Original Title: Confidential Letter to Dimension Limited Partners: The Cross-Pacific AI Chessboard
Original Author: Dimension Capital

Editor's Note: When the U.S. leads in advanced models and high-end chips while China has more abundant electricity and stronger cost constraints, how will the global AI competition unfold?

Dimension's internal memo attempts to answer this question. The author breaks down the competition into multiple interrelated levels, including computing power, energy, open-source ecosystem, application revenue, and capital market pricing.

The author's key insight is that export controls have not simply severed the ties between the U.S. and Chinese AI industries but have instead driven Chinese labs to emphasize engineering efficiency and open-weighted technical paths. The U.S. models, NVIDIA's computing power, Chinese open-source models, and cross-border data services are already embedded in the same industrial chain: technical capabilities can flow across borders, but income, costs, and bargaining power may not necessarily transfer synchronously. The future outcome may therefore depend not only on model performance but also on who can address their own scarcest elements and establish a more sustainable business loop.

This also explains why the article discusses both the U.S. power bottleneck and the high valuation of Chinese AI companies. The structural issues faced by the two regions are diametrically opposite: the U.S. has a more mature commercialization capability but is constrained by energy and infrastructure; China has rapid electricity expansion and an open ecosystem but still faces constraints such as advanced chip shortages, limited willingness to pay, and small revenue scales. As a result, there is a clear mismatch between technical capabilities, business income, and market valuations.

It should be noted that this letter carries a distinct investment perspective, and some of the income, model sources, and industry penetration data still need further validation. However, it provides a noteworthy analytical framework: the U.S.-China AI competition is no longer easily explained by mere technical leadership or policy blockade; what truly determines the industry landscape may be the constantly changing balance among computing power, energy, open-source diffusion, and capital allocation.

The following is the original text:

Dear Partners,

A small Dimension team visited China last week, visiting multiple leading labs and engaging with local investors, entrepreneurs, and technical experts. This trip was largely a revisit and a reassessment of our judgment formed after our 2025 Shanghai trip. At that time, we wrote and discussed that China's rising importance has surpassed the hardware domain. Now, as we summarize the gains from this trip, we believe it is necessary to take a fresh look at this trans-Pacific AI chess game. Below is a sharing of some of our current thoughts. It is important to note that these views are still evolving, and with new data emerging, they may change as soon as tomorrow.

1. Scarcity Will Drive Evolution

The original intention of export controls was to slow down the development of AI in China. However, the result was a unique evolutionary pressure that led to the birth of a distinctive laboratory in China.

With limited computing power, Chinese teams often focus on engineering optimizations at a more foundational level of abstraction, while American labs tend to overlook these aspects. When DeepSeek released V3 in March, they had previously conducted engineering at the PTX level of CUDA programming on an NVIDIA H800 GPU, shifting the most energy-intensive normalization operations from the GPU compute cores to 20 cores dedicated to serial interconnects.¹

We believe that this reflects not so much a shortage of talent in the U.S., but rather a difference in incentive mechanisms: the marginal benefit in American labs comes from adding more computing power, while the marginal benefit in Chinese labs comes from finding optimization space for compilers and chips.

Ultimately, Chinese labs have formed a full-stack efficiency culture that spans the kernel, optimizer, service system, and even chip design. The compounded effects of this culture far exceed the impact of mere point optimizations. Scarcity—of all things—has unexpectedly become the mother of invention; here, it has nurtured systems, assembly, and compiler engineering.

2. The Current State of Chinese Software Has Surpassed What Western Enterprise Software Could Achieve Two Years Ago: Infiltrating the Western Production Technology Stack

Chinese vendors have transitioned from zero penetration not even 18 months ago to now supporting over 25% of the OpenRouter token supply.²

Qwen has surpassed Llama in blind benchmarking and continues to release new models frequently. In less than two years, Chinese models have evolved from being the "fifth-best open-source model globally" to "at least 80% of AI startups in the U.S. use at least one Chinese open-source model in production environments."³ Open weights bypass procurement barriers—free software has no vendor to be vetted.

In essence, China is rapidly acquiring Western workloads, not Western revenue. As before, inference profits temporarily flow to the service providers actually delivering the tokens.

Whether intentional or the emergent outcome of a complex system at a global scale, cutting-edge Chinese labs are commodifying the revenue layers that U.S. competitors rely on, primarily occurring on the open-source end. This, in our view, is not coincidental, nor is it merely an eccentric experiment.

These companies aim to build large-scale, stable, low-revenue operational systems. Therefore, our mid-term assessment is that China is driving the formation of an ecosystem reliant on open weight models and rapid follow-ups, squeezing the revenue of leading U.S. labs; simultaneously, the revenue is directed towards massive cloud service providers and new inference infrastructures. We believe this trend is slowing down, even hindering the entry of truly advanced technologies into the current playing field.

3. The Competition Inside China's Cutting-Edge Labs Is Far More Intense Than the West Typically Realizes

Our assessment is that the current cutting-edge competition includes DeepSeek, Alibaba's Qwen, Dark Side of the Moon's Kimi, ByteDance's Bean, and Intelpath's GLM, with the difference between them fluctuating in ranking by approximately one quarter. It is worth noting that although Tencent has a huge existing user base, it operates outside of the ecosystem; this in itself is a problem worth paying attention to.

These labs compete head-on in research and development speed and model capability while also open-sourcing the model weights, forming a competitive ecosystem that is starkly different from the two closed labs in the United States.

Once again, we believe that the focus of competition in the next one to two years, rather than being between nations, is more of a contest between different incentive mechanisms. The Chinese labs we have interacted with often focus more on their competition with each other rather than the cross-border competition hyped by Western media.

4. Today, Model Capability Can Make Two Round Trips Across the Pacific and Still Reach End Users

The specific path is as follows: a cutting-edge lab in the United States releases a higher-performing model; a Chinese lab distills it and releases the model weights; U.S. application companies then use these open-source Chinese models to build products and sell them back to U.S. enterprise customers.

Cursor's internally developed Composer 2 is said to be built on top of Kimi K2.5; Cognition's SWE-1.5 seems to be built on a customized version of GLM. In short, as articulated by a representative of a Chinese cutting-edge lab we spoke to, the U.S. has a faster iteration speed while China has a faster release speed.

About a year ago, these companies would have done well to reach the level of Llama or even Mistral. Today, cutting-edge AI from the United States has integrated into China's open weight system, with cutting-edge intelligence from the U.S. also hidden within.

5. The Data Layer Seems to Be Following a Similar Trajectory

American data providers—Mercor, AfterQuery, Turing—have established commercial partnerships with companies like Anthropic and OpenAI. Meanwhile, in China, a group of startups is beginning to build evaluation and verification infrastructure to support annotation services.

For example, UniPat, founded by a Peking University doctoral student and backed by Mercedes, opened an office in Seattle in December 2025, reportedly using doctoral annotators from China working on these projects through a direct route from the American Midwest.

A Chinese lab posed a simple yet important question: **how can they ensure the quality of human oversight of the computational process, in an environment where computations are constrained not to exceed a certain threshold, while remaining consistent with a model that has adaptive capabilities and near-infinite resources?**⁴

We have observed that several teams in the ecosystem have started building corresponding products, with at least two of them rapidly approaching $100 million in revenue.

6. The U.S. and China Face Different, Opposing Constraints

In 2024, China is adding approximately 429GW of new generating capacity, while the U.S. is only adding about 51GW. By 2030, China's existing under-construction power projects may bring about an additional supply of over 400GW. In contrast, in the U.S., data center expansion is expected to account for more than one-fifth of the future incremental electricity demand.⁵

As we have written before, China's constraint stems from a lack of computational infrastructure. For example, the K3 new model on the dark side of the moon, although cutting-edge in capability, faced delays in deployment due to inadequate inference capacity.

The U.S. is rich in chip resources but increasingly constrained by power, while China is rich in power resources but increasingly constrained by chips.

Our assessment, which is also widely agreed upon in the industry, is that the future will be determined by the element that is currently in the shortest supply: the U.S. depends on how quickly data center space, power distribution, and transmission lines can be built; while China depends on the yield and packaging capabilities of SMIC and Huawei.

7. The Disparity in AI Revenue between the U.S. and China is Huge, Approaching Two Orders of Magnitude

As of the end of July, Anthropic's annualized revenue has exceeded $6.5 billion; as of August, OpenAI's annualized revenue reached $40 billion.⁶

In comparison, the highest-earning AI model in China is operated by ByteDance, with annualized revenue of approximately $200-300 million for its video model; the second-highest earner is the dark side of the moon, with revenue of around $1 billion. DeepSeek reportedly has an annualized revenue of less than $50 million.

On the consumer end, DòuBāo has over 345 million monthly users, more than three times that of DeepSeek. Both offer various sources of revenue, including commercial commissions, and based on our calculations, the Renminbi's average monthly realized revenue is around 1 yuan.

Historically, Chinese consumers are not accustomed to paying for software. We also do not believe that the Chinese economy has yet reached a point where the middle class is willing to rapidly change this habit. Insiders at companies have stated that they are deploying resources to commercialize and try to change this trend; there has indeed been a change in Chinese consumer internet habits ten years ago.

Of course, the Chinese lab may eventually monetize through commercial advertising, as American labs have been reluctant to do.

There is also a shameless pragmatism at play here: entrepreneurs are often unconcerned about using the wrong term or adopting a stigmatized expression, fearing censorship or backlash. This was evident as we walked the streets of Kowloon in Hong Kong: vendors listened to Ben Thompson's analysis of the American lab's irrational aversion to advertising and business models.⁷

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

8. Hardware is decoupling from the semiconductor layer, software layer, and data layer at a speed intertwined to the point where any government cannot keep up

American applications run on top of the Chinese model; Chinese labs, on the other hand, are trained on NVIDIA compute rented in Malaysia and Thailand, while distilling the output of the American model; expert data also flows in both directions.⁸

We find that the question of "who is writing the problem" is no longer reliable.

Likewise, systems are now deeply interdependent, embodying not only the "rhetorical nationalism" but also realizing a true transnationalism in practice—unless Recursive Self-Improvement (RSI) fundamentally alters the current game structure.

9. We discover that some top Chinese researchers believe they have achieved a weaker form of RSI

Generally speaking, RSI can be divided into three types: time-based, compute-based, and boundary RSI. The earliest form of RSI can be understood as researchers involving the model in its own iteration, manually assisting the model to expedite parts that would otherwise require the researcher's compute or time, or both. It is crucial to note that the "take-off" in this stage may not manifest as an exponential curve over timescales of decades to 2100, nor as a combined human effort surpassing the capacity of any individual or machine.

10. A Vital Forking Point

If indeed first-order self-improvement exists, then in terms of compute, America's advantage seems quite robust: by 2030, each lab operates around 2GW of compute and has over 5GW in contracted projects.

In theory, these projects are supported by substantial revenue and a strong capital market; the path to scaling computation, lateral thinking, the "self-improvement" loop, and learning how to effectively wield these capabilities seem endless.

However, looking at it the other way around, if the real constraint is on the validation quality, or if the timeline exceeds the curve's intended range, then the notion of "capacity optimization and domestic chip redundancy" — which may sound somewhat familiar — may continue to accelerate over the next 18 months:

The U.S. front-modeling capability accelerates commercialization;

China opens up unique model capabilities used for weighted models;

Bidding AI companies for inference layers gain more bargaining power;

Compute service providers gain greater volume and profit as foreign IP weights are included;

Ultimately, these effects will also cascade to inference service providers, such as hyperscale cloud providers and emerging cloud manufacturers.

This implies a significant mismatch in power infrastructure investment.

What we need to stress is: our analysis of U.S. and China infrastructure is perhaps the highest-potential return analysis we have ever done.

About the Future

11. A Future Scenario

In the future, U.S. front-end labs may no longer open APIs to next-generation cutting-edge models. In fact, for the vast majority of programmatic enterprise applications, the current technological level is already sufficient. High-end models of the future will be priced based on value rather than quantity, while also increasingly benefiting China from front-end capabilities.

As a result, U.S. labs may prioritize API revenue to support higher-priced, lower-volume models; this, in turn, will sever an important competitive channel crossing from the U.S. East Coast to the Pacific.

At the very least, this will compel more data to be processed domestically in China using pre-trained models, adding pressure to existing compute bottlenecks.

From what we understand, some sizable U.S. multinational corporations are piloting the relocation of engineering and math tasks to China.

12. The AI Landscape in China and the U.S.

We believe that AI has deeply intertwined the technological systems of the United States and China to the point where regulatory measures only severing one trans-Pacific link seem to have, at most, a temporary effect.

Technologists from the U.S. and China are, in a sense, in dialogue with each other; this exchange may occur at multiple levels such as model distillation, open weights, data labeling, inference, and software increasingly leveraging intelligent agent-based search infrastructure.

The practice of directly banning currency circulation, entity listing, custody prohibition, and procurement ban almost inevitably suppresses U.S. innovation, relinquishes the leadership position in relevant fields, and ultimately drives funds to Chinese competitors.

Despite the Western discussion about the AI valuation bubble, the atmosphere in the Chinese market is noticeably more enthusiastic.

Looking at the revenue multiple, the transaction valuation of Chinese labs is 5 to 10 times that of U.S. frontier labs; in terms of revenue scale, the gap between China and the U.S. is close to two orders of magnitude.

The most recent financing round of Lunar DarkSide was valued at $3.5 billion, while its July annualized revenue was approximately $100 million, resulting in a valuation of about 35 times revenue; the upcoming financing round is reportedly set at a pre-money valuation of $5 billion, equivalent to about 50 times revenue.

Intelligence Spectrum, when listed in Hong Kong in January this year, had a market capitalization of approximately $10 billion, roughly equivalent to a multiple of 20 times Anthropic's valuation multiple. The company's early trading on the open market was volatile: starting from a market capitalization of around $1.3 billion at the time of listing, it once rose to about $10 billion in January of this year.

Kong AI Knowledge Atlas Technology HK (2613) also saw its market capitalization rise from around $1.3 billion in January of this year to about $10 billion; Luna AI (HKG:100) similarly surged from around $15 billion all the way up to $33 billion, then fell back to about $10 billion.

Both of these stocks experienced a sharp decline before the end of the lock-up period. Our assessment is that this is mainly due to the immature market structure: decades of experience in consumer software investment, only now gradually accumulating the ability to prudently evaluate around enterprise financial disclosure and open-source frameworks.

This situation is not significantly different from what we have seen in the biotechnology sector: in the Hong Kong public market, there is almost no regulatory risk arising from cross investors or professional tech and biotech investors.

For the right investors, this presents a tremendous opportunity: they can enter the market and establish a leading position in early-stage public market investments in these categories.

Your Partners,

Nan, Adam, Zivain

Footnotes

DeepSeek V3 Technical Report: Utilizing 2048 H800 GPUs; PTX-level programming; about 30 million messages dedicated to inter-node communication.

In March 2026, OpenRouter's weekly token share was around 42%, with a fluctuation of about 5%; Qwen accounted for approximately 18%. It is worth noting that Hugging Face's model downloads only represent about 80% of China's open-weight model usage, with the remaining 20% distributed through other channels.

Anthropic had previously accused the Dark Side of the Moon, DeepSeek, and MiniMax of collecting millions of Claude dialogue data points; however, this claim was later denied and refuted by several companies.

Regarding comments on the accounting treatment differences between various throughput methods, this article will temporarily set that discussion aside.

Alibaba and ByteDance accelerated their data center training in Southeast Asia following restrictions imposed on April 20, 2026; these companies leased computing power in facilities outside of China in Thailand, Malaysia, and Japan. The OSTP directive from the U.S. White House in March 2026 introduced NVIDIA's export-grade GB300 servers to Thailand for K3 training; current export controls prioritize chip ownership over remote access. The leasing costs for computing power in Thailand, Malaysia, and Indonesia are about 30% higher, while operating costs are approximately 10% lower.

Prior to the end of the lock-up period, institutions holding more than a 5% stake cannot cash out; regular shareholders are also unable to sell before the lock-up period expires.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia