The cabinet had almost no cables running across its front.
Inside were 128 graphics processors sharing a single pool of bandwidth, close enough that none of them had to queue to talk to another.
The rack's sparseness is deliberate: the company is selling system design, not card count.
What was shown, and where
At the ninth China International Import Expo, held in Shanghai from November 5 to 10, 2026, MetaX (沐曦股份, 688802.SH) displayed two system-level products in the national exhibition's China Pavilion. Xinhua Daily and the Shanghai science and technology commission described the company as the only domestic GPU enterprise selected for the pavilion.
The first is the Xijing S600 (曦景S600) super-node (超节点). Built around the company's Xiyun C600 (曦云C600) all-domestic flagship GPU, the MXMACA software stack and the MetaXLink interconnect, a single rack supports 128 cards with full switching and no bandwidth convergence. The system can scale out to a cluster of ten thousand cards.
The second is a heterogeneous dual-engine rack developed with Orient Alpha (东方算芯). It pairs the C600 with an Orient Alpha DF1000 chip. The division of labor is deliberate: the C600 handles compute-intensive work such as long-context attention, while the DF1000 takes memory-intensive jobs such as sparse activation. The companies describe the design using the Roofline model, a standard way of asking whether a workload is limited by calculation or by memory movement.
Why the rack, not the chip
For years the GPU conversation was about single cards: peak floating-point performance, memory size, memory bandwidth. As models grew, a different bottleneck appeared. When a model is split across hundreds of cards, the time cards spend waiting for each other can dominate the time they spend calculating. The industry calls it the communication wall.
A super-node pushes that wall outward. By grouping many chips into one large, flat scale-up domain, communication stays inside a rack rather than crossing a network. "128 cards full exchange, no bandwidth convergence" means that, within the cabinet, any two GPUs can communicate without contending for a shared link.
The rack also uses what MetaX calls a "three zeros" design: no cables at the compute nodes, no cables at the switching nodes, and no cables between compute and switching nodes. Combined with a modular architecture, the claim is that a data center can deploy the system without rebuilding its server room. For a customer buying thousands of cards, cabling and maintenance are not small costs.
The software is half the argument
Hardware spec sheets are easy to copy into a table. Software is where domestic GPUs have historically struggled.
MetaX says its MXMACA stack supports more than 6,000 CUDA applications and over 1,000 AI models, works with more than 40 AI frameworks, covers all 2,410 operators in PyTorch 2.8, and has more than 500,000 registered developers. Those are company figures reported by Xinhua Daily. They matter because a GPU that runs only a narrow set of models becomes a maintenance burden rather than a platform.
The company also says a ten-thousand-card "token factory" built with a partner has entered production, with the first clusters moving from readiness to a running business model in one or two days. Again, this is the company's account, not an audited deployment record.
The business behind the demo
The expo display landed shortly after MetaX published its third-quarter report. For the first three quarters of 2026, the company reported revenue of RMB 2.824 billion (≈ US$393M / HK$3.0B), up 128.45 percent year on year, with net profit attributable to shareholders of RMB 639 million (≈ US$89M / HK$685M) and non-recurring-adjusted profit of RMB 259 million (≈ US$36M / HK$278M). Both profit measures turned positive from a year earlier. The figures were reported consistently by Shanghai Securities News, China Daily and Economic Information Daily.
The third quarter alone brought revenue of RMB 1.5 billion (≈ US$209M / HK$1.6B), more than the first half combined, with adjusted profit of RMB 308 million (≈ US$43M / HK$330M) and a profitable quarter. The nine-month adjusted profit was RMB 259 million, meaning the first half was loss-making overall and Q3 more than offset that earlier loss. The company attributes the growth to stronger customer purchasing and higher GPU shipments, with the C600 entering mass production in May 2026.
Demand is not in doubt. Chinese media have reported that daily token calls in China exceeded 500 trillion by mid-2026, up sharply from the end of 2025 — figures this article could not verify against a primary source. The harder question is whether system-level efficiency can turn that demand into durable margins.
Key takeaways
- MetaX showed the Xijing S600 super-node at the China Import Expo in November 2026, with 128 cards per rack, full switching and no bandwidth convergence, scalable to ten thousand cards.
- The system pairs the domestically produced Xiyun C600 GPU with the MXMACA software stack and MetaXLink interconnect, and uses a "three zeros" cabling design.
- A second product combines the C600 with Orient Alpha's DF1000 in a heterogeneous rack aimed at cloud inference.
- For the first three quarters of 2026, MetaX reported revenue of RMB 2.824 billion (≈ US$393M / HK$3.0B), up 128.45 percent, with both profit measures turning positive.
Honest limitations
- The 128-card, ten-thousand-card, "three zeros" and software-compatibility figures are company claims reported by state and financial media. No independent benchmark of the S600, or audited record of the token factory, was available here.
- The "only domestic GPU enterprise in the pavilion" description is attributed to Xinhua Daily and the Shanghai science and technology commission. It describes a specific exhibition selection, not a market ranking.
- This article looks at one vendor's super-node and one quarter's results. It does not compare the S600 with competing domestic or international systems on price, power consumption, software maturity or delivered performance.
What readers can do now
- If you evaluate AI infrastructure, ask for performance per rack and per watt, not just per chip. Cards that wait for each other are expensive idle capacity.
- Read software compatibility claims as a starting point. Test your own framework, operator set and inference engine before committing.
- For anyone tracking the domestic chip sector, watch the next quarterly report for whether revenue growth holds after the initial mass-production ramp.
