由 Allen Zeng Huawei's developer conference HDC 2026 opens today in Dongguan's Songshan Lake. HarmonyOS 7.0 is the headline. But the real story happened a week ago, 1,200 kilometers away in Shanghai, where Huawei Cloud dropped a barrage of AI agent infrastructure products that nobody in the English-language press paid enough attention to.
The company isn't building another chatbot. It's building the road, the power grid, and the traffic control system — then handing out free construction permits to anyone who wants to build on top.
This is a fundamentally different bet than what Alibaba, Tencent, or ByteDance are making. And it might actually work.
What Huawei Actually Launched
On June 5 at the INSPIRE conference in Shanghai, Huawei Cloud introduced something it calls "Agentic Infra" — a new paradigm for AI infrastructure. The pitch sounds like standard cloud marketing until you look at the numbers.
The AICS cluster delivers 200 EFLOPS of computing power across 100,000 interconnected cards. Token generation latency under 10 milliseconds. Throughput of 5 million tokens per second per thousand cards. These aren't demo numbers — they're the kind of specs you need when thousands of AI agents are running autonomous multi-hour tasks simultaneously, each one consuming tokens at rates that would bankrupt most cloud deployments.
Then there's the memory layer. The AMS storage solution provides petabyte-scale context memory with NPU-direct access, enabling agents to maintain state across day-long tasks. Most agent frameworks today crash or lose context after a few hours. Huawei is engineering for the scenario where an agent needs to remember what it was doing three days ago.
The scheduling engine — Volcano Next — pools training and inference workloads together, boosting resource utilization by over 30%. The security sandbox spins up in 100 milliseconds and handles 100,000 batch creations per minute.
None of this is flashy. It's the unglamorous plumbing that determines whether AI agents actually work at scale or remain impressive demos that break under real load.
The OS Play Nobody Noticed
HDC 2026 runs June 12-14, and HarmonyOS 7.0 is confirmed — the agenda leaked on QQ News shows sessions on "spatial aesthetics" and new UI components like "Flash Control Ball" and "Flash Control Window." But buried in the official description is the phrase everyone should be watching: "HarmonyOS AI core capabilities."
HarmonyOS already runs on over 900 million devices — phones, tablets, cars, TVs, watches, and an expanding universe of IoT endpoints. If Huawei embeds an agent runtime directly into the operating system, every one of those devices becomes a potential agent endpoint. Not an app you install. Not a web interface you visit. An agent that lives in the OS, with access to your calendar, your files, your connected devices, and the entire Huawei cloud backend.
This is the same play Microsoft is attempting with the Windows Agent Framework. But Huawei has one advantage Microsoft doesn't: it controls the entire stack from silicon to OS to cloud. The Kunpeng CPU, the Ascend NPU, HarmonyOS, and Huawei Cloud all report to the same product roadmap. Microsoft has to negotiate with Intel, AMD, and Qualcomm. Huawei just builds it.
The Open-Source Gambit
Then there's openJiuwen — the open-source version of Huawei's AgentArts enterprise agent platform, with over 90% code parity with the commercial version. This is a calculated move that deserves more analysis than it's getting.
Open-source agent frameworks like OpenClaw captured the developer imagination in early 2026, hitting 350,000 GitHub stars. But OpenClaw's momentum has cooled — the DIGITIMES report on HDC 2026 explicitly frames the conference as a test of whether AI agents have staying power "after OpenClaw hype loses steam."
Huawei's answer: give developers a production-grade agent platform for free, with near-identical code to what enterprise customers pay for. The strategy is transparent — grow the ecosystem on openJiuwen, then convert serious users to the managed AgentArts platform on Huawei Cloud. It's the Red Hat playbook applied to AI agents, and it's worked before.
The catch: Huawei Cloud is a distant third in China's public cloud market behind Alibaba Cloud and Tencent Cloud. Developers are already comfortable on Alibaba's ecosystem. openJiuwen might be free, but moving your entire agent infrastructure to a new cloud provider is not.
Why This Is a Different Game
Here's what separates Huawei's approach from the competition.
Alibaba is betting on vertical integration — Qianwen connects to Taobao's supply chain, letting users order coffee and laptops with natural language. Tencent is betting on distribution — WeChat's 1.3 billion users are the ultimate agent distribution channel. ByteDance is betting on the commerce loop — Doubao's 345 million monthly active users plugging into Douyin's content-to-transaction pipeline.
All three are fighting over the agent "entry point." Who owns the interface where users talk to the AI.
Huawei is fighting over the agent "road." Who owns the infrastructure that agents run on, the memory they store, the security they trust, and the operating system they live inside. If you control the road, you don't need to own every car that drives on it — you collect tolls from all of them.
The real question is whether Huawei can build enough traffic on that road to justify the toll booths. 200 EFLOPS of computing power is impressive. A petabyte-scale memory layer is impressive. But if only a handful of enterprise customers are actually deploying production agents on Huawei Cloud, the infrastructure is a cost center, not a moat.
This infrastructure-first play also aligns with what Chinese tech companies are genuinely good at. Building massive, reliable, boring backend systems is a Chinese engineering strength. The flashy consumer-facing AI apps get the headlines — Doubao's 345 million users, Qianwen's 166 million — but the infrastructure that makes those apps work is where the sustainable competitive advantage lives.
The One Company That Can Do End-to-End
There's something unique about Huawei that's easy to miss if you're not looking at the full picture. It's the only Chinese tech company — possibly the only company globally — with a complete AI agent stack spanning three layers.
At the bottom: silicon. The Ascend NPU series, designed specifically for AI inference workloads, not repurposed from gaming GPUs. In the middle: cloud infrastructure. Agentic Infra, ModelArts Next, AgentArts, all running on Huawei's own data centers. At the top: the operating system. HarmonyOS, running on 900 million devices, with AI core capabilities baked into the next version.
Nvidia has the silicon but not the cloud platform. AWS has the cloud but not the silicon or the OS. Apple has the OS and soon the silicon, but not the enterprise cloud infrastructure. Google has the cloud and the models but not the device OS at scale.
Whether this full-stack advantage translates into market share is an open question. Huawei's track record of turning technical integration into commercial dominance is mixed — strong in telecom infrastructure, still building in cloud services. The HDC 2026 announcements will tell us more. But the architecture is there. The pieces are on the board.
Meanwhile, China's enterprise AI agent market hit 21.2 billion yuan in 2025 and is expected to reach 44.9 billion yuan this year. Whoever owns the infrastructure layer of that market owns the most defensible position in China's AI industry. Huawei has spent two decades building infrastructure for telecom. Now it's doing the same for intelligence.