After a U.S. VC's China Trip: Dimension's AI Observations for LPs

十字路口Crossing·October 24, 2026·0·1

One AI Brain, Two Blood Supplies

One AI Brain, Two Circulatory Systems

👦🏻 Author: Dimension Capital

🥷 Translated by: Simon

Crossing Editors' Note (Koji)

We've been tracking how the United States views China. For me, outside perspectives can be uniquely valuable: the things we take for granted may, in others' eyes, be precisely the changes most worth interrogating.

"You cannot see the true face of Mount Lu, for you are standing in the midst of it." Being inside, we know the details intimately, yet easily lose sight of the full picture. Seeing ourselves through different eyes often reveals things obscured by daily experience.

Previously, we translated An American Engineer's 11 Days in China. This time, I'd like to recommend another American investor's observations on China.

This is an internal letter Dimension Capital sent to its LPs in August. After visiting AI labs in China and speaking with investors, founders, and technical experts, the team restructured its thinking. The letter later circulated publicly online.

Re-examining the Trans-Pacific AI Chessboard

Dear LPs:

A small team from Dimension visited China last week, calling on multiple leading large-model AI labs and meeting with local investors, founders, and technical experts. The trip was largely a return visit, a reassessment of the conclusions we formed after our 2025 Shanghai trip.

At that time, we wrote and discussed how China's rising importance had already extended beyond hardware. Now, in sorting through what we learned this trip, we believe it's necessary to re-examine this trans-Pacific AI chessboard.

Below we share some of our current thinking. A caveat: these views remain in flux; they may shift tomorrow as new data emerges.

Scarcity Drives Evolution

Export controls were designed to slow Chinese AI development. Instead, they created a unique evolutionary pressure that produced distinctly Chinese AI labs.

With limited compute, Chinese teams tend to focus on engineering optimization at deeper levels of abstraction — work that American AI labs easily overlook. When DeepSeek released V3 in March, it had already pushed optimization down to the PTX layer of CUDA programming on NVIDIA H800 GPUs, moving the most power-hungry normalization operations from GPU compute cores to twenty specialized cores handling serial interconnects.¹

We see this less as a shortage of talent in the US than as a difference in incentives: American AI labs gain marginal returns from adding compute; Chinese AI labs gain marginal returns from finding optimization space in compilers and chips.

Ultimately, Chinese AI labs developed a full-stack efficiency culture spanning kernels, optimizers, serving systems, and even chip design. The compounding effects of this culture far exceed what any single-point optimization could achieve.

Scarcity became the mother of invention; here, it gave birth to systems, assembly, and compiler engineering.

Chinese Models Enter Western Production Stacks

Chinese software has already advanced beyond where Western enterprise software was two years ago: it is entering Western production stacks.

Less than 18 months ago, Chinese vendors had zero penetration; now they support over 25% of token supply on OpenRouter.²

Qwen surpassed Llama in blind-test benchmarks and continues to release new models frequently. In under two years, Chinese models went from "the fifth-best open-source model globally" to "at least 80% of US AI startups use at least one Chinese open-source model in production."³

Open weights bypass procurement barriers — free software has no vendor to vet.

Essentially, China is rapidly capturing Western workloads without capturing Western revenue. As before, inference profits temporarily flow to the service providers actually delivering tokens.

Whether by design or as an emergent property of complex global systems, China's frontier AI labs are commoditizing the business layer that American competitors rely on for revenue, primarily on the open-source side. In our view, this is neither coincidental nor merely some eccentric experiment.

These companies aim to build massive, stable, lower-revenue operations. Thus our medium-term view: China is pushing an ecosystem built on open-weight models and fast follow-on, squeezing revenue for top American AI labs; meanwhile, revenue flows to hyperscale cloud providers and emerging inference infrastructure.

We believe this trend is slowing, even blocking genuinely advanced technology from entering the current competitive arena.

Competition Among Chinese AI Labs

Competition within China's frontier AI labs is far more intense than Western observers typically recognize.

Our assessment: the current frontier includes DeepSeek, Alibaba's Qwen, Moonshot AI's Kimi, ByteDance's Doubao, and Zhipu AI's GLM, with their relative rankings shifting on roughly quarterly timescales. Notably, Tencent, despite its massive existing user base, sits outside this ecosystem; that itself is worth watching.

These AI labs compete head-on in R&D speed and model capability while simultaneously open-sourcing model weights, forming a competitive ecology quite different from America's two closed-source AI labs.

We reiterate: competition over the next one to two years will occur less between nations than between different incentive structures. The Chinese AI labs we engaged with were often more concerned with competing against each other than with the cross-border rivalry portrayed in Western media.

Model Capabilities Cross the Pacific

Today, model capabilities can cross the Pacific twice and still reach end users.

The specific path: an American frontier AI lab releases a stronger model; a Chinese AI lab distills it and releases the model weights; an American application company then builds products using these Chinese open-source models, selling to American enterprise customers.

Reportedly, Cursor's internally developed Composer 2 is built on Kimi K2.5; Cognition's SWE-1.5 appears built on a customized version of GLM. In short, as one representative from a Chinese frontier AI lab told us, America iterates faster, but China releases faster.

About a year ago, these companies would have been doing well to reach Llama's level, even Mistral's. Now, American frontier AI has been absorbed into China's open-weight ecosystem, with frontier intelligence from America embedded within it.

The Data Layer Follows a Similar Path

American data providers Mercor, AfterQuery, and Turing have established commercial partnerships with Anthropic, OpenAI, and others. Meanwhile, in China, a crop of startups is beginning to build evaluation and verification infrastructure to support data labeling services.

For example, UniPat — founded by a Peking University PhD student and backed by Mercedes — established a Seattle office in December 2025. It reportedly uses PhD-level annotators from China, participating in these projects through a direct channel from the American Midwest.

One Chinese AI lab posed a simple but important question:

In an environment where compute is capped below a certain threshold, how do you ensure quality human oversight of the computational process while maintaining alignment with an adaptive, near-unlimited-resource model? ⁴

We observed several teams in the ecosystem already building corresponding products, with at least two seeing revenue rapidly approaching $100 million.

America Lacks Power, China Lacks Chips

The constraints facing China and America differ, even opposing each other.

In 2024, China added approximately 429GW of new power generation capacity; the United States added only about 51GW. By 2030, China's currently under-construction power projects may bring over 400GW of additional supply. By contrast, in the US, data center expansion is projected to account for over one-fifth of future new power demand.⁵

As we've written previously, China's constraint comes from insufficient compute infrastructure. For example, Moonshot AI's new model K3, though frontier-capable, had its deployment delayed due to insufficient inference capacity.

America is chip-rich and increasingly power-constrained; China is power-rich and increasingly chip-constrained.

Our view — widely shared in the industry — is that the future will be determined by whichever resource is currently scarcest: for the US, how quickly data center space, power distribution facilities, and transmission lines can be built; for China, the yield and packaging capabilities of SMIC and Huawei.

Revenue Gap Approaching Two Orders of Magnitude

The AI revenue gap between China and America is enormous, approaching two orders of magnitude.

As of end-July, Anthropic's annualized revenue exceeded $6.5 billion; as of August, OpenAI's annualized revenue reached $4 billion.⁶

By comparison, China's highest-revenue AI model is operated by ByteDance, with its video model generating roughly $200–300 million annualized; second is Moonshot AI, with approximately $100 million in revenue. DeepSeek reportedly generates less than $50 million annualized.

On the consumer side, Doubao has over 345 million monthly active users, more than triple DeepSeek's. Both have multiple revenue streams including commercial commissions; by our calculation, average monthly monetization is roughly RMB 1 per user.

Historically, Chinese consumers have not been accustomed to paying for software. Nor do we believe China's economy has developed to the point where the middle class would rapidly change this habit. Company insiders say they are investing resources to advance commercialization and trying to shift this trend; Chinese consumer internet users' habits did change about a decade ago.

Of course, Chinese AI labs may eventually monetize through commercial advertising, something American AI labs have been reluctant to do.

Here too there's an unapologetic pragmatism: founders often don't mind using wrong terminology or adopting stigmatized expressions, also touching on concerns about censorship or backlash. We saw this walking the streets of Kowloon, Hong Kong: vendors listening to Ben Thompson analyze American AI labs' irrational aversion to advertising and business models.⁷

By the way, this was the first Ben heard of it — Ben is one of the finest analysts of all time.

Decoupling and Entanglement Simultaneously

Hardware is decoupling at the semiconductor layer; software and data layers are entangling at a pace too fast for any government to track.

American applications run on Chinese models; Chinese AI labs train using NVIDIA compute rented in Malaysia and Thailand while distilling American model outputs; expert data also flows bidirectionally.⁸

We find that "who is writing the questions" is no longer a reliable guide.

Similarly, the systems are now so deeply interdependent that they demonstrate not merely "performative nationalism" but genuine cross-border collaboration in practice — unless RSI (recursive self-improvement) fundamentally alters the current game structure.

Weaker Forms of Self-Improvement

We found that some top Chinese researchers believe they have achieved some weaker form of RSI.

Generally, RSI can be divided into three types: time-based, compute-based, and boundary-type. The earliest form can be understood as researchers having models participate in their own iteration, with human assistance, accelerating work that would otherwise consume researcher compute, time, or both.

It's worth emphasizing that "takeoff" at this stage need not appear as an exponential curve extending to 2100 across decades, nor as a human collective intelligence surpassing any individual or machine capability.

Whether AI Can Accelerate Its Own Progress Will Determine Who Has Advantage

A critical question in the coming competition: Can AI help researchers develop the next generation of AI faster?

For example, models can participate in research, write code, analyze experimental results. If this work can sustainably accelerate model development, a loop may form: stronger AI helps develop stronger AI.

Whether this loop functions effectively will send competition down two different paths.

Scenario One: AI can sustainably accelerate R&D, and more compute yields clear progress.

In this case, America's advantage may be greater. American AI labs have stronger revenue generation and easier access to capital markets, allowing continued construction of large compute facilities.

We estimate that by 2030, each American frontier AI lab will operate compute facilities of roughly 2GW, with over 5GW of contracted projects. GW here refers to power scale, describing how large these facilities are, not directly measuring how smart the models are.

If more compute can support faster R&D, America has the opportunity to convert its capital and infrastructure advantages into further model capability advantages.

Scenario Two: AI self-improvement is slower than expected, and more compute cannot solve the most critical problems.

For example, AI may propose many new approaches, but researchers still struggle to determine which actually work. Here, the limiting factor may be verification capability, not compute.

In this case, improving efficiency of existing compute and expanding domestic chip supply may become more important over the next 18 months.

China and America may continue a division of labor: American AI labs push frontier models and their commercialization; Chinese teams continue providing open-weight models; application companies use these models to develop products.

As model choice increases, enterprises using models may gain greater bargaining power. Companies running models and providing inference services to customers may also gain more business and profit, including large cloud providers and emerging cloud vendors.

These two scenarios demand different infrastructure. If investors misjudge the development path, the compute facilities and supporting power infrastructure they build may not match actual demand.

Therefore, analyzing what each of China and America lacks, and which facilities truly need expansion, may yield highly valuable investment opportunities. Our analysis of China-US infrastructure may be the highest-potential-return analysis we've ever done.

On the Future: One Possible Scenario

In the future, American frontier AI labs may no longer open APIs for next-generation frontier models. Indeed, for the vast majority of programmatically accessed enterprise applications, current technology levels are already sufficient.

Future premium models will be priced by value, not by usage volume, while China will increasingly benefit from frontier capabilities.

Thus, American AI labs may prioritize API revenue to support higher-priced, lower-volume models; this in turn would sever a major competitive channel stretching from America's East Coast across the Pacific.

At minimum, this would force more data processing using pre-trained models within China's borders, further intensifying pressure on existing compute bottlenecks.

We understand that some larger American multinational enterprises are piloting the transfer of engineering and mathematical tasks to China.

The China-US AI Landscape and Capital Markets

We believe AI has so deeply entangled the technical systems of China and America that regulatory measures severing any single trans-Pacific link seem capable of producing, at most, temporary effects.

In a sense, Chinese and American technologists are in dialogue with each other. This exchange may occur at multiple levels: model distillation, open weights, data labeling, inference, and increasingly software using agentic search infrastructure.

Direct bans on capital flows, entity listings, hosting prohibitions, and procurement bans almost inevitably suppress American innovation, cede leadership in relevant domains, and ultimately drive capital toward Chinese competitors.

While the West discusses an AI valuation bubble, sentiment in China's market is noticeably more exuberant.

By revenue multiples, Chinese AI labs trade at 5–10x the valuation multiples of American frontier AI labs; yet in revenue scale, the China-US gap approaches two orders of magnitude.

Moonshot AI's most recent funding round valued the company at $3.5 billion, with July annualized revenue of roughly $100 million, or about 35x revenue; an upcoming round is reportedly set at a $5 billion pre-money valuation, roughly 50x revenue.

Zhipu AI (here the original text writes "Intelligence Spectrum") listed in Hong Kong this January at roughly a $10 billion market cap, corresponding to roughly 20x Anthropic's valuation multiple. The company saw significant early public-market volatility: starting from roughly $1.3 billion at listing, it rose to roughly $10 billion this January.

The original text also states that Kong AI Knowledge Atlas Technology HK (2613) rose from roughly $1.3 billion this January to roughly $10 billion; Luna AI (HKG:100) similarly climbed from roughly $15 billion to $33 billion before falling back to roughly $10 billion.

Both stocks fell sharply before their lock-up periods ended. Our assessment: this stems mainly from immature market structure — decades of consumer software investment experience, yet only now gradually building capacity for prudent evaluation around enterprise financial disclosure and open-source frameworks.

This differs little from what we see in biotech: in Hong Kong's public markets, there is virtually no regulatory risk from cross-domain investors or specialized tech and biotech investors.

For the right investors, this is a massive opportunity: they can enter the market and establish early leadership in public-market investments in these categories.

Your partners,

Nan, Adam, Zivain

Original Footnotes

DeepSeek V3 Technical Report: using 2,048 H800 GPUs; programming at PTX layer; approximately 30 million messages dedicated to inter-node communication.
March 2026, OpenRouter's weekly token share roughly 42%, fluctuating roughly 5%; Qwen roughly 18%. Notably, Hugging Face model downloads represent only roughly 80% of Chinese open-weight model usage, with the remaining 20% distributed through other channels.
Anthropic previously accused Moonshot AI, DeepSeek, and MiniMax of collecting millions of Claude conversation data points; however, this claim was later denied and refuted by multiple companies.
Differences in accounting treatment for different throughput methods are not discussed in this article.
After April 20, 2026 restrictions, Alibaba and ByteDance accelerated training activities in Southeast Asian data centers; these companies rented compute in facilities outside China located in Thailand, Malaysia, and Japan.
Before lock-up periods end, institutions holding over 5% cannot cash out; ordinary shareholders similarly cannot sell before lock-up periods end.

This article's author: Dimension Capital (signed at letter's end: Nan, Adam, Zivain)

Original: Dimension Research on China's AI: Open Models, Compute Bottleneck, and Commercialization Gap (Full Article)

References

[1] An American Engineer's 11 Days in China: https://mp.weixin.qq.com/s/h23dzzfXFEH4hZ2E1ZbTgw

[2] Dimension Research on China's AI: Open Models, Compute Bottleneck, and Commercialization Gap (Full Article): https://en.theblockbeats.news/news/63565