← Back to Library

Chinese AI in 2025, wrapped

Jordan Schneider delivers a year-end reckoning that defies the usual Western narrative of Chinese technological stagnation, arguing instead that 2025 was the year Beijing turned constraints into a catalyst for open-source dominance. The piece is notable not for predicting the future, but for documenting a bizarre, real-time pivot where Chinese models went from being pariahs to powering Silicon Valley's latest startup gold rush. For the busy listener, this is essential context: the global AI landscape is no longer a unipolar American hegemony, but a fractured, hyper-competitive ecosystem where the rules are being rewritten daily.

The DeepSeek Shockwave

Schneider anchors the year's narrative in the "DeepSeek Moment," a January release that shattered assumptions about compute scarcity. He writes, "The cost-efficient LLM, which uses a Mixture-of-Experts (MoE) architecture, caused many in Silicon Valley to re-evaluate their bets on scaling — and on unfettered American dominance in frontier models." This is a crucial distinction. The author isn't just reporting a model release; he is highlighting a fundamental architectural shift that allowed Chinese labs to punch above their weight class despite hardware embargoes. By leveraging MoE—a technique that activates only specific parts of a neural network for a given task—DeepSeek proved that efficiency could trump raw brute force.

Chinese AI in 2025, wrapped

The commentary notes that this success was not an isolated fluke. "Nearly every notable model released by Chinese companies in 2025 has been open source," Schneider observes, attributing this cultural shift directly to DeepSeek's example. This is a profound strategic move. By open-sourcing, Chinese firms bypassed the need for expensive, proprietary distribution channels and instead flooded the global market with high-quality tools, effectively making their technology the infrastructure for the next generation of startups. Critics might argue that open-sourcing frontier models invites security risks or theft of intellectual property, but the author suggests the opposite: it was a defensive maneuver to ensure relevance in a market where access to American chips is unpredictable.

Nearly every notable model released by Chinese companies in 2025 has been open source.

The Great Rebranding of Manus

Perhaps the most striking case study in the piece is the trajectory of the startup Manus. Schneider describes a company that launched as a "world-first general-purpose AI agent" in March, only to undergo a radical identity crisis by July. "Soon after, Manus didn't want to be Chinese anymore," he writes, detailing how the company scrubbed its domestic presence, moved to Singapore, and laid off its Beijing staff to secure American venture capital. This narrative arc is a stark illustration of the geopolitical friction tearing the tech world apart. The company's pivot wasn't about product failure; it was a survival calculation in the face of US Treasury scrutiny and investment bans.

Schneider frames this as the start of a broader trend: "Manus was the start of a wave of Chinese AI companies aggressively pursuing international expansion in the second half of this year." The author's analysis here is sharp, pointing out that while DeepSeek proved the technology was viable, the geopolitical reality forces companies to choose between their roots and their growth potential. The piece notes that whether these rebranded entities can truly shed their origins is a question for regulators in 2026. This raises a critical counterpoint: can a company truly decouple from its national ecosystem when its talent, supply chain, and data are inextricably linked to the state? The author leaves this tension unresolved, suggesting it will be the defining legal battle of the coming year.

The Chip War Whiplash

The coverage of the semiconductor conflict is a masterclass in documenting bureaucratic chaos. Schneider does not shy away from the absurdity of the situation, noting, "For most of the year, we waited with baited breath for the Trump administration to decide whether to export advanced AI chips to China — and for Beijing to make up its mind on whether it wants them after all." The timeline provided is dizzying: emergency rules, closed loopholes, secret tracking devices, and sudden revenue-sharing deals where the US government takes a 15% cut of chip sales to China.

The author highlights the specific friction point of the H20 chip, a downgrade version of Nvidia's hardware designed to skirt export controls. "The Cyberspace Administration of China (CAC) summoned Nvidia's representatives over risks of Nvidia being able to control H20s remotely, accusing them of having a 'kill switch'," Schneider reports. This accusation underscores the deep mistrust that now permeates the supply chain; even when a deal is struck, the hardware is viewed as a potential Trojan horse. The piece also touches on the rise of Huawei as a champion for an alternative ecosystem, noting how Beijing is working to bypass the CUDA moat. This is not just a trade dispute; it is a race to build a parallel internet of things, one that operates independently of American standards.

The CAC summoned top Chinese tech firms to pressure them to reduce H20s orders and supplant with domestic alternatives.

The AGI Debate and the State's Vision

Moving beyond hardware, Schneider probes the philosophical underpinnings of China's AI strategy. He questions whether the leadership truly believes in Artificial General Intelligence (AGI) or if it is merely a policy tool. The author points to a "vibe shift" in Chinese tech, where industry leaders are beginning to see themselves as pivotal to the nation's destiny. "Alibaba, whose family of Qwen models gained particular prominence in the latter half of this year, held its annual Yunqi Conference in September, and CEO Eddie Wu delivered a landmark speech sketching out his vision for transformative AI," Schneider writes. This aligns with the Politburo's recent study sessions, where experts were invited to discuss AI in terms that mirror the ambitions of Western pioneers like Yoshua Bengio and Geoffrey Hinton.

However, the piece also warns of the disconnect between grand visions and economic reality. The State Council's "AI+ Plan" aims to integrate AI into every sector of the economy, but Schneider asks, "Will the state be able to keep frustrations around unemployment at bay amid deflation?" This is a vital counterpoint often missed in techno-optimist circles. The rapid adoption of AI could exacerbate social instability if the productivity gains do not translate into broad-based economic growth. The author suggests that Beijing may need to adjust its expectations, acknowledging that AI capabilities are "jagged" and will not solve every problem instantly.

Robotics and the Bubble Risk

Finally, the commentary addresses the surge in embodied AI and humanoid robots. Schneider notes that the Government Work Report explicitly mentioned embodied AI for the first time, signaling a top-down push to leverage China's manufacturing dominance. "At least ten companies released humanoid robot models," he reports, with some competing on price while others pursue specialization. Yet, the author remains skeptical, warning that "some industry observers in China are worried that humanoids, and embodied AI in general, will turn out to be a bubble, given the sudden rush of investment and a lack of obvious business models."

This caution is well-placed. While the technology is advancing, the path to profitability remains murky. The piece highlights the tension between Western academia's reliance on affordable Chinese hardware and American policymakers' growing fear of Chinese robotics firms capturing global market share. It is a classic case of technological capability outpacing commercial viability, a pattern that has played out in previous tech bubbles.

Bottom Line

Jordan Schneider's analysis is strongest in its refusal to accept the binary narrative of American victory or Chinese collapse; instead, it reveals a messy, adaptive struggle where open-source strategies and geopolitical maneuvering are reshaping the global order. The piece's biggest vulnerability lies in its reliance on rapidly shifting policy details, which may date quickly as the executive branch continues to flip-flop on export controls. The reader should watch for the 2026 regulatory decisions that will determine whether Chinese AI firms can truly operate as global entities or if they remain trapped in a fragmented, politicized market. The era of easy access to frontier models is over; the era of strategic adaptation has begun.

Deep Dives

Explore these related deep dives:

  • Mixture of experts

    The article repeatedly references DeepSeek's Mixture-of-Experts (MoE) architecture as a key innovation enabling cost-efficient training. Understanding this machine learning technique would help readers grasp why this architectural choice was significant for Chinese AI development under compute constraints.

  • Export control

    The US-China chip war and export restrictions on Nvidia H20s/H200s are central to the article's narrative about Chinese AI development. Understanding the history and mechanisms of export controls provides essential context for the ongoing technology rivalry.

Sources

Chinese AI in 2025, wrapped

by Jordan Schneider · ChinaTalk · Read full article

A year for the history books for the Chinese AI beat. We began the year astonished by DeepSeek’s frontier model, and are ending in December with Chinese open models like Qwen powering Silicon Valley’s startup gold rush.

It’s a good time to stop and reflect on Chinese AI milestones throughout 2025. What really mattered, and what turned out to be nothingburgers?

This piece recaps:

The biggest model drops of the year

China’s evolving AGI discussion among Alibaba leadership and the Politburo

The biggest swings in the US-China chip war

Beijing’s answer to America’s AI Action plan and the MFA’s

Robots

Models.

The DeepSeek Moment.

Liang Wenfeng lit the fire

DeepSeek-R1 came out on January 20, thwarting everyone’s Chinese New Year plans. The cost-efficient LLM, which uses a Mixture-of-Experts (MoE) architecture, caused many in Silicon Valley to re-evaluate their bets on scaling — and on unfettered American dominance in frontier models. DeepSeek is powered by domestically trained Chinese engineering talent, an apparent belief in AGI, and no-strings-attached hedge fund money (it is owned by High-Flyer 幻方量化, a Hangzhou-based quantitative trading firm). There were initial concerns that such a recipe could not be replicated by more capital-constrained Chinese tech startups, but Kimi proved that wrong with K2 in July; Z.ai, Qwen, and MiniMax followed.

We translated Chinese tech media 36Kr’s interview with DeepSeek CEO Liang Wenfeng back in November 2024, and spent much of January 2025 on the DeepSeek beat (see Jordan’s conversations on DeepSeek with Miles Brundage here and with Kevin Xu of Interconnected here). Over at the newsletter, we covered how China reacted to DeepSeek’s rise, its secret sauce, and concerns around open-source as a strategy.

DeepSeek continues to be a big deal. For one, it paved the way for an open-source race dominated by Chinese models. Nearly every notable model released by Chinese companies in 2025 has been open source. In public blog posts, social media discussions, and private conversations, Chinese engineers and tech executives repeatedly attribute their open-source orientation to the example set by DeepSeek.

On the technical end, despite some remaining mystery surrounding the exact cost of training R1, DeepSeek’s viability was a shot in the arm for Chinese labs working under compute constraints. Going into 2026, with restrictions on H200s loosened and reporting that DeepSeek is still training on smuggled Nvidia, easier access to TSMC-fabbed Nvidia chips may be just what DeepSeek needs to get their ...