文 | 大模型之家
当大洋彼岸的Anthropic带着上万亿美元的估值预期叩响华尔街大门时，以智谱为代表的中国大模型厂商，正低头在财报里一分一毫地核算着API生意的毛利。
10月26日，大模型独角兽Anthropic公开了上市招股书，准备在11月中旬正式挂牌。这份招股书把公司冲刺IPO的家底摊在了台面上：今年上半年营收162.9亿美元，净亏损高达807亿美元。在此之前，市场对它的最高估值讨论一度达到2万亿美元。
就在Anthropic叩响华尔街大门的前两个月，中国的大模型公司也交出了一份成绩单。8月底，智谱披露中报，上半年收入9.54亿元，亏损20.72亿元，其中API业务的毛利率首次由负转正。
可以看出，大模型行业正在上演两种截然不同的叙事：大洋彼岸在用上万亿美元的估值给未来的技术假设买单，而国内厂商正忙着用毛利率证明大模型本身是一门能持续赚钱的生意。
过去，行业里常有一种误解，认为海外公司靠技术和资本讲未来，中国公司还在为商业化发愁。然而只依靠资本，会虚弱市场对于“未来”定价的信心。就在最近的投资者日上，Anthropic的估值讨论已经从2万亿美元缩水到了1.5万亿美元。
“万亿估值”赌的是未来，不是现在
在大模型之家看来，资本市场向来不是为一家公司今天的价值出价，而是为未来的增长空间买单。
Anthropic的招股书，把这一逻辑体现得淋漓尽致。
表面上看，Anthropic上半年净亏损807亿美元，像是在“流血上市”。但拆开财报就会发现，其中有724亿美元是因为公司估值暴涨，导致可转债在账面上计提的公允价值损失。如果只看实际业务运营，它上半年的经营亏损其实已经收窄到了6.93亿美元。
真正让投资人兴奋的是它的营收增速。2025年全年，Anthropic收入46亿美元，同比增长约12倍（同期经营亏损扩大至80.6亿美元，算力与基础设施支出达73.3亿美元）。到了今年上半年，它的收入直接冲到162.9亿美元，是去年同期的约14倍。
而且，在2025年的46亿美元营收里，约38亿美元来自按量付费，固定订阅收入只有7.89亿美元。这意味着企业客户不是一次性买断某个软件，而是在持续地按Token调用量掏钱。今年第二季度，Anthropic单季收入达到115亿美元，首次反超同期收入67亿美元的OpenAI。
这恰好证明了，企业如今已经认可大模型的商业价值，愿意为代码生成、模型调用和复杂任务买单。
但撑起Anthropic上万亿美元估值的，显然不只是现有的几百亿收入，而是收入往上冲的“加速度”。而这正是Anthropic估值暴跌的根源。
财务数据显示，Anthropic二季度营收116亿美元，环比翻了一倍；但到了三季度，营收为168亿美元，环比增速放缓到了45%。虽然整体还在高速增长，但只要增速不及华尔街的预期，估值逻辑就会立刻动摇。随后，关于它估值的讨论便从2万亿美元回落到了1.5万亿美元。
仅仅因为一个季度的增速放缓，估值预期就凭空蒸发了5000亿美元。这就是靠“买未来”定价的代价：投资人掏钱买的是增长的预期，一旦预期的斜率出现松动，价格就会被立刻清算。
“国模”的商业变现之路
“API业务有一个致命风险：可以一个晚上就被全部切走。”金沙江创投主管合伙人朱啸虎在今年评价国内大模型行业时，泼了这样一盆冷水。作为投出过滴滴和饿了么的投资人，他对生意模式的脆弱性有着近乎本能的警觉。
但从账面上看，国内大模型公司这半年的转型确实相当凶猛。
尤其曾被行业内称为“中国版Anthropic”的智谱，随着Anthropic的IPO在即，其锚点的定位效应正在削弱，而智谱也在探索自己的商业路径。今年上半年收入9.54亿元，同比增长399.7%，半年就超过了去年全年。比增速更重要的是收入结构的彻底翻转：上半年其开放平台及API业务收入达到8.25亿元，同比暴增2735.7%，占总收入的比例从去年同期的15.2%直接拉升到86.5%；而过去挑大梁的本地化部署收入，占比则从84.8%跌到了13.5%，企业级通用大模型收入同比下滑了54.6%。
简单来说，智谱在半年时间里换了一种活法：过去靠一单一单谈项目、做本地化交付的“卖盒子”生意在快速萎缩，按Token调用量计费的云端服务成了绝对主力。
更关键的是，这种增长并不是靠“价格战”打出来的。
上半年，智谱MaaS平台的Token调用量比年初增长了40倍以上，付费日活用户增长603%，而API的平均售价反而提高了约101%。智谱CEO张鹏把这套商业逻辑总结为一个公式：“AGI商业价值=智能上界×Token消耗规模”，其中，模型能力决定定价权，调用规模决定基本盘。今年上半年，这两条曲线第一次实现了同步向上，也让智谱的API毛利率从负值转正，达到了24.6%。
从“做项目”转向“卖API”确实已经是国内行业的共识。不只是智谱，MiniMax今年上半年收入1.17亿美元，同比增长283.1%，其开放平台及企业服务收入占比也从去年的30.3%升至63.4%，8月份的年化经常性收入（ARR）突破8亿美元，其中八成来自企业端。中国的大模型公司不是某一家在换赛道，而是集体在进行商业模式的切换。
不过，随着AI技术的普及，API会变成像水电一样的公用事业，毛利率终将回落到10%至20%的水平。朱啸虎甚至直言，Anthropic之所以急着在今年冲刺上市，也是因为心里清楚，等将来算力普及，靠前沿模型和编程工具赚60%到70%毛利的日子根本不可持续。
其言外之意，便是API业务今天能赚钱，全部得益于头部模型还有技术壁垒；将来难赚钱，是因为技术壁垒迟早会被摊平。
资本市场对这种终局风险的反应比谁都快。9月10日，DeepSeek发布了V4.1 Flash模型并宣布降价，当天智谱的股价还是下跌了10.34%，投行杰富瑞随后也将智谱API业务的估值倍数从50倍下调至30倍。
大模型的竞争从收入转向成本
提起中国大模型市场，外界最直观的印象就是出色的“性价比”。数据显示，Silicon Data的Token价格指数从今年5月底的2.04美元一路跌到9月初的0.97美元，首次跌破每百万Token 1美元的关口，背后的主要推手正是中国的开源模型。
但摩根士丹利的调研数据却指出：中国大模型API的实际平均价格非但没跌，反而在涨：2026年第二季度，国内模型平均输入价格为4.9元/百万Token，输出价格为21.9元/百万Token，比2025年第一季度分别上涨了约48%和80%。具体到各家产品，智谱Coding Plan的三档订阅价格分别上调了141%、261%和130%；月之暗面的Kimi K3价格涨了3倍多；就连一向带头降价的DeepSeek，在高峰时段缓存命中的调用价格也大幅抬高了12倍。
为什么一边是开源模型带头降价，一边是商业API均价大涨？
国家数据局的数据显示，国内日均Token调用量已经从2024年初的约1000亿，暴增到今年3月的140万亿。当池子足够深、应用场景足够复杂时，厂商就不再需要靠粗暴的一刀切降价来抢客户，而是学会了峰谷定价、缓存定价和分层订阅。
与此同时，国内厂商还在通过技术，压低底层的推理和训练成本。智谱上半年的单位Token推理成本比年初下降了80%；瑞银的报告也估算，中国头部模型的训练成本可能只有海外同行的十分之一。
如果这时候再去翻看Anthropic的招股书，会发现在S-1文件中，其未来十年的算力采购承诺高达5180亿美元，其中约80%不可取消，无论用不用都得掏钱。此外，它去年有47%的收入是通过亚马逊和谷歌的云市场完成的，每收到1美元就要向云厂商交16美分的分销费。
换句话说，在成本结构这件事上，中美厂商的角色反而对调了。被认为在“讲未来故事”的Anthropic，用一纸十年期合同把自己的算力成本死死锁在了高位；而被认为在“苦苦找变现”的中国厂商，则通过工程优化写出了一条持续下探的成本曲线。
写在最后
无论是冲刺万亿估值的巨头，还是精打细算熬毛利的创业公司，最终都要回到同一个商业闭环里交卷：技术创造客户价值，客户付费验证价值，收入反哺下一代研发。
再回到开头的问题：到底是谁在给大模型行业定价？
如今的答案已经愈发清晰：给行业定价的，从来不是路演大厅里的承销商，而是每天真实的Token调用量、每个季度的毛利率变化，以及成本下降与收入增长那两条曲线的斜率。无论故事怎么讲，当财务数据公布之时，便能看到谁在裸泳。
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。
Wen, LLM House.
China ' s LLM manufacturer, represented by Zhipu, was bowing down in the financial paper to account for every penny of the API business, when Anthropic on the other side of the ocean was bowing to the Wall Street gate with a trillion-dollar valuation.
On 26 October, LLM unicorn Anthropic made public the listing book, which was to be officially boarded in mid-November. The book spread the company ' s base on the IPO: in the first half of the year, $16.29 billion was collected, with a net loss of $80.7 billion. Prior to that, its maximum valuation was discussed at a time of $2 trillion.
Just two months before Anthropic bowed at the Wall Street gate, the Chinese company LLM also handed over a report card. At the end of August, Zhipu disclosed that it had earned $954 million in the first half of the year, at a loss of $2,072 million, of which the Māori ratio for API operations was reversed for the first time.
As can be seen, two distinct narratives are taking place in the LLM industry: on the other side of the ocean, a multi-billion-dollar valuation is being used to pay for future technical assumptions, and domestic manufacturers are busy using Māori rates to prove that LLM itself is a sustainable business.
In the past, there was often a misconception in the industry that overseas companies depended on technology and capital for their future, and that Chinese companies were still worried about commercialization. However, reliance on capital alone weakens market confidence in “future” pricing. On the latest investor day, the valuation discussion for Anthropic has shrunk from $2 trillion to $1.5 trillion.
It's the future, not now.
In the view of LLM House, capital markets have not always offered for the value of a company today, but rather for future growth space.
Anthropic's book of admission reflects this logic.
On the face of it, the net loss of Anthropic during the first half of the year was $80.7 billion, as if it were “blooded for the market”. However, when the financial statements were opened, it was found that $72.4 billion of that amount was due to a surge in corporate valuations, which resulted in a write-off of the fair value loss charged on the books. If only operational operations were to be seen, their losses in the first half of the year had actually narrowed to $693 million.
What really excites investors is its growth rate. Throughout 2025, Anthropic earned $4.6 billion, about 12 times the increase over the same period (8.06 billion in operating losses and $7.33 billion in computing and infrastructure expenditures). By the first half of the year, its income had hit straight at $16.29 billion, about 14 times that of the same period last year.
Moreover, of the $4.6 billion received in 2025, approximately $3.8 billion came from paid, with only $789 million from fixed subscriptions. This means that business clients do not buy a particular software at once, but pay for it on an ongoing basis at Token's call. In the second quarter of this year, Anthropic received $11.5 billion in single-season revenues, the first reverse of the $6.7 billion in revenues over the same period.
This is just proof that the business has now recognized the commercial value of LLM and is willing to pay for code generation, modeling and complex tasks.
But what sustains the trillion-dollar valuation of Anthropic is clearly more than the “acceleration” of income upwards than the billions of dollars available. And that's why Anthropic's valuation collapsed.
Financial data indicate that the Anthropic battalion collected $11.6 billion in the second quarter, which doubled the ring ratio; however, by the third quarter, the battalion received $16.8 billion, which slowed to 45 per cent. While the overall growth rate is still high, as long as it does not meet Wall Street expectations, the valuation logic will shake immediately. The discussion on its valuation then fell back from $2 trillion to $1.5 trillion.
The valuation was expected to evaporate $500 billion in space simply because of a slowdown in growth over a quarter. This is the cost of pricing by “buying the future”: investors pay for growth expectations, and prices are liquidated as soon as the expected slopes loosen.
Road to commercial realization of the “country model”
“API has a fatal risk: it can be cut off in one night.” This year, Zhu Zhu Zhu Zhuo, the chief partner of Kim Shajiang, poured such a cold water into the country ' s LLM industry. As an investor who had invested in dripping and hungry, he was almost instinctively alert to the fragility of business models.
However, on the books, the transformation of LLM in the country over the past six months has been ferocious.
In particular, Zhipu, known as the “China version” of Anthropic, which is exploring its own business path, is diminishing as the IPO of Anthropic is in place. In the first half of this year, revenues amounted to 954 million yuan, an increase of 399.7 per cent over the previous year, six months ahead of the year as a whole. More important than growth is the radical reversal of the income structure: In the first half of the year, income from its open platform and API operations amounted to 825 million yuan, an increase of 2735.7 per cent over the same period, and the share of total income rose directly to 86.5 per cent, from 15.2 per cent during the same period last year; while income from localized deployments, which had targeted large sorghums in the past, fell from 84.8 per cent to 13.5 per cent, while income from general LLM at the enterprise level decreased by 54.6 per cent.
In short, Zhipu has changed a living method over a six-month period: the “sell box” business, which used to be delivered locally by a single project, is shrinking rapidly, with cloud-based services, which are billed at Token, becoming an absolute major force.
More crucially, this growth was not driven by “price wars”.
In the first half of the year, the number of ZhipuMaas platform token calls increased more than 40 times from the beginning of the year, paid daily users increased by 603 per cent, while the average API sales price increased by about 101 per cent. ZhipuCEO Zhang Peng summed up this business logic as a formula: “AGI Business Value = Smart Upper & Token Consumption Scale”, where model capabilities determine pricing rights and call size determines the base disk. For the first half of this year, for the first time, the two curves were synchronized upwards, and the API ratio of Zhipu went positive from negative to 24.6 per cent.
The shift from “doing projects” to “selling API” is indeed a consensus in the domestic industry. Not only Zhipu, Mini Max earned $117 million in income in the first half of the year, an increase of 283.1 per cent over the previous year, but his share of income from open platforms and business services also rose from 30.3 per cent to 63.4 per cent last year, with annualized recurrent income (ARR) reaching $800 million in August, of which 80 per cent came from the corporate side. China ' s LLM company is not one of its own, but is collectively engaged in a business model switch.
However, with the spread of AI technology, API will become a utility like electricity and water, and the Māori rate will eventually fall to between 10 and 20 per cent. The reason why Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu Zhu is even to say Zhu
The implication is that API operations make money today, thanks to head models and technical barriers; it will be hard to make money in the future because the technological barriers will sooner or later be distributed.
Capital markets react more quickly than anyone to this final risk. On 10 September, DeepSeekpublish took the V4.1 Flash model and announced a price reduction. On the same day, the share price of Zhipu fell by 10.34 per cent and the investment bank Jeffrey subsequently doubled the value of ZhipuAPI from 50 to 30 times.
LLM competition shifted from income to cost
The most intuitive impression of China's LLM market was the excellent “price for money”. The data show that Silicon Data’s Token price index dropped from $2.04 at the end of May to $0.97 at the beginning of September, for the first time breaking the $1 per million Token threshold, with China’s open source model behind it.
However, Morgan Stanley's research data point out that China's real average price of LLMAPI, rather than falling, is rising: in the second quarter of 2026, the domestic model averaged an input price of 4.9 yuan/million Token and an output price of 21.9 yuan/million Token, an increase of about 48 per cent and 80 per cent, respectively, from the first quarter of 2025. For each product, ZhipuCoding Plan ' s third-class subscription price was increased by 141 per cent, 261 per cent and 130 per cent, respectively; the price of Kimi K3 on the dark side of the month was increased more than threefold; even DeepSeek, which had been leading the price reduction, was raised 12 times at peak times.
Why do open-source models lead price reductions, while commercial API prices increase so much?
According to data from the National Data Agency, Token ' s daily average has increased from about 100 billion in early 2024 to 1.4 trillion in March of this year. When the pool is deep enough and the application scene is complex enough, the manufacturer is no longer required to rob the client at a harsh one-size-fits-all price, but has learned to price the Peak Valley, cache and tiered subscriptions.
At the same time, domestic manufacturers are also using technology to lower the underlying reasoning and training costs. The unit cost of Token ' s reasoning in the first half of Zhipu was 80 per cent lower than at the beginning of the year; the Swiss Bank ' s report also estimated that the training costs of China ' s head model could be only one tenth of those of its overseas counterparts.
If you look at Anthropic's book again at this time, you will find in document S-1 that its firm procurement commitment for the next 10 years amounts to $518 billion, of which about 80 per cent cannot be cancelled and has to be paid for. In addition, 47 per cent of its income over the past year was channelled through the cloud market in Amazon and Google, with a 16-cent distribution fee paid to cloud manufacturers for each dollar received.
In other words, in the matter of cost structure, the role of the Central American manufacturer has been reversed. Anthropic, believed to be a “story of the future”, has locked himself to death with a 10-year contract, while the Chinese manufacturer, believed to be a “suffering search for cash”, has written a cost curve through engineering optimization.
At the end.
Whether it be the giants of the trillions of valuations, or the well-calculated start-ups of the Māori, they end up in the same business ring: technology creates customer value, customer paid validates value, and income feeds the next generation of research and development.
Back to the question at hand: Who is pricing the LLM industry?
The answer is now clearer: the industry has never been priced as a distributor in the road hall, but rather as a daily real Token call, a quarterly change in the Māori rate, and the slope of the two curves of cost decline and income growth. Whatever the story is, when financial data is published, you can see who's swimming naked.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.