall technology stack.
- Smuggling of AI chips from the US to China is probably a bit overblown in terms of the scale and impact, and it will likely be a less prominent issue in 2026 given approval of the H200s. Note as well that some Chinese tech companies will be able to legally access US compute, including the Blackwell series, through cloud providers.
23:38 Outlook for China’s Domestic AI Chip Sector
- Visibility into the capabilities and roadmaps for Chinese chipmakers is decreasing because of the political sensitivity.
- China’s chip sector will continue to face constraints producing advanced AI chips at scale, but some of these will ease over the course of the year. Huawei is developing its own foundry, which will free up some manufacturing capacity at SMIC. Huawei and SMIC are pushing the capabilities of DUV (deep ultraviolet) lithography to its limits to produce chips at the 7nm level, and the yields are still fairly low. Chinese toolmakers outside of lithography are advancing quickly due to the vacuum left by US export controls.
- Lack of access to high bandwidth memory (HBM) chips is another bottleneck for China, but progress in memory is moving faster than for logic chips. CXMT, which is preparing for an IPO, will likely be able to produce HBM by the end of this year or early 2027 that Huawei can use in its AI hardware packages. That is important because Huawei is running down its inventory of HBM chips previously purchased from South Korean firms.
33:28 China’s EUV Capabilities Are At Least Several Years Away
- Developing domestic EUV (extreme ultraviolet) lithography is a key piece of China’s “Project Mount Everest,” which aims for self-sufficiency in advanced chips. Leaked media reports about progress in EUV are likely efforts to boost morale. The reality is that China likely will not achieve a high-volume EUV system until 2029 or 2030.
- China has enormous engineering resources but lacks ASML’s experience as a systems integrator. Financing EUV and related technologies is another constraint, as even government investors in China want a return on investment, and these are projects with uncertain timing and payback.
48:35 Meta’s Acquisition of Manus is Not a Template
- Manus is fairly unique, in that it targeted the market outside of China, and sought Western capital, from an early stage and moved its operations to Singapore to enable this. Most other Chinese AI companies with global aspirations will want to use the domestic market as a springboard, and are seeking capital through IPOs in Hong Kong.
- The Chinese government and US government will both look at this deal, but it will likely go through. Beijing will want to at least put other Chinese companies on notice that if they are considering a similar route, they will need to obtain permission from regulators and not just move their core team to Singapore.
55:00 Outlook for Chinese AI Models and Applications
- Chinese AI companies will probably not close the cap with US frontier models in 2026 when it comes to benchmarks, but are also unlikely to fall further behind thanks to access to Nvidia’s H200 and clever engineering such as demonstrated by DeepSeek.
- Assessing capabilities only through benchmarks is becoming less relevant in a world of agentic AI, where the capability to deploy AI matters more than how advanced the model is. Chinese firms will be very competitive in agentic AI, and will push the technological and regulatory envelope, as demonstrated by ByteDance’s recent pilot of an agentic smartphone.
- Over the longer term, China’s ability to keep pace with the US in AI will depend on the ability of the domestic tech sector to produce hardware that is “good enough” to meet China’s needs, and that in turn will partly depend on how US export controls evolve.
