Weekly Dose of China Tech #04

Chinese AI models Kimi K3 and DeepSeek are forcing Silicon Valley to rethink assumptions about the AI race, with Moonshot pausing new Kimi K3 subscriptions less than 48 hours after launch due to overwhelming GPU demand. DeepSeek's open-source strategy and focus on AGI over commercialization, as described by founder Liang Wenfeng, challenge the industry's emphasis on scale and short-term profits.

Weekly Dose of China Tech 04 China's Two AI Shocks, Inside DeepSeek's Mind, AI Infrastructure Bottlenecks, Token Factories, AI Agents + One More Thing Hi friends, Welcome back. I have been thinking about a notable transition in China’s AI race this week. For the past two years, the biggest question was whether Chinese AI companies could build models good enough to compete with Silicon Valley. That question is starting to look different. Kimi K3 became one of the most talked-about AI releases of the year. Chinese models now dominate OpenRouter’s global usage rankings. DeepSeek has already changed how the industry thinks about open-source AI. The debate is no longer only about capability. It is about scale. But scale is exactly where the cracks appeared. Less than 48 hours after launch, Moonshot paused new Kimi K3 subscriptions because demand was overwhelming its GPU capacity. Building a frontier model was the first challenge. Turning it into a global-scale product is the next one. And that requires something very different: the GPUs, data centers, optical networks, and hidden supply chains that make AI possible at volume. And underneath all of this is a bigger question: if China’s AI companies can no longer be dismissed as followers, what kind of companies are they trying to become? Let’s get into it. This Week’s Features… Why Kimi K3 And DeepSeek Are Keeping Silicon Valley Up at Night ? For months, the AI race looked easy to understand. American companies built the frontier models, Chinese companies followed, and the competition was about who had more GPUs, bigger training runs, and deeper pockets. Then DeepSeek arrived. Then Kimi K3. And suddenly, the conversation changed. The interesting question is not why these models performed well. China has no shortage of serious AI labs — Z.ai, MiniMax, Alibaba, Tencent, and ByteDance have all built competitive systems. The question is why only two of them forced the industry to rethink its assumptions. DeepSeek challenged the belief that frontier AI progress would always require unlimited compute. Moonshot, the company behind Kimi, challenged the assumption that the hardest AI capabilities would remain concentrated inside American closed-source labs. The market reaction revealed the difference. Some models compete inside the existing AI race. Others force everyone to question whether the race itself has been defined correctly. This week’s piece explores why DeepSeek and Moonshot created a reaction that dozens of other Chinese AI companies did not .And why the biggest challenge to Silicon Valley may not simply be better models, but a different understanding of what makes an AI company valuable. The DeepSeek Doctrine Every technology company in the AI boom seems to be chasing the same things. More users. More revenue. More products. More markets. DeepSeek is doing something much stranger. It keeps saying no. No to chasing consumer traffic. No to building the next super app. No to maximizing short-term profits. No to closing its models. In a four-hour conversation with investors, founder Liang Wenfeng described a company that appears almost out of place in today’s AI race. While everyone else is rushing toward commercialization, DeepSeek seems focused on something much harder to measure: increasing its chances of eventually reaching AGI. At first glance, the strategy sounds irrational. Why release models openly? Why lower prices? Why ignore obvious business opportunities? But the logic becomes clearer when viewed through DeepSeek’s own framework. Lower costs expand access. Open source creates distribution. Efficiency matters when resources are limited. And a smaller organization may have advantages precisely because it cannot afford to chase everything. This week’s piece looks at the philosophy behind DeepSeek’s rise, and why its biggest competitive advantage may not be a model architecture — but a completely different definition of what winning looks like in the AI era. The News… i Kimi K3 Hit the World Stage. Then Reality Hit Back. Moonshot AI’s Kimi K3 launch created one of the strongest reactions to a Chinese AI model since DeepSeek. Within 48 hours, the company paused new subscriptions after demand overwhelmed its available GPU capacity. Existing users were unaffected, but new signups had to wait while Moonshot expanded infrastructure. The irony was difficult to miss. Kimi K3 is a 2.8 trillion-parameter model that reached the top tier of global benchmarks, including strong performance on AI Arena’s frontend coding rankings, and Moonshot plans to release open weights that would make it one of the largest open-weight frontier models ever released. The model itself was exactly the kind of breakthrough Chinese AI companies had been trying to prove was possible. But deployment exposed a different challenge. Building a powerful model and serving that model at global scale are two separate problems. Several users noted that Kimi K3 felt slower than leading US alternatives — the issue is not only model quality, but inference capacity, latency, and access to advanced computing resources. China’s AI labs have become much better at extracting more intelligence from limited resources, but export controls on advanced chips and semiconductor equipment mean the next challenge is no longer only building smarter models. It is building enough infrastructure to let those models run everywhere. ii Chinese AI Models Are Dominating Developer Usage. The Harder Question Is Retention. According to OpenRouter’s latest leaderboard, the five most-used AI models globally last week were all Chinese. Tencent’s Hy3 processed 11.8 trillion tokens, Xiaomi’s MiMo-V2.5 reached 9.37 trillion, DeepSeek V4 Flash recorded 5.34 trillion, Zhipu’s GLM 5.2 reached 3.57 trillion, and MiniMax M3 processed 3.46 trillion. Chinese models have now remained at the top of the ranking for 12 consecutive weeks. The numbers are striking, but they need context. OpenRouter measures developer usage, not consumer adoption, and high token volume can reflect free trials, experimentation, or developer testing rather than long-term commercial demand,Tencent Hy3, for example, is still in a free trial phase. The more interesting signal is not simply that Chinese models are being used, but that developers globally are increasingly willing to experiment with them. For years, the assumption was that Chinese AI models would mainly serve domestic users. OpenRouter suggests the conversation is becoming more global. The next question is whether usage can turn into durable ecosystems. iii The Hidden Winner of the AI Boom May Be the Company Testing the Connections. A Chinese optical testing equipment company just reported explosive growth. Lianxun Instrument, which develops testing systems for high-speed optical modules, expects first-half 2026 net profit to increase 800% to 900% year over year — and the reason is straightforward: AI data centers need faster connections. As AI clusters become larger, the bottleneck is no longer only computing power. Moving data between thousands of chips has become equally important, and optical modules operating at 400G, 800G, and the emerging 1.6T standard are becoming critical infrastructure, with every generation requiring more advanced testing equipment. Lianxun is only the second company globally capable of providing full testing coverage for 1.6T optical modules, behind US instrumentation giant Keysight. The company already holds the leading position in China’s optoelectronic testing equipment market and ranks third globally in optical communications testing. Q1 revenue grew 142% year over year at a 66.76% gross margin, and orders on hand exceeded RMB 3 billion. The AI boom is often described as a battle between GPU companies. But every GPU cluster needs an entire ecosystem around it, and sometimes the biggest winners are the companies building the tools that make the AI factories possible. iv China’s “Token Factories” Are Becoming the Next AI Infrastructure Race The next phase of China’s AI competition may not be about building another giant model. It may be about building the factories that produce AI output. Chinese AI infrastructure companies are increasingly describing their businesses as “token factories” — facilities designed to generate massive amounts of AI inference capacity at scale. One example is z.ai, which is building a 1GW-class AI data center powered entirely by domestic chips and has acquired heterogeneous-computing software company XCore Sigma. The acquisition reveals the real challenge facing Chinese AI infrastructure players: the problem is not simply getting chips, but making different chips work together. Because of export restrictions, Chinese companies cannot always rely on a single cutting-edge GPU platform. Instead, they must combine different hardware systems and optimize software across heterogeneous architectures — which makes utilization efficiency a genuine strategic advantage. A data center filled with less powerful chips can still compete if those chips are coordinated effectively. This is becoming one of the defining engineering challenges for China’s AI industry. The country may not be able to immediately replicate America’s hardware advantage, but it is being forced to develop a different skill set: how to extract maximum intelligence from every available piece of compute. That is why companies like Infinigence AI have attracted more than $325 million in investment as they move toward an IPO. The AI infrastructure race is expanding beyond chips. The next battlefield may be the software that makes imperfect hardware perform like a unified machine. v China’s Space Race Is Moving From Experiments to Operations For years, China’s commercial space industry was mostly defined by demonstrations, successful launches, experimental rockets, technology validation. That phase may now be ending. Dongfang Space’s Gravity-1 Y4 rocket successfully launched nine satellites into orbit from a sea platform near Shanghai, marking a shift from technology testing toward regular commercial operations. The company plans three additional Gravity-1 launches including missions related to large-scale satellite internet deployment, while its liquid-fuel Gravity-2 rocket is already undergoing ground testing and is expected to attempt its first flight in Q4 2026. The significance is not only the rocket itself. Commercial space companies are trying to build something closer to an operating system for the space economy: reusable launch capability, reliable schedules, and satellite deployment at scale. The same pattern has appeared across China’s technology sectors — the first challenge is proving something can work, and the second is turning it into a repeatable business. Commercial rockets are now entering that second stage. vi Alibaba Consolidates Its Agent Sprawl Alibaba is preparing to launch Qianwen Office, a unified AI agent platform for enterprise productivity, by merging three existing products — QoderWork, Wukong, and MuleRun — under DingTalk’s new CEO Chen Yusen. The move comes shortly after Tencent launched WorkBuddy, a desktop enterprise agent supporting MCP and more than 20 skill packages. The timing is not a coincidence. Alibaba had been running parallel agent products without a clear flagship, a strategy that made sense when the market was still being mapped but starts to look like fragmentation once a direct competitor ships a unified product. Consolidation is the easy part of this announcement. The harder question is whether QoderWork’s existing user base converts into genuine enterprise stickiness — the kind that survives a competitive market rather than just an open one. vii Xiaomi Raises Its Shipment Target. The Memory Market Notices. Supply chain sources say Xiaomi has quietly raised its 2026 smartphone shipment target from roughly 90 million to 110 million units, with most of the increase concentrated in lower-end models. The driver is a bet on the memory chip cycle turning favorably. That bet is already meeting resistance: OPPO and vivo have reportedly rejected Samsung’s third-quarter price quotes despite only modest increases, a signal that not everyone believes the memory boom has more room to run. Xiaomi’s revised target is a demand signal. The pushback from its competitors is a supply signal. Together, they suggest the memory market may be closer to its ceiling than Xiaomi’s optimism implies. viii A Chinese AI Model Was Used to Analyze an OpenAI Agent Security Incident Hugging Face disclosed a security investigation involving an OpenAI-powered autonomous AI agent that had been able to carry out a multi-step intrusion. To understand how the attack worked, Hugging Face researchers needed to analyze real attack commands and vulnerability payloads generated during the incident. But when they initially tested commercial frontier models, built-in safety guardrails blocked some of the analysis because the materials involved real offensive security techniques. The team then deployed GLM-5.2, a open model developed by Chinese AI company Z.ai, on its own infrastructure. Because the model could run locally, researchers were able to analyze sensitive security data without sending it to an external AI provider. Using GLM-5.2, Hugging Face analyzed more than 17,000 event records and reduced a process that would normally take days to several hours. For security teams, the ability to control where a model runs, how data is handled, and whether the system can be integrated into existing workflows may become just as important as benchmark performance. GLM-5.2 did not prove that it is the best AI model. But this case showed another dimension of competition in AI: not only intelligence, but deployability and control.