Inside China's AI Machine: Models, Chips, and Strategy

Digital in Asia — April 2026

China’s daily AI token usage surpassed 140 trillion in March 2026. That’s up from 100 billion at the start of 2024 — a more than 1,000-fold increase in two years. On OpenRouter, the world’s largest AI model API aggregation platform, Chinese models accounted for 61% of total token consumption among the platform’s top ten models in February 2026. During the week of 16–22 February, Chinese models processed 5.16 trillion tokens on the platform, compared to 2.7 trillion for US models. Four of the top five most-used models globally were Chinese.

These aren’t projections. They’re usage metrics published by China’s National Data Administration at the China Development Forum in March 2026, and confirmed by third-party platform data. China’s AI industry has moved past the demonstration phase and into a deployment pattern that, by volume at least, now exceeds the United States.

The question that everyone outside China keeps asking is the same one: how? How did a country cut off from the world’s most advanced AI chips build models that compete with — and in some cases outperform — the products of labs spending ten to fifty times more? And what does it mean that the answer isn’t a single company or a single breakthrough, but an entire industrial system designed to turn constraint into advantage?

This is an attempt to answer that question comprehensively. Not the DeepSeek headline. The whole machine.

The Model Landscape: It’s Not Just DeepSeek

The most common mistake in Western coverage of Chinese AI is treating the story as a single-company narrative. DeepSeek is important — its R1 model genuinely changed global expectations about what’s possible at what cost. But the Chinese AI model ecosystem is now a multi-front operation involving at least seven major players with distinct strategies, commercial models, and competitive positions.

DeepSeek remains the frontier research lab. Funded by the quantitative hedge fund High-Flyer and led by co-founder Liang Wenfeng, DeepSeek operates without public shareholders or external investor pressure. That independence lets it prioritise pure capability advancement over monetisation. The R1 model — trained for approximately $6 million compared to roughly $100 million for GPT-4 — used mixture of experts (MoE) architecture to activate only 37 billion of its 671 billion parameters per inference, slashing compute requirements while maintaining performance. DeepSeek’s V4 model launched in February 2026, reportedly optimised for coding, and the company received conditional government approval to purchase NVIDIA H20 chips for its next generation of training. But it’s also been encouraged by authorities to adopt Huawei’s Ascend chips — a tension that illustrates the broader dynamics at play. R2, the anticipated successor to R1, has been delayed, reportedly due in part to training difficulties on Ascend hardware.

Alibaba’s Qwen is the volume leader. The Qwen model family has generated over 100,000 derivative models on Hugging Face — the largest open-weight ecosystem on the platform, surpassing every Western counterpart including Meta’s Llama. Qwen claims more than 100 million monthly active users. Alibaba’s commercial strategy is to integrate AI into its existing e-commerce infrastructure: the Qwen app was updated to enable shopping, food ordering, and payments without leaving the interface, connecting directly to Taobao. The Qwen 3.5 series is imminent, with improved maths and coding performance. Even US companies are adopting Qwen — Airbnb uses it for customer service chatbots, drawn by comparable capability at lower cost.

ByteDance controls the most popular AI app in China. Its Doubao chatbot has 155.2 million weekly active users (QuestMobile). ByteDance’s Seedance 2.0 video generation model drew praise from Chinese state media, and its TikTok parent is investing $8.8 billion in regional data centre infrastructure across Southeast Asia. ByteDance is simultaneously a model developer, a distribution platform (through TikTok/Douyin), and an infrastructure provider — a vertical integration that no Western AI company currently matches.

Baidu pivoted from closed-source to open after DeepSeek’s breakout forced the issue. CEO Robin Li had argued closed models would dominate; within weeks of R1’s release, Baidu opened portions of its own models. Ernie 4.5 and the X1 reasoning model are the current flagships. Baidu’s particular strength is in AI-powered marketing through its Qingduo creative platform — which went from producing roughly 20 ad creatives per hour to more than 2,000 after integrating DeepSeek technology — and its dominant search platform processing over 6 billion daily queries in China.

MiniMax listed on the Hong Kong Stock Exchange in January 2026, with its share price doubling on the first day of trading. Its M2.5 model, released in February, topped the OpenRouter weekly call volume list in less than a week — contributing 1.44 trillion tokens in a single week to the platform’s total. MiniMax was founded by former engineers from AI surveillance firm SenseTime and has built strong video generation capabilities.

Moonshot AI revealed Kimi K2.5 in January 2026, claiming video-generation and agentic capabilities outperforming the three leading US models. Kimi’s cumulative revenue in less than a month after the K2.5 release exceeded its total earnings for all of 2025. The model can dispatch up to 100 “agent avatars” to work in parallel, increasing complex task processing efficiency by three to ten times.

Zhipu AI (the GLM model family), Tencent (Hunyuan), and a growing roster of university-affiliated labs round out the ecosystem. The depth matters. This isn’t a one-horse race or even a two-horse race. It’s a field with enough players that the competitive dynamics themselves drive rapid improvement.

The Cost Advantage: Structural, Not Accidental

Chinese AI models run at roughly one-sixth to one-quarter the cost of comparable American systems, according to a RAND report published in early 2026. DeepSeek’s pricing on its API sits at approximately $0.028 per million tokens — roughly 1/180th of equivalent GPT pricing.

This cost advantage isn’t a temporary subsidy or a pricing war tactic. It’s structural, rooted in three reinforcing factors.

First, algorithmic efficiency. DeepSeek’s core innovations — MoE architecture, multi-head latent attention for memory compression, multi-token prediction for faster inference — dramatically reduce how much compute each query requires. These techniques have been adopted across the Chinese model ecosystem, creating a baseline of efficiency that Western frontier labs are still integrating.

Second, infrastructure economics. China’s central and local governments subsidise electricity costs for data centres, with provinces like Gansu, Guizhou, and Inner Mongolia offering to slash cloud providers’ power bills by up to 50%. The “East-West Computing Resource Transfer” project moves compute-intensive workloads from expensive eastern cities to energy-rich western provinces. The economics of running AI in China are fundamentally different from the economics of running AI in Virginia or Oregon.

Third, the open-source feedback loop. When models are open-weight, improvements compound across the ecosystem. A USCC report published in March 2026 identifies this as the first of “two loops” driving China’s AI strategy: an iterative open collaboration where frontier labs refine each other’s base models, enterprises adapt them for niche applications, and the resulting deployment data feeds back into capability improvements. DeepSeek’s R1-Distill-Qwen-32B — a Qwen derivative that claims to outperform the original base model on certain benchmarks — illustrates the dynamic: companies building on each other’s work rather than competing purely from scratch.

The Chip Question: Building a Parallel Stack

US export controls have cut China off from NVIDIA’s most advanced AI chips. The H100, A100, and their export-variant successors are all restricted. Even the cut-down H800 and A800 are off-limits. The B200 is entirely unavailable to Chinese buyers.

China’s response has been to build a parallel hardware ecosystem from the ground up, anchored by Huawei’s Ascend line of AI accelerators.

Huawei plans to produce approximately 600,000 Ascend 910C chips in 2026 — roughly double the 2025 output. Total Ascend line production could reach 1.6 million dies in 2026 across all models. The chips are manufactured by SMIC on an enhanced 7nm process, compared to TSMC’s 4nm process used for NVIDIA’s latest chips. The performance gap at the individual chip level is real — Chinese chip industry leaders have publicly acknowledged a five-to-ten year lag behind US-manufactured AI chips.

Huawei’s solution is scale and system-level architecture. The Atlas 950 SuperPoD, expected in late 2026, links 8,192 Ascend chips. Huawei claims it will deliver 6.7 times the compute of NVIDIA’s planned NVL144 cluster, with 15 times the memory and 62 times the interconnect bandwidth. The roadmap runs through Ascend 960 (Q4 2027) and Ascend 970 (Q4 2028), each doubling compute capacity over its predecessor.

The software stack is the harder problem. NVIDIA’s competitive advantage isn’t just hardware — it’s CUDA, the programming framework that fifteen years of AI research has been built on. Huawei’s alternative is CANN (Compute Architecture for Neural Networks). Training on Ascend/CANN requires different workflows, different optimisations, and in some cases different architectural choices than CUDA-native development. DeepSeek’s R2 delays have been partly attributed to training difficulties on Ascend hardware, with the company reportedly reverting to NVIDIA H800s for critical training runs. The hardware works for inference; full pre-training of frontier models is where the gap remains substantial.

Smaller domestic players are also entering the market. Cambricon Technologies is manufacturing AI accelerators at domestic fabs, with production expected to reach 300,000–350,000 units in 2026. Alibaba’s T-Head semiconductor unit and Baidu-backed Kunlun Tech have introduced chips compatible with NVIDIA’s CUDA ecosystem, making adoption easier for developers. Over 60 semiconductor companies are now backed by Huawei’s investment arm Hubble, building an ecosystem spanning design tools, packaging materials, photoresists, and optical interconnects.

China’s AI hardware stack is forking from the global one. For anyone deploying AI in or for the Chinese market, this means a separate hardware and software ecosystem with different performance characteristics, different toolchains, and different optimisation requirements.

The State Strategy: AI+ and the 15th Five-Year Plan

China’s 15th Five-Year Plan (2026–2030), passed by the National People’s Congress in March 2026, embeds AI as a foundational national priority comparable to defence and economic growth. The plan’s ambition is explicit: China expects its AI-related industries to exceed 10 trillion yuan ($1.4 trillion) in value by 2030.

The “AI+” initiative, released by the State Council in August 2025, sets clear penetration targets. AI adoption in key sectors must reach 70% by 2027 and 90% by 2030. By 2035, China envisions what the plan describes as “a fully intelligent economy and society.” The initiative prioritises AI deployment across six domains: science and technology, industrial development, consumer services, public welfare, governance, and international cooperation.

This isn’t aspirational language in the way Western government strategies often are. China’s implementation architecture includes subsidised electricity for AI data centres, a national “East-West Computing Resource Transfer” infrastructure project, and direct state procurement of AI products and services. Local governments compete to attract AI investment with tax incentives and dedicated computing clusters. The Cybersecurity Law was amended in October 2025, effective from January 2026, to explicitly include AI — bringing algorithm development, AI infrastructure, and AI ethics into national law for the first time.

The AI Safety Governance Framework 2.0, adopted in September 2025, takes a full lifecycle approach to risk management, covering model design through deployment. China has also committed to developing over 20 international AI standards, positioning itself to shape the governance frameworks that other countries adopt.

The Carnegie Endowment describes China’s current position as a “Crossroads Era” — the fourth distinct phase of Chinese AI policy since 2017. The first three: a “Go-Go Era” (2017–2020) of minimal regulation and massive investment; a “Crackdown Era” (2020–2022) when the CCP reasserted control over tech companies; and a “Catch-Up Era” (2022–2025) that loosened restrictions to boost growth. The current phase, triggered by DeepSeek’s success, combines accelerated adoption with renewed confidence — and the open question of whether the state will reassert ideological control now that it perceives technological strength.

The Deployment Story: Applied AI at Scale

The most underreported aspect of China’s AI trajectory isn’t models or chips — it’s deployment. While the US debate focuses on the race to artificial general intelligence, China is running what Brookings describes as “multiple AI races” simultaneously, with the deployment race arguably further advanced than any other country’s.

About 34% of job functions at Chinese companies are already fully integrated with AI tools, compared with 30% globally (AlixPartners 2026 Disruption Index). Platforms like Alibaba’s DingTalk and ByteDance’s Feishu function as AI-native operating systems for office work — not add-ons but the primary interface through which work gets done. AI is ambient, embedded in the workflow rather than activated as a separate application.

Embodied AI is a target industry in the new Five-Year Plan. Robotaxi services from WeRide, Baidu’s Apollo Go, and Pony.ai are expanding across Chinese cities and internationally. Autonomous delivery vehicles and drones are commonplace in Shenzhen and Shanghai. Robotics firms Unitree, UBTech, and AgiBot are racing to mass-produce humanoid robots by the thousands. China’s advantage here is its overlapping hardware manufacturing ecosystem — the same industrial base that produces smartphones and electric vehicles can produce AI-enabled physical systems at scale and at cost.

The token economy itself is becoming a commercial infrastructure. The National Commission for Terminology in Science and Technology issued a formal Chinese name for “token” — “ciyuan” (word element) — in March 2026, reflecting how central the concept has become to Chinese economic planning. Some model developers reported that revenue over a 20-day period since late January 2026 exceeded their total earnings for all of 2025. The token-based pricing model is creating a new service economy where compute and intelligence are metered, priced, and traded as commodities.

The Open-Source Strategy: Ecosystem as Geopolitics

China’s commitment to open-source AI is strategic, not ideological. The government has explicitly positioned China as the global provider of open AI development — a framing designed to capture developer mindshare, establish Chinese models as industry standards, and create dependency relationships that mirror what NVIDIA’s CUDA ecosystem achieved in hardware.

Premier Li Qiang declared at the 2025 World Economic Forum that China’s innovation is “open and open-source” and the country is “willing to share indigenous technologies with the world.” The AI+ Initiative calls for government support of AI products with open-source development as a priority. More than 90% of Chinese enterprises use open-source technologies.

The distribution numbers tell the story. An Andreessen Horowitz partner estimated that 80% of US startups use Chinese base models for derivative development. Chinese models’ weekly token consumption on OpenRouter surpassed US models in February 2026 and the gap has widened since. In Southeast Asia, adoption is accelerating — Singapore’s OCBC bank runs 30+ internal tools on DeepSeek and Qwen, Indonesia’s Indosat partnered with an AI firm building on DeepSeek, and Malaysia launched a sovereign AI ecosystem on Huawei hardware.

The open-source strategy creates a second feedback loop identified in the USCC report: global diffusion that generates adoption data, developer contributions, and commercial relationships, reinforcing China’s position as the default AI infrastructure layer for cost-sensitive markets worldwide. The strategic calculation is that even if US frontier labs produce more capable individual models, the Chinese open ecosystem — cheaper, more accessible, more adaptable — becomes the foundation that most of the world actually builds on.

What’s Coming Next

Several developments will shape the next 12 months.

Huawei’s Ascend 950PR shipped in Q1 2026. Whether it meaningfully closes the gap with NVIDIA’s B200 will determine how quickly China’s hardware ecosystem can support frontier-model training without relying on stockpiled NVIDIA chips.

DeepSeek’s R2 remains the most anticipated model release in the Chinese ecosystem. If the Ascend training difficulties are resolved, R2 could demonstrate that frontier models can be developed entirely on domestic hardware — a milestone with enormous strategic significance. If it requires NVIDIA chips, the dependency narrative persists.

The token economy is scaling exponentially. Xiaomi’s head of large models predicted token consumption could increase 100-fold in 2026 alone, driven by agent-based applications. OpenClaw — an open-source autonomous agent framework — contributed 20% of OpenRouter’s token consumption in a single week in March 2026, illustrating how agent workloads are fundamentally different from chatbot interactions in their compute intensity.

Embodied AI is moving from demonstration to deployment. The Five-Year Plan’s explicit targeting of this sector, combined with China’s manufacturing advantages, suggests that AI-enabled robotics and autonomous systems could become a major Chinese export category within the plan period.

Global regulatory dynamics will shape how Chinese AI models operate internationally. Vietnam’s AI Law, the EU AI Act, and various ASEAN governance frameworks all create compliance requirements that Chinese model providers must navigate. The degree to which open-weight Chinese models are treated differently from closed Western APIs in regulatory frameworks could significantly affect adoption patterns.

What This Means for Asia

The question isn’t whether China’s AI industry is significant. The usage data, the model performance, and the deployment metrics make that self-evident. The question is what the rest of Asia does with this reality — because the implications are different depending on whether you’re in Singapore, Jakarta, Tokyo, or Mumbai.

Southeast Asia is already choosing. In practice, if not always in policy statements, the region is adopting a dual-stack approach — US hyperscaler infrastructure layered with Chinese open-source models. Singapore’s OCBC runs DeepSeek and Qwen alongside Google’s Gemma for different internal functions. Malaysia launched a sovereign AI ecosystem on Huawei GPUs. Indonesia’s Indosat is building telco AI solutions on DeepSeek. Thailand is accepting billions from both Microsoft and ByteDance for data centres in the same provinces.

This isn’t fence-sitting. It’s pragmatism in a region where the priority is deployment speed and cost, not geopolitical alignment. A Vietnamese startup training a local-language model cares about GPU access and API pricing, not about whether the base model was trained in Hangzhou or San Francisco. An Indonesian e-commerce platform evaluating AI for product recommendations will pick whatever performs best at the lowest cost per query. Right now, Chinese models win that evaluation more often than the Western AI industry would like to admit.

The dependency question is real, though. If the open-weight model ecosystem becomes predominantly Chinese in origin — and the Hugging Face data says it’s heading there — then the foundational AI layer for applications built across Southeast Asia increasingly originates from Chinese research. That’s different from buying Chinese smartphones or using TikTok. It means the reasoning architecture, the training biases, the optimisation priorities, and potentially the values embedded in training data all flow from decisions made by Chinese labs and, ultimately, by a government that has explicitly integrated AI into its national security framework. For most Southeast Asian businesses, that’s an abstraction. For Southeast Asian governments building sovereign AI strategies, it’s a structural consideration that doesn’t have an easy answer.

Japan and South Korea face different calculations. Both are US security allies with their own semiconductor industries, their own AI strategies, and their own reasons to be cautious about Chinese technology dependency. Japan’s AI Promotion Act and South Korea’s 2026–2028 AI Action Plan are both designed to build domestic capability rather than rely on imported models.

But even in these markets, the Chinese efficiency breakthrough creates pressure. Japanese AI adoption is strikingly low — 26.7% among the general public, compared to 81.2% in China. If Chinese open-source models are the fastest way to close that gap, the policy question gets uncomfortable. South Korea’s semiconductor companies — Samsung and SK Hynix — sell memory chips to Chinese AI firms while simultaneously being critical to the US AI supply chain. The hardware interdependency runs deeper than the model competition.

India is playing its own game. The IndiaAI Mission’s sovereign LLM push, the country’s massive engineering talent pipeline, and its digital public infrastructure (Aadhaar, UPI) give India more leverage than most countries to build domestically rather than adopt from either the US or China. But India’s AI startup ecosystem is intensely cost-sensitive, and Chinese models’ price advantage is directly relevant. The question for India is whether it can build a domestic AI industry fast enough that its developers don’t default to Chinese open-source models as their foundation layer — not because of policy, but because the economics are overwhelming.

What This Means for the West

The hardest thing for Western observers to accept about China’s AI trajectory is that it doesn’t require being “better” in the way Silicon Valley defines the term. China doesn’t need to build the most capable frontier model. It needs to build models that are good enough, cheap enough, and open enough to become the default for most of the world’s AI deployment. That’s a different competition, and it’s one where China’s structural advantages — cost, scale, manufacturing, state coordination — are arguably stronger than America’s.

The spending asymmetry is becoming a liability, not an advantage. America’s Big Four hyperscalers — Alphabet, Amazon, Meta, and Microsoft — have collectively announced AI spending exceeding $650 billion in 2026 alone, with total US AI compute infrastructure spending projected to surpass $2.8 trillion by 2029 (Brookings). China’s AI ecosystem is producing comparable or competitive models at a fraction of that investment. A RAND report found Chinese models cost one-sixth to one-quarter of US equivalents. If the US is spending more to produce less widely-adopted products, the spending itself becomes a strategic question rather than a strategic answer.

The US theory of AI dominance has always been: build the most powerful chips, train the most capable models, and the world will buy access. That theory assumed compute was the binding constraint. DeepSeek demonstrated that algorithmic efficiency can substitute for brute-force compute. The follow-on effects — Alibaba investing 380 billion yuan ($52 billion) in AI and cloud, MiniMax doubling on IPO, four of the top five global models by usage being Chinese — suggest the substitution isn’t a one-off. It’s a systemic shift.

The open-source dynamic creates a different kind of lock-in. Western AI strategy has largely assumed that proprietary models would capture most commercial value, the way proprietary software did in the 1990s and 2000s. China’s open-source push inverts that logic. If the foundational models are free and open, value migrates to deployment, integration, and the applications built on top — areas where China’s vast domestic market and manufacturing base create natural advantages. Even in the US, an estimated 80% of startups are building on Chinese base models. The foundations of Western AI products are, in a growing number of cases, Chinese.

European companies face an especially awkward position. The EU AI Act creates compliance requirements that favour large, well-resourced model providers — which increasingly means Chinese companies with the scale and legal infrastructure to navigate multi-jurisdictional regulation, alongside the US incumbents. European AI companies, caught between US frontier ambitions and Chinese cost efficiency, risk being squeezed out of both the model layer and the deployment layer.

The real competition isn’t models — it’s infrastructure. China’s 15th Five-Year Plan treats AI as a foundational utility, comparable to high-speed rail or the electrical grid. The “AI+” initiative targets 70% sectoral penetration by 2027 and 90% by 2030. That’s not a technology strategy. It’s an infrastructure transformation with AI embedded into manufacturing, logistics, agriculture, healthcare, and governance at a level that no Western economy is attempting.

The Brookings framing is useful here: the US is running a race to AGI while China is running multiple races simultaneously — models, deployment, embodied AI, the token economy, and physical AI integration. If AGI arrives and delivers the transformative capabilities its advocates predict, the US bet pays off. If AI development follows a more incremental trajectory, China’s “good enough and everywhere” approach could generate more aggregate economic value, faster, across more of the global economy.

What This Means for Investors and Operators

For companies already in Asian markets: Chinese AI models and infrastructure are part of your technology stack whether you’ve chosen them or not. Your employees may be using DeepSeek. Your cloud provider may be running Qwen derivatives. Your competitors almost certainly are. The right approach isn’t to pick a side — it’s to build evaluation pipelines that can accommodate models from multiple providers, compare them honestly on performance and cost for your specific use cases, and maintain the flexibility to shift as capabilities change. Model loyalty is a luxury that Asian market economics don’t support.

For companies looking at Asian markets from outside: The AI landscape you’re entering is fundamentally different from what you’re used to. The assumption that enterprise AI means OpenAI or Anthropic or Google doesn’t hold in a region where cost-per-query drives adoption decisions and where Chinese models handle local languages better. If you’re building products for Southeast Asian consumers or enterprises, you need to evaluate Chinese models alongside Western ones — and you may find they outperform on the metrics that matter most in these markets.

For investors: The Chinese AI ecosystem is generating real commercial returns. MiniMax’s Hong Kong IPO doubled on day one. Moonshot’s Kimi K2.5 earned more revenue in its first month than the prior year’s total. Some model providers saw 20 days of post-January revenue exceed their entire 2025 earnings. The token economy is creating measurable, transactable value at unprecedented scale — 140 trillion daily tokens isn’t a vanity metric, it’s a commercial infrastructure.

But the risks are equally real. Regulatory stability in China oscillates between promotion and control on cycles that even experienced China operators find hard to predict. The chip supply question — whether Huawei’s Ascend line can actually support frontier training, or whether the ecosystem remains quietly dependent on stockpiled NVIDIA hardware — is unresolved. The energy infrastructure challenge mirrors Southeast Asia’s: you can’t run 140 trillion daily tokens without electricity, and China’s power grid is already under strain. And an ecosystem with this many well-funded competitors may ultimately produce returns for the sector without producing returns for individual companies.

The bottom line for global markets: China isn’t building the same AI future as the United States. It’s building a different one — cheaper, more distributed, more integrated into physical infrastructure and daily commerce, and increasingly present in the markets where most of the world’s population actually lives and works and buys things. The Western assumption that AI leadership means frontier model capability is being tested by a Chinese ecosystem that’s prioritising deployment breadth, cost efficiency, and physical-world integration.

That doesn’t mean the US approach is wrong. If AGI arrives, compute maximalism may prove visionary. But it does mean that the competition isn’t the one most Western commentators think they’re watching. The race for the smartest model is one competition. The race to make AI the operating system for the global economy is another. Right now, China is further ahead in the second race than most people outside Asia realise.

And if you’re operating in Asian markets, or selling into them, or investing in them — the second race is the one that affects your business.

Tom Simpson is the founder of Digital in Asia. He has worked across gaming, advertising, and digital markets in Asia for twenty years.

Inside China’s AI Machine: Models, Chips, Strategy, and What Comes Next