单季归母净利润212亿元的公司，两个交易日股价下跌近15%；业绩预增近6倍的公司，跌停之后第二天股价翻红。同一段行情里，决定股价的已经不是增长有多快，而是盈利的质量。
宁德时代近日披露三季报显示，第三季度宁德时代营收1652.21亿元，同比增长58.58%，创单季历史新高，前三季度合计4421.37亿元。
问题出在利润端，综合毛利率连续三个季度下滑，从一季度的24.82%到二季度的23.15%，再到三季度的20.69%，创2022年四季度以来新低；单季归母净利润212.46亿元，同比增长14.54%，低于财报披露前所有主流机构的预测下限，伯恩斯坦预计229亿元，高盛240亿元，中信里昂250亿至260亿元。
管理层在业绩交流中把利润压力归为三项，前两年低价周期签订的储能长协订单集中在三季度交付，9月起开征2%的锂电池消费税，4月起电池出口退税率由9%下调至6%；第一项被定性为一次性的成本错配，后两项是长期变量。
这些解释没能止住抛售，10月21日，宁德时代A股收跌12.03%，成交324亿元，创上市以来第二大单日跌幅，单日市值蒸发约1400亿元；10月22日继续下跌至220.41元，两个交易日累计近15%，较5月7日467.35元的年内高点近乎腰斩。
同一天，另一类资产走出相反的行情，移远通信两连板，恒玄科技、瑞芯微涨逾5%，端侧AI是资金最集中的承接方向。上涨的驱动来自两方面，DeepSeek前一日开源推理框架DeepInfer，市场由此形成推理成本下降的预期；Counterpoint则预测2027年生成式AI手机的渗透率将达到52%。
科创50ETF华夏当日再获近15亿元净流入，指数下跌的同时，资金维持净申购。
10月12日收盘后，光芯片公司源杰科技发布业绩预告，预计第三季度归母净利润4.03亿元至4.53亿元，同比增长578.85%至659.70%，单季营收增速在264%以上。此前其股价在10月8日20%跌停；预告披露后的首个交易日低开6.15%，收涨3.64%。同日晚间，联讯仪器预告前三季度净利润19亿至21亿元，同比增长1866%至2073%，次日收涨4.37%，重夺A股股王。业绩高增长的公司，披露后仍获得资金认可。
调整已持续近一个月，9月22日至10月22日，沪指累计下跌7.39%，创业板指下跌24.20%，通信、电子板块跌幅都超过20%；10月13日至16日当周，科创50下跌6.77%，周线四连阴，创业板指单周跌8.90%，中证红利当周下跌0.37%，银行上涨1.52%。
10月上旬风格分化一度非常清晰，截至10月12日，中证红利上涨2.29%，银行上涨2.28%，多只银行股创历史新高；但这轮防御行情未能延续超过两周，中证红利10月14日见顶后回吐全部涨幅，10月累计转为下跌1.42%，防御方向的资金也开始松动。
业绩大超预期的公司被资金买入，板块估值却持续下移，这轮调整在杀什么，决定了接下来一个季度如何研判AI板块。背离指向的答案大体清楚，被压的是估值，算力链已兑现的业绩并未恶化；发生动摇的，是市场对增长可持续性的信心，以及此前靠AI叙事给板块估值的逻辑。
跌下去的是估值
10月13日至16日当周，上证指数跌2.56%，沪深300跌3.40%，深证成指跌6.64%，科创50跌6.77%，科创综指跌8.26%，创业板指跌8.90%。跌幅与成长属性呈正相关，风格分化清晰。
第一根大阴线出现在10月8日，比美国9月CPI公布早一周，触发因素是摩根士丹利10月1日的光模块调研纪要，其推演自3.2T代际起，物料清单中美系价值占比不低于65%才可能获得美国市场豁免，对当前800G、1.6T产品影响有限。情景本身偏温和，光芯片环节应声重挫，源杰科技、长光华芯20%跌停，科创50单日跌近5%。
10月13日，美国FCC 26-50新规生效，设备搭载管制清单内企业生产的逻辑承载硬件组件，将无法获得新的设备认证。核对规则原文与清单，中际旭创、新易盛、源杰并不在其列，直接压力落在华为、中兴生态的设备与路由器、穿戴、无人机等品类。市场定价的，是清单会不会扩容的预期。一周后公布的9月CPI超预期、加息预期升温，成为调整期内的第二层压力。
分歧由此产生，贴现率上行是分母端的压力，研报推演与FCC清单扩容预期动摇的是海外收入的可持续性，那是分子端的事。公司层面的基本面目前没有恶化，源杰科技说明数据中心CW产品出货持续增长、毛利率维持高位，管理层称未来两年需求能见度很高；联讯仪器的弹性同样来自真实订单。同一家公司，一边承受海外监管的估值压制，一边在加深海外绑定，两个方向合在一起，指向的是估值。
估值维度上的分化也明显，截至10月16日，创业板指PE-TTM约32倍，较6月约54倍回落四成；科创50在110倍上方，较6月约264倍的高点回落近六成，绝对水平不低。调整让AI板块的估值从高位回落，但算力核心资产的估值消化尚未完成。算力链核心公司中，尚未见到业绩指引下修的公开案例。
科创50前十大权重合计59.66%，覆盖芯片设计、晶圆制造与半导体设备，寒武纪权重约10%，唯一的光芯片公司源杰科技位列第六，它实质上是一篮子AI算力资产的影子指数，跌6.77%，等于几十只算力股按同一个逻辑同时下跌。市场的担忧，一半来自贴现率重估下的仓位再平衡，一半来自海外监管对收入预期的威胁。
产业景气数据仍在向上，SIA数据显示，8月全球半导体销售额1597亿美元，同比增长144.3%；海关口径下，前三季度我国集成电路出口增长108.1%。但两组数据的增量主要来自存储涨价与再出口，机构复盘也把存储供给格局的担忧列为这轮回调的导火索之一。多空双方引用的是同一组数据，分歧在未来订单与需求结构。
综合这些证据，这一轮的主基调仍是杀估值，已兑现的盈利没有恶化，调整的底部由业绩支撑；估值何时修复，取决于贴现率与风险偏好，也就是叙事和资金。而海外监管正把分歧从分母端引向分子端，这是后续最大的变数。
交易盘撤出，配置盘承接
资金面常被误读，直觉上股价下跌对应资金流出，数据给出的却是相反的答案。
Wind数据显示，9月28日以来股票型ETF连续8个交易日净流入，截至10月14日累计947.53亿元，其中科创50ETF华夏净流入154.63亿元。到10月21日，科创50ETF华夏规模达1060.92亿元，近一个月份额增长35.25%、创历史新高，新晋千亿ETF。申购力度随跌幅扩大而增强。10月24日的动作更直接，多只宽基ETF的放量从往日的尾盘提前到上午，创业板ETF易方达全天成交197.38亿元，创2024年10月9日以来新高，沪深300ETF华泰柏瑞、科创50ETF华夏分别成交154.23亿元和151.97亿元。
ETF份额在指数下跌中持续增长，说明抛压与承接来自不同的资金。交易盘对贴现率和拥挤度敏感，在涨幅累积后获利了结，在利空出现时止损离场；配置盘依据估值分位和产业趋势布局，估值越低增配力度越大。两类资金同场交易，指数连跌与份额新高才能够并存。
同样的现象出现在创业板，截至10月15日，创业板相关ETF近一周净流入超150亿元，创业板ETF华夏连续11日净流入。银河证券指出，红利与防御配置需求升温，国资增持、被动资金回流和险资布局支撑银行股估值。银行端有被动资金托底，成长端在下跌中持续获得增配，下方承接力量逐步增强。
避险端同样不稳，前文提到的红利见顶回吐就发生在这两周，仅上证50连续四周保持周度最抗跌。资金既未留在AI板块，也未停留于红利资产，而是在观望，等待新的主线形成。
开头那两笔10月21日的行情，最能说明当前市场如何定价，盈利遭质疑的公司被抛售，订单兑现的公司被买入。当日1.43万亿元成交额说明市场流动性充裕，缺乏的是一致预期。
一级市场：定价权转向产业资本
二级市场分歧最大的问题，一级市场给出了更清晰的答案，AI资产的定价权，正在从财务投资者转向产业资本。
DeepSeek第二轮融资已确认完成，据报道及工商登记信息，规模约1000亿元，融资目标在洽谈期间至少两次上调，从最初的500亿元提到800亿元，最终逼近1000亿元，距首轮约500亿元的外部融资仅过去四个月。腾讯与宁德时代是出资最多的投资方，投前估值维持约5000亿元不变，新老股东出资各占约一半。
名单的变化传递信号，腾讯、IDG等老股东以财务投资为主，新进的石溪资本、中芯聚源、CPE源峰、君联资本与吉利以产业方为主，其中宁德时代的出资从首轮约50亿元增至约100亿元。绑定产业链的那一端，权重明显上升，新增资金的投向也集中在产业链上，7月参与长鑫科技打新，8月以1.41亿元战略配售宇树科技，10月计划领投一家机器人初创公司。
产业资本的估值框架不同于市盈率定价，更看重供应链卡位、算力协同与订单绑定，估值逻辑从财务回报转向产业绑定，因而比二级市场情绪更稳定。这也解释了一级市场为何能在二级暴跌时维持估值不变。
智谱较6月盘中高点回落逾八成，但这个八成是回撤幅度，部分反映其上市初期自由流通股不足总股本3%、7月解禁与配售让稀缺性溢价到期，不等于二级市场在给AI统一打折。尽管如此，一级约5000亿元的投前估值与二级的深度回撤并存，价差已经足够显著。
据《财经》等媒体报道，DeepSeek自8月下旬起接触多家投行，已聘请中信证券保荐，最快年底递表、2027年挂牌；为规避股东人数200人的监管红线，部分专项基金改为盲池备案，9月下旬原高瓴创投合伙人严文韬出任CFO，结束公司三年没有CFO的局面。商业化层面的数据也开始兑现，梁文锋在近期投资者会议上披露，年化营收运行率已突破10亿美元，较数月前不足5亿美元翻倍有余，API提价后客户未出现流失。
IPO窗口与深度回撤同时出现，窗口在收窄，先上者获得定价权，后上者承受更苛刻的定价。
类似的重估也发生在其他头部公司身上，月之暗面最新一轮融资估值约500亿美元，各报道对投前投后口径存在分歧；其ARR从6月的约3亿美元增至8月后的约10亿美元，对应约50倍ARR，定价依据是未来收入的增长速度，而不是当期收入。就在截稿当天，MiniMax跌破165港元发行价，为上市以来首次破发，是同一逻辑的又一例证。
对二级市场而言，一级市场以实际出资为AI核心资产提供了产业定价，这一定价由产业资本的长期利益决定，不随单周行情变动，为二级市场提供了相对稳定的参照。
技术目录修订：开源豁免，本地化部署登记
10月16日，商务部、科技部就修订《中国禁止出口限制出口技术目录》公开征求意见，拟新增禁止出口技术14项、限制出口技术26项，其中人工智能领域23项，覆盖大模型、智能体、算力集群、信息处理等环节。同一天，国务院常务会议听取“人工智能+”行动进展情况汇报。收紧与鼓励同日出台，政策并非单边转紧。
豁免安排比列管清单的影响更直接，商务部发言人明确，已通过开源、发表学术论文等方式进入公共领域的技术，不纳入出口管理范围；人工智能类技术开源、境外自建自营算力中心、赴境外开展芯片流片制造等情形，适用自动许可等便利措施，基本“即报即出”；企业利用大模型对外提供API调用、App/网页交互、SaaS应用、本地端侧轻量化部署，不属于大模型技术出口管理范围。对纳入限制类的大模型，海外本地化部署属于管理范围；部署在海外自有云等风险可控情形下，可适用自动许可。
落到产业层面，条款若按征求意见稿落地，对DeepSeek为代表的开源路线是明确倾斜。开源被排除在管制之外，等于确认开源是受鼓励的路线；依赖闭源模型海外部署的厂商，境外交付进入管理范围，新增合规成本。公告尚在征求意见，以正式文本为准。
目录上一次因AI引发关注还是2020年，新增个性化信息推送技术，直接改变了字节跳动出海交易的谈判格局。这次选择在大模型资本运作最密集的窗口期列管，并同步启动出口前登记，规则先行于交易。征求意见期本身就是企业调整商业模式的窗口。
政策的另一面，也在为二级市场的分化定价提供依据，开源生态的厂商获得了一层政策保障，闭源出海的厂商则要承担额外的合规成本。当产业资本在一级市场主导定价、监管在政策层面划出边界，AI公司之间的估值分化将不再由板块整体行情决定，而由公司自身的差异决定。
三季报是最近的验证点
几条线索在此汇合，估值回到相对合理的区间，配置资金在逆势买入，一级市场给出了更稳的产业定价，政策明确了技术路线的边界。剩下的变量只有两个，贴现率何时回落，新叙事何时凝聚。
中信证券10月9日的研报判断，下一轮科技行情的窗口大概率在四季度，需要三个条件验证，加息预期边际逆转、新叙事凝聚共识、资金面完成再平衡。三个条件中，前一个是外生变量，取决于美国的通胀路径；后两个，三季报就是最近的检验。
资金面的再平衡走得最靠前，ETF逆势申购已持续三周以上，产业资本在一级市场维持估值，政策边界已经划出。
悬而未决的是叙事，上一轮共识是算力军备竞赛，微软将2026年资本开支的账面口径由约1900亿美元下调至约1750亿美元，Meta收窄了指引区间，海外对AI投资回报的审视收紧，旧叙事的力量在衰减；端侧AI放量、推理成本下降、国产算力闭环，均未凝聚到足以接力的程度；10月24日人形机器人板块全天上涨7.98%，宇树科技录得上市以来首次涨停，是新的候选之一，但单日行情尚不足以构成叙事。叙事真空、海外监管冲击，叠加加息预期反复，是这轮调整的成因。
A股已进入三季报密集披露期，要观察三件事，算力链订单能见度能否延伸到2027年，毛利率在产能爬坡期是升是降，端侧AI放量是脉冲还是趋势。
第一问在10月23日至24日有了答案，中际旭创公告海外主要客户订单已排至2027年，部分重点客户给出2028年需求的明确指引；新易盛说明1.6T光模块进入持续放量阶段，并抛出8亿至10亿元的回购方案；天孚通信、光迅科技、源杰科技同步公告在手订单充足、行业景气度没有变化。后两问需等待更多财报披露。
三者兑现，本轮调整就是新叙事的起点；出现下修，说明杀估值只是提前反映了对2027年的疑虑。
最后回到开头那组矛盾，业绩大超预期的公司被资金买入，板块估值却仍在下移；一级市场维持估值不变，二级市场则对流动性稀缺的已上市公司反复定价。
这组矛盾在10月24日出现了解开的迹象，当天早盘受隔夜纳指下跌2.6%拖累，沪指一度跌至3574.27点，跌破3600点；午间五家光通信公司的订单公告落地，午后市场完成V型拉升，创业板指从跌超3%到收涨4.16%，科创50上涨3.04%，全天成交1.96万亿元，较前一交易日放量4276亿元。同一天，银行板块下跌1.30%，中国银行跌超4%，资金从防御方向撤出，回流成长。
周线仍是阴线，沪指本周跌0.69%，创业板指跌3.18%，科创50跌0.97%，单日反弹未能收复此前一个月的跌幅。但从ETF连续两周的净申购，到午间那五份订单公告，资金的倾向已经明确。（本文首发钛媒体APP，作者 | AGI-Signal，编辑 | 焦燕）
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
A single-season net profit of $21.2 billion fell by almost 15 per cent for two trading days; a company whose performance had increased by a factor of almost six rose the following day. In the same line of action, stock prices are determined not by how fast they grow, but by the quality of profits.
The recent disclosure in the Ningde Times of the tri-quarters shows that the Ningde era received $16,5221 million in the third quarter, an increase of 58.58 per cent over the same period, a single-season record high of $44,237 million in the first three quarters.
At the end of the profit line, the combined Māori rate declined for three consecutive quarters, from 24.82 per cent in the first quarter to 23.15 per cent in the second quarter and 20.69 per cent in the third quarter, which is a new low since the fourth quarter of 2022; net profit from single-season maternity benefits of $21,246 million, an increase of 14.54 per cent over the same period, is lower than the forecast floor for all mainstream institutions before the financial disclosure. Bernstein expects $22.9 billion, Goldman Sachs $24 billion and Central Sinion $25 billion to $26 billion.
Management classified the profit pressure in the performance exchange as three, with the first two-year low-priced CAC orders concentrated on three quarters, a 2 per cent lithium battery consumption tax introduced in September and a 6 per cent reduction in battery export refund rates from 9 per cent to April; the first was characterized as a one-time cost mismatch, and the second two were long-term variables.
These explanations did not stop sales, and on 21 October, the Ningde-era A stock fell by 12.03 per cent to $32.4 billion, the second-largest single-day fall since it was listed, evaporating the market value on a single-day basis by about $140 billion; on 22 October, it continued to fall to $220.41, with two trading days aggregating almost 15 per cent, almost a little higher than the annual figure of $467.35 on 7 May.
On the same day, another type of asset moved out of the opposite direction, moving away from two lines of communication, with Hengentech, Sichuan, slightly increasing by more than 5 per cent, with the end side AI being the most centralized recipient. The driving forces behind this increase are two-fold, DeepInfer, the previous day ' s open-source reasoning framework in DeepSeek, which led to the market ' s expectation of decreasing the cost of reasoning; and Counterpoint, which projected 52 per cent penetration in 2027.
The 50 ETF net inflow was nearly $1.5 billion a summer, while the index fell while maintaining a net requisition.
After closing off on 12 October, the performance forecast of the light chip company ' s source publish technology forecast a net profit return of between $403 million and $453 million to the mother in the third quarter, an increase of 578.85 per cent to 659.70 per cent over the same period, with a single-season increase of more than 264 per cent. Its share price had previously fallen by 20 per cent on 8 October; the first trading day following the announcement had been lower by 6.15 per cent, resulting in an increase of 3.64 per cent. During the evening of the same day, the UNISPAL instrument forecasted a net profit of 1.9 billion to 2.1 billion yuan in the first three quarters, an increase of 1866 to 2073 per cent over the same period, an increase of 4.37 per cent the following day and the re-emergence of the A equity owner. Companies with high performance growth continue to receive financial recognition after disclosure.
The adjustment lasted for almost a month, from 22 September to 22 October, with a cumulative decline of 7.39 per cent, a decline of 24.20 per cent in the entrepreneurship plate, and a drop of more than 20 per cent in communications and electronic plates. During the week of 13 to 16 October, the number of science cases fell by 6.77 per cent, the number of weekdays fell by 8.90 per cent, the number of weekdays fell by 0.37 per cent and the number of banks increased by 1.52 per cent.
The pattern of split in the first half of October was clear, with a 2.29 per cent increase in the value of the card, a 2.28 per cent increase in banks and an all-time high in the number of bank shares as at 12 October; however, the line of defence did not last for more than two weeks, with a full increase in the value of the card on 14 October and a cumulative decrease of 1.42 per cent in October, and the funds in the defence direction began to loosen.
Firms with high performance expectations were bought into the funds, while the valuation of the block continued to move downwards, and the round of adjustments was killing what, determining how the AI block would be studied in the next quarter. The answer to the deviation is generally clear: it is the valuation that is pressed and the performance of the calculus chain that has been realized has not deteriorated; it is the confidence of the market in the sustainability of growth that has been shaken, and the logic of previous valuations by AI narratives.
The downfall is the valuation.
During the week of 13-16 October, the index dropped by 2.56 per cent, depth by 3.40 per cent, depth by 300, depth by 6.64 per cent, science by 50 to 6.77 per cent, science by 8.26 per cent and entrepreneurship by 8.90 per cent. The drop is positively related to the growth properties, and the style is distinct.
The first large-coloured line appeared on 8 October, one week before the announcement of the United States CPI in September. The trigger factor was Morgan Stanley ' s photomodule survey of 1 October, which was extrapolated from 3.2T generations, with a percentage of the US dollar value on the material list not less than 65 per cent to obtain an exemption from the United States market, with limited impact on the current 800G, 1.6T products. The scenario itself is mild, with the photochip retweeting, with a 20 per cent drop in source-tech, long-light core and a 50-day drop of nearly 5 per cent.
On 13 October, the new United States FCC 26-50 regulations came into force, and the logical carrying hardware components produced by the enterprises on the control list will not be certified. Checking the original text and list of rules, which are not included in the list, directly pressurized by Chinese-made, ecologically viable equipment and routers, clothing, drones, etc. Market pricing is the expectation that the list will be expanded. A week later, CPI was released in September with an expected increase in interest rates, which became the second layer of pressure during the adjustment period.
Disagreements have arisen, with the upper discount rate being a parent-end pressure, and it is the sustainability of income abroad that is expected to be shaken by the development and expansion of the FCC list, which is a molecular matter. The basics at the corporate level have not deteriorated at this time, and sourcer technology suggests that CW products from the data centre continue to grow and have a high Māori rate, and management claims that demand is high over the next two years; the elasticity of telecommunication instruments also comes from real orders. The same company, while under pressure from overseas regulation, is deepening its overseas binding in both directions, pointing to valuation.
Differences in the dimensions of the valuation were also evident, with the entrepreneurship board representing about 32 times the PE-TTM as at 16 October, 40 per cent more than in June, and 50 times more than 110 times more than in June, almost 60 per cent more than in June, with an absolute level. The adjustment allowed the valuation of the AI plate to fall from a high position, but the absorption of the valuation of the Nucleus core asset has not yet been completed. No public cases of corrections under performance guidance have been seen in Nucleus.
A total of 59.66 per cent of the top 10 weights, covering chip design, crystal round manufacturing and semiconductor equipment, about 10 per cent of the weight of the cold martial arts, and the sixth largest source of technology for the sole photochip company, are essentially the shadow index of a basket of AI-calculable assets, down by 6.77 per cent, equal to the simultaneous decline of dozens of computing units in the same logic. Market concerns come in half from rebalancing the position at the revaluation of discount rates and in half from the threat to revenue expectations from overseas regulation.
The industry is still moving upwards, with SIA data showing that global semiconductor sales in August amounted to $159.7 billion, an increase of 144.3 per cent over the same period, while our exports of integrated circuits increased by 108.1 per cent in the first three quarters, under customs standards. However, the increase in two sets of data is mainly due to storage price increases and re-exports, and concerns about storage supply patterns have been identified as one of the triggers for this rotation. Both sides cited the same set of data, with differences over future order and demand structures.
Combining these evidences, the main tone of this round is still to be valued, the realized profits have not deteriorated and the bottom of the adjustment is supported by performance; when valuations are repaired depends on discount rates and risk preferences, i.e. narratives and funds. Overseas regulation, on the other hand, is leading differences from the denominator to the molecular end, the biggest remaining variable.
Disengage, configure the disk.
The financial context is often misread, the intuitive decline in equity prices corresponds to the outflow of funds, and the data provide the opposite answer.
Wind data show that since 28 September there have been eight consecutive days of net inflows of ETF equities, which as of 14 October totalled $94,753 million, of which $50 ETF net inflows of $15,463 million were recorded. By 21 October, 50 ETF was worth 10,092 million yuan, and its share had increased by 35.25 per cent in the past month, reaching a record high of 100 billion dollars. Requisitions increased with the fall. The 24 October movement was even more direct, with the release of many wide-based ETFs moving forward from the tails of the previous day to the morning, and the ETFs of the start-up board reaching $19,738 million a day, which is high since 9 October 2024, with 30 ETF Tabri and 50 ETF, respectively.
The ETF share has continued to increase in the index decline, indicating that the push-off and the take-over come from different sources of finance. Dealboards are sensitive to discount rates and congestion, benefit from the build-up of the increase and stop losses and losses when space is available; the layout is based on valuation rankings and industry trend layouts, with the valuation increasing at a lower level. The two types of funds are traded side by side, and the index drops and shares rise.
The same phenomenon occurred on the entrepreneurship board, where, as of 15 October, there had been net inflows of more than $15 billion to the EF and 11 consecutive days to the TF. Galaxy securities indicate that the demand for dividends and defence configurations is warming, and that the valuation of the bank unit is supported by additional State resources, the flow of passive funds and the venture capital allocation. The bank end has a passive capital base, the growth end has been continuously enriched in decline and the downside is gradually gaining strength.
The risk-averse is equally volatile, with the aforementioned dividends at the end of the week and only 50 consecutive weeks remaining the most resistant. The funds were neither on the AI board nor on the dividend assets, but were waiting for a new lead line to be formed.
The first two events of 21 October illustrated how the current market was priced, companies whose profits were questioned were sold, and companies whose orders were honoured. The sum of 143 trillion yuan that day indicated that the market was sufficiently liquid and that what was lacking was consistent expectations.
Level I markets: shift of pricing rights to industrial capital
The issue of the greatest divergence in the secondary market, where the first market provides a clearer answer, is shifting the pricing power of AI assets from financial investors to industrial capital.
The second round of DeepSeek financing was confirmed, with business registration information reported to be about $100 billion, and the target was raised at least twice during the negotiations, from the initial $50 billion to $80 billion, and eventually close to $100 billion, with external financing about $50 billion in the first round just four months. Tencent and the Ninde era were the most funded investors, with the pre-investment valuation maintained at around $500 billion, with both old and new shareholders contributing about half.
The changes in the list signal that old shareholders, such as Tencent and IDG, are dominated by financial investments, with the newly imported Rock Creek capital, the central core concentration, the CPE peak, the Kings' Union capital and the Jili dominated by industry, with funding from the Ningde era increasing from about $5 billion in the first round to about $10 billion. At the end of the chain, the weight of power has risen markedly, and the investment of additional funds has been concentrated in the chain. In July, it participated in the re-establishment of long-term technology, in August, when it was allocated 141 million yuan in a strategy to sell tree technology, and in October, it was planned to take over a robotic pioneer company.
The valuation framework for industrial capital is different from market gain pricing, with more emphasis on supply chain rankings, calculators and orders tied, and the logic of valuation shifting from financial returns to industry binding, thereby stabilizing the mood of secondary markets. This also explains why the first-level market can keep the valuation unchanged in the event of a second-level collapse.
Zhipu fell by more than 80 per cent from the June mid-point, but this is an 80 per cent retreat, partly reflecting its initial listing of less than 3 per cent of the total equity, the release of the ban and distribution in July, which resulted in the expiration of the scarce premium, and does not mean that the secondary market is discounting AI. Nevertheless, the pre-investment valuation of approximately $500 billion in level I combined with the depth retreat of level II is already sufficiently significant.
According to media reports, including the Financial Affairs Institute, DeepSeek has been in contact with a number of investment banks since late August and has been engaged to offer Chinese-intelligent securities, which will be submitted at the end of the year at the latest and registered in 2027; in order to circumvent the red line of supervision of 200 shareholders, some of the special funds were turned into blind pools for filing; and in late September, the former Takashi Founding Partner, Zhengwei, took office as CFO, ending three years without CFO. Data at the commercial level are also beginning to materialize, and Liang Wenfeng disclosed at a recent investor conference that the annualized operating rate had exceeded $1 billion, more than double that of less than $500 million a few months ago, and that there had been no loss to clients since API price increases.
The IPO window appears at the same time as the deep retreat, and the window is narrow, the first one gets pricing rights and the last one is subject to more stringent pricing.
A similar revaluation took place with other front companies, with the latest round of financing under the moon valued at approximately $50 billion, with reports of differences over the calibre of post-investment; its ARR increased from about $300 million in June to about $1 billion after August, corresponding to about 50 times ARR, based on the growth of future revenues rather than current revenues. On the same day, Mini Max broke the HK$165 distribution price for the first time since he was listed, another example of the same logic.
For the secondary market, the first market provides industrial pricing of AI core assets at actual cost, which is determined by the long-term interests of industry capital and does not vary from one week to another, providing a relatively stable reference for the secondary market.
Technical Directory Revision: Open Source Exemption, Localized Deployment Register
On 16 October, the Ministry of Commerce and the Ministry of Science and Technology issued a public consultation on the revision of China ' s Directory of Export-Restricted Technologies and proposed the addition of 14 new export-restricted technologies and 26 additional export-restricted technologies, of which 23 are in the area of artificial intelligence, covering the areas of LLM, AI agent, Calculator Cluster, Information Processing, etc. On the same day, the Standing Committee of the State Council was briefed on the progress of the AI.+ operation. Tightening and encouraging the same day were not unilateral.
The exemption arrangements are more direct than the impact of the list, and the spokesperson for the Ministry of Commerce has made it clear that technologies that enter the public domain through open sources, the publication of academic papers, etc., are not included in export management; facilities such as the use of automatic permits are applied in the case of artificial intelligence technology open sources, self-contained self-accounting centres outside the country and the production of chips outside the country; enterprises use LLM to provide external API calls, App/web interfaces, SaaS applications, light quantization deployments on the local side, which are not part of the export management of LLM technology. For LLM, which is included in the restricted category, overseas localization is regulated; automatic licences may be applied in situations where the risk of deployment is manageable, such as the presence of a cloud abroad.
At the industrial level, the clause, if written for comment, has a clear preference for the open source route represented by DeepSeek. The exclusion of open sources from control is tantamount to recognizing open sources as an encouraged route; relying on closed-source models for overseas deployment of manufacturers, offshore deliveries enter the regulatory domain and add additional compliance costs. The bulletin is still being consulted and is based on the official text.
The last time the catalogue was concerned about AI was in 2020, with the addition of individualized information dissemination technologies, which directly changed the negotiating pattern of byte jump-off transactions. This time, the option was to be placed in the window with the most intensive capital operation in LLM and to simultaneously initiate pre-export registrations, with the rule going ahead of the transaction. The consultation period itself is a window for business models.
The other side of the policy is also providing a basis for differentiated pricing for secondary markets, with open-source eco-producers receiving a layer of policy safeguards and closed-source operators bearing additional compliance costs. When industrial capital is priced at the first level of the market and regulations are drawn at the policy level, valuation differentiation between AI companies will no longer be determined by the overall size of the plate, but by differences among companies themselves.
The Three Seasons are the latest validation point.
Several threads converged here, valuations returned to relatively reasonable ranges, allocated funds were bought in reverse, first-tier markets gave more stable industrial pricing, and policies defined the boundaries of technical routes. There are only two remaining variables, when the discount rate will fall and when the new narrative will converge.
On October 9th, the Chinese-Chinese securities study concluded that the next window for science and technology performance would be roughly four quarters, requiring three conditions for validation, an expected marginal reversal of interest rates, a convergence of new narratives, and a rebalancing of funding. Of the three conditions, the former is an external variable, depending on the inflation path in the United States; the latter two, the three-quarters, are the most recent tests.
The rebalancing of the financial landscape was the most advanced, with the reverse requisition of the ETF having lasted more than three weeks, industrial capital being valued at the first level of the market, and policy boundaries being drawn.
What is pending is the narrative, the last round of consensus is a calculus arms race, Microsoft downgraded its 2026 capital expenditure account from approximately $19 billion to about $175 billion, Meta narrowed the directional zone, the offshore scrutiny of AI's return on investment tightened, and the power of old narratives fell; the side AI release, the reduced cost of reasoning, the closure of national production, did not congregate enough for relay power; the human robotic plate rose by 7.98 per cent all day on 24 October, and for the first time since Okitech was listed, it was one of the new candidates. This round of adjustments is the result of a narrative vacuum, an overseas regulatory shock, and expectations of double interest rates.
Unit A has entered an intensive three-quarter disclosure period to observe three things: whether the visibility of the chain orders can be extended to 2027, whether the Māori rate rises or falls during the capacity-crawling period, and whether the end-side AI discharges are pulses or trends.
The first question was answered from 23 to 24 October, when the Chinese-intelligent announcement of the order of the main overseas customers was scheduled to 2027, with some of the key clients giving clear guidance on the needs of 2028; the new euphoria showing that the 1.6T-ray module was entering the continuous release phase and throwing 800 million to 1 billion yuan in buy-back programmes; and the telecommunications, light-for-sight technology, and the simultaneous announcement of source-jet technology were well-placed and the industry landscape remained unchanged. The latter two questions require further disclosure.
In three cases, the adjustment for the current round was the starting point for the new narrative; the downward revision suggested that the valuation of the killing was only an early reflection of doubts about 2027.
Finally, back to the initial set of contradictions, in which large-performing, over-expected companies were bought and the valuation of the plate was still moving downwards; first-tier markets maintained the same valuation, while secondary markets repeatedly priced already-listed companies with scarce liquidity.
On 24 October, the paradox showed signs of understanding, with the early drive being dragged down by 2.6 per cent by the overnight finger, which fell to 3574.27 points at a time and fell to 3600 points; the five light communications companies at midday, where the order was announced, and the market completed the V lift after midday, with the start-up board ranging from a drop of over 3 per cent to a gain of 4.16 per cent, with a 50 per cent increase of 3.04 billion yuan, and a full-day deal of 19.6 trillion yuan, which was set at 42.7 billion on the previous day. On the same day, the bank block fell by 1.3 per cent, the Chinese bank fell by more than 4 per cent, and the funds were withdrawn from defensive directions and grew back.
The weekly line remains a vaginal line, with a drop of 0.69 per cent this week, a drop of 3.18 per cent in the entrepreneurship plate, a 50-0.97 per cent decline in science and a single-day rebound that failed to recover the decline of the previous month. However, from the two-week net requisition of the ETF to the midday announcement of the five purchase orders, the financial trend has been clear. (This article starts with titanium media app, by | AGI-Signal, editor | Joo Yon)
Auto-fetched and translated by AI Atlas. All rights belong to the original author.