Meta与Google AI代理模型对现金流的影响分析
Making sense of Muse, the agentic model, and evaporating cash flow at META and GOOG
Meta has been spending a lot on the AI buildout. It generated over $40 billion in FCF TTM, but less than $2 billion in Q2 2026.
Meta 在 AI 基础设施建设上投入了大量资金。其过去十二个月自由现金流（FCF TTM）超过 400 亿美元，但 2026 年第二季度仅为不到 20 亿美元。
With Muse, since it is mostly agentic, it seems it will consume a lot more tokens than a simple LLM. That means it should come with a much higher variable cost than typical LLMs.
由于 Muse 主要是代理型（agentic）产品，它似乎会比简单的 LLM 消耗更多的 token。这意味着它的可变成本应远高于典型的 LLM。
I have been trying Muse and it seems to do a comparable job to Claude on many agentic tasks that I want to do. It will work for many minutes on complex tasks that requires searching multiple sites and compiling results.
我一直在试用 Muse，它在许多我想做的代理型任务中表现与 Claude 相当。对于需要搜索多个网站并汇总结果的复杂任务，它会运行数分钟。
However Anthropic charges $100/month for Claude Max, and even then on some complex tasks I hit the token usage limit, say when I am searching across several websites.
然而，Anthropic 对 Claude Max 收取每月 100 美元的费用，即便如此，在一些复杂任务中我仍会遇到 token 使用量限制，例如在跨多个网站进行搜索时。
Muse is being offered for free and it seems to go for very long times, and I have not hit a usage wall yet.
Muse 目前免费提供服务，且似乎能运行很长时间，我尚未遇到使用上限。
How is this possible? It seems like the token usage will be very large.
这怎么可能？token 使用量似乎会非常大。
Zuckerberg has said he plans to take a cut of transactions done with Muse. But I am also seeing many people on social media questioning whether anyone trusts giving access to payments or email to Muse.
扎克伯格表示他计划从通过 Muse 完成的交易中抽取分成。但我也在社交媒体上看到许多人质疑是否有人愿意将支付或电子邮件的访问权限交给 Muse。
Personally, I do not feel comfortable giving access to those to Muse, at least currently.
就个人而言，目前我并不放心将此类权限授予 Muse。
And you need enough monetizable tasks to pay for all the non-monetizable tasks that Muse does.
你需要足够多的可变现任务，来为 Muse 所做的所有不可变现任务买单。
Consumer trust is generally higher for Google over Facebook, and Google already has access to most people’s email and credit card numbers, so it seems like it would be easier for them to launch a successful agentic product. However Google’s free cash flow already turned negative in Q2 2026, and the launch of a pure agentic product would push that even further into the red.
消费者对 Google 的信任度通常高于 Facebook，而且 Google 已经能够访问大多数人的电子邮件和信用卡号，因此他们推出成功的代理型产品似乎更容易。然而，Google 的自由现金流在 2026 年第二季度已转为负值，而推出一款纯粹的代理型产品会使这一情况进一步恶化。
All of these agentic products seem like they would drive a lot more token demand but all that token usage should continue to drive FCF at these firms lower.
所有这些代理型产品似乎都会带来大量的 token 需求，但这些 token 的使用量应会继续压低这些公司的自由现金流（FCF）。
I am still not sure of the right business model that justifies this large of a variable cost of usage. As these products expand into emerging markets, they will carry the same variable cost with lower incremental revenues. So it’s hard for me to see how you could offer a free product and try to take a cut of transactions.
我仍不确定何种商业模式能证明如此高昂的可变使用成本是合理的。随着这些产品向新兴市场扩展，它们将面临相同的可变成本，但增量收入却较低。因此，很难想象如何提供免费产品并试图从交易中抽取分成。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力