Shares of Moore Threads tumbled 20% following the expiration of lock-up periods, hitting an all-time low since its market debut and dragging its total market capitalization below the 200 billion yuan threshold. In response, the company stated that its operations continue to improve, with first-half revenue reaching 1.736 billion yuan, a year-on-year surge of 147%, while directing investors to official disclosures for detailed financials. Meanwhile, ILUVATAR COREX and MetaX (沐曦股份) also experienced declines of 11.42% and 9.67%, respectively, whereas Cambricon Technologies Corporation Limited bucked the trend with a modest 1.9% gain.
The lock-up expiration merely served as a trigger; the core reason for the downturn lies in premium valuations attached to ambitious narratives despite a lack of sustainable profitability. However, investors focused on domestic AI computing power need not despair, as the sector's full potential has yet to be unleashed, and it would be premature to draw final conclusions. This raises pertinent questions: where exactly do domestic GPU manufacturers currently stand, and how should valuations for this industry be assessed?
Turning Profitable: Easier Said Than Done
Domestic GPU makers have now entered a phase transitioning from product validation to large-scale commercialization. While leading players have begun demonstrating profitability, the industry as a whole has yet to achieve stable and consistent earnings. The 2026 interim results from four major domestic GPU manufacturers paint a superficially similar picture: revenue growth doubling and accelerated domestic substitution. From a profitability standpoint, Cambricon Technologies Corporation Limited, MetaX (沐曦股份), and ILUVATAR COREX reported net profits attributable to shareholders of 2.311 billion yuan, 612 million yuan, and 106 million yuan, respectively, for the first half of 2026, while Moore Threads, Biren Technology, and Enflame Tech continued to post losses.
Notably, MetaX's (沐曦股份) turnaround was driven by an 887 million yuan gain from fair value changes, which the company explicitly flagged as non-recurring in its interim report. Similarly, ILUVATAR COREX's swing to profitability stemmed largely from a 760 million yuan fair value gain on its equity investment in Shenghe Jingwei. Both companies' turnaround efforts were not attributable to core operating profits. This leaves Cambricon Technologies Corporation Limited as the sole GPU maker whose profitability carries substantial weight, while others still need to achieve operating-level turnaround. Cambricon Technologies Corporation Limited had already achieved operating profitability as early as Q4 2024, and the pivotal shift occurred that year when its largest customer became an unnamed long-term partner contributing 930 million yuan in sales, accounting for 79.15% of annual revenue. Given that Cambricon's total 2024 revenue was merely 1.174 billion yuan, the transition of its primary client base from government entities to major internet firms and telecom operators became a crucial factor driving its core business profitability. In essence, major clients and substantial orders now underpin both Cambricon's performance and its 600 billion yuan valuation.
Why, then, did Cambricon Technologies Corporation Limited emerge as the frontrunner? On one hand, major internet companies could no longer procure NVIDIA's GPUs. After the bans on A100 and H100, H20 became the primary AI inference solution in China in 2024, but even this "special edition" chip faced a sales ban in April 2025, opening a window of opportunity for domestic GPU manufacturers. On the other hand, Cambricon capitalized on its first-mover advantage. When demand exploded in 2024, its Siyuan 590 chip, which had been delivery-ready since 2023, was perfectly positioned to capture the surge in major internet orders, whereas competitors lacked comparable offerings at the right time. Furthermore, Cambricon's Siyuan 590 outperformed most domestic rivals in FP16 computing power (256-345 TFLOPS), memory capacity, and single-card power consumption during that period. By comparison, Moore Threads' S4000 launched only in December 2023 with a mere 100 TFLOPS FP16 performance, and its H100-comparable S5000 did not achieve volume delivery until June 2025. Enflame Tech similarly lagged, having released the T20 in 2021 with only 80 TFLOPS FP16 output and debuting the S60 inference card only in July 2024. MetaX (沐曦股份), meanwhile, adopted a low-cost entry path prioritizing inference before training, with its Xis N100 inference chip reaching mass production in April 2023 and its Xiyun C500 training-inference integrated chip following in February 2024, though yield ramping and customer validation processes meant it too failed to fully catch the wave. Biren Technology and ILUVATAR COREX also missed the opportunity window due to product gaps, delivery capability shortfalls, misaligned customer structures, or the impact of US sanctions.
The influence of major orders and key clients on domestic GPU makers' path to profitability cannot be overstated. For instance, Moore Threads announced on March 31, 2026, that it had signed a product sales agreement with a client worth a total of 660 million yuan, enabling the company to achieve profitability in a single quarter and significantly narrow losses in the first half. This begs the question: when will the next time window for domestic GPUs arrive, and will more Cambricon-like success stories emerge?
The Supply Chain Is the Crux
The vacuum left by NVIDIA's retreat must inevitably be filled by domestic chips. In 2025, NVIDIA still commanded an absolute dominant 55% share of China's accelerator market, while domestic players held a comparatively low and fragmented overall share. NVIDIA's financial reports explicitly record China's data center computing revenue (H100/H200/H20) as zero. Should NVIDIA's market share fall to 8% in 2026 (with consumer-grade gaming cards remaining unrestricted), over 40 percentage points of market space would be collectively released, presenting a significant substitution window for the domestic AI computing industry. (Data source: Zheshang Securities)
Beyond substitution, the domestic AI computing sector remains an incremental growth track. IDC forecasts that 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 growing, and more Cambricon-like companies will inevitably emerge, yet opportunities will not be evenly distributed among players. Success depends not only on securing major orders but also on supply chain management capabilities. Morgan Stanley research indicates that HBM3E supply will become the biggest bottleneck for China's AI GPU shipments in 2027, with supply to Chinese manufacturers tightening further as the global trio pivots to HBM4. "Supply chain is the hardest challenge right now. Most capacity goes to domestic giants, leaving limited allocation for smaller players like us," a GPU industry insider revealed.
GPU manufacturers' balance sheets corroborate this trend. Cambricon Technologies Corporation Limited holds inventories of 8.248 billion yuan and prepayments of 2.914 billion yuan; Moore Threads carries inventories of 3.55 billion yuan and prepayments of 1.344 billion yuan; and MetaX (沐曦股份) maintains inventories of 1.432 billion yuan with prepayments of 2.196 billion yuan. Inventories typically comprise wafers, work-in-progress items (masks, bare dies), raw materials (HBM, substrates), and finished goods (packaged boards), while prepayments represent "capacity reservation fees" to foundries, "packaging prepayments" to assembly and test houses, and "memory prepayments" to HBM suppliers. An analysis of Cambricon's 8.248 billion yuan inventory structure reveals raw materials of 5.749 billion yuan, outsourced processing goods of 2.537 billion yuan, and finished goods of merely 85 million yuan, demonstrating that its chips sell as soon as they are manufactured, with no demand concerns whatsoever. Moore Threads and other GPU manufacturers exhibit remarkably similar inventory structures.
Companies like Cambricon Technologies Corporation Limited, Moore Threads, and MetaX (沐曦股份) are actively building stockpiles, leveraging prepayments to secure capacity and raw materials. Compared to A-share listed peers, Hong Kong-listed GPU companies employ more flexible capacity-locking mechanisms through deeper interest alignment. Under the HKEX 18C/Chapter 17 rules, ILUVATAR COREX and Biren Technology can now expand their incentive programs to include supply chain suppliers such as wafer foundries, assembly and test facilities, and EDA providers, ensuring stable access to advanced process capacity and toolchains. For instance, ILUVATAR COREX's H-share equity incentive plan covers service providers across wafer manufacturing, packaging and testing, and EDA, with a ten-year validity period and a minimum vesting period of twelve months. Everyone is going all out to secure inventory.
The potential of domestic AI computing power has yet to be fully realized. Once supply chain bottlenecks are alleviated, domestic AI chip shipment progress is expected to accelerate, with visible results anticipated in 2027 and 2028. In the near term, valuations may remain suppressed by factors such as lock-up expirations, but over the long haul, GPU manufacturers are poised for valuation recovery driven by earnings growth.
Comments