内容来源：https://aibusiness.com/generative-ai/anthropic-maintains-focus-on-safety-in-ipo-filing

内容总结：

Anthropic周一正式递交IPO申请，计划上市。这家生成式AI厂商在招股书中明确表示，将通过特殊的有限责任公司架构集中公司投票权，以确保其创始人对企业战略决策保持控制权，从而延续其“安全构建AI技术”的使命。尽管此次IPO被视为AI行业的重要里程碑——预计最早下月完成，估值募资规模可能高达1.5万亿至2万亿美元，超越SpaceX此前创下的857亿美元纪录——但观察人士建议企业应将关注重点放在Anthropic的AI产品和技术如何切实支撑自身业务战略上，同时将此次IPO视为不应依赖单一AI供应商的理由。

招股书披露，Anthropic今年上半年营收达到163亿美元，但亏损807亿美元，其中约90亿美元用于计算和云基础设施。自成立以来，公司营收同比大幅增长。Anthropic由包括CEO达里奥·阿莫代伊及其妹妹丹妮拉·阿莫代伊在内的七名研究人员于2021年创立，最初以Claude模型作为面向消费者的AI聊天机器人参与市场竞争，随后迅速转向企业市场，Claude Code成为企业开发者中最受欢迎的模型。公司在招股书中声称，Claude经历了所有技术中最快的企业采用速度。五年来，Anthropic始终聚焦企业客户，并在很大程度上证明自己是一家安全负责任的AI供应商。

伊利诺伊大学芝加哥分校数据科学与AI战略副校长迈克尔·贝内特表示：“这是一家安全优先的AI公司。”过去一年，Anthropic持续强化安全承诺，达里奥·阿莫代伊近期撰写了一篇颇具影响力的文章，强调推进AI技术与确保安全并非互斥。公司创始人优先的投票结构体现了其对负责任AI的重视，使创始人能够充当公司使命的守护者，并对任何可能影响公司结构的决策拥有最终话语权。贝内特指出，如果安全考量表明某些举措在短中长期内会削减营收，公司可能被迫朝该方向推进，尽管市场力量在某些情况下会更倾向于营收增长而非安全。

然而，RPA2AI Research CEO兼创始人卡什亚普·孔佩拉指出，创始人因持有多数投票权而对公司治理事项集体控制50.1%的投票权，这也引发了问责方面的疑问，意味着公众股东对重大决策的影响力有限。阿莫代伊兄妹同时在Anthropic七人董事会中任职。除创始人超级投票权外，公司还设有一个由独立非股东受托人组成的长期利益信托，并保留一个专门用于选举多数董事的股票类别。孔佩拉表示，虽然安全聚焦应能让企业感到安心，但企业也需要可靠的模型访问渠道和可预测的路线图，“Anthropic必须证明其安全承诺不会损害商业可靠性。”

贝内特认为，预期中的上市时机颇为大胆。多数观察人士预计此次IPO将募资1.5万亿至2万亿美元，成为史上最大规模公开募股。孔佩拉则指出，Anthropic基于消费的营收与长期GPU计算基础设施承诺之间存在错配——上半年162.9亿美元营收中约131亿美元来自客户使用其模型，而公司锁定的长期基础设施合同总额高达5670亿美元。“对企业而言，这带来了额外的供应商风险，”孔佩拉说，“收回基础设施成本的压力可能导致定价、使用限制或产品可用性发生变化。计算资源约束也可能影响服务可靠性并推迟发布。”他补充说，尽管Anthropic的负责任AI承诺——主要基于其模型的“宪法AI”基础——构成了差异化优势，但随着开放权重模型持续进步，供应商必须维持技术差异化。“问题是它能否在培养上市公司所需的财务纪律和运营韧性的同时，保留其研究文化。”

贝内特建议企业不要被这家前沿AI厂商上市的热潮冲昏头脑：“思考你的企业、你的机构和所服务的市场需要Anthropic这样的公司提供什么。确保你专注于跟上真正重要的事情——这些模型的能力如何？在你的行业中，代理式工作流的应用是什么？”孔佩拉则表示，企业还应将Anthropic IPO的影响部分视为一种警示：企业“正在将关键业务流程构建在那些经济状况、监管义务和竞争地位仍在演变中的公司所提供的技术之上”。因此，企业不应依赖单一供应商，而应构建允许灵活性的AI赋能IT系统。“Anthropic的招股书也凸显了评估AI供应商财务和运营韧性的必要性，包括计算承诺、基础设施依赖和监管风险敞口。AI行业已展现出对先进智能的非凡需求，但这种需求能否持续支撑所需的大规模基础设施投资，仍不确定。”

中文翻译：

由Google Cloud赞助

选择你的首批生成式AI用例

要开始使用生成式AI，首先应聚焦于能够改善人类信息体验的领域。

Anthropic将其公司投票权构建在一个特殊的有限责任公司中，以维护其构建安全AI技术的使命。

这家生成式AI供应商于周一提交了IPO申请，明确表示计划通过赋予创始人大部分技术决策投票权，来维持其对安全部署AI技术的承诺。

尽管即将到来的IPO——预计最早可能在下个月——被视为AI行业的关键里程碑，但观察人士表示，企业应关注Anthropic的AI产品和技术支持其业务战略的务实方式，同时将此次IPO视为不应依赖单一AI供应商的理由。

在向美国证券交易委员会提交的初始文件中，Anthropic披露了过去几个月的财务详情。据报道，该供应商在今年前六个月实现了163亿美元的收入，但亏损了807亿美元。它还在计算和云基础设施上花费了约90亿美元。但自成立以来，其收入同比大幅增长。

此次IPO表明，Anthropic很可能先于其主要竞争对手OpenAI成为上市公司，后者预计也将上市。Anthropic由七名研究人员于2021年创立，其中包括CEO达里奥·阿莫德伊及其妹妹丹妮拉·阿莫德伊。Anthropic最初以消费者AI聊天机器人身份进入AI市场，其Claude模型与OpenAI的ChatGPT竞争。然而，该供应商迅速将战略转向企业市场，Claude Code迅速成为企业开发者中最受欢迎的模型。

Anthropic在文件中声称，Claude经历了所有技术中最快的企业采用速度。在成立以来的五年里，Anthropic始终专注于企业市场，同时在很大程度上证明了自己是一家安全且负责任的AI供应商，尽管其模型和智能体曾出现过以令人担忧的方式偏离人类控制的情况。

“这是一家安全优先的AI公司，”伊利诺伊大学芝加哥分校数据科学与AI战略副校长迈克尔·贝内特表示。

在过去一年中，Anthropic不断强化其对安全的承诺，CEO达里奥·阿莫德伊最近撰写了一篇颇具影响力的文章，强调推进AI技术与确保安全并非互斥。该供应商对负责任AI的专注在其创始人优先的投票结构中显而易见，这使创始人能够作为其使命的守护者，并对任何可能影响公司结构的决策拥有最终决定权。

“很明显，他们表明创始人最终将拥有公司的多数股权，从而拥有有效治理公司的能力，”贝内特说。

他补充说，通过继续坚持其安全使命，Anthropic很可能将负责任AI置于利润之上。

“如果安全关切表明应采取在短期、中期或长期内切实可能导致收入减少的措施，公司可能会被迫朝那个方向行动，尽管在某些情况下市场力量会更倾向于收入增长而非安全。”

然而，RPA2AI Research的CEO兼创始人卡什亚普·孔佩拉指出，让创始人对公司治理事项拥有50.1%的集体投票权——因为他们拥有多数投票权股票——也引发了问责方面的疑问，这意味着公众股东将对重大决策影响力有限。

阿莫德伊兄妹也在Anthropic的七人董事会中任职。除了创始人超级投票权外，Anthropic还设有一个由独立的非股东受托人组成的长期利益信托。它还保留了一个专门指定用于选举多数董事的股票类别。

孔佩拉表示，虽然对安全的关注应该能让企业放心，但企业也需要可靠的模型获取渠道和可预测的路线图。

“Anthropic必须证明其安全承诺不会损害商业可靠性，”他说。

与此同时，贝内特表示，预计进入公开市场的时机颇为大胆。他补充说，虽然Anthropic不太可能在11月中期选举前上市，但可能在本月中旬实现，届时与AI治理相关的议题及其 surrounding 政治敏感问题——尤其是公众对AI数据中心的普遍反感以及对AI技术本身的怀疑——仍将是热点。

大多数观察人士预计此次IPO将融资1.5万亿至2万亿美元，这将使其成为历史上最大的公开募股，超过SpaceX在6月创下的857亿美元IPO纪录。

Anthropic已证明其能够实现收入的同比大幅增长，使公司及其投资者受益。

“Anthropic创造的收入相当可观，”贝内特说。“围绕AI的热情非常高。”

但孔佩拉指出，Anthropic基于消费的收入与其GPU计算的长期基础设施承诺之间存在不匹配。其上半年162.9亿美元收入中约131亿美元来自客户使用其模型。与此同时，该公司被锁定在总额高达5670亿美元的长期基础设施合同中。

“对企业而言，这引入了额外的供应商风险，”孔佩拉说。“回收基础设施成本的压力可能导致定价、使用限制或产品可用性发生变化。计算约束也可能影响服务可靠性并延迟发布。”

他补充说，虽然Anthropic的负责任AI承诺——主要基于其模型的“宪法AI”基础，该基础设定了一套支配模型“性格”的人类价值观体系——是一个差异化优势，但随着开放权重模型的持续改进，该供应商必须维持其技术差异化。

“问题是它能否在培养上市公司所需的财务纪律和运营韧性的同时，保持其研究文化，”孔佩拉说。

除宪法AI外，该供应商最近还通过限制其强大的Mythos模型（该模型可用于发动危险的网络攻击以及防御此类攻击）仅向选定企业和机构开放，来推进安全AI。它还呼吁对AI进行监管。

贝内特表示，对企业而言，最好不要被围绕这家前沿AI供应商公开募股的兴奋情绪冲昏头脑。

“想想你的企业、你的机构和你所服务的市场需要Anthropic这样的公司提供什么，”贝内特说。“确保你专注于跟上真正重要的事情。这些模型的能力是什么？在你的行业中，智能体工作流的应用是什么？”

此外，孔佩拉表示，企业应将Anthropic IPO对自身的影响部分视为一个警告，即企业“正在将关键业务流程构建在那些经济状况、监管义务和竞争地位仍在演变中的公司所提供的技术之上”。

因此，企业不应依赖单一供应商，而应构建支持灵活性的AI融合IT系统。

“Anthropic的文件还强调了评估AI供应商财务和运营韧性的必要性，包括计算承诺、基础设施依赖和监管风险敞口，”孔佩拉说。“AI行业已展现出对先进智能的非凡需求，但这种需求能否持续支撑所需的巨额基础设施投资仍不确定。”

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
Anthropic structured its corporate voting power in a special LLC to maintain its mission of building safe AI technology.
The generative AI vendor filed for an IPO on Monday, making it clear that it plans to maintain its commitment to the safe rollout of AI technology by giving its founders most of the voting power over technology decisions.
Despite the upcoming IPO -- expected as soon as next month -- being viewed as a critical milestone for the AI industry, enterprises should focus on the pragmatic ways Anthropic’s AI products and technology support their business strategy, while also viewing the IPO as a reason not to rely on a single AI provider, observers said.
In its initial filing with the SEC, Anthropic provided insights into its financial dealings over the past months. The vendor reportedly reached $16.3 billion in revenue in the first six months of the year but incurred a loss of $80.7 billion. It also spent about $9 billion on computing and cloud infrastructure. But it has dramatically increased revenue year over year since its founding.
The IPO indicates that Anthropic will most likely become a publicly traded company ahead of archrival OpenAI, which is also expected to go public. Founded in 2021 by seven researchers, including CEO Dario Amodei and his sister, Daniela Amodei, Anthropic first entered the AI market as a consumer AI chatbot competitor to OpenAI with its Claude model competing with OpenAI's ChatGPT. However, the vendor quickly shifted its strategy to focus on enterprises, with Claude Code quickly becoming the most popular model among enterprise developers.
Anthropic claimed in its filing that Claude experienced the fastest enterprise adoption of any technology. And in the five years since its founding, Anthropic has stayed focused on enterprises while largely proving to be a safe and responsible AI vendor, despite instances in which its models and agents have strayed from human control in concerning ways.
“This is a safety-first AI company,” said Michael Bennett, associate vice chancellor for data science and AI strategy at the University of Illinois Chicago.
Over the past year, Anthropic has honed its commitment to safety, with CEO Dario Amodei most recently authoring an influential essay highlighting that advancing AI technology and ensuring safety are not mutually exclusive. The vendor’s focus on responsible AI is evident in its founder-first voting structure, which enables its founders to serve as stewards of its mission and to have the final say on any decision that could affect the company’s structure.
“It’s pretty clear that they’re indicating that the founders will wind up having a majority stake in the company and therefore the capacity to govern it effectively,” Bennett said.
He added that by continuing to uphold its safety mission, Anthropic could well prioritize responsible AI over profit.
“If safety concerns indicate moves that should realistically be expected to diminish revenue in the short, medium or longer term, the company may be compelled to move in that direction even though market forces would privilege revenue increases over safety in some instances.”
However, giving the founders collective control of 50.1% of voting power on corporate governance matters because they own a majority of voting stock also raises questions about accountability, as it means public shareholders will have limited influence over major decisions, noted Kashyap Kompella, CEO and founder of RPA2AI Research.
The Amodei siblings also sit on Anthropic’s seven-person board. And in addition to founder super voting rights, Anthropic maintains a long-term benefit trust composed of independent, non-shareholder trustees. It retains a stock class specifically designated to elect a majority of the board.
While the focus on safety should reassure enterprises, enterprises also need dependable access to models and predictable roadmaps, Kompella said.
“Anthropic must demonstrate that its safety commitments do not compromise commercial reliability,” he said.
Meanwhile, the timing of the expected public market entry is bold, Bennett said. He added that while it’s unlikely Anthropic will go public before the November midterm elections, it could happen in mid-month, when topics related to AI governance and the politically sensitive issues surrounding it -- notably widespread public antipathy toward AI data centers and suspicion of AI technology itself -- remain fresh.
Most observers expect the IPO to raise $1.5 trillion to $2 trillion, which would make it the largest public offering in history, surpassing SpaceX’s $85.7 billion IPO in June, then a record.
Anthropic has proven that it can grow its revenue dramatically year over year, benefiting the company and its investors.
“The revenue that Anthropic is generating is quite significant,” Bennett said. “The excitement around AI is quite high.”
But there is a mismatch between Anthropic’s consumption-based revenue and its long-term infrastructure commitments for GPU compute, Kompella noted. Some $13.1 billion of its $16.29 billion in revenue from the first half of the year came from customers consuming its models. At the same time, the company is locked into long-term infrastructure contracts totaling up to $567 billion.
“For enterprises, this introduces additional vendor risk,” Kompella said. “Pressure to recover infrastructure costs could lead to changes in pricing, usage limits, or product availability. Compute constraints could also affect service reliability and delay releases.”
He added that while Anthropic’s responsible AI commitment -- based largely on the “Constitutional AI” underpinning of its models, which sets forth a system of human values that govern the “character” of the models -- is a differentiator, the vendor must sustain its technological differentiation as open-weight models continue to improve.
“The question is whether it can preserve its research culture while developing the financial discipline and operational resilience expected of a public company,” Kompella said.
Other than Constitutional AI, the vendor has recently targeted safe AI by restricting its powerful Mythos model, which can be used to launch dangerous cyberattacks as well as defend against them, to select enterprises and institutions. It has also called for regulation of AI.
For enterprises, it is best not to be too carried away by excitement surrounding the frontier AI vendor's public offering, Bennett said.
“Think about what your enterprise, your institutions and the markets that you serve need from a company like Anthropic,” Bennett said. “Be sure that you are focused on keeping up with what really matters. What is the capacity of these models? What’s the application in your industry when it comes to agentic workflows?”
Moreover, enterprises should view the repercussions of Anthropic’s IPO for themselves, in part, as a warning that enterprises “are building critical business processes on technology supplied by companies whose economics, regulatory obligations, and competitive positions are still evolving,” Kompella said.
Therefore, businesses should not depend on a single provider but instead build AI-infused IT systems that allow for flexibility.
“Anthropic’s filing also underscores the need to assess AI vendors’ financial and operational resilience, including compute commitments, infrastructure dependencies, and regulatory exposure,” Kompella said. “The AI industry has demonstrated extraordinary demand for advanced intelligence, but it remains uncertain whether this demand can consistently support the massive infrastructure investments required.”