IN Brief:
- MetaX says its C600 general-purpose GPU entered mass production in May 2026.
- The OAM 2.0 accelerator has a maximum board power of 1,000W and supports both air and liquid cooling.
- MetaXLink-E supports supernodes containing up to 128 GPUs, extending the C600 from individual accelerators into larger AI computing systems.
MetaX has moved its C600 general-purpose GPU into mass production, turning its latest accelerator architecture into a commercial product as increasing GPU shipments support rapid revenue growth at the Chinese semiconductor developer.
The C600 entered volume production in May 2026, according to MetaX’s first-half results. Revenue for the six months reached CNY1.324 billion, an increase of 44.67% year on year, with the company attributing much of the improvement to higher shipment volumes as its GPU products gained further customer adoption.
The headline profit figure requires rather more qualification. Net profit attributable to shareholders reached CNY612 million, reversing a CNY186 million loss during the equivalent period last year, but MetaX still recorded a CNY49 million loss after non-recurring items were excluded. On that measure, the loss narrowed 75.83% year on year, while the second quarter produced an underlying profit of CNY54 million.
The more useful electronics milestone is therefore the C600 production ramp rather than the accounting swing. Moving a large accelerator from design and qualification into sustained manufacture places new demands on wafer supply, packaging, high-bandwidth memory, power delivery, cooling, production test, and the software environment needed to turn the completed module into a usable computing platform.
MetaX describes the C600 as a new-generation general-purpose GPU based on its proprietary architecture and a domestic advanced process. The accelerator is supplied as an OAM 2.0 module with a maximum board power of 1,000W and supports both air-cooled and liquid-cooled systems.
The power figure alone illustrates how far accelerator design has moved from the conventional add-in graphics card. A 1kW module places substantial requirements on the server’s upstream power conversion, board and busbar design, connectors, thermal interfaces, and cooling system before the processor has performed a useful calculation.
MetaX pairs the accelerator with high-capacity, high-bandwidth GPU memory protected by end-to-end ECC. The company does not disclose the memory capacity or bandwidth on its public C600 specification page, but error protection across the memory path is important in systems expected to run long training, inference, data-processing, and general-purpose computing workloads without silent data corruption.
Multi-GPU communication is handled through MetaXLink, with the MetaXLink-E interface supporting supernodes containing as many as 128 GPUs. That scaling capability becomes increasingly important as model size grows beyond the memory or processing capacity available from a single accelerator.
At that point, raw processor throughput is only part of system performance. GPUs have to exchange parameters and intermediate data frequently enough that communications overhead does not leave expensive compute hardware idle. Interconnect bandwidth, latency, topology, memory movement, and software orchestration can therefore become limiting factors even when individual accelerator performance continues to increase.
MetaX also offers a dual-socket C600 server developed with its partners, carrying eight C600 OAM modules in each system. The server supports both air and liquid cooling and is designed for large-language-model workloads, training, high-concurrency inference, and intelligent-computing centres.
The hardware still depends heavily on the software stack surrounding it. MetaX’s MXMACA environment is designed to provide compatibility with mainstream GPU computing ecosystems, reducing the amount of application redevelopment required when customers move existing AI and general-purpose computing workloads onto its processors.
That compatibility work is a substantial part of accelerator development. Production silicon is of limited value if frameworks, compilers, libraries, distributed-computing tools, drivers, and model-serving software cannot use it efficiently, particularly where customers already operate large bodies of code developed around established GPU platforms.
MetaX spent CNY525 million on research and development during the first half, equivalent to 39.65% of revenue. The scale of that spending reflects the requirement to develop silicon, systems, and software in parallel rather than treating the GPU die as a self-contained product.
The company’s first-half report also indicates that the core MXC600 chip has progressed through domestic security and reliability assessment, another requirement for deployments in customers where component origin, software control, and supply-chain continuity form part of procurement decisions.
Mass production now moves the C600 into a different phase of that programme. MetaX has to maintain accelerator and package yield, secure memory and substrate supply, support 1kW thermal designs, and keep its software stack aligned with rapidly changing AI frameworks while shipment volumes increase.
China’s demand for locally supplied computing hardware gives the company a substantial prospective market, but domestic availability alone will not sustain adoption. Accelerator customers ultimately buy usable system performance, which depends on reliable hardware, scalable interconnects, mature software, and predictable supply operating together.
The first-half figures show that more MetaX hardware is reaching customers. The next measure is less flattering but more informative: whether C600 production can continue scaling while 128-GPU systems, software compatibility, cooling, and supply consistency hold up outside the comparatively controlled environment of early deployments.
