By Eduardo Baptista

April 29 (Reuters) - Beijing just sent a message to every Chinese AI startup dreaming of a Western exit — and it could not be clearer.

China’s state planner has ordered the unwinding of Meta’s META.O acquisition of Chinese AI agent startup Manus, once touted as the country’s next DeepSeek. The landmark decision is a warning shot to every Chinese tech founder with Western ambitions.

Manus sought a global profile by first marketing aggressively on X and then relocating core technical staff to Singapore after closing shop on the mainland.

The downfall of this deal could be faster than the origin story. Manus’ founders are restricted from leaving China while the company risks being hollowed out as Chinese regulators assert sovereignty over the intellectual property and talent Meta acquired. An M&A deal is being dismantled in real time.

In recent years, China has banned companies from exporting user data, algorithms and other digital assets without regulatory approval. Neither Meta nor Manus sought Chinese regulatory approval for the deal or its relocation to Singapore, sources told Reuters. Meta did not respond to a request for comment.

Beijing’s message is clear: AI is key geopolitical leverage in its Cold War with Washington. Everything else – the interests of private Chinese tech firms, their investors, and multinational corporations – takes a backseat to national security.

The Manus decision shows how quickly the AI world is splitting into rival camps.

In this week’s issue, we look at how China is exercising its power and corporate alliances to hit back at U.S. export controls in another way. Forward this newsletter to a friend who follows AI, chips or geopolitics. They can also subscribe here.

OUR LATEST REPORTING IN TECH AND AI

• Blocking of Meta's AI startup buy raises risk for cross-border China tech deals

• Exclusive: US State Dept orders global warning about alleged AI thefts by DeepSeek, other Chinese firms

• DeepSeek unveils new AI model tailored for Huawei chips as China pushes for tech autonomy

• DeepSeek-V4, the Chinese AI model adapted for Huawei chips

• DeepSeek's new AI model does not wow markets in fast-changing industry

PARALLEL AI ECOSYSTEMS

DeepSeek’s preview of its long-awaited V4 model failed to tank global markets this time. But the launch introduced a different kind of reveal.

Among the limited information the small Chinese startup disclosed was how the new model has been optimized for Huawei chips, one of the clearest signs yet of China’s push to build a self-sufficient AI ecosystem. Huawei said the two companies coordinated closely to maximize performance on its Ascend 950 chips, something that had never happened with prior DeepSeek model releases.

That is precisely the nightmare Nvidia NVDA.O CEO Jensen Huang had warned about earlier this month: “The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation,” he said on the Dwarkesh Patel podcast.

For DeepSeek, the collaboration with Huawei might have been years in the making. The firm’s founder Liang Wenfeng said in his only media interview to date in 2024 that domestic chipmakers needed “someone in China to step up” and build a “supporting tech community” if the country was to build an equivalent to Nvidia.

DeepSeek did not disclose what chips it used to train V4. That omission matters. Liang warned two years ago that DeepSeek’s challenge “has never been money but the embargo on high-end chips.”

Last year, skepticism centered on its claim of low compute usage, specifically a cluster of 2,048 Nvidia H800 chips, with training costs lower than $6 million.

This year, the disclosure about its training hardware is more opaque after officials in Washington alleged the Hangzhou-based startup had acquired Nvidia chips under U.S. export controls to power its model training. Nvidia has said compliance is a top priority. Beijing has repeatedly said it opposes expanding the concept of national security to politicize trade and technology issues.

That leaves a key unknown: how much of V4 was trained on Nvidia hardware, and how much, if any, on Huawei.

The direction, however, is clear. After its breakout moment, DeepSeek is now prioritizing adaptation to domestic infrastructure. If Huawei can supply enough high-end chips, DeepSeek could become the flagship model in China’s push for AI self-reliance.

It’s a big “if.” The most powerful version of V4 is far more expensive to run due to the limited supply of Huawei’s best chips, but costs will fall once the Shenzhen-based tech giant scales production, according to DeepSeek.

Until then, the startup’s competitiveness will depend not only on model innovation but on whether China’s chip industry can keep pace, making Huawei as critical to the next phase as DeepSeek itself.

CHART OF THE WEEK:

What a difference a year makes in the geopolitical battle over AI.

On January 27, 2025, a few days after DeepSeek released its R1 AI model, its claim to have used a fraction of the computing power and cash that leading U.S. AI firms spend on model development triggered a global tech share selloff. Nvidia lost about 17%, or close to $593 billion in market value - a record one-day loss for any company on Wall Street. Meanwhile, shares of semiconductor, power and infrastructure companies exposed to AI collectively shed more than $1 trillion.

But when markets opened last Friday, several hours after DeepSeek-V4 was released, Nvidia shares rose, while Chinese ‌chipmakers rallied ⁠on expectations for wider use of homegrown chips, with Huahong Semiconductor and SMIC surging 15% and 10%, respectively.

DeepSeek's latest model may not have triggered another Nvidia rout, but it points to something more durable: the emergence of parallel AI ecosystems, one led by the United States and one increasingly built around China's own models, chips and infrastructure.