Dimension Investigates China's AI: Open Models, Compute Bottlenecks, and the Commercialization Gap

China lacks chips, the US lacks power—US-China AI enters a new chess game

Original title: Confidential Letter to Dimension Limited Partners: Trans-Pacific AI Chess
Original author: Dimension Capital

Editor's note: When the US holds leading models and high-end chips, while China has more abundant electricity and stronger cost constraints, how will global AI competition unfold?

Dimension's internal letter attempts to answer this question. The author breaks down the competition into multiple interconnected layers: compute, energy, open-source ecosystems, application revenue, and capital market pricing.

The author's key judgment is that export controls have not simply severed US-China AI industry ties; instead, they have pushed Chinese labs to form a technical route that emphasizes engineering efficiency and open weights. US models, Nvidia compute, Chinese open-source models, and cross-border data services are already embedded in the same industrial chain: technical capabilities can flow across borders, but revenue, costs, and bargaining power do not necessarily transfer in sync. Future victory may therefore depend not only on model performance, but also on who can solve their own scarcest factor and build a more sustainable commercial closed loop.

This also explains why the article simultaneously discusses the US power bottleneck and the high valuations of Chinese AI companies. The structural problems faced by the two regions are exactly opposite: the US has more mature commercialization capabilities but is constrained by energy and infrastructure; China has rapidly expanding power supply and an open ecosystem, but still faces constraints such as advanced chip shortages, limited willingness to pay, and small revenue scale. Technical capability, commercial revenue, and market valuation are thus clearly misaligned.

It should be noted that this letter carries a distinct investment institution perspective, and some of the revenue, model source, and industry penetration data still require further verification. But it provides an analytical framework worth attention: US-China AI competition can no longer be explained by simple technological leadership or policy blockade; what truly determines the industrial landscape may be the ever-shifting balance among compute, energy, open-source diffusion, and capital allocation.

The following is the original text:

Dear Partners:

A small team from Dimension traveled to China last week, visiting several leading labs and exchanging views with local investors, entrepreneurs, and technical experts. This trip was largely a revisit, and a re-examination of the judgments we formed after our 2025 Shanghai trip. At that time, we wrote and spoke about China's growing importance extending beyond hardware. Now, as we summarize the gains from this trip, we believe it is necessary to re-examine this trans-Pacific AI chess game. Below are some of our current thoughts. It should be noted that these views are still evolving and may change as new data emerges, perhaps even tomorrow.

1. Scarcity breeds evolution

The original intent of export controls was to slow China's AI development. But the result is that they created evolutionary pressure, giving rise to distinctive Chinese labs.

With limited compute, Chinese teams often perform engineering optimization at lower abstraction levels that US labs can often ignore. When DeepSeek released V3 in March, it used PTX-level CUDA programming on an Nvidia H800 GPU to move the most energy-intensive normalization operations from GPU compute cores to 20 cores dedicated to serial interconnect.¹

We believe this reflects less a shortage of US talent than a difference in incentives: US labs' marginal returns come from adding more compute, while Chinese labs' marginal returns come from finding optimization space for compilers and chips.

Ultimately, Chinese labs have formed a full-stack efficiency culture covering kernels, optimizers, serving systems, and even chip design. The compounding effects of this culture far exceed the impact of individual optimizations. Scarcity—who would have thought?—turned out to be the mother of invention; here, it has nurtured systems, assembly, and compiler engineering.

2. The current state of Chinese software has already surpassed what Western enterprise software could achieve two years ago: penetrating the Western production tech stack

Chinese vendors have gone from zero penetration less than 18 months ago to now supporting over 25% of OpenRouter token volume.²

Qwen has surpassed Llama in blind arena tests and continues to release new models frequently. In less than two years, Chinese models have evolved from "the world's fifth-best open-source model" to "at least 80% of US AI startups using at least one Chinese open-source model in production."³ Open weights bypass procurement barriers—free software has no vendor to vet.

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

Whether intentional or an emergent result of a complex system at global scale, Chinese frontier labs are commoditizing the revenue layer that US competitors depend on, primarily on the open-source side. This does not appear to us to be accidental, nor merely a curious experiment.

These companies' goal is to build large-scale, stable, low-revenue operating systems. Therefore, our medium-term judgment is: China is driving the formation of an ecosystem that relies on open-weight models and rapid follow-up, squeezing the revenue of US leading labs; at the same time, it is directing revenue toward hyperscale cloud providers and new inference infrastructure. We believe this trend is slowing, even hindering, the entry of substantively advanced technology into the current chess game.

3. Competition within Chinese frontier labs is far more intense than the West generally realizes

Our judgment is that the current competitive frontier includes DeepSeek, Alibaba's Qwen, Moonshot AI's Kimi, ByteDance's Doubao, and Zhipu's GLM, with gaps of only about one quarter in ranking fluctuations. Notably, Tencent, despite its massive existing user base, is outside this ecosystem; that in itself is worth attention.

These labs compete head-to-head on R&D speed and model capability while simultaneously opening their model weights, forming a competitive ecosystem completely different from the two closed US labs.

Again, we believe the competitive focus over the next one to two years is less about nation-vs-nation competition than about competition between different incentive structures. The Chinese labs we met tend to focus more on competition with each other than on the cross-border competition that Western media loves to hype.

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

The specific path: US frontier labs release more powerful models; Chinese labs distill them and release model weights; US application companies then use these open-source Chinese models to build products and sell them back to US enterprise customers.

Cursor's self-developed Composer 2 is reportedly built on Kimi K2.5; Cognition's SWE-1.5 appears to be built on a customized GLM. In short, as we heard one Chinese frontier lab representative say, the US iterates fast, China releases fast.

About a year ago, these companies would have been considered good if they reached Llama or even Mistral levels. Today, US frontier AI has been integrated into China's open-weight system, and US frontier intelligence lurks within it.

5. The data layer seems to be repeating a similar trajectory

US data vendors—Mercor, AfterQuery, Turing—have established commercial relationships with Anthropic, OpenAI, and others. Meanwhile, in China, a group of startups has begun building evaluation and verification infrastructure to support annotation services.

For example, UniPat was founded by a Peking University PhD student and is backed by Mercedes-Benz. The company opened an office in Seattle in December 2025 and reportedly uses PhD annotators from China, working on these projects via direct US Midwest transit routes.

Chinese labs raise a simple but important question: **How can they ensure the quality of human oversight over computational processes, in environments where computation is constrained and must not exceed a certain threshold, remains aligned with models that have adaptive capabilities and near-unlimited resources?**⁴

We have observed that several teams in the ecosystem have begun building corresponding products; at least two of them are rapidly approaching revenue above $100 million.

6. The US and China face constraints that are different in nature and opposite in direction

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

As we have written before, China's constraint comes from an insufficient compute base. One example is that Moonshot AI's new K3 model, while at the frontier in capability, had its release delayed due to insufficient inference capacity.

The US has abundant chips but compute increasingly constrained by power; China has abundant power but compute increasingly constrained by chips.

Our judgment—and this is the industry consensus—is that the future will be determined by the currently scarcest factor: for the US, it depends on how quickly data center land, transmission and distribution, and power lines can be built; for China, it depends on SMIC and Huawei's yields and packaging capabilities.

7. The gap between US and Chinese AI revenue is very large, close to two orders of magnitude

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

In contrast, the highest model revenue in China is ByteDance, whose video model annualized revenue is about $200-300 million; the second highest is Moonshot AI, at about $100 million. DeepSeek reportedly has annualized revenue below $50 million.

On the consumer side, Doubao has 345 million monthly users, more than three times DeepSeek's. Both offer multiple revenue streams including commercial commissions; by our calculations, monthly monetization revenue per user is about RMB 1.

Historically, Chinese consumers are not accustomed to paying for software. We also do not think Chinese economics has become neutral enough for the middle class to rapidly change this habit. Company insiders say they are deploying resources to attempt commercialization; Chinese consumers' internet usage habits did change a decade ago.

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

There is also an unabashed pragmatism factor here: entrepreneurs generally do not worry about being censored or attacked for using the wrong words or expressions considered stigmatized. This was very evident as we walked the streets of Kowloon, Hong Kong: vendors were listening to Ben Thompson's analysis of US labs' irrational avoidance of advertising and business models.⁷

By the way, this is the first time Ben has heard of this—Ben is one of the greatest analysts in history.

8. Hardware is decoupling from the semiconductor layer, software layer, and data layer, and their interweaving is accelerating faster than any government can respond

US applications run on Chinese models; Chinese labs train on Nvidia compute rented in Malaysia and Thailand while distilling US model outputs; expert data flows in both directions.⁸

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

Similarly, systems are now deeply interdependent, fitting both the imagination of "rhetorical nation-state supremacy" and achieving genuine transnationalism in actual operation—unless recursive self-improvement, or RSI, substantively changes the current game structure.

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

Broadly speaking, RSI can be divided into three types: time, compute, and boundary RSI. The earliest RSI can be understood as researchers involving models in their own iteration, manually assisting models to accelerate steps that would otherwise consume researchers' compute or time, or both. It should be noted that the "takeoff" at this stage likely will not appear as an exponential curve over long timescales like decades to 2100; nor will it be a case where human collective effort exceeds what any single person or machine can achieve.

10. A key bifurcation point

Assuming first-order self-improvement does exist, then in terms of compute, the US advantage seems very solid: by 2030, two labs each operate about 2GW of compute, and each has more than 5GW of contracted projects.

In theory, these projects are funded by substantial revenue and strong capital markets; there seems to be no end to scaled computing, lateral thinking, "self-improvement" loops, and learning how to effectively use these capabilities.

But conversely, if the real constraint lies in verification quality, or if the timeline exceeds what the curve should allow, then the phrase "capacity optimization and domestic chip redundancy"—which may sound familiar—could continue to accelerate over the next 18 months:

Accelerated commoditization of US frontier model capabilities;

Increased unique model capabilities used by Chinese open-weight models;

Higher bargaining power for bid-based AI companies at the inference layer;

Greater transaction volume and profits for compute service providers as foreign-owned weights are incorporated;

Ultimately, these effects will also transmit to inference service providers, such as hyperscale cloud providers and new cloud vendors.

This means there will be a massive misallocation of investment in power infrastructure.

We need to emphasize: our analysis of US and Chinese infrastructure may be the highest potential return category among all the analyses we have ever done.

About the future

11. A possible future world

In the future, US frontier labs may no longer open APIs to the next generation of most advanced models. In fact, for the vast majority of programmatic enterprise applications, current technology levels are already sufficient. Future high-end models will be priced by value rather than usage, and will increasingly drive China to benefit from frontier capabilities.

Therefore, US labs may raise the priority of API revenue to support higher-priced, lower-usage models; this in turn will cut off an important competitive pipeline from the US East Coast across the Pacific.

At the very least, this will force more data to be processed within China using pre-trained models, thereby increasing pressure on the existing compute bottleneck.

From what we have learned, some sizable US multinational companies are conducting pilots to migrate tasks such as engineering and mathematics to China for processing.

12. The US-China AI landscape

We believe AI has already deeply intertwined the US and Chinese technology systems to the point where regulatory measures that merely sever one trans-Pacific link seem at best temporary.

US and Chinese technologists are, in a sense, talking to each other; this exchange may occur at multiple levels such as model distillation, open weights, data annotation, and inference, while software is increasingly leveraging agentic search infrastructure.

Direct bans on currency flows, entity listings, custody bans, and procurement bans will almost certainly suppress US innovation and cede leadership in related fields, ultimately driving capital to Chinese competitors.

Despite the loud Western discussion about an AI valuation bubble, the atmosphere in the Chinese market is clearly much hotter.

If you look at revenue multiples, Chinese labs trade at 5 to 10 times the valuation of US frontier labs; by revenue scale, the US-China gap is close to two orders of magnitude.

Moonshot AI's latest funding round valued the company at $3.5 billion, with July annualized revenue of about $100 million, a valuation of about 35 times revenue; the new round is reportedly being conducted at a $5 billion pre-money valuation, equivalent to about 50 times revenue.

Zhipu's market cap at its Hong Kong listing in January this year was about $10 billion, roughly 20 times Anthropic's valuation multiple. The company's early public market stock price was highly volatile: from a market cap of about $1.3 billion at listing, it once rose to about $10 billion in January this year.

Kong AI Knowledge Atlas Technology HK (2613) also rose from a market cap of about $1.3 billion in January this year to about $10 billion; Luna AI (HKG:100) went from about $15 billion all the way up to $33 billion, then fell back to about $10 billion.

Both stocks fell sharply before the end of their lock-up periods. Our judgment is that this mainly stems from immature market structure: decades of consumer software investment experience are only now gradually accumulating the ability to conduct prudent evaluations around corporate financial disclosures and open-source frameworks.

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

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

Your partners,

Nan, Adam, Zivain

Notes

DeepSeek V3 technical report: used 2048 H800 GPUs; PTX-level programming; about 30 million messages dedicated to cross-node communication.

In March 2026, OpenRouter's weekly token share was about 42%, fluctuating by about 5%; Qwen accounted for about 18%. It is worth noting that Hugging Face model downloads account for only about 80% of Chinese open-weight model usage, with about 20% distributed through other channels.

Anthropic has formally accused Moonshot AI, DeepSeek, and MiniMax of collecting millions of Claude conversation data; this claim was subsequently denied and rebutted by several companies.

Comments on accounting treatment differences among different run-rate methods are omitted for now.

Alibaba and ByteDance's data center training in Southeast Asia accelerated after restrictions were introduced on April 20, 2026; these companies rent compute in non-Chinese-owned facilities in Thailand, Malaysia, and Japan. The US White House's March 2026 OSTP directive has brought Nvidia export-grade GB300 servers into Thailand for K3 training; current export controls place more emphasis on chip ownership than remote access. Compute rental costs in Thailand, Malaysia, and Indonesia are about 30% higher, while operating costs are about 10% lower.

Before the lock-up period expires, institutions holding more than 5% cannot cash out; ordinary shareholders also cannot sell before the lock-up period ends.