China's GPU-LLM Breakthrough: Moore Threads Turns Profit as DeepSeek V4 Powers OpenClaw

The “Double Helix” Breakthrough in China’s AI Infrastructure: Strategic Synergy Between Moore Threads’ Q1 Profitability and DeepSeek V4’s Adoption as OpenClaw’s Default LLM

In Q1 2025, China’s foundational AI technology development reached a landmark inflection point: Moore Threads achieved a staggering 155.35% year-on-year revenue growth—and, for the first time, posted positive net profit for a single quarter. Almost simultaneously, DeepSeek’s newly released V4 Flash model was officially adopted by OpenClaw—the global open-source AI infrastructure project—as its default large language model. At first glance, these appear to be two independent news items. In reality, they form a clear technology–commercial closed loop: China’s domestically developed, full-featured GPU chips and indigenous large language models have jointly completed their first large-scale, real-world mutual validation and positive feedback cycle. This substantive “hardware–software co-optimization” goes far beyond incremental product iteration—it signals that China’s drive toward self-reliance in AI computing infrastructure is accelerating from “functional usability” into a new phase of “excellent usability, rapid deployment, and scalable adoption.”

I. Moore Threads’ Path to Profitability: A Critical Threshold for Commercialization of Domestic GPUs

Moore Threads’ Q1 financial results carry exceptional indicator value. Its 155.35% revenue surge did not stem from conceptual orders or government subsidies, but rather from bulk deliveries of its MTT S4000-series GPUs to domestic AI server clusters, edge inference nodes, and industry-specific intelligent computing centers. Industry supply-chain research indicates its core customers now include the intelligent computing cloud platforms of China’s three major telecom operators, AI training clusters onboard vehicles at leading new-energy automakers, and multiple provincial government LLM training bases. Crucially, achieving net profitability confirms that its per-chip cost—including wafer fabrication, packaging, and verification—has fallen below the market’s acceptable average procurement price, granting it initial price competitiveness against international Tier-2 vendors (e.g., AMD Instinct MI250X) in targeted application scenarios.

This milestone reflects a pragmatic technical strategy: Rather than blindly chasing NVIDIA H100-level peak performance, Moore Threads focused on the “H20 replacement window,” making targeted optimizations in mixed-precision (FP16/BF16) computation, PCIe bandwidth utilization, and memory bandwidth compression algorithms. As a result, the S4000 delivers 1.8× the inference throughput of the H20 for models ranging from 7B to 70B parameters—while reducing power consumption by 22%. This “scenario-driven customization” enabled rapid market penetration in latency-sensitive domains with moderate peak compute requirements—such as real-time financial risk control and multimodal industrial quality inspection. Profitability is thus not merely a financial milestone; it marks the watershed moment when domestic GPUs transitioned from “lab validation” to “commercial sustainability.”

II. DeepSeek V4 Flash’s Adoption by OpenClaw: A Substantive Leap in Open-Source Ecosystem Influence

Concurrently, DeepSeek V4 Flash’s selection as OpenClaw’s default model carries profound strategic implications. Supported by the Linux Foundation, OpenClaw aims to build a standardized, cross-hardware inference framework for AI. Its choice of a default model directly shapes downstream developer toolchains, quantization strategies, and service deployment templates. V4 Flash’s victory stemmed not only from parameter count or benchmark scores—but more importantly, from its engineered native compatibility with China’s domestic hardware stack:

• Model weight formats are deeply compatible with Moore Threads’ MUSA architecture INT4 quantization instruction set;
• Its inference engine includes AVX-512 optimization paths specifically tuned for domestic server CPUs (e.g., Hygon C86);
• It provides a complete ONNX Runtime–Moore Threads plugin, slashing end-to-end deployment cycles to one-third of conventional approaches.

This decision means developers no longer need weeks of debugging to “get the model running.” Within the OpenClaw ecosystem, a single codebase delivers near-identical inference performance across Moore Threads GPUs, Ascend 910B accelerators, and even Cambricon MLUs. Indigenous models are no longer merely “usable”—they have become the “adhesive” bridging domestic chips and the developer ecosystem. Furthermore, its Apache 2.0 open-source license and open-weight policy effectively mitigate regulatory, security-audit, and private-deployment risks associated with proprietary models—fully aligning with the Central Committee and State Council’s Opinions on Strengthening Algorithm Governance for Internet Platforms, which emphasize “algorithm transparency” and “security assessment.”

III. The “Chip + Model” Positive Feedback Loop: Reshaping AI Infrastructure Replacement Logic and Industrial Chain Effects

The synergy between Moore Threads and DeepSeek is catalyzing a novel industrial feedback mechanism:

• Downstream Drive: Rising GPU shipments → higher utilization rates at SMIC’s N+2 process nodes → improved yield ramp-up for advanced packaging (e.g., JCET’s 2.5D CoWoS-like solutions) → enabling OEMs (e.g., Inspur, Sugon) to launch more cost-competitive, fully domestic intelligent computing servers;
• Upstream Empowerment: Widespread adoption of V4 Flash within OpenClaw → attracts millions of developers building applications atop China’s domestic stack → generates richer vertical-domain training data → feeds back to optimize Moore Threads’ MUSA Compiler for specific operators;
• Horizontal Expansion: Together, the duo has effectively established a de facto “China-native AI reference architecture,” now being incorporated into technical white papers for intelligent computing center tenders nationwide—accelerating the replacement of restricted NVIDIA H20/H800 solutions. Per IDC projections, the share of China’s AI server market adopting purely domestic chip + model solutions will rise from 8% in 2024 to 22% in 2025.

IV. Global Capital Revaluation: A Paradigm Shift from “Point Solutions” to “Systemic Capability”

This synergistic breakthrough may trigger a fundamental re-evaluation by overseas capital of investment logic in China’s AIGC sector. Historically, international investors applied valuation discounts to Chinese AI players citing “technology gaps” or “ecosystem fragmentation.” Today, however, the “Moore Threads + DeepSeek + OpenClaw” triad demonstrates conclusively that China possesses the capacity to define next-generation AI infrastructure standards—not by replicating the CUDA + PyTorch paradigm, but by forging a new triangle: MUSA + DeepSeek + OpenClaw. Companies with such deep “hardware–software co-optimization” capabilities—including Huawei’s Ascend/Pangu (with full-stack chip–model–framework R&D) or Baichuan Intelligence (deeply integrated with domestic hardware ecosystems)—will see their valuation anchors shift away from isolated technical metrics toward systemic dimensions: ecosystem control, developer penetration rate, and speed of commercial closed-loop execution.

While global developments—including Iran’s ceasefire negotiations and the U.S. Federal Reserve chair nomination confirmation—command broad attention, China’s autonomous breakthrough in AI infrastructure is reshaping the technological geopolitical landscape in quieter yet more resilient ways. When a domestically designed GPU powers millions of servers—and an open-source model becomes the default starting point for developers worldwide—true technological sovereignty takes root, silently and firmly.