Why China Just Started Hitting Back in the AI Race

In Brief

X. Eyeé, CEO of AI consulting firm Malo Santo, came on CNBC's Squawk on the Street with a sharp framing: the AI rivalry between the United States and China just turned bilateral. For the past few years the United States has been the side making moves: forcing the sale of TikTok, restricting chip exports, scrutinizing Chinese-linked deals. This week, China made two of its own.

Beijing blocked Meta's $2 billion acquisition of Manus, a Chinese-founded AI agent startup that had moved to Singapore. And DeepSeek released V4, a free open-source model trained on Huawei chips instead of NVIDIA's. Each move would matter on its own. Together they signal something larger: China is no longer just reacting to American AI moves. It is making its own.

What just happened with Manus

Meta announced its acquisition of Manus in December 2025 for around $2 billion. It was Meta's first major step into agentic AI, software that doesn't just answer questions but takes actions. Manus had been founded inside Chinese startup Butterfly Effect, then spun out and moved its headquarters to Singapore. By the time Meta bought it, the company had reached around $100 million in annual revenue and was one of the most-watched AI agent startups in the world.

This week, the Manus website was updated to say Manus has joined Meta. Within hours, Beijing replied with a one-line statement: the deal cannot move forward. No reasoning. No context. Just a flat refusal.

Eyeé points to the parallel with TikTok:

"What they did not allow the sale and transfer of was the underlying recommendation algorithm itself. So they sold the name, the IP, the platform. But the U.S. corporations that purchased it had to have their own recommendation algorithm."

The pattern is consistent. What Beijing decides is sensitive, it keeps. The most advanced models, the most valuable algorithms, the engineering teams behind them — these are not for sale, even after a deal closes.

The complication for Meta is that the technology has already been integrated. Engineers have moved over. The website says they have joined Meta. Untangling that, after-the-fact, is going to be expensive and messy, and Eyeé suggests the next few weeks will reveal whether Beijing meant it as a hard rule or a negotiating posture.

What just happened with DeepSeek V4

A year ago, DeepSeek shocked the AI industry by shipping a reasoning model that matched American frontier systems while using far less compute (the raw computing power needed to train and run AI). That release prompted Anthropic and others to publicly question whether massive American AI spending was actually necessary.

This week, DeepSeek did it again. Not the next reasoning model. The simpler general-purpose one: DeepSeek V4.

What V4 is, according to Eyeé:

• Better than every other open-source competitor
• Tracks alongside Claude Opus 4.6, which Anthropic touted as the best model in the industry when it shipped
• Beats Google Gemini Pro on several benchmarks Google itself uses
• Free to use

But the most important detail is the hardware. V4 was trained and optimized to run on Huawei Ascend chips, not on NVIDIA's. Inference (the step where the model actually answers a question, the part most companies pay for) was specifically targeted at Chinese chips.

The U.S. has spent years restricting NVIDIA exports to China to slow Chinese AI progress. V4 is the answer: build a competitive model that doesn't need NVIDIA chips at all.

NVIDIA wants in anyway

Within days of the V4 release, NVIDIA CEO Jensen Huang said NVIDIA had found ways to optimize DeepSeek to also run on NVIDIA hardware. That is not a coincidence. It is an admission that whatever Huawei is offering, NVIDIA cannot afford to ignore.

Eyeé reads this as the AI race heating up on two layers at once: software (the models themselves) and hardware (the chips they run on). For the past year, the public story was mostly about software. Now China is contesting the hardware layer too.

The real shift

The convenient American narrative has been: China copies us, releases cheaper versions, and we keep the lead. That narrative is harder to defend after this week.

When Beijing forces Meta to undo a $2 billion deal that already closed, that is not a defensive move. It is a signal that the most valuable AI assets are not allowed to leave China, even after they have apparently been sold.

When DeepSeek releases a model that runs on Chinese chips, prices it at a fraction of the U.S. competition, and publishes the weights for free, that is also not defensive. It is an attempt to build a parallel AI ecosystem, with chips, models, and distribution that don't depend on American technology.

Two events. Same direction. The cold war Eyeé describes isn't a metaphor about who has the better demo. It's about which country gets to decide where the most important AI runs, and on whose hardware.

Glossary

Sources and resources