Phoenix 24
Speed now beats spectacle in AI.
Beijing, February 2026
China’s AI sector treated the run up to Lunar New Year as a strategic deadline, not a marketing season. A cluster of releases from Alibaba, ByteDance, and Zhipu AI arrived in close succession, signaling a shift from chat centric novelty to operational systems designed to execute tasks. The central bet is that the next competitive moat will be “agents” that can plan, navigate tools, and complete workflows end to end. In that framing, model quality matters, but the decisive advantage comes from distribution, infrastructure, and the ability to industrialize iteration under real user load.
Alibaba’s Qwen line was positioned as a move toward multimodal, multilingual execution at scale, with claims spanning text, image, and video handling across roughly two hundred languages. The more revealing part was not the breadth but the emphasis on agent speed and workflow execution, including form completion and web style navigation. Those are the unglamorous tasks that sit inside enterprise operations, where adoption depends on reliability, latency, and governance rather than clever replies. Alibaba’s pricing posture also matters, because cost curves determine whether AI becomes a universal layer or a premium add on.
The infrastructure story behind that pricing is more important than any single benchmark. Alibaba has signaled a major multi year investment push into cloud and AI, a classic play to control the environment where models run and where enterprise customers enforce policy. Cloud scale is not just compute, it is compliance tooling, audit trails, identity management, and predictable service levels. If agents become the interface to procurement, HR, finance, and analytics, the firm that owns the runtime can win by being the default rather than the flashiest. This is where the competitive map starts to look less like a model race and more like a platform war.
ByteDance is pushing the same direction, but through consumer distribution surfaces that can be converted into enterprise entry points. Its chatbot updates were described as adding stronger reasoning and multi step task capability, and its generative video model, SeeDance, widened the battlefield into synthetic media production. That is strategically coherent for a company that already lives inside attention markets, because creator tooling and advertising workflows are natural ramps into business spend. When a firm controls both the model and the distribution channel, feedback loops tighten and iteration accelerates. Over time, that advantage can outweigh small differences in raw model capability.
Generative video also pulls the sector into a legal and reputational storm that is becoming inevitable. The Motion Picture Association criticized SeeDance as enabling large scale use of protected works, effectively framing the model as a rights evasion tool rather than a neutral technology. ByteDance’s response has been to stress safeguards, but the broader issue is structural: the stronger the output, the higher the liability, and the more governance becomes a competitive variable. The next stage of the market will likely reward companies that can demonstrate provenance, enforce policy, and survive litigation pressure without breaking the product. In other words, capability without control is a short lived advantage.
Zhipu’s GLM 5 release adds another axis that is less about features and more about strategic autonomy. The model is positioned as open source and oriented toward agent intelligence, long session work, and tool usage, which aligns with the global push toward systems that do tasks rather than only generate content. More consequential is the claim that the model was trained entirely on Huawei Ascend chips, presented as a step toward reducing dependence on US made semiconductor hardware. In a world shaped by export controls and sanctions risk, compute is no longer an input, it is leverage. Hardware independence, even partial, changes the credibility of a country’s AI roadmap.
DeepSeek remains a market moving presence in this landscape, partly because its rapid adoption has been treated as proof that capability can scale fast when cost and accessibility align. Expectations of an imminent new release around Lunar New Year were reinforced by reporting that the company could refresh its model stack, potentially replacing earlier versions associated with its assistant. Reports of an expanded context window also matter because context length is becoming a proxy for usable agent workflows, especially for long documents, planning tasks, and multi step reasoning. What makes DeepSeek strategically significant is not only performance claims, but the way it compresses timelines for competitors and triggers institutional reactions abroad. When a single model brand can move equity narratives and procurement anxiety, it becomes a geopolitical object whether it intends to or not.
Europe’s response highlights the emerging logic of AI blocs built on trust regimes rather than on raw capability. Several European governments have reportedly restricted public sector use of DeepSeek models due to data security and cybersecurity concerns, with additional cautionary moves in other capitals. This is not simply fear of foreign tech, it is a procurement reality in a world where models touch sensitive data and influence administrative decisions. Once governments treat AI tools as potential security liabilities, adoption shifts from individual departments to national policy. That creates friction for cross border model diffusion and pushes the global market toward jurisdictional segmentation.
The deeper pattern is that China’s pre holiday launch wave is a signal about the next phase of competition. Agents will be evaluated on reliability, cost, governance, and integration into institutional workflows, not on cleverness alone. The winners will be those who can bundle models, compute, distribution, and compliance into one durable operating layer, while managing legal exposure and export control risk. The contest is still technical, but the decisive terrain is institutional, and the strategic prize is becoming the default layer of execution inside the economy. In that environment, speed is not just a feature, it is a form of power.
Lo visible y lo oculto, en contexto. / The visible and the hidden, in context.