Morgan Stanley: China's AI Differentiates Through Cost-Efficiency and Multimodal Capabilities
0xBroomberg
Morgan Stanley lays out China's AI trajectory: rather than chasing the U.S. across every dimension, China is carving a distinct position on cost-performance and multimodal capabilities, with domestic AI chip self-sufficiency projected at 86% by 2030.
How much has Chinese AI actually closed the gap?
Of the world's top ten foundation models, the U.S. and China each account for roughly half — but the top three remain American. Chinese models cluster from fourth to tenth.
On an intelligence timeline, Chinese models trail the leading U.S. models by roughly 3 to 6 months. This means → China is not a generation behind — it is half a step behind, closing fast but still short of the very top.
The core selling point is not "smarter" but "equally smart at a fraction of the price" — high intelligence, low per-token cost. The discount to U.S. models has narrowed but still sits at roughly 15–20%.
Multimodal capabilities — where does China's edge come from?
Many leading U.S. foundation models remain text-centric. Chinese AI labs, by contrast, built in video and audio generation from the start.
The result: the world's top three video-generation models all come from China — ByteDance, Alibaba, and Kuaishou.
In plain terms = the U.S. leads in "text chat"; China is ahead in "write and shoot." This reflects fundamentally different resource-allocation logic between the two paths.
What does it take to build a frontier model? China is weak on one factor, level on one, and strong on one
Compute: clearly constrained. Export controls put a hard ceiling on China's access to top-end chips.
Data: no material gap with the U.S. China's internet ecosystem is large enough that training data is not the bottleneck.
Algorithms: potentially ahead. Chinese universities produce a massive pipeline of math and engineering talent. Architecture innovations — MoE (Mixture of Experts, splitting one large model into cooperating "small specialists"), various attention mechanisms — frequently originate from Chinese research teams.
Chips are constrained — now what? From "transistor density" to "system scale"
Morgan Stanley projects the Chinese AI accelerator chip market will reach roughly $67 billion by 2030, with a domestic self-sufficiency rate of roughly 86%.
Advanced-node capacity rises from about 8,000 wafers per month in 2025 to about 50,000 by 2030; yields climb from roughly 20% to roughly 50%.
This means → the process gap persists, but China is taking a detour: multi-chip packaging at the chip level, massive rack-and-cluster scaling at the system level, and ecosystem-wide capacity coordination at the manufacturing level. The contest shifts from "how strong is one chip" to "how large is the whole system."
Which domestic GPU makers can break out?
China's GPU ecosystem now includes over a dozen vendors. Major buyers prioritize strategically important domestic suppliers to secure self-sufficiency channels.
Morgan Stanley's call: before 2027, the addressable market for third-party vendors is limited. After 2028, supply opens up — and that is when domestic GPU companies will truly have to prove themselves.
In plain terms = the current phase is policy-driven "secure supply." Real market competition starts in two to three years, once capacity catches up.
How much is going into AI infrastructure? China's CAPEX is just one-tenth of America's
Morgan Stanley projects Chinese hyperscaler cumulative CAPEX at roughly $70 billion through 2027, versus roughly $1 trillion in the U.S. over the same period — China's spend is about 10% of America's.
Even so, AI already contributes roughly 10% of cloud revenue, projected to rise to about 40% by 2029. Future incremental demand will tilt toward inference rather than training.
This means → China's AI infrastructure ceiling is far from reached — if demand overshoots, CAPEX still has significant room to rise. And an inference-heavy demand mix plays directly to the cost-performance playbook.
Content is for reference only, not financial advice.