MetaX C-588 is all about strategy

blog29042026-4

One card. Not the fastest GPU — but perhaps the most practical for real-world inference of large models.

The Battle for Memory, Not FLOPS

Shifting the focus from FLOPS to memory is not just a matter of speed, but of operational economics. Distributed inference, when a single model is spread across multiple cards, requires complex software to synchronize data between GPUs. This inevitably generates massive overhead:

Latencies for data transfer over the bus (NVLink/NVSwitch);
The need for load balancing;
Synchronization of states (KV-cache) between nodes.

As a result, to get a result from two cards, additional effort and resources are often required. A card with 128 GB of memory, like the MetaX C-588, nips this problem in the bud. It allows for a simple and efficient 1 card = 1 instance architecture, which radically reduces the skill requirements for support engineers and simplifies service scaling.

For three years, the AI GPU market moved according to the logic of «more TFLOPS — more money.»

H100, MI300X, Blackwell — each new chip promised to double peak performance. But practice has shifted priorities.

For inference, which now accounts for over 50% of AI spending in China, peak compute speed is not critical; memory capacity is.

A 70B parameter model in FP16 weighs ~140 GB. In GPTQ-4bit, ~35 GB. But you also need room for KV-cache, batches, and system overhead.

On an 80 GB H100, this becomes a headache: either aggressive quantization with quality loss, or distributed inference with increased latency.

128 GB changes the equation. A model up to 120B parameters (in a quantized format) fits on a single card.

One GPU = one instance + minimum latency + minimum complexity. This is precisely the package offered by the MetaX C-588 (曦云 C588).

What is the MetaX C-588?

MetaX (沐曦集成电路) is a Shanghai-based company founded in 2020 by former AMD engineers. Their portfolio includes three product lines:

C-Series (AI training and inference);
N-Series (Inference and video);
G-Series (Graphics).

The company’s revenue grew from 426,000 CNY to 743 million CNY in two years.

IPO in December 2025 on the Shanghai Stock Exchange (SSE: 688802) — shares soared 693% on the first day.

The MetaX C-588 is the flagship C-Series model.

The card is built entirely on its own GPU architecture (not a licensed copy of AMD/NVIDIA), the MXMACA software stack with CUDA compatibility, and MetaXLink technology for multi-GPU scaling.

Key Specifications:

Video Memory (VRAM): 128 GB.
Type: HBM3e (High Bandwidth Memory).
This is the card’s main feature. This volume allows running models up to 120 billion parameters (in a quantized format) on a single accelerator, eliminating the need for complex and slow distributed inference.
Compute Performance:
FP32 (Vector): 60 TFLOPS;
FP16 (Vector): 480 TFLOPS;
TF32 / FP / BF16 (Tensor): 240 / 480 TFLOPS;
FP8 / INT8 (Tensor): 960 / 960 TOPS.
Memory Bandwidth: 1200 GB/s.
Ensures fast data delivery to the compute cores, preventing stalls.
Interconnect and Scalability:
Interface: PCIe 5.0 x16;
Proprietary Interconnect: MetaXLink.
MetaXLink bandwidth reaches 896 GB/s, allowing up to 8 cards to be combined into a single compute pool with a «full-flow» topology for maximum efficiency.
Power Consumption and Form Factor:
TDP: 850 W.
Form Factor: OAM 2.0 (OCP Accelerator Module).
Cooling: Passive (designed for airflow in a server chassis).

Use Cases:

Inference and training of large language models (LLM, e.g., DeepSeek V4, GLM-5.1);
High-Performance Computing (HPC);
Big Data processing.

Software Stack

The card runs on the proprietary MXMACA stack, which ensures compatibility with major AI ecosystems, simplifying migration from competitor solutions.

Its key positioning is between A100 and H100 analogues in terms of compute power, but with 128 GB HBM — 60% more than both.

Parameter MetaX C-588 NVIDIA H100 NVIDIA A100
Memory 128 GB HBM 80 GB HBM3 80 GB HBM2e
Inference 70B (FP16): ❌ Doesn’t fit ❌ Doesn’t fit ❌ Doesn’t fit
Inference 70B (GPTQ-4bit): ✅ With room to spare ⚠️ Barely fits ⚠️ Barely fits
Inference 120B (GPTQ-4bit): ✅ With room to spare ❌ Doesn’t fit ❌ Doesn’t fit

For the Russian market, where the largest models are DeepSeek-V3/R1 (671B MoE), Qwen 72B, Llama 70B — it is precisely the memory capacity that determines how many cards are needed for deployment and at what price.

Three Reasons for the C-588’s Competitiveness

Inference has become the main expense item AI inference spending in China has surpassed model training spending for the first time. The logic has shifted: it is now more important not to train a model in 3 days, but to answer 100K requests per hour with acceptable latency. And for that, memory capacity and batching efficiency matter more than peak FP16 TFLOPS.
DeepSeek V4 confirmed the ecosystem On April 24, 2026, DeepSeek announced Day-0 adaptation of V4 for 8 Chinese GPU vendors. MetaX is on this list. This is not «we tried to run it,» but production-ready compatibility. For a market where the main question is «will my model run?» — this is a decisive argument.
MoArk (formerly Gitee AI) and the Partner Ecosystem Ecosystem development is no less important a factor than the hardware’s technical specifications. The success of any new chip depends on how easily it can be integrated into existing business processes. In this direction, MetaX and its partners have made significant progress.

The accelerator is available for pre-order in the Chaitex catalog.

MetaX, together with Gitee AI (now ARKA/MoArk), held the premiere launch of distilled DeepSeek R1 models on MetaX C-Series accelerators. This event was a signal to developers: the platform is ready to work with cutting-edge open models. The largest cloud provider, CE Cloud, has already integrated MetaX accelerators into its all-in-one servers designed for AI tasks. Furthermore, Chaixun Telecom ensures the seamless integration of these solutions into commercial environments and corporate networks.

The ecosystem is actively growing, and this is a key signal for enterprise buyers. For large businesses, not only performance matters, but also the availability of ready-made solutions, support, and proven integrations. This allows them to implement new technologies without the risk of being guinea pigs, and to be confident in the stability and scalability of the chosen platform.

MetaX C-588: An Alternative Strategy

For companies that need to inference large models with minimal infrastructure complexity, 128 GB on a single card is a serious advantage. Fewer cards per cluster, simpler deployment, lower latency.

For companies that need peak performance for training — the C-588 is not the best choice. The Moore Threads S5000 is likely preferable.

Chaitex Recommendation: if your task is inference of 30B — 120B parameter models (Qwen 72B, Llama 70B, DeepSeek-R1 distillates) and you are ready for the engineering effort to adapt to MXMACA — the C-588 deserves a POC.

The MetaX C-588 is not just hardware. It is a strategic product that solves the main problem of modern inference: memory shortage. The ability to place a model up to 120B parameters on a single card eliminates the need for complex and slow distributed inference, making it an ideal solution for the commercial application of AI.

Q&A

Q: Is the MetaX C-588 a «Chinese H100»? A: No. In terms of peak performance, the C-588 is closer to the gap between the A100 and H100. Its advantage is 128 GB of memory, not TFLOPS. It is a different architectural philosophy, tuned for inference, not training.

Q: What are MetaX’s long-term prospects? A: The company is public, has state support, and a roadmap for the C600 (mass production H1 2026, HBM3e and FP8 support). Risks exist (dependence on TSMC, software stack), but its market position looks stable.

Let's start a conversation

Start Scaling Today

Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.

Location: Moscow, ext.ter.g., sh. Ochakovskoe, 28, p.2
GPU Servers High-density compute
Облачная аренда Быстрый старт ИИ моделей
Enterprise SLA Guaranteed availability
Low Latency Optimal network routes