A cloud used to sell you compute by the hour. Huawei Cloud now wants to sell you an environment where autonomous agents can run, remember and recover — at scale, and safely. At HUAWEI CONNECT 2026 in Shanghai (17–19 September), Huawei Cloud CEO Zhou Yuefeng (周跃峰) laid out what the company calls "Agentic Infra," and the pitch is less about chips and more about the plumbing agents need to actually work.
The new paradigm
Huawei Cloud defines Agentic Infra around four ideas:
- Efficient tokens — making every generated token cheaper and faster.
- Enhanced memory — giving agents a large, fast place to remember.
- Unified general-and-AI scheduling — mixing ordinary and AI workloads on one fabric.
- Secure autonomy — letting agents run in production without blowing up.
The company says this framework has already served more than 3,500 customers.
The compute layer: AICS
The headline product is the AICS cluster service built on the Ascend 950 (昇腾950) base, branded "灵衢". According to Huawei's own launch materials:
- Supports 40 days of stable training on the cloud.
- Recovers from faults within 10 minutes via a five-level fast-recovery mechanism.
- Delivers about 20% higher token throughput than the previous generation.
- Goes commercial in China on 30 September and outside China on 30 November 2026.
The memory piece, CMS, is described as PB-level space with twice the capacity and 50% better performance than comparable products — built for the long-horizon tasks agents stumble on today.
The model and agent layers
On top of the hardware, Huawei Cloud pushed three things:
- Agentic MaaS (model-as-a-service): one-click, no-deploy calls to leading SOTA models. MiniMax demoed its multimodal model on stage.
- AgentArts, a commercial enterprise-agent platform, now serving 100+ organizations including the Shenzhen Longgang District Government (深圳龙岗政府), China Postal Savings Bank, China Southern Power Grid, Guangzhou Laboratory, the University of Science and Technology of China, and Kingsoft Office.
- openJiuwen, the open-source sibling, with 50,000+ community stars and 3.29 million downloads.
Why "commercial plus open-source"
Huawei Cloud is explicit about a dual-wheel strategy: a paid platform for enterprises, and an open-source track to pull in developers. It says it has opened 5,000+ general MCP assets and 1,000+ industry assets. The Industry AI Foundry (行业AI梦工厂) claims 1,000+ industry assets and 1,000+ deployed projects across healthcare, embodied AI (具身智能), smart manufacturing, AI for science and finance, with new Smart Government and AI Hardware zones added.
What this really is
Strip the branding and the move is a repositioning: from "we sell you GPUs" to "we are the substrate (硅基黑土地) where your agents live." As AI shifts from chatbots to autonomous workflows, the bottleneck is no longer just raw FLOPS but memory, scheduling and safe execution — exactly the layers Huawei is packaging.
The competitive backdrop
Huawei Cloud is not alone in pitching agent infrastructure. Alibaba Cloud, Baidu AI Cloud and Tencent Cloud have each pushed agent and model-service platforms through 2026, and the real contest is who converts demos into paid, in-production agents. Huawei's differentiator is the tight coupling of its own Ascend (昇腾) silicon with the software stack — a vertical story that resonates with customers who need supply-chain certainty under chip export controls.
The partner layer
Alongside the products, Huawei launched "SCALE," a partner support system covering scenario solutions, co-innovation, marketing, local services and collaboration. The company says partners have already built 100+ scenario solutions across eight industries and that it has set up 130+ showcases globally. The bet is that agents scale through industry partners, not just Huawei's own sales force.
Honest limitations
Every product name, customer count and performance claim above comes from Huawei Cloud's official 18 September 2026 news release and corroborating event coverage; we did not independently benchmark the 20% throughput or 40-day-stability claims. Customer names are Huawei's listed references, not a full client list. The "3,500 customers" and "100+ enterprises" figures are Huawei-reported. openJiuwen download and star counts are community metrics, not audited. We did not verify whether deployed-agent outcomes (ROI for those 100+ orgs) are published. Some earlier (June 2026) Huawei Cloud specs — such as a 200 EFLOPS / 100,000-card cluster figure — appeared at a different event and are not restated here to avoid mixing announcements.
What readers can do now
- If you run enterprise AI, compare Huawei Cloud's AgentArts against Alibaba's and Baidu's agent platforms on the one dimension that matters: deployed, in-production agents, not demo zones.
- Watch the 30 November overseas commercial date — it is a clean test of whether this stack sells outside China.
- Track openJiuwen's star/download trend as a proxy for developer pull versus marketing noise.