Lunar New Year: DeepSeek and China's tech comeback

The Lunar New Year began with an unexpected shift in the global tech race. US export controls on advanced chips forced DeepSeek’s engineers to squeeze everything they could out of their algorithms, and the result made up for their lack of hardware. Chinese tech shares, held down for years by regulatory crackdowns and Western containment, were sharply re-rated by institutional investors.
How the discount built up
Chinese technology stocks have spent four years in the penalty box. The crackdown began in November 2020, when regulators halted Ant Group’s stock market listing days before it was due, and widened in 2021 to Didi, to online tutoring firms and to the platform giants’ lending and data businesses. Washington added its own pressure through export controls, first imposed on advanced chips and chipmaking tools in October 2022 and tightened in each of the two years since. Global funds cut their exposure, and valuations fell well below those of American peers with similar growth.
Brains over brute force
Western AI companies have spent two years on a brute-force capex strategy, assuming that access to leading-edge foundry capacity would keep rivals out. DeepSeek’s mixture-of-experts design showed that better use of memory bandwidth can stand in for top-tier chips. By publishing its model weights, it gave developers an open alternative to Silicon Valley’s closed, paid-for models.

DeepSeek’s background explains some of its approach. It was founded by Liang Wenfeng and is funded by High-Flyer, a quantitative hedge fund that had built up a stock of Nvidia chips for its trading research before the export controls took hold. The company does not depend on outside venture capital or on selling subscriptions, which let it publish its work openly and price access to its models far below American rivals.
The engineering choices are specific. A mixture-of-experts model activates only a small part of its parameters for each piece of text, which cuts the computation needed per answer. DeepSeek also compressed the memory its attention mechanism uses and trained on H800 chips, a version of Nvidia’s hardware with reduced interconnect bandwidth that was built to comply with American rules before those rules were tightened. Released under a permissive licence, the R1 model can be downloaded, modified and run by anyone, which spreads its techniques across the industry within weeks.
Sanctions that backfired
For global investors, the lesson is that technology sanctions can provoke the very competition they were meant to stop. Denied the best chips, Chinese engineers redesigned their models to get far more out of older ones. The pattern is familiar. In 2023 Huawei launched a smartphone powered by an advanced processor made domestically by SMIC, despite American efforts to cut it off, and that summer Beijing answered the chip controls with restrictions on exports of gallium and germanium.
Who stands to gain
The clearest beneficiaries are Chinese platform companies with large user bases and cloud businesses, such as Alibaba, Tencent and Baidu, which can deploy cheaper models at scale without depending on the newest American chips. Domestic chip designers gain an argument that good-enough hardware is sufficient. On the other side, the American labs that charge for closed models face pressure on their prices, and the chip suppliers lose some of the leverage that scarcity gave them. With mainland exchanges shut for the holiday, the first full test of the re-rating will come when trading resumes next week.

The market reaction outside China was just as telling. Hong Kong-listed technology shares had begun to recover last September, when Beijing announced a package of monetary easing and support for the stock market, and DeepSeek gave foreign investors a reason to add growth to that policy story. For American funds that have been underweight China for years, the cost of staying out rises if Chinese firms can compete in the one industry where valuations assumed they could not.
A problem for Washington
The policy question is now awkward for the United States. Export controls were designed on the assumption that computing power is the binding constraint on AI progress, and this month the outgoing administration issued a further rule limiting exports of advanced chips to most of the world. If efficiency can substitute for hardware, controls on chips slow Chinese labs without stopping them, and they push Chinese firms towards domestic suppliers such as Huawei, whose Ascend chips are the obvious alternative. The new administration may respond by extending controls to the H20, the weaker chip Nvidia still sells in China, which would hurt Nvidia’s revenue more than DeepSeek’s research.
The case against a re-rating
The sceptics have solid ground. Chinese AI firms still earn little from their models, and the domestic market for paid software is weaker than America’s. Regulatory risk has eased but not gone, since Beijing still vets generative AI services and can turn on its own champions as it did in 2021. Export controls also bite harder at the stage of deployment, when serving millions of users requires large numbers of chips, than at the stage of training a single model. A lead in efficiency is useful, but it can be copied by American labs as quickly as DeepSeek published it.
What would confirm it
The indicators to follow are the capital spending plans of the Chinese platforms in their next results, any further tightening of American export rules in response, and whether Western companies begin to build products on Chinese open-weights models. A sustained inflow of foreign money into Hong Kong-listed tech would show that the move is more than a short squeeze. China’s tech sector has shown that clever engineering can erode a lead built on expensive hardware. International investors will probably have to price Chinese AI firms as serious competitors, and American chip moats as narrower, for years to come, and the discount on Chinese platform stocks should shrink over the next year even if Washington tightens controls again.
Write to Arnav Goel at contact@thelombardreview.com
The Sunday Edition, in your inbox
India's economy, markets and money every Sunday, plus our analysis twice a week. Free.
Unsubscribe any time. See our Privacy Notice.



