文 | 听筒Tech（ID:tingtongtech），作者 | 杨 林，编辑 | 饶霞飞
巨头的数据摆在一起，就构成了行业刺眼的一组对照。
2026年10月26日，Anthropic向SEC递交了招股书，目标11月登陆纳斯达克，估值1.5万亿美元到2万亿美元。
如按上限定价，这是美股史上最大的IPO之一。但招股书迎面撞上的是一组让人需要揉眼睛的数字，2026年上半年营收162.86亿美元，同比增长约1300%，净亏损却高达807.48亿美元。
巨亏的背后，藏着AI行业底层逻辑的秘密，这从另一家巨头的财报中可窥一二。
在更早的10月20日，SpaceX发布了第三季度财报，营收145.75亿美元，净利润44.04亿美元，实现了上市后首次季度盈利。而撑起集团85%营业利润的，仍是星链通信业务。
一边是模型公司亏出天文数字，一边是“卖铲子”的公司开始赚钱。这背后，不在谁亏谁赚，而在成本结构。
诸如，SpaceX的盈利，始终靠一门走过资本开支高峰的通信业务给AI板块兜底。不过，遗憾的是，Anthropic没有这层缓冲带，每一美元收入，都要直面推理成本和刚性算力承诺的检验。
巨亏的Anthropic
要理解Anthropic的招股书，得先理解这家公司是怎么来的。
2021年初，达里奥·阿莫迪带着妹妹丹妮拉和另外五名OpenAI核心成员出走，创立Anthropic。达里奥深度参与了GPT-2和GPT-3的研发，是RLHF的共同发明人之一。
离开的原因不是钱，不是职位，是对OpenAI在微软入局后加速商业化的方向，产生了不可调和的分歧。达里奥越来越不相信“能力越强，安全自然跟上”这句话会自己成真。
2021年5月，Anthropic完成了1.24亿美元A轮。但在这笔钱到账之前，早期融资并不顺利，绝大多数机构没有出手，A轮资金甚至分三期才到账。
Anthropic的商业逻辑也另辟蹊径，从第一天起就没有选择和ChatGPT在消费者市场正面竞争。
它将筹码押在企业客户和API上。数据显示，2026年上半年，与OpenAI多数收入来自C端订阅不同，Anthropic按Token实际用量计费的API收入131亿美元，占总营收80%；订阅收入32亿美元，只占20%。
旗舰编程工具Claude Code也是增长的核心变量。上线后仅六个月，模型就达到10亿美元年化收入，到2026年初年化收入突破25亿美元，在AI编程工具市场拿下约54%的份额。
这支撑Anthropic快速增长。2023年，公司商业化第一年营收为4100万美元。接下来的2024-2025年，这一数据分别为3.86亿美元和45.85亿美元。到了2026年上半年，直接跳到162.86亿美元，同比增长约1300%。
但企业客户的好处和代价是一体两面的。
诸如，付费能力强、切换成本高、收入可预测是企业客户的属性；代价却是，销售周期长、单客获取成本高，以及当对手愿意用折扣敲门的时候，“切换成本”并没有想象中那么高。
不过，这次让市场震惊的，是亏损数额。
招股书显示，2026年上半年，Anthropic净亏损807.48亿美元。不过，仔细来看，这807亿美元里，有723.88亿美元来自可转换票据、衍生负债和SAFE工具的公允价值变动，占净亏损的89.6%。
机制不复杂。因为Anthropic在上市前经历了多轮融资，估值从数百亿美元攀升至H轮投后的9650亿美元。
按照GAAP要求，向早期投资者发行的可转换票据和SAFE工具需要以公允价值计量，公司估值每上升一次，负债端就相应增值一次，在利润表上表现为亏损。
这更像是一套会计准则与私募估值机制碰撞出的技术性产物，而非经营恶化的证据。数据显示，2025年全年，同一项会计处理已经制造了343.08亿美元的账面损失，2026年上半年只是延续了这个逻辑。
也即是说，公司越值钱，账面“亏损”就越大。
亚马逊持有的62亿美元可转债是最典型的一笔。到了2026年6月底，这些票据的账面价值已经从2025年底的400亿美元，重估直接到了1085.66亿美元。
亏损背后
抛开财务图景，Anthropic运营层面的问题，则要微妙得多。
数据显示 ，在2026年第二季度，Anthropic实现了9.74亿美元营业利润，核心驱动力是算力及基础设施成本，占收入的比例从2025年第一季度的251%，骤降至2026年第二季度的48%。
当然，规模效应确实在发生。当收入基数从十亿级跳升至百亿级，前期锁定的算力资源被摊薄，单位经济模型出现了方向性的改善。
但改善的代价，则出现在算力成本表上。
数据显示，Anthropic锁定的算力基础设施承诺已突破6000亿美元，若正在谈判的额外协议落地，总额可能达到7500亿至8100亿美元。
其中，与谷歌云的合同约1110亿美元，与AWS约1100亿美元，与微软约310亿至449亿美元，另外还有与博通关联的TPU租赁义务逾1600亿美元，以及与Nscale、Lambda合计最高860亿美元的芯片租赁安排。
更关键的是，合同结构的变化。其中大量协议，被标注为“无论实际使用量如何均须支付”，期限延伸至2033年至2034年。
这也意味着，Anthropic需要用未来七到十年的确定性支出，换取了当前算力供给的确定性。而纯模型公司的估值，在很大程度上，仍寄生在基础设施供应商的现金流之上。
诸多分析认为，在收入增速仍高的阶段，这套结构运转良好；但一旦增速放缓，刚性合同就会从规模杠杆变成利润黑洞。
事实上，2026年三季度，Anthropic的环比增长约45%，相比上半年动辄翻倍以上的环比增速，明显踩了一脚刹车。
不过，增速放缓本身并不意外，任何千亿级营收基座上的公司都不可能长期维持三位数增长。问题是，Anthropic的合同义务不会因增速放缓而相应缩减。这也是“以时间换空间”战略最脆弱的地方，赌的是收入跑得比账单快。
Anthropic高度集中的收入结构，也为运营敲响了警钟。
数据显示，今年上半年Anthropic的营收，约20%来自仅两家客户，对云平台渠道依赖度高。集中的客户基础，叠加尚未长期锁定的合同关系，使得Anthropic的高增长，在质量的维度上，仍存在需要验证的空白。
当然，转机仍出现在模型代际上。
9月下旬，Claude Opus 5.5的发布，让运行成本比上一代低40%，Anthropic在聚合平台上的调用热度迅速回升。这也意味着，“模型能力领先”仍然能带来定价权，但这个窗口期的长度，正在被对手的折扣和跟进速度持续压缩。
Rippling的案例就是一个切片。
2026年5月，开发者使用Anthropic模型处理编码工作的比例是57%，OpenAI只有16%。几个月后，OpenAI用正常价格20%到40%的折扣加赠送额度抢单，比例逆转为OpenAI 52%、Anthropic 18%。
而Replit，在OpenAI给出年额超1亿美元、折扣超30%的方案后，将主力模型切了过去。
业内人士认为，本质在于Anthropic的定价策略偏强硬，客户一旦触及合同用量上限，折扣直接取消。但当对手愿意用负利润率换份额的阶段，Anthropic就没了优势。
AI行业的未来，模型还是算力？
相对于Anthropic的巨额亏损，SpaceX三季度财报则提供了另一个参照样本。
2026年10月20日，SpaceX发布了第三季度财报，营收145.75亿美元，净利润44.04亿美元，实现了上市后首次季度盈利。
不过，利润的来源与AI模型公司形成反差。
诸如，撑起集团85%营业利润的，是星链所在的通信业务；而收入占比首次过半的超级智能板块，当季经营利润仅4.31亿美元，对应资本开支高达164亿美元。
换言之，SpaceX的AI业务增长凶猛，但它赚钱的方式是出售算力基础设施，而不是承担模型推理的边际成本。
星链的通信业务之所以贡献绝大部分利润，恰恰因为它是一门已经走过资本开支高峰、边际成本相对可控的生意。
模型公司面对的成本结构则完全不同。
诸多分析指出，传统互联网的边际成本趋零逻辑在AI推理面前失效了。每一次用户调用都对应真实的GPU消耗、电力支出和基础设施运维。这种“规模不经济”，在用户量高速增长阶段会被放大，而非缓解。
摩根大通在这一轮下调中的判断值得细读，其将国内某模型公司2030年预期市盈率从30倍下调至20倍，理由是“前沿模型的领先位置正在快速轮换，单一厂商长期保持技术领先的确定性低于此前预期”。
这句话，点破了AI模型公司估值中脆弱的一环，技术领先是一种会随时间自动贬值的资产。当市场愿意为稀缺性支付溢价时，模型公司的估值可脱离财务基本面运行；但当稀缺被竞争稀释，定价的锚就会迅速回落。
在AI业内人士廖哥看来，行业接下来的分水岭，不太可能由模型排行榜决定。
“谁能将智能卖得更贵，同时将算力成本压得更低，谁才有机会把收入增长转化为可持续的现金流。”廖哥表示。
纽约大学Aswath Damodaran也提供了一个量化参照，如Anthropic的IPO估值达到2万亿美元，公司若要在十年后进入成熟期，需要做到接近1.2万亿美元的年收入；若成熟期拉长到十五年，需要的收入是接近2万亿美元。
这个数字本身不构成对错判断。但它意味着，留给“先烧钱、后证明”的时间窗口正在收窄。
Anthropic的管理层显然清楚这一点。今年第二季度的单季营业利润转正，主要是其向市场表示，模型公司可以在经营层面盈利，规模效应是真实存在的。
只是，盈利的代价是未来十年数千亿美元的刚性算力承诺。
“事实上，Anthropic的亏损，以及SpaceX的利润多数来自基础设施侧，都指向同一个结论，AI行业的竞争，正在从谁更会讲故事，转向谁的单位经济模型先跑通。”廖哥坦言。
同时，廖哥仍认为，巨亏本身并不可怕，问题在于，当亏损中的会计成分被剥离之后，剩下那部分真正需要被填补的缺口，到底需要多长时间、多少收入、多高的毛利率，才能被填上。
Anthropic的招股书，则把这道题摊开在了桌面上。接下来，市场可能会用同一把尺子，去衡量每一家AI公司。
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。
Tech (ID:tingtongtech), by Yang Lin, editor
The data of the giants together constitute an eye-opening set of industry comparisons.
On 26 October 2026, Anthropic submitted a equity book to the SEC with a target landing at NASDAQ in November, valued at $1.5 trillion to $2 trillion.
This is one of the largest IPOs in the history of the United States stock, if priced at the ceiling. However, the book calls were met with a set of figures requiring eye rubbing, with $16,286 million collected in the first half of 2026, an increase of about 130 per cent over the same period, with a net loss of $80,748 million.
Behind the loss lies the secret of the bottom logic of the AI industry, which can be seen in the financials of another giant.
Earlier, on 20 October, SpaceXpublish issued the third quarter of the financial paper, with $14,575 million in revenue and a net profit of $4,404 million, achieving its first quarterly profit after being listed. In contrast, 85 per cent of the group ' s business profits continue to be held up by satellite chain communications.
On the one hand, model companies lose astronomical figures, and on the other hand, companies selling shovels start making money. Behind this is not who earns what they earn, but what they cost.
For example, SpaceX ' s profits have always been driven under the AI Board by communications that go through the peak of capital expenditure. Unfortunately, however, Anthropic does not have the buffer zone, and each dollar of income must face up to the cost of reasoning and the test of hard-core commitments.
Losing Anthropic.
To understand Anthropic's equity book, you have to understand how this company came from.
In early 2021, Dario Amody left with her sister, Daniela, and five other Open AI core members to create Anthropic. Dario was deeply involved in GPT-2 and GPT-3, one of the co-inventors of RLHF.
The reason for leaving was not money, not jobs, but an irreconcilable disagreement over the direction of Open AI to accelerate commercialization after Microsoft entry. Daario is becoming increasingly unconvinced that the words “the more capable he is, the safer he is” will come true.
In May 2021, Anthropic completed round A of $124 million. However, the early financing had not been smooth until the money had arrived, the vast majority of the agencies had failed to do so and the A round had not even arrived in three instalments.
Anthropic’s business logic is also different, with no choice from day one and ChatGPT competing positively in the consumer market.
It puts chips on corporate customers and API. Data show that in the first half of 2026, in contrast to most of Open AI ' s income from C-end subscriptions, Anthropic earned $13.1 billion, or 80 per cent of the total, from API based on Token ' s actual usage; and $3.2 billion, or 20 per cent, from subscriptions.
The flagship programming tool Claude Code is also the core variable of growth. Only six months after it went online, the model reached $1 billion in annualized revenues, which exceeded $2.5 billion by early 2026, with an estimated 54 per cent share in the AI programming tool market.
This supports the rapid growth of Anthropic. In 2023, the first year of commercialization earned $41 million. For the next 2024-2025, the figures are $386 million and $4.585 billion, respectively. By the first half of 2026, it had jumped directly to $16,286 million, an increase of approximately 1.3 per cent over the same period.
But the benefits and costs of business clients are two-sided.
For example, the ability of paid, the high cost of switching, and the predictability of income are attributes of the enterprise's customers; at the expense of long sales cycles, the high cost of individual access, and when opponents are willing to knock on the door with a discount, the “turn cost” is not as high as expected.
This time, however, the market was shocked by the amount of losses.
According to the equity book, in the first half of 2026, Anthropic had a net loss of $80,748 million. However, of the $80.7 billion, $72.388 billion, or 89.6 per cent of the net loss, came from changes in the fair value of convertible instruments, derivative liabilities and SAFE instruments.
Mechanisms are not complex. Because Anthropic had experienced multiple rounds of financing prior to listing, the valuation climbed from tens of billions of dollars to $965 billion after round H.
As required by GAAP, convertible instruments and SAFE tools issued to early investors need to be measured at fair value, and each increase in the company ' s valuation adds a corresponding value to the liability and shows a loss on the profit statement.
This is more like a technical product of a set of accounting standards colliding with private valuation mechanisms than evidence of a deterioration of operations. The data show that, throughout 2025, the same accounting treatment had resulted in a book loss of $34,308 million, a logic that was only continued in the first half of 2026.
That is, the more valuable the company is, the greater the “losses” on the books.
The $6.2 billion in revolving debt held by the Amazon is the most typical. By the end of June 2026, the carrying value of those instruments had been revised directly to $108,566 million, from $40 billion at the end of 2025.
Behind the loss.
Leaving aside the financial picture, the problem at the operational level of Anthropic is much more delicate.
Data show that in the second quarter of 2026, Anthropic achieved a business profit of $974 million, with the core driving force being capacity and infrastructure costs, which fell sharply from 251 per cent of revenues in the first quarter of 2025 to 48 per cent in the second quarter of 2026.
Of course, the scale effect does occur. When the revenue base jumped from a billion to a billion, the calculator resources targeted in the preceding period were strained and the unit economic model improved in direction.
The cost of improvement, however, appears on the cost tables.
The data show that the targeted computing infrastructure commitments of Anthropic have exceeded $600 billion, and that if the additional agreements under negotiation land, totals between $75 billion and $810 billion.
Of that amount, contracts with Google Clouds were estimated at $111 billion, with AWS at $11 billion, with Microsoft at $31 billion to $44.9 billion, plus TPU lease obligations linked to Chase at over $160 billion, and chip lease arrangements totalling up to $86 billion with Nscale and Lambda.
More crucial is the change in contract structure. A large number of these agreements, labelled as “payable regardless of actual use”, extend to 2033 to 2034.
This also means that Anthropic would need to spend the next seven to ten years of certainty in exchange for certainty in the current calculus supply. To a large extent, pure model company valuations remain embedded in the cash flow of infrastructure suppliers.
Many analyses suggest that this structure works well at a time when income growth is still high; but once growth slows down, rigid contracts shift from big leverage to profit black holes.
In fact, in the third quarter of 2026, Anthropic ' s ring grew by about 45 per cent, more than double the ring than in the first half of the year, clearly stepping on a brake.
However, the slowdown is not surprising in itself, and no company on the base of a trillion-degree base is likely to sustain triple-digit growth in the long term. The problem is that Anthropic ' s contractual obligations will not be reduced by slowing the rate of increase. This is also where the “time-for-space” strategy is most vulnerable, betting on income running faster than bills.
The highly concentrated revenue structure of Anthropic also sounded alarms for operations.
The data show that about 20 per cent of Anthropic collections in the first half of the year came from only two clients, with high dependence on cloud platform channels. The concentration of the customer base, which is compounded by contractual relationships that have not been locked in for long periods, has led to a high growth of Anthropic, and there are still gaps in quality dimensions to be verified.
Of course, the transition still occurs in the model generations.
In late September, Claude Opus 5.5, publish, reduced the running cost by 40 per cent over the previous generation, and Anthropic rapidly recovered the call heat on the polymer platform. This also means that “model capacity leads” can still bring pricing rights, but the length of the window period is being consistently reduced by discounts and speed of follow-up by the counterparty.
The Ripping case is a slice.
In May 2026, 57 per cent of developers used the Anthropic model for coding and only 16 per cent of Open AI. A few months later, Open AI gave quota purchase orders at discounts ranging from 20 to 40 per cent of the normal price, which turned back to Open AI 52 per cent and Anthropic 18 per cent.
And Refrit, after giving Open AI a programme of over $100 million per year and a discount of over 30 per cent, cut off the dominant model.
According to industry sources, the essence of this is that Anthropic ' s pricing strategy is rigid and the discount is cancelled as soon as the client touches the contract ceiling. But when the opponent is willing to trade negative margins for shares, Anthropic has no advantage.
The future of the AI industry, models or numbers?
Compared to the large losses of Anthropic, SpaceX provided another reference to Z111Z for the third quarter.
On 20 October 2026, SpaceXpublish issued the third quarter of the financial paper, which generated $14,575 million in revenue and a net profit of $4,404 million, achieving the first quarterly profit since its listing.
However, the source of the profits contrasted with AI Model.
For example, 85 per cent of the group's operating profits are in communications, where the chain is located; while income accounts for more than half of the first super-smart block, with only $431 million operating in the current quarter, with a corresponding capital expenditure of $16.4 billion.
In other words, SpaceX ' s AI business has grown dramatically, but it earns money by selling computing infrastructure rather than by bearing the marginal costs of model reasoning.
The vast majority of profits are contributed by the communications business of the chain precisely because it is a business that has passed the peak of capital spending and relatively manageable marginal costs.
Model companies face completely different cost structures.
Numerous analyses have pointed out that the marginal cost logic of the traditional Internet is no longer valid for AI reasoning. Each call by a user corresponds to real GPU consumption, electricity expenditure and infrastructure. This “uneconomic” dimension is magnified rather than mitigated at a time of high user growth.
The judgement of Chase Morgan in the downward revision of this round is worthy of reading, as it reduces the expected market gain of a model company in the country from 30 to 20 times in 2030, on the grounds that “the lead position of the front model is rapidly rotating, and the single manufacturer is maintaining a longer-term technical lead certainty than previously anticipated”.
This sentence breaks the weak link in AI Models ' valuation, and technological lead is an asset that automatically depreciates over time. When the market is willing to pay a premium for scarcity, the model company ' s valuation can operate away from the financial fundamentals; however, when the scarcity is diluted by competition, the anchor of the price will quickly fall back.
It is unlikely that the next watersheds in the industry will be determined by model rankings in the eyes of IAI.
“Who sells intelligence more expensive and lowers the cost of calculation, and who has the opportunity to turn income growth into a sustainable cash flow.” Ciao says.
Aswath Damodaran at New York University also provides a quantization reference, as in the case of Anthropic, whose IPO valuation reached $2 trillion, companies would need close to $1.2 trillion in annual revenues if they were to reach maturity in 10 years; if maturity reached 15 years, it would require close to $2 trillion.
This figure does not in itself constitute a false judgement. But it means that the window left for “burning money first, then proving” is narrowing.
The management of Anthropic is clearly aware of this. The single-season profits of the second quarter of this year have improved, mainly because they indicate to the market that model companies can make profits at the operational level and that the scale effect is real.
The price of profit is simply a firm commitment of hundreds of billions of dollars over the next decade.
“In fact, the losses of Anthropic, as well as the profits of SpaceX, are mostly from the infrastructure side, pointing to the same conclusion that competition in the AI industry is going from who is better at telling stories to whose unit economic model runs first.” Liagotan said.
At the same time, León continues to believe that the huge loss itself is not a terrible problem. The problem is that when the accounting component of the loss is removed, the remaining part of the gap that really needs to be filled, how long, how much income and how high the Māori rate is to be filled.
Anthropic's equity book spreads the subject over the desktop. Next, the market could measure every AI company with the same ruler.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.