摩尔线程交出了第一份单季扣非净利润转正的财报。
10月28日晚，摩尔线程发布2026年第三季度报告，该季度实现营业收入15.52亿元，同比增长1773.92%，环比增长55.4%；归母净利润4926.04万元，扣非归母净利润2007.24万元，均同比扭亏为盈。
收入创下单季度新高，增长开始反映在利润上，对于仍需持续投入研发的GPU企业，扣非转正是一个值得关注的变化。它意味着，在剔除相关非经常性损益后，摩尔线程当季仍实现了盈利。当然，第三季度扣非净利率约为1.29%，利润空间还不厚。
把时间线延长到前三季度，摩尔线程累计营收32.89亿元，归母净利润3769.73万元，扣非归母净利润仍亏损约1.31亿元。
这使得摩尔线程成为观察国产GPU商业化的一个样本，当销售规模扩大，市场会进一步关注收入如何转化为利润，以及增长需要付出多少资金成本。
盈利前移一年，毛利率居国产GPU第一梯队
摩尔线程此前表示，预计2027年实现盈利，本份财报将盈利时间表大幅提前。其提到，收入增长是因为人工智能产业对算力及全功能GPU的需求，以及以MTT S5000为代表的智算产品加速落地。
其披露，客户包括互联网及云服务商、词元服务商、AI创新企业、运营商及智算中心等。9 月，摩尔线程与京东云宣布深度合作，拟建设十万卡全功能 GPU 集群，面向大模型训练、推理和具身智能。运营商侧，中国移动披露移动云北京 Token 中心顺义节点部署了 S5000。10 月，无锡惠山国产智算中心一期满载运行。基于 S5000 的 PD 分离异构推理方案，已经承接了国内头部模型厂商的真实词元流量。
通常而言，算力芯片一旦进入客户的生产环境，后续的扩容要考虑已有的软件适配、运维经验和迁移成本，存量客户有机会带来持续采购。
前三季度，摩尔线程研发投入13.40亿元，同比增长55.72%，已经超过2025年全年13.05亿元；第三季度研发投入5.71亿元，同比增长87.89%。公司表示，围绕新一代全功能GPU“花港”架构的产品等项目，研发进度正在加快。
比利润更需要耐心观察的，是现金回流。前三季度，摩尔线程经营活动现金流量净额为-33.09亿元。公司解释，为应对需求增长、保障供应链，公司扩大生产规模并实施战略性备货，采购支出随之增加。
截至9月底，公司存货约33.63亿元，去年末为13.32亿元；预付款项从去年末约17.82亿元增至25.99亿元；应收账款从约4.34亿元增至9.94亿元。
从综合毛利率看，摩尔线程在国产AI芯片厂商中处于较高水平。2026年上半年，其综合毛利率为56.95%，与沐曦的约57.2%、寒武纪的约55.3%接近；壁仞科技、燧原科技、天数智芯分别约为42.7%、30.4%和17.2%。各家公司业务结构存在差异，综合毛利率不能直接等同于GPU产品毛利率。
摩尔线程这份财报呈现出的事实是，收入扩张的同时，单季扣非利润跨过了盈亏线，但前三季度累计扣非亏损、较大的经营现金净流出，以及持续增长的研发投入，说明公司仍处在扩张与盈利相互磨合的阶段。
国产算力芯片的竞争，进入交付阶段
摩尔线程不是唯一一家带来好消息的国产GPU公司，10 月 23 日晚，寒武纪披露三季报，前三季度营收 95.26 亿元，同比增 106.76%，归母净利润 37.15 亿元。再往前两天，海光信息发布三季报，前三季度营收 149.55 亿元，同比增 57.58%，归母净利润 28.47 亿元。沐曦股份的业绩预告给出前三季度 25.41 亿到 31.06 亿元的区间，同比扭亏。
这是行业景气程度的变化，主要发生在需求结构层面，过去两年国产算力采购方集中在几家大模型公司和各地智算中心，现在推理的占比在快速上升，客户名单里多出了词元服务商，它们的生意直接挂钩 Token 调用量，调用量越大，对 GPU 的需求越刚性。
推理场景的国产替代进度要快于训练，寒武纪、昆仑芯的部分产品已能平替 H20 级别的小卡，大参数训练仍主要依赖海外卡，国内卡多用于 Agent 推理与平台调度。
HBM 与代工产能仍然是制约增长的重要因素，各家厂商都在预防性增加存货，直接诱因源于上游的供给紧缺和涨价。
例如，台积电已确定将按不同制程上调晶圆价格 3% 到 6%；三星上调代工订单价格，先进制程涨幅最高 15%；中芯国际、华虹、晶合集成等内地代工厂也陆续提价。存储侧，TrendForce 预计 2026 年第四季度一般型 DRAM 价格季增 10% 到 15%，整体维持供不应求。
缺口主要卡在 HBM 上。全球 HBM 产能长期集中在 SK 海力士、三星电子和美光三家，中国本土供应链最薄弱的一环正在这里。
长鑫存储的 DRAM 月产能预计到 2026 年底达到约 35 万片，2028 年的目标是 50 万片，但它在 HBM 上走到的是另一条时间线：2026 年 9 月初开始试产 HBM3，年底前后才有小批量量产，月产能目标约 6 万片晶圆，HBM4 仍处于预研阶段。
长鑫在 HBM 技术上落后头部厂商约三到四年，彭博行业研究的一项调查估算，2026 年中国国产 AI 芯片需求约 420 万颗，供给约 260 万颗，160 万颗的缺口主要受制于 HBM 配给，而非芯片设计或逻辑制程产能。
业内普遍认为供应紧张会持续至少 24 个月，2026 到 2027 年仍是卖方市场，一位买方人士表示，各家最新一代产品的技术能力正在接近，差别落在适配、交付和商务环节，能不能持续规模化交付才是分水岭。（本文首发于钛媒体APP，文 | TechPulse）
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The Moor route handed over the first one-season debarment of non-net profits.
On the evening of 28 October, the third quarter of publish2026, Moor Line, reported that it had achieved operating income of $1,552 million, an increase of 1773.92 per cent over the same period and an increase of 55.4 per cent over the ring; the net profit from the mother ' s return of $4926.04 million, excluding the net profit from the non-mother ' s return of 2007.24 million, was in excess of the same amount.
Income has reached a new high in the single quarter, and growth has begun to be reflected in profits. For GPU enterprises that still need sustained investment in R & D, withholding is a notable change. It means that, after removing the related non-recurring gains and losses, the Moor line was profitable during the season. Of course, in the third quarter, the deduction of non-net interest rates was about 1.29 per cent, leaving little room for profit.
The extension of the time line to the first three quarters resulted in a cumulative profit of RMB 3,289 million for the Moor route, resulting in a net profit of RMB 3,769,733 for the mother, and a loss of some RMB 131 million for the non-mother ' s net profit.
This makes the Moor route a sample for commercialization of GPU, an observer country, and, when sales expand, the market will pay more attention to how revenue translates into profits and how much it costs to grow.
One year before the profit, the Māori rate was the first GPU.
The Moor line previously indicated that it was expected to be profitable in 2027 and that the present financial statement would be well ahead of the profit schedule. It was mentioned that the increase in income was due to the demand of the artificial intelligence industry for computing and full-capacity GPUs and the accelerated landing of smart products represented by MTT S5,000.
Its disclosures include Internet and cloud service providers, terminologists, AI Innovation Enterprises, Operators and IQC, among others. In September, Moor Line announced an in-depth collaboration with Kyoto Dongyun to build a 100,000-calorie fully functional GPU cluster for LLM training, reasoning and intelligence. On the side of the operator, S5,000 was deployed at the Suyi node of the Token Centre for Mobile Disclosure in China. In October, the Centre de télécniques de l ' économiques et de l ' économiques nationale de Côte d ' Ivoire (CIF) in Sehueshan expired. The S5,000-based PD separation isomer reasoning programme has been carried over the real word flow of the domestic head modeler.
Typically, once the calculator chip enters the customer ' s production environment, subsequent expansion takes into account the suitability of the available software, the operational experience and the costs of migration, and the opportunity for the inventory customer to bring about continuous procurement.
In the first three quarters, Moor line research and development invested RMB 1,340 million, an increase of 55.72 per cent over the same period, exceeding RMB 1,305 million in the full year of 2025; R & D investment in the third quarter increased by 87.89 per cent over the same period. The company indicated that research and development was accelerating around projects such as a new generation of full-purpose GPU “hwa Harbour” structures.
What is more patient than profit is the return of cash. In the first three quarters, the net cash flow from the Moor line activities was $3.309 billion. Companies explained that, in response to increased demand and secure supply chains, they had increased their procurement expenditures by scaling up production and implementing strategic readiness.
By the end of September, the company ' s stock stood at approximately $3,363 million, compared with $1,332 million at the end of last year; advances increased from approximately $1,782 million at the end of last year to $2,599 million; and receivables increased from approximately $434 million to $994 million.
In terms of the combined Māori rate, the moor line is at a high level among the national AI chip producers. In the first half of 2026, the combined Māori rate was 56.95 per cent, which was close to approximately 57.2 per cent of the anointed and 55.3 per cent of the cold war, and about 42.7 per cent, 30.4 per cent and 17.2 per cent, respectively, of the mural technology, the tiara technology and the day core. There are differences in the business structure of companies and the combined Māori rate cannot be directly equal to that of GPU products.
The financial report of the Moor line shows that while income expansion goes hand in hand with single-season non-profit deductions crossing the profit-loss line, the cumulative deduction of non-losses, large net operating cash outflows, and growing R & D inputs over the first three quarters indicate that firms are still at a stage of expansion and profit-making convergence.
Competition for the national production chip, at the delivery stage
The Moor line is not the only national GPU company with good news. On the night of October 23rd, Cold War revealed that the three-quarter editions of the newspaper had received $9,526 million, an increase of 106,76 per cent over the same period and a net profit of $37.15 billion for the mother. In the first two days, the publish tri-quarterly newspaper Hairay received $144.5 billion for the first three quarters, an increase of 57.58 per cent over the previous year, to the net profit of 28.47 billion yuan for the mother. The performance forecast for the musaka shares for the first three quarters was 25.41 billion to 3.106 billion yuan, a loss on the same basis.
This is a change in the extent of the industry, mainly at the demand-structure level, where the share of reasoning is now rising rapidly in the last two years, with a concentration of domestic purchasing power in a few LLM companies and local centres of wisdom, with more word-servers on the client list, whose business is directly linked to Token’s call, the higher the call, and the more demanding the demand for GPU.
The national production substitution of the reasoning scene is faster than the training, some of the products of the cold martial arts, Kunlen cores have been able to replace the H20-grade card, large-parametric training remains largely dependent on overseas cards, and domestic Cardo is used for Agent reasoning and platform movement.
HBM and surrogate capacity continue to be important constraints on growth, with manufacturers increasing stocks in a preventive manner, directly as a result of upstream supply shortages and price increases.
For example, it has been determined that the price of the round will be adjusted from 3 to 6 per cent at different production scales; the order price of the three-star shifter will increase by the highest 15 per cent for the advanced process; and the price will be raised from time to time in the interior for factories such as CGI International, Zhuan, Crystal Integration, etc. On the side of storage, TrendForce projected a quarterly increase of 10 to 15 per cent in average DRAM prices in the fourth quarter of 2026, with overall supply not needed.
The gap is mainly stuck on HBM. The global HBM capacity has long been concentrated in the SK Hercules, Samsung Electronics and American Lights, and the weakest link in China's indigenous supply chain is here.
The long-stored DRAM monthly capacity is expected to reach about 350,000 by the end of 2026, with a target of 500,000 by 2028, but on HBM it has reached another timeline: 2026, early in September, when the commissioning of HBM3 began, and only after the end of the year there was a small batch of production, the monthly energy target was about 60,000 crystals, and HBM4 was still in the preparatory stage.
A study by the Bloomberg industry estimated that, in 2026, China’s national production of AI chips required about 4.2 million, providing about 2.6 million, and that the gap of 1.6 million was mainly subject to HBM rations rather than chip design or logical process capacity.
It is widely believed that supply constraints will last for at least 24 months, and between 2026 and 2027 it will remain the seller ' s market, with one buyer saying that the technological capabilities of the newest generation of products are approaching, that the difference is between fit, delivery and business, and that sustainable scale delivery is the watershed. (This post was launched on titanium medium APP, Man, TechPulse)
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.