WAIC 2026 Shifts Gears: Major LLM Players Pivot to Commercial Deployment as China's Domestic AI Chips Steal the Spotlight

The July heatwave in Shanghai was matched by the bustling crowds inside the exhibition halls of the 2026 World Artificial Intelligence Conference (WAIC). Now in its ninth year, WAIC surpassed 100,000 square meters of exhibition space for the first time, with over 1,100 leading enterprises and tech startups showcasing more than 3,000 cutting-edge products. However, unlike previous years when LLM companies dominated the spotlight during the "war of a hundred models," the winds have shifted subtly but profoundly: the once-noisy battle for positioning is giving way to a sober race for real-world deployment. AI agents, robotics, and computing chips have moved to the forefront, while the former "Six Little Dragons" of China's LLM sector are charting their own paths of advance, retreat, and transformation.

The LLM "Six Little Dragons": Some Are Absent, Some Sell Phones, Others Dig Deep into Verticals

Inside and outside the exhibition halls, the participation strategies of LLM companies have become a barometer of their strategic intent. The most notable absentee was Zhipu AI, which has recently been on a tear in the secondary market. The company explained to Boot Camp that it chose not to exhibit at this year's WAIC because it is in a quiet period ahead of its planned A-share listing. Another high-profile company, DeepSeek, remained "invisible" at WAIC as usual, never setting up an independent booth. Yet its models were ubiquitous in third-party products throughout the venue, evoking the sense that while it was absent from the scene, its legend was everywhere.

In contrast, fellow "Six Little Dragons" members MiniMax and Moonshot AI both chose to exhibit, albeit with very different postures. MiniMax's booth mainly showcased its multimodal creation agent, MiniMax Hub, and its AI coding tool, MiniMax Code, alongside a dedicated area displaying the company's cutting-edge research papers. Moonshot AI was remarkably low-key, with a booth featuring only a table, a few chairs, and several posters. A staff member said, "We just released K3 recently, and we still prefer to let the product speak for itself." The Kimi K3 model, released by Moonshot AI on the eve of WAIC, boasts 2.8 trillion parameters, making it the world's first open-source model at the 3-trillion-parameter level. It topped the global leaderboard on the third-party evaluation platform Arena AI's Code Arena with a score of 1,679.

The most eye-catching product at the conference belonged to StepFun. Its AI smartphone, the StepX Neo, released just before WAIC, was a massive hit on the show floor. Even though visitors could only experience it under the guidance of staff, it drew a constant stream of crowds throughout the three-day event. This marks a decisive move for on-device AI hardware from the demo phase into reality.

Furthermore, two companies that pivoted to vertical sectors in the later stages of the "war of a hundred models"—Baichuan AI and 01.AI—both returned to WAIC this year after being absent last year. Baichuan AI showcased for the first time the real-world deployment results of its Baichuan-M4 model in specialized medical scenarios such as oncology and pediatrics, signaling a complete shift to the medical track. 01.AI brought B2B products like the "Boss Agent," "Top Sales AI," and "Investment Officer AI," targeting business leaders and frontline staff. Its on-site staff were busy juggling a book signing for founder Kai-Fu Lee with livestreaming to connect with potential clients.

Computing Power Hall Debuts: SuperPoD Clusters Grab Center Stage, Challenging Nvidia's Moat

If the booths of LLM companies revealed a sobering focus on commercialization, the "Chip-Compute Fusion Hall," making its debut as an independent pavilion this year, was filled with the combative spirit of hardcore competition. Nearly all mainstream Chinese GPU manufacturers—including Huawei Ascend, Enflame Technology, Moore Threads, and MetaX—were present. The centerpiece of their booths was no longer individual chips but massive, all-black cabinets: SuperPoDs.

The Huawei Ascend 950 SuperPoD (Atlas 950 SuperPoD) made its debut as a physical machine, interconnecting 1,024 Ascend cards to deliver 1 EFLOPS of FP8 computing power. Enflame's CloudBlazer ESL64-O SuperPoD, MetaX's newly launched XiJing S600 SuperPoD, and Moore Threads' KUAE 10,000-card cluster—these "big beasts" dominated the entire hall. ZTE also partnered with Lightelligence, Biren Technology, and others to build a domestic Matrix SuperPoD. Baidu's Kunlun Core Tianchi SuperPoD has also completed adaptation for mainstream models like ERNIE and DeepSeek, boosting inference efficiency by 50%. Huatai Securities has dubbed 2026 the "Year One for China's SuperPoD," forecasting the market could reach 341.4 billion yuan (approximately $50.4 billion) by 2028.

The SuperPoD craze fundamentally represents a qualitative shift in how computing power is organized. The competition used to be about single-card performance; now, it is about making hundreds or thousands of chips work together as seamlessly as a single chip. A staff member at MetaX's booth told Phoenix New Media Technology that its SuperPoD "has 64 cards in a single cabinet and can scale horizontally to 10,000 cards. The biggest difference is its 'three-zero' design—no cables"—drastically reducing transmission loss. A staff member from Enflame Technology was blunt, stating that its clients include internet companies like Tencent and Meitu, and "in the end, everyone comes back to the issue of cost-effectiveness."

However, even as the hardware specifications of China's domestic computing power race to catch up, the moat Nvidia has built with its CUDA software ecosystem remains formidable. A relevant person in charge at TsingMicro told Phoenix New Media Technology that they have joined forces with nine entities, including Huawei, Cambricon, and Moore Threads, to build a joint ecosystem with the Beijing Academy of Artificial Intelligence (BAAI), with the goal of replacing CUDA at 80% of the functional level this year. Moore Threads is a radical player in the "domestic chips, domestic models" movement. Its KUAE 10,000-card intelligent computing cluster has been successfully deployed and has completed the full training of a MoE-236B foundational model from scratch. Its estimated revenue for the first half of the year is 1.65 billion to 1.75 billion yuan (approximately $243.7 million to $258.5 million), 2.35 to 2.49 times the figure from the same period last year.

"Token Factories" Proliferate, Supply Chain Synergies Emerge

"Token Factory" was the most frequently mentioned term in the exhibition halls this year. Tsinghua University-affiliated company Qingcheng Jizhi partnered with computing power providers to create an interactive "Token Factory" space at the World Expo Exhibition Hall, showcasing its self-developed Chitu inference engine, natively optimized for domestic chips. The company's co-founder, Tang Shizhi, gave an example: new models like DeepSeek natively use FP4 precision for training, but most domestic chips do not support this in hardware. Traditional solutions require precision conversion, which doubles video memory usage. Chitu, however, uses a self-developed software floating-point simulation computing architecture, reducing the number of servers needed to deploy the full version of DeepSeek from four to one, drastically cutting hardware costs.

This deep coupling—from models to chips to inference engines—is outlining a clear path for synergy across China's domestic AI industry chain. Upstream, models like Kimi K3, DeepSeek V4, and Zhipu GLM-5.2 are catching up to global leaders in performance, while their API pricing is only a fraction of that of their overseas counterparts. In the midstream, these models are proactively adapted to domestic chips from Huawei Ascend, Moore Threads, and MetaX on the very day of their release, achieving "Day 0 adaptation." Downstream, domestic chipmakers are pivoting to compete on SuperPoDs, system efficiency, and the cost per token for deployment. As one industry insider remarked at the conference, "It's actually not that complicated. In the end, the product has to speak for itself."

As the tide recedes, the players still at the table are no longer competing on buzz and parameters, but on who has truly found the right application scenarios and created economic value. From the pragmatic pivot of LLM companies to the collective show of force by China's domestic computing power clusters, WAIC 2026 sent a clear signal: the "mid-game battle" for China's AI industry has begun. A full-scale, supply-chain-wide encirclement, centered on cost-effectiveness and real-world application, is now well underway.