분석가들이 기반 인프라 구축 사이클을 과소평가했다고 지적하는 광범위한 매도세로 국내 컴퓨테이션 관련 주식들이 20% 급락했다.

Due to lock-up expiration pressures, Moore Threads’ stock plummeted 20%, hitting its lowest point since its IPO and pushing its total market capitalization below 200 billion yuan. In response, Moore Threads told Miaotou that the company’s operations and production continue to show positive momentum, with first-half revenue reaching 1.736 billion yuan, a 147% year-on-year increase. Specific operational and financial details will be subject to the company’s publicly disclosed information and subsequent formal announcements. Meanwhile, Tiangu Zhixin and MetaX also faced downward pressure, declining 11.42% and 9.67%, respectively. Cambricon, however, remained unaffected, rising 1.9%.

While the lock-up expiration served as the immediate trigger, the core driver of the decline lies in the valuation premium attached to grand narratives despite the sector’s lack of sustained profitability. Nevertheless, investors tracking domestic computing power should remain cautious rather than pessimistic. The potential of China’s domestic GPU ecosystem has yet to be fully realized, and definitive conclusions remain premature.

Where do current domestic GPU manufacturers stand? How should the industry’s valuation be assessed?

Turning losses into profits is both straightforward and challenging. Domestic GPUs have now transitioned from product validation to large-scale commercialization. Leading vendors are beginning to validate profitability, though the industry as a whole has not yet entered a stable profitable phase. Mid-year reports for four domestic GPU makers covering 2026 have been released. Superficially, Cambricon, Moore Threads, MetaX, and Tiangu Zhixin all advance the same narrative: revenue doubling and accelerated domestic substitution.

From a profitability standpoint, Cambricon, MetaX, and Tiangu Zhixin reported net profits attributable to shareholders of 2.311 billion yuan, 612 million yuan, and 106 million yuan, respectively, in the first half of 2026. Moore Threads, Biren Technology, and Enflame remain unprofitable. Notably, MetaX’s turnaround was driven by an 887 million yuan gain from fair value changes, a figure the company explicitly noted lacks sustainability. Similarly, Tiangu Zhixin’s return to profitability stemmed primarily from a 760 million yuan unrealized gain on its equity investment in Shenghe Jingwei. For both companies, the turnaround did not originate from core business operations. Consequently, Cambricon stands out as the only GPU vendor with high-quality, operationally driven profitability; others still need to achieve sustainable turns in their main businesses.

Cambricon actually achieved core business profitability as early as Q4 2024, marking a pivotal turning point. Its largest customer became an unnamed long-term partner, contributing 930 million yuan in sales, accounting for 79.15% of its annual total. Given that Cambricon’s full-year 2024 revenue was merely 1.174 billion yuan, this single partnership was transformative. The shift of Cambricon’s downstream client base from government entities to major internet corporations and telecom operators proved instrumental in driving its operational turnaround. Ultimately, large clients and substantial orders became the foundation supporting Cambricon’s performance and its 600 billion yuan valuation.

Why did Cambricon emerge first? Partly because major internet corporations lost access to Nvidia GPUs. Following bans on the A100 and H100, the H20 became China’s primary AI inference chip in 2024. By April 2025, even the “special edition” H20 was banned, creating a window for domestic GPU vendors to accelerate. Additionally, Cambricon possessed a first-mover advantage and capitalized on surging demand. During the 2024 demand explosion, Cambricon’s Siyuan 590 had already been deliverable since 2023, perfectly aligning with the surge in corporate orders. Competitors lagged in readiness. At the time, the Siyuan 590 outperformed most domestic rivals in FP16 compute (256–345 TFLOPS), memory capacity, and per-card power efficiency.

In contrast, Moore Threads’ S4000 launched only in December 2023, delivering just 100 TFLOPS in FP16. Its H100-rival S5000 would not reach scale delivery until June 2025. Enflame faced similar delays: its T20, released in 2021, offered only 80 TFLOPS in FP16, with its S60 inference card arriving in July 2024. MetaX adopted a low-cost “inference-first, training-later” strategy, mass-producing its N100 inference chip in April 2023 and its C500 training-inference chip in February 2024. However, post-launch yield ramping and customer validation prevented it from fully capturing the initial wave. Biren Technology and Tiangu Zhixin also missed the timing window due to product roadmaps, delivery capabilities, misaligned customer structures, and U.S. sanctions.

Large contracts and key accounts undeniably drive profitability for domestic GPU makers. For instance, on March 31, 2026, Moore Threads announced a sales agreement with an undisclosed client totaling 660 million yuan. This deal enabled the company to turn profitable in Q1 alone, significantly narrowing its first-half losses.

When will the next inflection point arrive for domestic GPUs? Will more Cambricons emerge? Supply chain management will be decisive.

As Nvidia retreats, domestic chips must fill the void. In 2025, Nvidia still dominated China’s accelerator card market with a 55% share, while domestic vendors held fragmented, lower positions. Nvidia itself confirmed in its earnings reports that revenue from China’s data center segment (H100/H200/H20) would be recorded as zero. If Nvidia’s market share falls to 8% in 2026 (excluding unrestricted consumer gaming cards), over 40 percentage points of market space will open, creating a massive substitution window for domestic computing infrastructure. (Data source: Zheshang Securities)

Beyond import substitution, domestic computing remains a high-growth sector. IDC projects China’s accelerated server market will expand from $37.7 billion in 2025 to $150.1 billion by 2030, representing a compound annual growth rate exceeding 30%. The pie is expanding, meaning more Cambricons will likely emerge. However, opportunities will not be distributed equally. Success will depend not only on securing large orders but also on supply chain execution.

According to a Morgan Stanley research report, HBM3E supply will become the primary bottleneck for China’s AI GPU shipments in 2027. As global leaders pivot to HBM4, supplies of HBM3E to Chinese vendors will further tighten. “Currently, the most difficult challenge in the industry remains the supply chain. Most capacity is allocated to leading domestic enterprises, leaving limited availability for smaller firms like ours,” a GPU industry insider told Miaotou.

Balance sheets across GPU vendors corroborate this trend. Cambricon holds 8.248 billion yuan in inventory and 2.914 billion yuan in prepayments; Moore Threads carries 3.55 billion yuan in inventory and 1.344 billion yuan in prepayments; MetaX reports 1.432 billion yuan in inventory and 2.196 billion yuan in prepayments. Inventory typically comprises wafers, work-in-progress (masks/dies), raw materials (HBM/substrates), and finished goods (packaged boards). Prepayments function as “capacity reservation fees” for foundries, “advance payments” for packaging and testing, and “memory advances” for HBM suppliers.

Analyzing Cambricon’s 8.248 billion yuan inventory breakdown reveals 5.749 billion yuan in raw materials, 2.537 billion yuan in entrusted processing materials, and only 0.085 billion yuan in finished goods. This indicates that once Cambricon’s chips are manufactured, they sell immediately—demand is never in question. Moore Threads and other GPU vendors exhibit highly similar inventory structures. Miaotou observes that Cambricon, Moore Threads, and MetaX are aggressively stocking up, using prepayments to lock in capacity and raw materials.

Compared to mainland A-share listed companies, Hong Kong-listed GPU firms employ more flexible mechanisms to secure capacity through deep interest alignment. Under Hong Kong’s new Rule 18C/Chapter 17 regulations, Tiangu Zhixin and Biren Technology can include foundry partners, packaging and testing facilities, and EDA providers within their incentive frameworks to ensure stable advanced-node capacity and toolchain access. For example, Tiangu Zhixin’s H-share equity incentive plan covers suppliers across the wafer fabrication, packaging/testing, and EDA sectors, with a 10-year validity period and a minimum vesting horizon of 12 months.

Clearly, every player is deploying maximum effort to secure inventory and capacity. Miaotou concludes that the potential of domestic computing power has not yet been fully unleashed. Once supply chain constraints ease, shipment progress for domestic AI chips is expected to accelerate, with tangible results materializing in 2027 and 2028. In the short term, valuations for domestic computing stocks will face pressure from factors such as lock-up expirations. Over the long term, however, sustained earnings growth will drive valuation recovery for GPU vendors.