WAIC 2026: China's AI Industrial Strategy
“Ecosystem” has become a cliché in China tech. The problem is, after another WAIC, it’s still the most accurate description of what’s happening.
Things you might have missed
WAIC has taken over the feed, so here are three things happening elsewhere in China tech.
FROM OUR RECENT COVERAGE
BYD thinks it can overtake Toyota without entering the US. BYD sold 4.5 million vehicles last year versus Toyota’s 10.5 million. Its overseas chief says the company can close that six-million-vehicle gap within five years through Europe, Southeast Asia and Latin America.
Sam’s Club may reach RMB 200 billion in China revenue this year. That would put it nearly RMB 100 billion ahead of Hema, while warehouse clubs, instant retail and discount chains all compete for the same household spending.
Before we dive in, apologies for the delayed WAIC write-up. For the past two weeks, I’ve been helping run the inaugural Youth Tech China Trek, a separate project loosely affiliated with Tech Buzz China. We welcomed more than 50 parents and students from six countries to China to explore some of the country’s most advanced technology firsthand.
One of the highlights was simply watching 10- to 18-year-olds react to humanoid robots, autonomous robovans and factories. Seeing China through fresh eyes was a great reminder of just how quickly frontier tech is moving.
We’ll be running more cohorts later this year, so if you’re interested, you can learn more on our website.
Now, onto our biggest takeaways from WAIC—including some fascinating survey results shared by our partner, the Shenzhen Robotics Association, which surveyed industry participants during the conference. They also just joined X and will be posting more there, so please give them a follow.
If there was a single takeaway from WAIC, it was that China increasingly measures AI progress at the level of the system rather than the model. Its competitive advantage is not expected to come from winning every benchmark or releasing the highest-scoring foundation model. Instead, it lies in integrating manufacturing, energy, computing infrastructure, semiconductors, models and deployment into a single industrial ecosystem that can scale more rapidly and at lower cost.
In this framework, the frontier model is an input into the industrial system, not the end product of it.
China is attempting to integrate five existing strengths into that ecosystem:
How do you deploy AI at scale? Embodied intelligence.
China’s manufacturing ecosystem, spanning robots, sensors, motors, batteries and supporting hardware, enables AI advances to move more quickly from models into physical products.How do you power and scale AI? Energy and infrastructure.
Continued investment in power generation, intelligent computing infrastructure and industrial capacity provides the physical foundation needed to support AI deployment at scale.How do you reduce dependence on foreign technology? Domestic chips and system engineering.
Rather than relying solely on the best individual accelerator, China is investing in domestic chips, large-scale clusters, networking and hardware-software co-design to maximize the performance of the systems it can build.How do you maximize adoption? Open models and low-cost deployment.
Open-weight models, aggressive pricing and broad cloud availability encourage rapid domestic adoption while making Chinese AI easier to distribute internationally.How do you create demand beyond the technology itself? State-supported markets and international partnerships.
Procurement, technical standards, financing and international partnerships help accelerate deployment at home while expanding markets abroad.
Let’s look at each in turn.
How do you deploy AI at scale? Embodied intelligence.
If there was one thing that dominated WAIC, it was robots. Officially, it remained a general AI conference. In practice, the exhibition floor often felt like a robotics conference. More than 200 embodied AI companies exhibited this year, up from just over 80 last year, with more than 300 robots on display. They attracted the largest crowds and produced most of the demonstrations visitors stopped to watch. A frontier model, coding agent or inference architecture is difficult to demonstrate in a booth. A humanoid robot sorting packages, making coffee or walking through a crowd is not.
The atmosphere also felt different from last year. In 2025, robots and their teleoperators wandered throughout the exhibition halls, and simply seeing a humanoid move through a crowded public space without falling over felt like progress. This year, that was no longer enough. The emphasis shifted from proving robots could move to proving they could work. Nearly every major robotics company built some version of a real-world scenario: warehouse logistics, battery assembly, precision screwdriving, retail reception, hotel service, household assistance or factory material handling. The question had shifted from Can the robot walk? to Can it create economic value?
Whether the industry has answered that second question remains much less clear.
The demonstrations were more practical than last year’s, but they were also narrow, slow and highly structured. The robots showed real progress, particularly in manipulation, yet few demonstrations felt transformative. The industry has moved beyond proving that robots can move, but it has not yet shown they can perform broad categories of work reliably, economically and at commercial scale.
Interestingly, the people inside the industry largely agreed. A survey conducted by our collaborator Shenzhen Robotics Association during a side event at WAIC, covering 119 industry participants, found that 88% believed embodied AI models had improved, and nearly two-thirds said their confidence in the sector had increased. Yet only 39% thought the exhibition contained many meaningful breakthroughs, while 57% described most demonstrations as fairly ordinary. More than half believed product homogeneity had become worse, and nearly three-quarters said the biggest challenge was not hardware itself, but integrating hardware and AI models.
The bottleneck is no longer the building but the deploying.
One final theme came up repeatedly: hardware and software cannot be developed separately. Companies consistently argued that embodied AI requires co-design. Xynova, a dexterous-hand company developing both hardware and software, told us that today’s software processes only a small fraction of the tactile information future robotic hands will need. Improving dexterity is therefore not simply a matter of training a better model. Sensor design, data bandwidth, onboard compute and learning algorithms all have to evolve together. Unlike language models, which can often run on standardized servers, a robot’s intelligence is fundamentally constrained by the body it inhabits.
How do you power and scale AI? Energy and infrastructure.
Perhaps the least glamorous (but arguably one of the most important) themes at WAIC was not a new model or robot, but the continued build-out of the physical infrastructure behind China’s AI ambitions. If the conference’s robotics halls showcased what AI might eventually do, the investment announcements highlighted what will be required to support that future.
Perhaps the clearest signal came not from a product launch but from where governments and infrastructure providers said they were putting capital. Shanghai announced 32 major AI projects with planned investment exceeding RMB 40.9 billion (approximately US$6.0 billion), spanning intelligent computing infrastructure, embodied intelligence, AI agents and AI for science. Separately, China Unicom unveiled its UniAI initiative, committing more than RMB 25 billion (approximately US$3.7 billion) to intelligent computing infrastructure and AI applications in Shanghai. These announcements were accompanied by procurement programs, enterprise matchmaking sessions and application showcases, underscoring that deployment—not just technological progress—is becoming a central priority.
The exhibition itself reflected the same emphasis. Intelligent computing had become one of WAIC’s largest industry tracks, with more than 200 exhibitors focused on computing infrastructure, cloud platforms and related technologies. The message was straightforward: AI at national scale requires much more than better models. It also requires power, data centers, networking, financing and industrial capacity.
How do you reduce dependence on foreign technology? Domestic chips and system engineering.
The most important computing announcements at WAIC—and one of the main reasons I attended—were the AI computing systems. Unlike the frontier model companies, these infrastructure providers are much harder to meet, and their products are far less visible to the public.
The underlying strategy was not new. Over the past several years, and especially since Huawei introduced CloudMatrix384, Chinese companies have argued that competitiveness would come not only from building better individual chips, but from designing better systems around the chips they could build. Larger clusters, faster interconnects, unified memory, hardware-software co-design and specialized inference software could all increase the useful work those accelerators perform.
What stood out at WAIC was not a change in strategy, but how broadly that strategy has now been adopted across the industry. Huawei publicly demonstrated its 1,024-chip Atlas 950 SuperPoD, while outlining an architecture that scales to an 8,192-chip single logical computer and, ultimately, Atlas 950 SuperClusters with more than 500,000 Ascend chips. ZTE unveiled its OEX supernode architecture, paired with its dOCS Matrix optical switching fabric, as an open platform supporting accelerators from Biren, MetaX, Enflame, Iluvatar CoreX and Lightelligence rather than a single proprietary chip. Sugon showcased its Dawning 8000 (Dengfeng) AI supercluster, while Lenovo emphasized infrastructure optimized for producing “high-value tokens” rather than simply maximizing peak FLOPS.