A clear inflection point in AI adoption
On a March morning in Singapore, the country’s legions of technology enthusiasts and beginners flocked to a Tencent Cloud OpenClaw event, waiting to install a new AI agent platform called OpenClaw. Essentially, the attendees were there to “raise the lobster”, a rather affectionate term referring to the training of OpenClaw agents to automate tasks. Unlike conventional reactive AI models that respond to explicit instructions, agentic AI refers to autonomous AI systems capable of planning, reasoning and executing multi‑step tasks independently to achieve a defined objective with minimal human supervision.
Scenes of such “lobster farming” are not unique to Singapore; China has been feverishly busy “raising lobsters”, driven by a powerful mix of fear of missing out and a government push to accelerate AI adoption. While OpenClaw might have been a short-lived catch given how Beijing wasted no time in stepping in with official warnings over cybersecurity concerns, the lobster frenzy reveals something deeper: China’s growing strength in transforming AI research and capabilities into scalable and market-viable applications.
To help quantify and understand the seismic shift in China’s AI landscape, it is instructive to look at Table 1 below showing global token consumption on a given week. Tokens are the basic and smallest units of data in large language models and consumption is structurally higher in China due to the rapid expansion of AI agents and multi-step workflow, which have a prodigious appetite for data.
Table 1: Top 10 weekly ranking in terms of token consumption by model
Source: OpenRouter, Jefferies. April 2026
Little wonder that leading Chinese cloud service providers including Alibaba Cloud and Baidu Intelligent Cloud have raised prices for AI computing power and storage services, in response to massive jumps in daily token consumption, high demand for AI models and rising supply chain costs for hardware such as AI chips and memory.
China’s policy tailwinds in AI development
To get a grip on how Beijing is betting big on technology and innovation to drive its next phase of economic growth, it is useful to consider the terms that made the most appearances in this year’s government work report: artificial intelligence, high-quality development, and scientific and technological innovation. Interestingly, this year’s report also featured AI agents for the very first time.
For those less inclined to read between the lines, the 2026 Government Work Report also spelt out the country’s broader ambition clearly: “we will advance and expand the AI Plus Initiative”. This involves a focus on promoting faster AI application adoption and supporting open-source AI communities. First announced in 2024 by Chinese Premier Li Qiang, the AI Plus initiative includes a target of integrating AI across all sectors of China’s economy to drive productivity and technological self-sufficiency.
Beyond this, China’s 15th Five Year Plan offered few material surprises, instead underscoring continuity in the Chinese Communist Party’s strategic ambitions—namely its aim to achieve technological supremacy and self-reliance. Against this backdrop, we see favourable structural growth opportunities in AI emerging across China’s technology ecosystem, with leading firms increasingly aligning their business strategies with national priorities.
A different vision for AI from the US
Currently, China is probably the only competitor to match the US in the AI development race—so how does the country stack up?
Capital expenditure (capex) patterns differentiate China’s AI trajectory from that of the US, with industry research forecasts projecting US AI capex by the biggest hyperscalers to reach a staggering USD 700 billion. Thus, the US continues to retain an edge in frontier compute model and research depth. China, by contrast, is compressing commercialisation cycles as it remains constrained by limited access to high‑end chips. This reinforces China’s AI champions’ strategic emphasis on efficiency and self‑reliance in computing technologies and is reflected in the successful IPOs of firms such as MiniMax and Zhipu, alongside ongoing listings by memory manufacturers Kunlunxin, ChangXin Memory Technologies and Yangtze Memory Technologies Corp. Minimax and Zhipu are part of China’s six “AI tigers”—companies building large language models to rival the likes of OpenAI and Anthropic.
China’s progress in AI is perhaps best shown through its strength in model optimisation, energy efficiency, inference costs and vertical integration. These advantages have allowed rapid translation of AI capability into everyday use cases. During the Lunar New Year period, for instance, China’s tech titans launched a large-scale marketing blitz to promote their AI apps (essentially handing out cash to lure in users). Alibaba offered 25 yuan vouchers that could be used on food, drinks and other purchases made through its AI platform, Qwen. Another tech giant, ByteDance, held a lucky draw during a televised gala where audiences could participate through the company’s Doubao chatbot app for a chance to win gifts like robots, drones and 3D printers, as well as red packets (gifts handed out during holidays and special occasions) of up to 8,888 yuan (eight is deemed as a particularly auspicious number) in cash. More recently, Tencent launched a tool to integrate its WeChat ecosystem with the OpenClaw agent, transforming the messaging platform into a command layer for task execution. The agentic transformation of commercial services enables the maximal integration of user services, reinforcing engagement and strengthening user retention.
What we are watching over the next two to three years
Over the next few years, we expect Chinese developers to further narrow the gap with their US peers amid continued progress in domestic large language model (LLM) development. In late April, Chinese AI upstart DeepSeek launched its much-anticipated new model, the V4. While the latest model is unlikely to cause the same type of stir in the markets as the launch of the R1 reasoning model in 2025, initial feedback has shown significant performance improvements over the previous iteration, V3.2. Tencent is also set to launch a new version, Hy3, which should include significant advances in complex reasoning and coding capabilities. We expect Tencent to focus on the integration of agent capabilities and coding, eventually leveraging Hy3 to facilitate its gaming production cycle.
Chart 1: Inference costs by intelligence class
Source: Artificial Analysis. https://artificialanalysis.ai/models/gpt-5-4#pricing. April 2026
At the same time, we also expect inference costs to continue trending down. AI inference—the process of running data through a trained model to generate outputs—is an important aspect of revenue generation with AI infrastructure. Currently, Chinese developers maintain a cost advantage, with their inference costs significantly below those of US peers. Over time, declining inference costs may expand the range of economically viable AI-driven applications and services, support faster deployment of their AI solutions and ultimately enhance businesses profitability.
Chart 2: Comparison of modalities between Chinese and global LLMs
Source: Artificial Analysis. https://artificialanalysis.ai/models/gpt-5-4#pricing. April 2026
Multimodal functionality is also becoming more of a “core feature” rather than an add-on or point of differentiation. Developers are releasing more models with multimodal capabilities that combine text, images, video and audio capabilities. Examples include Kuaishou’s KlingAI and ByteDance’s Seedance in the AI-generated video space, which have made waves among users with their abilities to transform script prompts into realistic movie scenes. It might not be obvious to some of us as we scroll through our phones, but don’t be surprised if that cinema-quality video on your screen is a made-in-China AI product.
Pragmatism is at times said to be the core of the Chinese DNA, and the Chinese AI ecosystem reflects this. Chinese companies are now focused more on monetisation, and they are more proactive in embedding agentic capabilities into their existing business models spanning areas such as advertising workflows, gaming production and customer engagement tools.
Upside to AI capex and R&D investment trajectory
While US firms’ capex and research and development (R&D) spending dwarfs those of even the most ambitious Chinese tech companies, we still see upside to China’s AI capex outlook and R&D intensity. Management commentary continues to suggest potential upgrades to already substantial multi-year AI investment plans, with Alibaba, for example, signalling the possibility of exceeding its three-year 380 billion yuan investment plan to advance its cloud computing capabilities and AI infrastructure. Meanwhile, Tencent and Baidu have continued to demonstrate commitment to AI-related capex. At the same time, IPO activity among domestic AI chipmakers could help strengthen R&D capabilities and support continued performance catch-up, alongside the ramping-up of domestic chip capacity. Improved availability of imported high-end chips such as NVIDIA H200 also provides incremental support to the country’s AI needs after the Trump administration gave the nod to China-bound sales of Nvidia’s H200 chips (under certain conditions, of course) earlier this year.
Table 2: Internet leaders vs emerging AI labs in China
Source: Amova Asset Management, April 2026
From a business model perspective, we believe these are exciting times and are keeping an eye on emerging AI incumbents. While established internet leaders continue to prioritise the domestic consumer market—by building AI chatbot apps or integrating AI chatbots into their super apps to optimise advertising traffic—emerging AI incumbents are pursuing a different path. These companies are more focused on enterprise services and overseas markets, supported by improving model performance and cost advantages. Over time, this positioning may allow them to capture more scalable and diversified revenue streams as AI adoption broadens beyond consumer-facing applications.