A company that spent four decades laying the pipes of mobile networks is now pitching something different: not just the connectivity, but the factories that make artificial intelligence run. At a Shanghai trade show this summer, ZTE showed how far it wants to stretch beyond its old comfort zone.
Key takeaways
- At MWC Shanghai 2026, ZTE unveiled a "TCO-optimal AI factory" ecosystem built around a new operating layer called NewStart AIOS.
- Its SuperPod system packs 128 GPUs in a single rack using an open, multi-chip design, and scales up to 16,000 GPUs.
- The Co-Claw enterprise agent platform is being built into ZTE terminals, with its skill marketplace already past 320,000 downloads and 1 million calls.
- ZTE's home and terminal business remains large: over 100 million home-terminal shipments for two straight years, and a global lead in 5G fixed-wireless access.
- The push reframes ZTE from a network vendor into a full-stack AI infrastructure supplier — a crowded, capital-heavy race.
From base stations to AI factories
ZTE Corporation (0763.HK / 000063.SZ) has always been a telecommunications equipment house. What changed at MWC Shanghai 2026 is the ambition. The company presented an end-to-end stack that spans computing, network, storage, energy and software, all aimed at what it calls the "token economy (Token 经济)" — the idea that the unit of value in the AI era is the token produced, scheduled and sold.
ZTE leans on 41 years of building ultra-large, complex systems as its credibility card — the same discipline, it argues, that lets it wire together thousands of GPUs without the system falling over. The centerpiece is NewStart AIOS, which ZTE positions as the "technological foundation in the AI era," handling token scheduling, orchestration and service delivery. On top of it sits Co-Claw, an enterprise-level agent platform ZTE is baking into its phones, cloud PCs and home screens so that businesses and households get an always-on assistant.
The hardware that has to back the story
Software needs silicon, and ZTE's answer is SuperPod. Using a multi-chip open-collaboration design and an Orthogonal Electrical eXchange (OEX) architecture, a single rack supports 128 GPUs and can scale to 16,000 — a configuration meant for both training and inference at data-center scale. Around it, ZTE showed an AI data-center design with 800V high-voltage direct-current power and full-stack liquid cooling, targeting lower energy use and a tighter total cost of ownership.
Earlier in the year, at its April 2026 partner conference, ZTE had already previewed pieces of this: a 576×800G port AI switch that it says cuts latency by about a third and trims optical and fiber counts by around 40%, plus a vector-capable version of its GoldenDB database. The through-line is consistency — ZTE wants to sell the whole rack, not just a box.
Not alone in the pivot
ZTE is not the only one making this turn. Huawei, Inspur and a string of specialist vendors are all racing to package Chinese-made AI clusters as turnkey "AI factories." What makes ZTE's version notable is that it arrives with a global telecom-installation footprint and a services organization already embedded with operators — a distribution advantage that pure hardware startups do not have.
Where the money still comes from
For all the AI theater, ZTE's base business is enormous and still growing. Its semi-annual report shows home-terminal shipments above 100 million units for two consecutive years, a global number-one position in PON customer-premises equipment and in 5G fixed-wireless access for five straight years, and over 3 million AI middle-screen units shipped. It also launched the nubia NaviX Ultra, an AI-agent phone, extending the assistant story into a pocket device. Those volumes give ZTE the manufacturing scale and channel reach most AI upstarts lack.
According to Li An, chief scientist at BrainNet (脑机网), the real differentiator for telecom vendors moving into AI hardware will be system-level integration and energy efficiency, not any single chip, because buyers increasingly purchase a working cluster, not a specification sheet.
Honest limitations
Product and architecture details come from ZTE's official releases and a PRNewswire/Yahoo Finance distribution of its MWC Shanghai 2026 statements, plus its semi-annual report carried by Shanghai Securities News. These are company claims, not independent benchmarks; we have not tested SuperPod's scaling or the liquid-cooling efficiency claims. ZTE's AI infrastructure revenue is not broken out separately, so the size of this bet versus its core network business is unclear. Advanced-chip supply remains constrained by US export controls, which affects every Chinese AI hardware vendor. The "token economy" framing is ZTE's marketing language, not an established market metric.
What readers can do now
- If you buy AI infrastructure in China, treat ZTE as a full-stack alternative to consider alongside Huawei and Inspur, but ask for third-party benchmarks before committing to a 16,000-GPU-scale deployment.
- If you follow ZTE, watch future filings for a separate disclosure of AI-infrastructure revenue; until then, the AI story is a direction, not a measured business.
- If you run data centers, the energy-efficiency claims (800V DC, liquid cooling) are worth a pilot, since power and cooling — not just chips — now drive total cost.