This year’s WAIC SAIL Star award once again went to a supernode, and once again to ZTE.

At WAIC this time, ZTE brought its own version: the OEX supernode.

The domestic high-performance Matrix supernode solution, built on the OEX plus dOCS architecture, also won this year’s SAIL Award at the conference, marking the second consecutive year ZTE has received the honor.

After NVIDIA charted the path with the GB200 NVL72, the supernode approach has become a direction every major player is rushing toward.

Once NVIDIA validated this path, China began turning it into infrastructure and engineering practice.

On the same booth, there was also a phone.

The nubia NaviX Ultra, an AI agent phone led end to end by ZTE’s nubia team, also made its first public appearance at the conference.

One is an interconnect architecture inside data center racks; the other is a phone that fits in a pocket. There is a lot between them.

ZTE put them in the same booth, with four exhibition zones placed side by side: AI infrastructure, AI applications, AI smart home, and AI agent phones.

The supernode in the rack, the power system in the data center, and the phone in your hand are all pointing to the same question:

how to take computing power from where it is produced and deliver it all the way to where users can actually touch it.

The Full AI Chain in One Exhibition Zone

At WAIC, ZTE set up four exhibition zones: AI infrastructure, AI applications, AI smart home, and AI agent phones.

In the AI infrastructure zone, the OEX supernode was the focus.

The same zone also featured ZTE’s AIDC compute-power coordination solution and computing power network architecture.

Together, these two systems are designed to move computing power out of the rack reliably.

In the AI applications zone, ZTE showed its “New Pivot AIOS,” a system covering three areas: industry scenarios, home products, and personal devices.

It is positioned as an intelligent ecosystem for device-edge-cloud collaboration. Through it, ZTE wants to move AI from isolated capabilities to system-level capabilities.

The same zone also displayed Co-Claw, an enterprise-grade agent platform.

Inside ZTE, it is used to build an “AI-driven organization,” focused on office work, R&D, and operations; externally, its AI capabilities can be replicated across vertical industries such as manufacturing, rail transit, mining, metallurgy, and power.

ZTE also showcased its AI healthcare solution and put forward several concrete figures: medical report generation efficiency improved by more than 10 times; an HTA, or health technology assessment, report can be produced in 1 to 5 minutes; and cardiovascular assisted diagnosis reached 96% accuracy across 10,000 clinical validations.

The zone also featured an embodied intelligence lineup made up of three humanoid robots.

ZTE Xingzai, a lightweight 30-kilogram-class bionic robot, is aimed at commercial service scenarios; ZTE Xingqing, a bipedal heavy-load robot with a payload of more than 30 kilograms, and ZTE Xingwang, a wheeled heavy-load robot, are targeted at industrial manufacturing scenarios.

They are equipped with ZTE’s self-developed Xingqiao series dexterous hands, with operating precision within 0.2 millimeters and payloads above 15 kilograms. Their controllers use the Xinghe automotive-grade main chip.

The AI smart home zone showcased ZTE’s home terminal product lineup, which has ranked first globally in market share for five consecutive years, with annual shipments exceeding 100 million units. Cumulative shipments of its AI mid-screen products have also surpassed 3 million units.

In the AI agent phone zone stood one of the highlights of the conference: nubia NaviX Ultra, billed as the world’s first AI agent phone.

The mass-produced flagship was led end to end by ZTE’s nubia team. Its core selling point is that “one sentence can make AI handle complex cross-app tasks.” In ZTE’s own wording: “Say one thing, and nubia takes care of it.”

Today, ZTE also held a “Full-Stack AI Computing Ecosystem Forum,” where the “ZTE OEX supernode” displayed in the AI infrastructure zone was formally launched.

A Supernode Is an AI Factory

ZTE was the first company in the industry to propose the design concept of an orthogonal-architecture supernode.

The product it brought this time is called the OEX orthogonal-architecture supernode.

A single cabinet of this supernode can hold 128 GPUs, giving it the highest integration density in the industry, and it can scale up to a 16,000-card system.

This scale is enabled by a hard architectural innovation.

Traditional cabinets often contain thousands of cables, creating high communication loss and difficult maintenance.

OEX’s answer is to connect compute trays and switching trays in a vertical cross pattern, which is where the word “orthogonal” comes from.

The overall design therefore achieves zero cabling and no backplane. Signal paths between GPUs inside the machine become shorter, signal loss and latency are lower, and interconnect costs fall sharply.

Cables and backplanes solve the hardware problem itself; how chips coordinate with one another is a separate issue.

ZTE’s coordination approach relies on collaborative sensing technologies across five types of chips: CPUs, intra-machine and inter-machine interconnect chips, network cards, and DPUs. It works closely with mainstream GPU makers to precisely match the optimal solution as needed.

As a result, components inside the OEX supernode can be decoupled from one another, flexibly replaced as needed, and combined for the best fit, avoiding lock-in to any single chip vendor.

ZTE has also developed its own intra-machine interconnect switching chip, compatible with both the CLink protocol led by China’s Ministry of Industry and Information Technology and Broadcom’s SUE protocol, effectively leaving a door open for GPUs from different camps.

Once architecture and coordination are connected, two more issues remain: deployment speed and operations efficiency.

To address this, ZTE uses a front-loaded development model. Before GPU chips are released, it works with chip vendors on simulation, architectural pre-adaptation, and early algorithm optimization.

By the time the chips enter mass production, the supporting full-system solution has already been validated in advance.

Beyond that, ZTE has unified cabinet form factors, interface specifications, and interconnect protocols, forming a replicable engineering model. As a result, product deployment cycles have been compressed from more than a year to less than six months.

Around this model, ZTE provides GPU vendors with a plug-and-play “computing power container”: vendors can install their chips and use it, without designing cabinets and interconnects from scratch.

The modular design brings another benefit: cabinet installation time has been cut from days to minutes.

In terms of scale, this architecture does not stop at a single cabinet.

Building on the single-cabinet foundation, ZTE uses electrical switching plus optical interconnects for flexible expansion, creating cluster supernodes that can scale up to 16,384 GPUs.

This entire approach was not developed by ZTE behind closed doors.

ZTE joined with chip makers including Xizhi Technology, Biren Technology, MetaX, Enflame Technology, and Iluvatar CoreX, adding the dOCS, or distributed optical switching, architecture on top of OEX.

This architecture can dynamically adjust the connection topology between nodes according to computing power tasks, without locking the system into a fixed set of routes.

It has helped ZTE and these vendors achieve better compatibility, enabling the Matrix supernode.

The joint solution won this year’s WAIC SAIL Award, marking the second consecutive year ZTE has received the honor.

Who Solves Power and Cooling?

Beyond the OEX supernode, ZTE has built a full AIDC solution for the era of large models and high-density AI computing, integrating compute-power coordination with source-grid-load-storage systems.

Its 800V high-voltage direct current solution will enter commercial use this September.

The solution mainly addresses two problems: power supply losses and cooling pressure.

Both problems stem from the same trigger: as demand for intelligent computing power grows, rack power density is also rising sharply.

The power-loss problem is constrained by the traditional power supply chain.

Traditional data centers use multi-stage AC-DC conversion, moving from utility power to UPS to server power supplies, with overall power supply efficiency of about 94%.

The higher the rack density, the more severe the losses.

For a 10-megawatt-class data center, under the traditional power supply model, power losses alone can add millions of yuan in operating costs each year.

Higher power density also creates another problem.

At the same voltage, higher power inevitably means higher current, which requires thicker busbars and cable cross-sections. Construction difficulty, floor space, and copper costs all rise together.

ZTE’s logic is straightforward: since power equals current multiplied by voltage, if you do not want current to get too large, you have to raise the voltage.

That is how ZTE arrived at its 800V HVDC solution.

The solution can reduce current by 50% to 80%, allowing thinner lines and saving data center space and construction investment.

At the same time, lower current cuts heat losses. Cable copper loss is reduced by 50% to 60%, and end-to-end power supply efficiency rises from 94% to 98%.

A single set of the system can reach up to 2 megawatts in capacity and support intelligent computing cabinets with per-cabinet power consumption above 200 kilowatts, a level at which traditional power supply solutions can barely hold up.

Cooling pressure, meanwhile, is running into the physical limits of air cooling.

As training clusters grow more power-dense, air cooling has become increasingly unable to remove heat efficiently, while the physical deployment and management of cabinets have become more complex.

ZTE addresses this with AI computing micro-modules and liquid cooling solutions.

The AI computing micro-module integrates local monitoring, fan walls, enclosed aisles, CDU liquid cooling distribution units, 800V HVDC, and intelligent busways, supporting deployments above 250 kilowatts per cabinet, with power usage effectiveness, or PUE, as low as 1.15.

Beyond more precise scheduling, the hardware itself also has to do more work.

ZTE’s cold-plate liquid cooling solution can expand single-cabinet cooling capacity from 1,500 kW to 2,400 kW. It uses a 1+1 redundant configuration and supports online coolant replacement without interruption.

The system also includes an AI-driven DCIM platform that dynamically adjusts the relationship between power supply and cooling based on load changes, performing predictive optimization instead of waiting for problems before making repairs.

Beyond power and cooling, real-world projects often face a third drag: delivery cycles.

Under the traditional model, computing power, power, cooling, and management systems are often planned separately and built independently. Only during final on-site integration and commissioning do teams discover incompatibilities. Rework is common, and schedules slip.

ZTE’s approach is to deeply couple four dimensions from the start of solution design: computing power, or Bit; power, or Watt; heat, or Heat; and management, or Management. It calls this “4T collaboration.”

Combined with 90% factory-prefabricated modular production, units for power distribution, computing power, and cooling are assembled in the factory first. Once shipped to the site, they only need to be put together, rather than built from scratch.

For a 20 MW AI computing campus, the overall delivery cycle can therefore be compressed to 4.5 months.

From Bit to Token

The OEX supernode concentrates and connects computing power, while AIDC solves the power and heat problems. Together, these two systems address the production stage of computing power.

After computing power is produced, it still has to be delivered all the way to application endpoints such as phones and robots before it is truly put to use.

ZTE calls this full chain from underlying computing power to terminal applications “from Bit to Token.”

A bit is the smallest unit of computing power and storage, while a token is the smallest unit of content generated by a large model. In the process between the two, every unit of electricity saved and every millisecond of latency reduced inside the rack ultimately lands in the words users see when they talk to AI.

ZTE started out as a communications equipment company.

The core problem communications networks solve is how to move massive amounts of data efficiently and reliably from one node to another.

The high-speed intra-machine interconnects and multi-protocol compatibility of the OEX supernode are, in essence, a reuse of that capability inside the cabinet.

The power transmission design of 800V HVDC also moves the communications industry’s accumulated experience in large-scale network scheduling onto the power line.

Together, these two solutions form ZTE’s dual engine of “connectivity plus computing power.”

NVIDIA has already taken one pass at integrating dozens or even hundreds of GPU chips into a supernode through high-speed, lossless interconnects with the GB200 NVL72.

Now, supernodes have become a core breakthrough direction for next-generation AI computing power architecture.

At this moment, ZTE is working with chip vendors to turn supernodes from a concept into products.