The Dragon Algorithm: China’s Conquest of the LLM Market

The prevailing assumption in artificial intelligence strategy is, on the surface, disarmingly simple: the organisations that are first to reach a critical mass of adoption will prevail. Lower costs, rapid uptake among developers and data feedback loops create a self-reinforcing dynamic in which usage generates data, data improves models, and improved models attract further users.

At first glance, recent evidence appears to confirm this logic. Global usage of large language models rose from around 2.4 billion to over 8 billion monthly visits between April 2024 and August 2025 — a 3.4-fold increase in sixteen months (Wang & Siler-Evans, 2026)[1]. During the same period, Chinese models gained market share at an unprecedented rate: from 3% to 13% of global usage within two months of the release of DeepSeek R1 (Wang & Siler-Evans, 2026)[1].

Yet this interpretation, however convincing, is incomplete. It explains how artificial intelligence spreads, but not how value is captured or defended over time. What is missing is the structural level underlying adoption: the geography of production, the distribution of capital and the configuration of the AI value chain (Agostini, 2025a; Agostini, 2025b)[2,3].

The competitive advantage in AI does not stem from the performance of a single model, but from control over the systems that enable its production, distribution and scalability.

A closer examination of the data reveals a more complex and asymmetrical picture. Chinese labs are gaining ground in the infrastructure and open-model layers of the LLM ecosystem, whilst US companies continue to dominate the enterprise and application layers. This is not a temporary divergence, but the result of profound structural differences in the way the two ecosystems are built and scaled.

The geography of AI production: where the winners emerge

The geography of AI production provides the first key insight. A comprehensive dataset of 1,246 AI companies spread across 32 economies shows that the development of artificial intelligence is highly concentrated: nearly 700 companies are based in the United States, and around 250 in China (Rishabh & Shreeti, 2026)[4]. No other region comes close to this scale.

This concentration is not merely descriptive: it reflects specific economic forces. The likelihood of an economy hosting AI companies is strongly correlated with GDP size, expenditure on research and development and, crucially, venture capital flows (Rishabh & Shreeti, 2026)[4]. Venture capital plays a disproportionately large role because it finances high-risk, intangible-asset-intensive businesses, a type that traditional banking systems are structurally unable to support.

AI ecosystems are therefore not merely technological phenomena: they are financial and institutional constructs. And this distinction is key to understanding who stands to gain — and on what terms.

The AI market is a multi-layered system, not a single market

Adoption alone cannot explain market leadership, because adoption follows production capacity. The ability to build, deploy and scale AI systems depends on the density of companies within the ecosystem, the availability of capital, and integration within a broader value chain.

Artificial intelligence is not a single market, but a multi-layered system comprising interdependent segments: computing, cloud infrastructure, data tools, models and applications. Most economies specialise in just one or two of these layers. Only a small number — primarily the United States and China — operate across the entire value chain (Gambacorta & Shreeti, 2025; Rishabh & Shreeti, 2026)[4].

Systemic advantage does not accrue to those who dominate a single layer, but to those who simultaneously orchestrate multiple levels of the value chain. This is the difference between a niche operator and a structurally dominant player.

Operating figures: when benchmarks don’t tell the whole story

At an operational level, this structural advantage becomes apparent when looking at actual usage — not benchmarks. OpenRouter, a multi-model routing platform that aggregates over 100 trillion tokens of productive activity, shows that seven of the ten most widely used models by weekly token volume come from Chinese developers: Alibaba Qwen, DeepSeek, Minimax, Stepfun, Xiaomi’s Mimo and Moonshot AI’s Kimi (OpenRouter, 2026)[5].

Collectively, these models account for the majority of token consumption among the most widely used systems. Qwen3.6 Plus alone generated over 5 trillion tokens in a single week (OpenRouter, 2026)[5]. Usage data reveals a shift in competitive power that benchmark rankings fail to capture.

Benchmarks capture performance in the lab. Tokens reveal who is building the industrial engine of AI.

This shift is amplified by the dynamics of open ecosystems. The rapid proliferation of open-source models — particularly within the Qwen ecosystem — has generated thousands of derivative models and integrations on developer platforms (Hugging Face, 2026)[6]. Each deployment creates a new node in a growing network of usage, reinforcing adoption through scale. Over time, this produces a self-reinforcing dynamic in which distribution expands more rapidly than centralised control.

The token economy: AI becomes an industrial system

The emergence of the token economy is further accelerating this transformation. Daily token consumption in China exceeded 140 trillion by March 2026, up from around 100 billion at the start of 2024 (Bloomberg, 2026; Fortune, 2026; Hello China Tech, 2026)[7,8,9]. Chinese policymakers and industry leaders have begun to view token production as a strategic industrial capability and a potential export sector (China Daily, 2026)[10].

Companies such as MiniMax, Moonshot AI, Zhipu AI and DeepSeek are increasingly being assessed on the basis of token throughput, utilisation and cost efficiency (Bloomberg, 2026; Frontier Wisdom, 2026)[7,11]. Chinese companies benefit from structural advantages rooted in lower-cost infrastructure, access to relatively cheap electricity and highly competitive developer ecosystems. The result: they are able to deliver output tokens at prices close to one dollar per million, compared to $15 or more for comparable US systems (Bloomberg, 2026; Yahoo Finance, 2026)[7,12]. The cost per token — not the sophistication of the model — is becoming the dominant competitive variable.

Path dependency: why it is difficult to recover

Investment patterns reinforce existing structural positions. AI companies exhibit a strong home bias: US and Chinese firms conduct 64% and 74% of their operations domestically, respectively (Rishabh & Shreeti, 2026)[4]. Investment flows also tend to reinforce specialisation, as firms invest disproportionately in companies operating at the same level of the value chain.

This creates path dependencies, making it increasingly difficult for new entrants to diversify or reposition themselves. AI ecosystems are not merely competitive: they are self-reinforcing and resistant to structural change. Those who have not yet established a foothold in the supply chain are losing ground with every passing year – ground that becomes harder to regain.

Nevertheless, US companies retain decisive advantages in other segments of the value chain. OpenAI, Google, Anthropic, Microsoft, Meta, xAI and Perplexity continue to dominate enterprise deployment through integrated ecosystems, reliable infrastructure and deep institutional relationships. Private investment in AI in the US remains significantly higher, reinforcing leadership in capital-intensive segments (Stanford HAI, 2026; Forbes, 2025)[13,14].

The true nature of competition: who controls the constraints?

Taken together, these data lead to a fundamental conclusion. Artificial intelligence is not a homogeneous market governed by performance, but a layered socio-technical system shaped by constraints. Infrastructure defines the physical limits of scalability. Capital determines which companies can survive and expand. Positioning within the value chain determines where value is created. Governance frameworks determine what is permitted (WEF, 2026)[15].

The apparent contradiction between Chinese and US dominance is resolved as soon as these layers are taken into account. Chinese labs are dominating the infrastructure and consumer layers through cost efficiency, scale and open ecosystems. US companies are dominating the enterprise and application layers through integration, monetisation and trust. The AI market is not a ‘winner-takes-all’ game: it is structurally segmented across different forms of control.

The key question is no longer how to maximise adoption, but what constraints adoption must meet — and who controls them.

Adoption can be replicated, pricing can be eroded, and model capabilities can converge. Constraints — such as access to capital, infrastructure availability and positioning within the value chain — are far more enduring.

It is the management of constraints — not the scale of adoption — that is the true source of long-term advantage in artificial intelligence.

In this sense, the global race for AI is not simply a competition over models, metrics or market share. It is a systemic contest over the very architecture of production. The organisations that will define the next phase of AI are not those that build the most advanced models, but those that control the bottlenecks through which all models must pass. (photo by Stone John on Unsplash)

References

[1] Wang, A. H.-E., & Siler-Evans, K. (2026). U.S.-China competition for artificial intelligence markets. RAND Corporation. https://www.rand.org/pubs/research_reports/RRA4355-1.html

[2] Agostini, M. (2025a). AI can’t personalize chaos: Why data governance is the real engine of personalization. Medium. https://medium.com/@tarifabeach/ai-cant-personalize-chaos-why-data-governance-is-the-real-engine-of-personalization-dc71c358e8e3

[3] Agostini, M. (2025b). AI isn’t killing search – it’s rebuilding the front door of the internet. Medium. https://medium.com/@tarifabeach/ai-isnt-killing-search-it-s-rebuilding-the-front-door-of-the-internet-a739ed92d8b4

[4] Rishabh, K., & Shreeti, V. (2026). The geography of AI firms. Bank for International Settlements. BIS Working Paper No. 1343. https://www.bis.org/publ/work1343.htm

[5] OpenRouter. (2026). State of AI and model usage data. https://openrouter.ai/state-of-ai

[6] Hugging Face. (2026). Model repository and download statistics. https://huggingface.co

[7] Bloomberg. (2026, 20 aprile). AI’s token economy revolution creates new China tech winners. https://www.bloomberg.com/news/articles/2026-04-20/ai-s-token-economy-revolution-creates-new-china-tech-winners

[8] Fortune. (2026, 12 aprile). China’s token economy AI boom reshapes big tech and startups. https://fortune.com/2026/04/12/china-token-economy-ai-boom-big-tech-startups/

[9] Hello China Tech. (2026). China’s AI token economy reaches 140 trillion daily tokens. https://hellochinatech.com/p/china-token-economy-140-trillion

[10] China Daily. (2026). China accelerates AI-driven economic transformation. https://global.chinadaily.com.cn/a/202604/06/WS69d3c692a310d6866eb41d92.html

[11] Frontier Wisdom. (2026). China AI token economy explained. https://frontierwisdom.com/china-ai-token-economy-explained-guide/

[12] Yahoo Finance. (2026, 20 aprile). AI’s token economy revolution creates new China tech winners. https://finance.yahoo.com/news/ais-token-economy-revolution-creates-new-china-tech-winners-035211861.html

[13] Stanford Institute for Human-Centered Artificial Intelligence. (2026). AI Index Report 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report

[14] Forbes. (2025). AI 50 list. https://www.forbes.com/lists/ai50

[15] World Economic Forum. (2026). Global Risks Report 2026. https://www.weforum.org/reports/global-risks-report-2026

ALL RIGHTS RESERVED ©