A player logs into a wuxia game and books a table at an inn — by talking to a non-player character that remembers their last quest. Across the city, a music fan hums a melody and watches it become a finished, distributable track. Neither experience begins with a search box.
NetEase (网易) rarely shows up in the same breath as the model labs that grab headlines. Yet the company has built one of China's broadest portfolios of applied AI, and almost none of it looks like a chatbot.
The game that writes its own stories
NetEase's most visible AI work lives inside 逆水寒 (Justice), a massively multiplayer title that became a testbed for generative characters. According to the company's WAIC 2026 showcase materials, players have created more than 40 million AI "门客" (retainers/companions) — non-player characters with their own personalities and memory — and produced nearly 20 million user-generated story scenes through an "AI film-crew" (AI剧组) mode that lets players direct rather than merely consume.
The significance is not the novelty but the scale. Tens of millions of persistent AI characters inside a single live game is a different order of deployment from a demo video. It forces hard engineering on memory, moderation, and cost that pure research releases never confront.
A music platform that composes for you
On NetEase Cloud Music (网易云音乐), the company rolled out 网易天音 (Tianyin), an AI music-creation platform that stitches together lyric writing, composition, arrangement, singing, and distribution. New features added dedicated voice training and one-click publishing, and the platform ran creation contests to seed a community.
The pitch is democratization with a business loop: lower the barrier to make a song, keep the creator inside NetEase's ecosystem, and turn the output into publishable catalog. It is AI as a creative co-pilot bound to a distribution network, not a standalone generator.
AI that goes to the office and the dock
NetEase's AI also leaves the screen. At WAIC 2026 the company demonstrated Youdao (有道) OpenPods AI earbuds that transcribe meetings, interpret across languages in real time, and draft structured notes and to-do lists — a wearable agent for mobile work. A separate office agent, LobsterAI, was shown breaking down tasks and handling cross-platform busywork.
The most physical example comes from 网易灵动 (Lingdong), NetEase's embodied-intelligence brand. Its intelligent hold-cleaning robot for bulk-cargo ships — a dirty, dangerous job — has been deployed commercially across more than ten provinces, with the company citing over 700,000 safe operational tasks under "human-robot crewed" and "lights-out worksite" modes. That is AI measured in injuries avoided, not prompts answered.
The research bet behind it all
Much of this traces back to NetEase Fuxi Lab (伏羲实验室), the company's long-running AI research organization. In 2025, Fuxi's work on key technologies for 3D digital humans, developed with Zhejiang University and others, won a second-class prize in the National Science and Technology Progress Award — a rare state-level endorsement for game-adjacent research.
The spending backs the ambition. NetEase states it has invested more than ¥80 billion (≈ US$11B / HK$88B) in R&D over the past five years, a sum that funds both the flashy product features and the less visible simulation and generation stacks underneath them.
Why embed instead of launch
NetEase's instinct is to dissolve the model into the product rather than ship a general assistant. A game character, a music tool, an earbud, a ship-cleaning robot — each is a narrower, better-defined problem than "be helpful." Narrow problems are easier to make reliable, and reliability is what turns a demo into revenue.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, NetEase illustrates a second Chinese AI path: not the open-weight race for the biggest model, but the quiet insertion of models wherever users already spend time and money. The risk is fragmentation — a dozen AI features that never consolidate into a platform — but the near-term traction is real.
Honest limitations
Most of the deployment figures here (40 million companions, 20 million scenes, 700,000 robot tasks) come from NetEase's own WAIC presentations and affiliated media, not from independent audits; treat them as company-reported scale, not verified counts. I have not seen published cost-per-interaction, retention, or moderation-failure rates for the in-game AI characters, which matter for whether this is sustainable or subsidized. The National Science and Technology Progress Award citation is a real honor but does not by itself validate commercial performance. Readers should view NetEase's AI as impressively broad applied engineering rather than a single breakthrough model.
What readers can do now
- If you build consumer products, study NetEase's "AI inside the workflow" pattern: attach generation to an existing creation-and-distribution loop instead of building a separate chat surface.
- Game and content studios should watch 逆水寒's companion model as a live template for persistent, memory-bearing NPCs at scale.
- Enterprises evaluating AI should note NetEase Lingdong's ship-cleaning robot as evidence that embodied AI is already commercial in narrow, high-risk, low-glamour jobs — a more useful proof point than most lab demos.