文 | 版面之外，作者｜画画、版君
硅谷的故事，往往从一次分道扬镳开始。
五年前，OpenAI 的一群核心研究员选择离开。他们对 AI 的发展方向有不同判断，也决定为此另起炉灶。
这次出走，后来催生了一家估值与营收都迅速攀升的 AI 公司：Anthropic。
2021 年，在旧金山先锋广场的旧厂房里，Dario Amodei 和妹妹 Daniela Amodei 正把自己的物品装进纸箱。
跟着他们一起离开 OpenAI 的，包括后来参与 GPT-4 研发的核心研究员、OpenAI 安全团队骨干在内。
离开 OpenAI 的原因并不复杂，CEO Sam Altman 正在推动 OpenAI 从一家非营利研究机构，转向一家追求高估值与商业化的公司。Dario 认为这种急速狂奔正在让安全失去控制。
他们给新公司取名 Anthropic（意思是以人为本），并制定了一套宪章，希望在硅谷的 AI 狂热中建立一座讲求克制与安全的实验室。
Dario 以为自己离开了 Sam Altman 的商业战场。但他很快发现，在大模型的技术路线下，对安全和性能的追求越深，对算力的需求就越紧张。
正是这份试图独立于 Sam Altman 的决心，把它推向了一条回不了头的资本游戏。
一、启动资金与被选择的筹码
要训练能够对抗 OpenAI 的前沿模型，算力是无法绕过的物理门槛，算力则需要巨额资金来支撑。
为了维持研究的独立性，Dario 早期刻意避免引入传统硅谷风投。
但命运，却在这里出现了环环相扣的戏剧性。2022 年，向 Anthropic 提供 5.8 亿美元早期资金的，是加密货币交易所 FTX 的创始人 Sam Bankman-Fried（SBF）。
然而短短几个月后，FTX 轰然倒塌，SBF 被带上法庭。
作为差点被认定清算资产的一部分，Anthropic 股权面临被司法拍卖的风险，公司一度处于生存边缘。
市场不等人，AI 竞争更是以天甚至小时来迭代。这场突如其来的资金断裂，使得 Dario 需要在极短时间内找到足够大且能提供稳定算力的接盘者。
嗅到机会的云巨头们顺势入场。
2023 年秋天，亚马逊宣布向 Anthropic 投资最高 40 亿美元，谷歌随后跟进 20 亿美元。这笔资金解了燃眉之急，但也确定了这场救援的真实代价。
为了落实这项协议，AWS 甚至为其打造了专用的 Trainium2 芯片大集群。Anthropic 也同意让亚马逊拥有对其架构改进的早期优先调优权，将芯片设计与算法优化深度粘合在一起。
二、拒绝王座
正当 Anthropic 在巨头的基础设施框架下艰难推进时，一个看似能够改变格局的转折点，忽然出现了。
2023 年 11 月 17 日傍晚，OpenAI 董事会突发解雇 CEO Sam Altman。在随后的 48 小时里，陷入混乱的 OpenAI 董事会秘密拨通了 Dario Amodei 的电话，提出一个让意想不到的方案：邀请 Dario 接任 OpenAI CEO，并将两家公司合并。
这是 Dario 距离重掌旧部，整合资源最近的时刻。
但他选择了拒绝。
他看清了那场治理危机背后的利益纠葛，最终回到 Anthropic 的实验室，继续推进 Claude 的研发。
这次拒绝，也让重回 OpenAI 的 Sam Altman 加速推进了数千亿美元级别的算力扩张计划。
擅长资本游戏的Sam Altman，不断地用攻势挤压着 Anthropic 的市场空间。为了不被落下，Anthropic 不得不在 Claude 3 与 3.5 系列的研发中，将训练集群规模成倍放大。
到了 Claude 3.5 Sonnet 迭代时，仅单次完整训练与对齐跑通所消耗的算力成本，就已经达到了上一代旗舰模型的数倍。
为了在模型迭代上不被 OpenAI 拉开代际差距，Dario 没有退路，只能跟着将 Claude 的训练规模和算力消耗推向更高的层级。
而要想支撑这笔空前的算力开支，Anthropic 就需要找到能够迅速变现的商业模式。
三、倒逼出来的商业模式
在算力成本按月翻倍的财务压力下，Anthropic 无法像 OpenAI 那样靠面向大众的通用聊天框慢慢培养 C 端付费习惯，它直接去找了愿意为效率付大钱的客户。
于是，Anthropic 做出了一个关键的战略转向，放弃与 ChatGPT 争夺 C 端流量，转向企业侧的战场。
2025 年，Anthropic 推出了直接嵌入终端命令行运行的编程工具 Claude Code。
这项突破来自于团队对底层 Prompt 与计算图架构的重构。
过去开发者需要手动把整个工程的代码文件打包粘贴进网页框，而 Claude Code 能够直接挂载本地文件系统，在命令行后台精准调用索引，在不泄露企业核心资产的前提下，自动完成万行级别的跨文件重构。
它没有把资源消耗在泛泛的 AI 对话上，让模型直接进入软件开发流程，替企业处理代码重构、Bug 排查和架构调试。工程师们不再将 Claude 当作简单的问答软件，而是作为可以量化投入产出比的工程生产力。
按实际计算用量（API）付费的收入随之爆发，2025 年 Anthropic 营收达到 45.85 亿美元；到了 2026 年上半年，这一数字直接冲到了 162.86 亿美元，年消费超 10 万美元的企业客户激增至 11900 多家。
然而，商业上的爆发并没有解开 Dario 的困境，反而被他推向了更紧迫的境地。
招股书里的细节揭示了这套商业模式的冷酷之处。2026 年第三季度，超过 80% 的收入来自发布不足六个月的新模型版本。
这意味着，模型能力的快速迭代，会导致旧版本的商业价值在短时间内衰减。
企业客户越依赖 Claude 处理复杂任务，Anthropic 就越需要持续投入资金采购更高性能的算力来训练新模型。
为了挣钱付算力费而开发的产品，最终要求公司投入更大的算力。
四、招股书里的隐藏条款
10 月 26 日，Anthropic 正式向美国证监会（SEC）递交招股书，拟登陆纳斯达克，股票代码 ANT。
文件披露了这几年环环相扣的经营结果。
2026 年上半年账面亏损高达 807.48 亿美元。这其中约 724 亿美元来自于估值上升导致的可转换票据公允价值变动。而在 Non-GAAP 口径下，其上半年经营利润已经达到了 16.18 亿美元，经营现金流也转正至 39.29 亿美元。
但真正为这五年故事画上句号的，是招股书附注里长达几百页的基础设施采购合同。
为了履行此前对巨头的算力承诺，并维持模型的迭代，Anthropic 签下了累计约 6370 亿美元的长期安排，其中包括：
若叠加微软、Lambda 等其他算力合约，整体长期承诺规模更高。
在博通高达 1612 亿美元的硬件合作背后，是针对自定义 ASIC 芯片与下一代光互联总线的定制化绑定。这意味着 Anthropic 未来的算力底层，在物理硬件层面上就已经与博通的供应链紧紧绑定在一起。
数据还显示，2025 年 Anthropic 有 47% 的销售额（约 21.6 亿美元）是通过亚马逊 AWS 和谷歌云市场完成的，为此支付了 3.51 亿美元的分销渠道费。
亚马逊与谷歌既是 Anthropic 的股东，又是其最大的算力供应商和分销渠道，同时还在研发各自的 Gemini 与硬件芯片。
五、硅谷的宿命
Anthropic 的这场 IPO，不是一家公司走向公开市场这样简单，它还是硅谷二十多年来罕见的一次集体下注。
自 2004 年 Google 上市、2012 年 Meta（Facebook）敲钟以来，硅谷已经很久没有见证过如此庞大、且股权与基础设施高度融合的上市事件。
但这不再是过去人们熟知的那个硅谷故事。
在过去的半个世纪里，硅谷引以为傲的神话，是车库创业与风险投资的组合。
两个年轻人在车库里写出代码，拿到几百万美元风投，靠着产品体验迅速颠覆掉庞大的行业巨头。
那是仙童半导体的时代，是苹果与微软的时代，也是早期谷歌与 Facebook 的时代。
而 Anthropic 的上市宣告了这种单打独斗模式的终结。
没有哪家初创公司，能够脱离巨头的云端基础设施独立跑完这场马拉松。
这场上市的本质，是硅谷巨头们用资本、芯片与数据中心，将一家原本试图挑战旧体制的叛逆者，重新招安到既有的硅谷产业秩序里。
亚马逊、谷歌、微软与博通们，不再只是坐在台下的财务投资人，它们是出资方、出租方、渠道商，也是这场技术革命最大的赢家。
甚至连曾经作为独立中立力量的顶尖大学实验室与研究机构，也因为买不起数万块 GPU 组成的计算集群，不得不将研究课题挂靠在巨头的算力云端。这种资源向少数巨头的集中，正在重塑整个硅谷的创新模式。
这就是硅谷在 AI 浪潮中的新位置，它依然是全球创新和人才的汇聚地，但技术创新的主导权，从分散的个人与实验室，逐渐聚拢到了极少数掌控着算力与资本的巨头手中。
Dario Amodei，那个当年走出旧金山旧厂房的理想主义者，用五年的时间完成了一场漫长的环行。
他逃离了 Sam Altman 的战场，却最终向整个硅谷的新秩序交出了答卷。
【版面之外】的话：
每一代技术革命，都会改变创新的门槛。
当代码就能启动一家公司的时代变得更难，拥有多少算力、能调动多少资本，也开始决定谁有资格参与竞争。
技术仍在向前，创新的门槛却在抬高。
硅谷下一轮洗牌，或许就从这里开始。
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。
In addition to the rest of the page, the author paints paintings and reviews them.
The story of Silicon Valley often begins with a split.
Five years ago, a core group of Open AI researchers chose to leave. They have different opinions about the direction of AI and have decided to set up a different course of action.
The departure was followed by the creation of an AI company, Anthropic, whose valuation and revenue were rising rapidly.
In 2021, at the old factory in San Francisco Pioneer Square, Dario Amodei and sister Daniela Amodei were loading their belongings into a cardboard box.
Those who left with them included the core researchers who later participated in GPT-4 development and the security team cadre of Open AI.
The reasons for leaving Open AI are not complex, and CEO Sam Altman is pushing Open AI to move from a non-profit research institution to a company seeking high valuation and commercialization. Dario thinks this rush is taking security out of control.
They named the new company Anthropic (means people-centred) and developed a charter to set up a laboratory for restraint and security in the AI fanaticism in Silicon Valley.
Dario thought he left Sam Altman's commercial battlefield. But he quickly found that the deeper the pursuit of safety and performance under the technical route of LLM, the more stressful the need for numeracy.
It's this determination to be independent of Sam Altman, to push it towards a capital game that can't go back.
I. Funding for start-up and leverage selected
To be trained to counter the forward model of Open AI, calculus is a physical threshold that cannot be bypassed, and calculus requires huge sums of money.
In order to maintain the independence of the research, Dario deliberately avoided the introduction of traditional Silicon Valley investments at an early stage.
But fate is here with a dramatic chain. In 2022, $580 million in early financing was provided to Anthropic, the founder of the encrypted currency exchange FTX, Sam Bankman-Fried (SBF).
But just a few months later, the FTX collapsed and the SBF was brought to court.
As part of the near-confirmation of assets, Anthropic equity was at risk of judicial auctions, and the company was once on the verge of survival.
Markets don't wait, AI competes day or hourly. The sudden break-up of funding has made it necessary for Dario to find in a very short period of time a sufficiently large connector to provide a stable calculator.
The clouds that smell the opportunity come in.
In the autumn of 2023, Amazon announced an investment of up to $4 billion to Anthropic, followed by Google with $2 billion. This funding has freed up the immediate need, but has also determined the real cost of this relief.
In order to implement this agreement, AWS even created a dedicated Trainium 2 chip cluster for it. Anthropic also agreed to give Amazon early priority to its architecture improvements, combining chip design with algorithm optimization depth.
II. Rejection of the throne
While Anthropic was struggling to advance within the infrastructure framework of the giant, a turning point seemingly capable of changing the pattern suddenly emerged.
On the evening of 17 November 2023, the Open AI Board of Directors summarily dismissed CEO Sam Altman. In the ensuing 48 hours, the confused Open AI board secretly called Dario Amodei, proposing an unexpected solution: inviting Dario to replace Open AI CEO and merging the two companies.
This is Dario's closest moment to recapturing the old, integrating resources.
But he chose to refuse.
He saw the interests behind the crisis, and eventually returned to Anthropic's lab, where he continued to advance Claude's development.
This refusal also allowed Sam Altman, who returned to Open AI, to speed up his multi-billion-dollar calculator expansion program.
Sam Altman, who's good at capital games, keeps squeezing the market space of Anthropic. In order not to be left behind, Anthropic had to multiply the training clusters in the research and development of Claude 3 and 3.5 series.
By the time Claude 3.5 Sonet was in it, the cost of calculating a single complete training and alignment had already reached several times the previous generation of flagship models.
In order not to open up the inter-generational gap on the model, Dario did not turn back, but followed Claude's training scale and computational consumption to a higher level.
And to sustain this unprecedented calculus, Anthropic needs to find business models that can be quickly realized.
III. EMPLOYED BUSINESS
Under financial pressure to double the monthly cost of computing, Anthropic cannot slowly develop C-end paid habits, like Open AI, with a general chat box for the general public, looking directly for clients willing to pay for efficiency.
And so, Anthropic made a key strategic shift, abandoning the fight with ChatGPT over C-end traffic and turning to the business side of the battlefield.
In 2025, Anthropic launched a programming tool for running directly embedded terminals command line, Claude Code.
The breakthrough came from the team's re-engineering of the bottom Prompt and the computing architecture.
Previously the developers had to manually package the entire project code files into the web box, while Claude Code was able to mount the local file system directly, fine-tuned the index in the backstage of command line, and automatically completed the cross-file reconstruction across the line without disclosing the enterprise ' s core assets.
It does not consume resources into a general AI dialogue, allowing the model to enter directly into the software development process, processing code re-constructing, bug sorting and architecture debugging for the enterprise. Engineers no longer use Claude as a simple question-and-answer software, but rather as the engineering productivity that allows quantization to input output.
Revenues from actual usage (API) Z105Z subsequently erupted, reaching $4.585 billion in 2025 at the Anthropic battalion; in the first half of 2026, this figure went directly to $1686 million, with business clients spending more than $100,000 a year surged to over 11,900.
However, the commercial outbreak did not solve Dario's predicament, but instead he pushed it into a more urgent situation.
The details in the book reveal the coldness of the business model. In the third quarter of 2026, more than 80% of the income came from the new model version of publish for less than six months.
This means that the rapid iterativeization of modelling capacity could lead to a decline in the commercial value of the old version in a short period of time.
The more business clients rely on Claude to handle complex tasks, the more Anthropic will need to continuously invest in better performance calculations to train new models.
Products developed to make money to pay for their energy will ultimately require companies to invest more in their calculations.
IV. Hidden clauses in equity book
On 26 October, Anthropic officially submitted to the United States Securities Commission (SEC) an offer to land in NASDAQ, stock code ANT.
The document discloses the results of operations that have been rounded up over the years.
In the first half of the year, there was a deficit of $80,748 million. Of that amount, some $72.4 billion was derived from changes in the fair value of convertible instruments resulting from increases in valuations. In the first half of the year, however, the operating profits of the Non-GAAP had reached US$ 1,618 million, and the operating cash flow had shifted to US$ 3,929 million.
But what really ends the five-year story is the hundreds of pages of infrastructure procurement contracts in the notes to the book.
In order to fulfil its previous calculus commitment to the giant and to maintain the iterative nature of the model, Anthropic signed a long-term arrangement totalling some $63.7 billion, including:
The overall level of long-term commitment is higher when other calculator contracts, such as Microsoft and Lambda, are folded.
Behind the $161.2 billion in hardware cooperation, the customisation of the custom-built ASIC chip to the next generation of light connects. This means that the future of Anthropic is at the bottom of the calculus, which is already closely tied to the supply chain at the level of physical hardware.
The data also show that in 2025 Anthropic had 47 per cent of sales (approximately $2.16 billion) made through the Amazon AWS and Google Clouds markets, for which $351 million was paid for distribution channels.
Amazon and Google are both shareholders in Anthropic and its largest calculator and distribution channels, while developing their respective Gemini and hardware chips.
The fate of Silicon Valley
This IPO of Anthropic, it's not as simple as a company going to the open market, it's also a rare collective bet in Silicon Valley for over 20 years.
Since Google was listed in 2004 and Meta (Facebook) struck the bell in 2012, Silicon Valley has not witnessed such a large listing with a high degree of integration of equity and infrastructure.
But this isn't the Silicon Valley story that people used to know.
In the past half century, Silicon Valley has been proud of its myth of a combination of garage entrepreneurship and venture capital.
Two young people wrote codes in their garages, received millions of dollars in windfall and experienced their products quickly subverting large industrial giants.
It was the age of fairies and semiconductors, the age of apples and Microsoft, and the age of early Google and Facebook.
And Anthropic's listing announced the end of this one-off mode.
No start-up company was able to run this marathon independently from the giant cloud infrastructure.
The essence of this listing is that Silicon Valley giants, using capital, chips and data centres, have re-engineered a rebel who had tried to challenge the old system to the established Silicon Valley industry order.
Amazons, Google, Microsoft and Chase are no longer just financial investors sitting under the table; they are the financiers, the lessors, the channel dealers and the biggest winners of the technological revolution.
Even the leading university laboratories and research institutes that used to be independent and neutral forces, because they could not afford to buy tens of thousands of GPUs, had to put the research topic on top of the giant's computing cloud. This concentration of resources towards a few giants is reshaping the innovative model of the entire Silicon Valley.
This is the new location of Silicon Valley in the AI wave, which remains the global convergence of innovation and talent, but the dominance of technological innovation, from dispersed individuals and laboratories, has gradually converged into the hands of a very small number of giants in control of computing and capital.
Dario Amodei, the idealist who left the old San Francisco factory that year, completed a long tour in five years.
He fled the battlefield of Sam Altman and eventually turned over his reply to the new order throughout Silicon Valley.
In the words:
Every generation of technological revolution changes the threshold of innovation.
The age when codes can activate a company becomes more difficult, with the ability to calculate, mobilize and decide who is eligible to compete.
Technology is still moving forward and the threshold for innovation is rising.
The next round of shuffles in Silicon Valley may start here.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.