Author: David, Schain TechFlow
DeepSight summary: As China’s large models surge forward, an increasing number of Western VCs are turning their attention eastward, traveling across the ocean to assess the true capabilities of China’s AI industry.
Previously, we published another article 《A Western VC’s Notes on Visiting China’s AI Industry: Shenzhen’s Hardware Shocked Me, I’m Bearish on Chinese Software》, authored by a partner at the renowned crypto firm Delphi Digital. Today, a radically different set of notes on China has gone viral across the English tech community and X (Twitter).
This time, the perspective is even more hardcore, led by Silicon Valley’s top frontier technology and biotech venture capital firm Dimension.
As a leading capital firm with significant influence in the hard tech sector, Dimension’s partners Nan, Adam, and Zavain led a week-long intensive tour in August to assess China’s top AI labs and infrastructure ecosystems in cities including Beijing, Shanghai, and Hong Kong.
Unlike the previous VC’s discussion on “consumer software and hardware manufacturing,” Dimension’s internal letter to LPs (limited partners) focuses directly on the core of the U.S.-China AI rivalry: foundational compute, open-source models, and technological decoupling.
Amid Western mainstream media and political circles continuously promoting the grand narrative of “complete decoupling between China and the U.S. in AI,” these Silicon Valley insiders witnessed a completely different, surreal reality:
America's chip export controls have not halted China's AI progress; instead, they've forced the creation of an incredibly powerful "bottom-layer performance extraction machine," much like a pressure cooker. Rather than breaking apart, the AI industries of China and the United States are now tightly welded together at the foundational level through distillation, open-source initiatives, and data.
Due to the length of this letter, we have compiled and distilled its key content, summarizing some interesting observations on China's AI industry from a Western perspective as follows.
Scarcity drives evolution; sanctions push to the limit
The original intent of U.S. export controls was to slow China’s AI progress, but the result was the opposite. It created immense evolutionary pressure, giving rise to a fundamentally different form of AI laboratory.
Due to a severe lack of high-end computing power, the Chinese team was forced to perform extreme engineering optimizations at the code level that American labs wouldn't even consider.
One extreme example is DeepSeek’s V3 model, which was trained on 2,048 H800 chips deliberately stripped of performance. To compensate for the deficient inter-chip connectivity, their engineers bypassed the mainstream CUDA framework entirely, writing code at the lower-level PTX layer and reallocating 20 (out of 132) streaming multiprocessors on each GPU exclusively for cross-node communication.
This is not because the U.S. lacks talented engineers, but because the incentive structures on both sides are completely different.
In U.S. labs, spending an extra dollar means buying more one-time computing power; in Chinese labs, hiring one more engineer means eliminating the need for computing power at the architectural level.
This extreme scarcity has fostered China’s unique full-stack efficiency culture—where they are pushing system performance to its absolute limit, from the kernel and optimizers to service systems and chips.
An intensely competitive arena, not only externally but also internally.
To Silicon Valley, Chinese AI seems merely to be competing with the United States. But after on-the-ground investigation, it becomes clear that internal competition within China’s cutting-edge AI sector is far more intense than the West imagines.
The five leading labs—DeepSeek, Alibaba (Qwen), Moonshot (Kimi), ByteDance (Doubao), and Zhipu (GLM)—are so closely matched in strength that their rankings nearly reset each quarter.
These five labs are not only fiercely competing on model capabilities and iteration speed, but are also aggressively open-sourcing their work.
This high-pressure internal “arena” environment is vastly different from the ecosystem dominated by only two closed-source giants in the United States (OpenAI and Anthropic), and it drives innovation no less than the transnational competition between China and the United States.
Emerging as a brand-new "AI data evaluation黑马"
U.S. data providers such as Mercor, AfterQuery, and Turing have documented business relationships with Ant Financial, Alibaba, and ByteDance. On the ground, they observe a new generation of Chinese elites building infrastructure for evaluation and validation, rather than data labeling services.
In Beijing and Shanghai specifically, they expected to see numerous low-cost "data labeling factories," but instead discovered a group of highly dynamic young startups focused on other aspects of AI.
For example, UniPat, an AI evaluation and prediction company founded by a Ph.D. graduate from Peking University, has seen its revenue surge to over $100 million in less than 24 months by compensating for its computational limitations with exceptionally high-quality human oversight.
Insufficient computing power? Compensate with human effort.
What surprised Silicon Valley investors the most was that during their visits, they encountered a group of top Chinese researchers implementing a very primitive but effective form of "Weak Self-Improvement (RSI)."
Due to a severe lack of computing power, researchers stepped in personally to accelerate model training through manual intervention and assistance.
Investors remarked that it was like the early days of "Centaur chess" (humans combined with machines defeating pure machines), manually bridging the computational gap with human effort and time.
Unburdened commercial monetization and pragmatism
There remains a gap of nearly two orders of magnitude in revenue scale between Chinese and U.S. AI companies. By August 2026, Anthropic’s ARR (annual recurring revenue) had surpassed $6.5 billion, while OpenAI reached $40 billion;
Meanwhile, China’s highest-revenue large model business (ByteDance’s video model) generates an annualized revenue of approximately $2–3 billion, while pure large model labs typically generate revenues in the hundreds of millions of dollars.
But the practicality and grassroots capabilities of Chinese companies are impressive.
As Dimension’s investors walked through the crowded, bustling streets of Hong Kong, listening to a popular Silicon Valley podcast analyzing why American labs proudly resisted advertising and e-commerce monetization, the reality before them in China presented a stark contrast.
DouBao by ByteDance has already reached 345 million monthly active users, surpassing the combined total of Qwen and DeepSeek.
Although its daily revenue currently falls short of 1 million RMB, nearly all of it comes from commission on e-commerce product promotions. Chinese entrepreneurs exhibit unapologetic pragmatism and couldn’t care less whether the monetization method is “lowbrow” or not.
The utterly fantastical "Pacific Data Circulation," a China-U.S. AI cycle
Although hardware is decoupled at the semiconductor level and below, the convergence of software and data layers is outpacing the ability of any government to respond. They believe the true technological cycle of current AI is in fact an extraordinarily magical “transoceanic technology circulation”:
Top models trained by U.S. frontier labs → Chinese teams distill and open-source the weights → U.S. vertical AI companies fine-tune the open-source Chinese models → Package and sell them to U.S. enterprises.
So, the so-called American vertical AI applications are increasingly running on Chinese open-source weights, which harbor the intelligence of America’s cutting-edge models.
For example, Cursor’s Composer 2, a popular coding tool, is built on top of Kimi’s K2.5; while Harvey, a rising legal AI startup, developed Harvey Tenet through fine-tuning Kimi’s open-source model K3.
Although China has not directly earned money from the West, it is aggressively capturing the West’s AI-era workflows. Open-source models bypass the strict corporate procurement barriers in Europe and the United States; when software is free and open, there is simply no reason to reject a vendor.
7. An even more疯狂的 valuation bubble
Western capital markets often talk about the AI bubble, but in China, this bubble is boiling far hotter than in Silicon Valley.
Given the significant revenue gap between Chinese and U.S. model companies, the valuation multiples of China’s top labs are often 5 to 10 times those of their U.S. counterparts. For example, Moonshot achieved a $3.5 billion funding round at a $35 billion valuation in late July—roughly 115 times its $300 million ARR—and is currently raising funds for a pre-IPO valuation of $50 billion (by comparison, Anthropic is valued at only 20 times its revenue).
The secondary market has become more frenzied and volatile, with the market capitalizations of companies such as Zhipu AI and MiniMax experiencing dramatic ups and downs before and after listing.
Finally, when this long letter was posted on overseas social media, a brilliant comment in the comments section captured the entire industry ecosystem under the U.S.-China AI competition:
Chips are decoupling, but intelligence can still flow.
In this hyper-competitive cycle, the so-called "technical decoupling" is entirely a false proposition.
As Dimension noted at the end of the letter, forcibly severing ties through administrative bans will likely only stifle the innovation speed of American companies themselves. The two countries have already become deeply intertwined at the levels of distillation, open source, data, and reasoning.
Instead of blindly cheering, truly understanding the real flow of chips on the table and the technical hand behind the scenes is what top capital and professionals should focus on.