The Great GPU Bypass: Morgan Stanley Breaks Down China’s "Agentic" AI Pivot
While Washington continues to tighten the silicon curtain, expecting chip export controls to strangle China’s artificial intelligence ambitions, a new report from Morgan Stanley suggests the world’s second-largest economy is executing a strategic pivot. The narrative found in the bank’s December 8, 2025, note titled "China AI: What's Going On" is clear: if you can’t out-muscle the US on raw hardware, you out-engineer them on software efficiency.
Morgan Stanley’s Asia team, led by Gary Yu, argues that China’s path to AI supremacy has shifted toward "scaling model capabilities via software/engineering efficiency to overcome GPU supply chain constraints." The focus now isn't just on training larger models, but on "Agentic AI"—autonomous agents capable of reasoning, coding, and navigating complex tasks.
The Rise of Efficiency: DeepSeek and Hunyuan
The hardware gap is being closed by algorithmic breakthroughs. Morgan Stanley highlights the "DeepSeek V3.2" series as a prime example of this efficiency-first doctrine. DeepSeek has implemented what the analysts call "DeepSeek Sparse Attention," a technological leap that significantly reduces computing power consumption when processing long-text content.
The results are showing up in the benchmarks. According to the report, DeepSeek V3.2 now ranks "second only to Gemini 3.0 Pro in the coding benchmark and leads the world's math benchmark."
Simultaneously, Tencent has entered the fray with its Hunyuan 2.0 model, launched just days ago on December 5, 2025. This is a massive proprietary beast with 406 billion total parameters. However, consistent with the efficiency narrative, only 32 billion are "activated" parameters, allowing for aggressive pricing.
The bank notes that Hunyuan 2.0 boasts "low API pricing" and currently leads in "math, agentic coding, and reasoning benchmarks." Both models are already being integrated into consumer-facing applications, signaling a rapid transition from lab theory to commercial deployment.
The "Super App" Civil War
However, the road to an AI-integrated economy is paved with internecine conflict. While the technology improves, China’s infamous "walled gardens" remain a significant barrier to the "Agentic AI" era.
Morgan Stanley highlights a telling setback for the owner of TikTok. When ByteDance released its "Doubao" smartphone AI assistant, it attempted to act as a universal layer over other applications. The reaction from incumbents was swift and brutal.
"Recently, WeChat, Taobao, and other apps blocked Doubao assistant access on sample phones," the analysts write. While the blockage was officially attributed to "security concerns," the report suggests the real motivation was "to prevent Doubao from taking user traffic."
This highlights a structural problem for Chinese AI agents: "The main hurdle for smartphone AI agents is building a multi-tool ecosystem and competing for main user entry points." In a market dominated by Super Apps, an AI agent that cannot cross application borders is severely handicapped.
The Video Frontier and Monetization Struggles
Beyond text and code, the battle for multimodal dominance—video and audio—is heating up, though monetization remains elusive.
Kuaishou has announced its Kling O1 and Kling 2.6 updates. The O1 is a "unified multimodal video model" that creates consistent character control, while the 2.6 update adds native audio generation for synchronized dialogue. Morgan Stanley sees the tech improving, but warns that "further upside will likely depend upon ongoing user engagement and monetization."
Similarly, Bilibili officially launched its "Huasheng" video generation model on November 24, 2025. The platform is attempting a direct subscription model, charging RMB 248 yuan (US$34) per year for the service. The bank remains cautious here, noting that "monetization remains key amid intense competition."
Investment Implications
Despite the supply chain headwinds and fierce domestic competition, Morgan Stanley holds an "Overweight" rating on both Tencent and Alibaba.
The bank’s valuation models for Tencent, based on a sum-of-the-parts approach, suggest resilience in its core gaming and social businesses, with "2C AI adoption ramp-up" listed as a key upside risk. For Alibaba, the analysts point to "stronger AI demand to drive cloud revenue" as a potential catalyst to re-accelerate growth.
The takeaway for investors is straightforward: The hardware embargo has forced Chinese tech giants to become leaner and smarter. By focusing on "sparse attention" architectures and efficient parameter activation, Beijing's tech leviathans are engineering a route around the GPU bottleneck. The primary risk to their success may no longer be Washington, but rather their own refusal to tear down the walls between their digital ecosystems.