Chinese AI Through the Eyes of Silicon Valley VCs: Sanctions Didn't Lock It Down, but Instead Brought Out a 'Monster'

Author: David, Deep Tide TechFlow

Deep Tide Overview: With the strong rise of China's large models, more and more Western VCs are turning their attention to the East, personally crossing the ocean to gauge the true level of Chinese AI.

Previously, we published another article titled "Western VCs' Observations on Chinese AI: Hardware Invincible, Software All Bubble," authored by a partner at the well-known crypto institution Delphi Digital. Just today, another completely different exploration note on China has gone viral in the English tech circle and on X (formerly Twitter).

This time, the perspective is more hardcore, crafted by the top-tier frontier technology and biotechnology VC Dimension.

As a heavyweight capital with significant influence in the hard tech field, Dimension's partners Nan, Adam, and Zavain led a team to deeply investigate the top AI laboratories and infrastructure ecosystems in cities like Beijing, Shanghai, and Hong Kong for a week in August this year.

Unlike the previous VC's discussion on "consumer-grade software and hardware manufacturing," this internal letter to LPs (limited partners) from Dimension directly targets the core of the Sino-American AI competition: underlying computing power, model open-sourcing, and technological decoupling.

Amidst the grand narrative propagated by mainstream Western media and political circles about the "complete decoupling of Sino-American AI," these Silicon Valley insiders witnessed a starkly different magical reality on the ground:

The U.S. chip export controls have not only failed to lock down China's AI progress but have instead, like a pressure cooker, forced out an extremely terrifying "underlying performance extraction machine." The AI industries of both countries have not only not broken apart but are tightly welded together at the foundational level through distillation, open-sourcing, and data.

Due to the length of the letter, we have compiled and distilled key content, organizing some interesting observations on the Chinese AI industry from a Western perspective as follows.

Scarcity Breeds Evolution, Sanctions Force Limits {#article-toc-33631-2}

The original intention of U.S. export controls was to slow down China's AI progress, but the result has been counterproductive. It has created immense evolutionary pressure, giving rise to a distinctly different form of AI laboratory.

Due to a severe lack of high-end computing power, Chinese teams have been forced to perform extreme engineering optimizations at the code level that American laboratories would not even consider.

One extreme example is DeepSeek's V3 model, which was trained on 2048 deliberately "castrated" H800 chips. To compensate for the deficiencies in chip interconnect performance, their engineers bypassed the mainstream CUDA framework and wrote code at the lower PTX level, reallocating 20 (out of 132) streaming multiprocessors for each GPU specifically for inter-node communication.

This is not due to a lack of genius engineers in the U.S., but because the incentive mechanisms on both sides are entirely different.

In American laboratories, spending more dollars means buying more disposable computing power; whereas in Chinese laboratories, hiring one more engineer is aimed at cutting down the demand for computing power from the foundational architecture.

This extreme scarcity has nurtured a unique full-stack efficiency culture in China—extracting system performance to the utmost from the kernel, optimizer, service system to chips.

An Extremely Competitive Arena, Not Just Externally, but Internally as Well {#article-toc-33631-3}

From Silicon Valley's perspective, Chinese AI seems to be merely competing with the U.S. However, after on-site investigations, it was found that the internal competition in China's cutting-edge AI field is far more brutal than the West imagines.

The five leading laboratories—DeepSeek, Alibaba (Qwen), Dark Side of the Moon (Kimi), ByteDance (Doubao), and Zhiyu (GLM)—are so tightly matched in strength that their rankings are almost reshuffled every quarter.

These five laboratories not only fight fiercely over model capabilities and iteration rhythms but are also crazily open-sourcing.

This high-pressure internal "arena" environment is starkly different from the ecosystem dominated by only two closed-source giants (OpenAI and Anthropic) in the U.S., and the innovative driving force it brings is no less than the transnational competition between China and the U.S.

Emerging as a New "AI Data Evaluation Dark Horse" {#article-toc-33631-4}

American data suppliers, such as Mercor, AfterQuery, and Turing, have already recorded commercial relationships with Ant Financial, Alibaba, and ByteDance. On-site, they saw a group of emerging Chinese new elites building evaluation and verification infrastructures rather than data labeling services.

Specifically in Beijing and Shanghai, they initially expected to see a large number of cheap "data labeling factories," but instead found a batch of explosive young startups focused on other aspects of AI.

For example, an AI evaluation and prediction company called UniPat, founded by a PhD from Peking University, has rapidly increased its revenue to over $100 million within less than 24 months by compensating for computing power disadvantages with extremely high quality of human supervision.

When Computing Power is Insufficient, Human Effort Fills the Gap {#article-toc-33631-5}

What surprised Silicon Valley investors the most was discovering a group of top Chinese researchers implementing "weak forms of self-improvement (RSI)" in a very primitive but effective way.

Due to the extreme lack of computing power, researchers personally intervened and assisted to accelerate model training.

Investors remarked that this was akin to the early "Centaur Chess" (a combination of human and machine defeating pure machines), where human effort and time were used to bridge the computing power gap.

Unburdened Commercial Monetization and Pragmatism {#article-toc-33631-6}

There remains a significant gap of nearly two orders of magnitude in revenue scale between Chinese and American AI. By August 2026, Anthropic's ARR (Annual Recurring Revenue) had surpassed $6.5 billion, while OpenAI reached $40 billion;

In contrast, the strongest revenue-generating large model business in China (ByteDance's video model) has an annualized revenue of about $2-3 billion, while purely large model laboratory revenues are generally in the hundreds of millions.

However, the pragmatism and down-to-earth capabilities of Chinese enterprises are impressive.

When Dimension's investors walked through the crowded streets of Hong Kong, listening to a well-known Silicon Valley podcast analyzing why American laboratories high-mindedly resist advertising and e-commerce monetization, the stark contrast with the Chinese reality before their eyes was striking.

ByteDance's Doubao has already amassed 345 million monthly active users, surpassing the total of Qwen and DeepSeek combined.

Although its daily revenue is currently less than 1 million RMB, almost all of it comes from e-commerce commissions. Chinese entrepreneurs exhibit unabashed pragmatism, showing no concern for whether the monetization method is "low" or not.

The Extremely Magical "Pacific Data Circulation," A Sino-American AI Cycle {#article-toc-33631-7}

Although hardware has decoupled at the semiconductor level and below, the speed of integration at the software and data layers is faster than any government can respond. They believe that the current real technological cycle chain of AI is actually an extremely magical "cross-ocean technology circulation":

American frontier laboratories train top models ---> Chinese teams perform distillation and open-source weights ---> American vertical AI companies fine-tune Chinese open-source models ---> Package and sell to American enterprises.

Thus, the so-called American vertical AI applications are increasingly running on Chinese open-source weights, harboring the intelligence of American frontier models.

For instance, the popular coding tool Cursor's Composer 2 is actually based on Kimi's K2.5; while the new legal AI player Harvey's Harvey Tenet is based on Kimi's open-source model K3 for post-training.

Although China has not directly earned Western money, it is crazily occupying the "workflow" of the Western AI era. Open-source models directly bypass the stringent corporate procurement review walls in Europe and America; when software is freely available, there is simply no reason to reject a supplier.

A More Frantic Valuation Bubble {#article-toc-33631-8}

Western capital markets often discuss the AI bubble, but in China, the boiling degree of this bubble far exceeds that of Silicon Valley.

Considering the huge revenue gap between Sino-American model companies, the valuation multiples of leading Chinese laboratories are often 5 to 10 times that of their American counterparts. For example, Dark Side of the Moon (Moonshot) completed a round of $3.5 billion financing at a valuation of $35 billion at the end of July, which is about 115 times its $300 million ARR, and is currently operating a financing for a $50 billion valuation before IPO (in contrast, Anthropic only has a 20 times multiple).

The secondary market is even more fervent and volatile, with companies like Zhiyu AI and MiniMax experiencing dramatic roller-coaster market fluctuations before and after their IPOs.

Finally, when this long letter was published on overseas social media, a comment in the comment section succinctly captured the entire industrial ecology under the Sino-American AI competition:

"Chips are decoupling, but intelligence can still flow."

In this extremely competitive cycle, the so-called "hard technological decoupling" is entirely a false proposition.

As Dimension stated at the end of the letter, forcibly severing ties with administrative orders will likely only stifle the innovation speed of American enterprises. The two countries have already deeply intertwined at the levels of distillation, open-sourcing, data, and reasoning.

Rather than blindly indulging, understanding the real flow of chips and technological cards on the table is what top capital and practitioners should truly focus on.