终身君主制：Class F 股本与不可稀释的创始人控制权
在剖析 Palantir 创纪录的第二季度财报之前，投资者必须首先理解其独特的公司治理结构，这一设计在主流科技巨头中是绝无仅有的。大多数硅谷公司（如 Google、Meta 和 Snap）通常采用双重股权结构（Dual Class Share Structure: 将股票分为普通投票权和高投票权两类，以保障创始人对公司长期战略的控制）。然而，Palantir 走向了更极端的第三条路：引入了专为三位创始人——首席执行官 Alex Karp、联合创始人 Peter Thiel 和 Steven Cohen 设计的 Class F 股票。这一机制会自动调整其投票权重，以确保无论创始人抛售多少股票，只要他们持有最低限度的基础股本，就能永久保留至少 49.999% 的绝对投票权。
与存在自然衰退周期的双重股权结构不同，Palantir 的 Class F 结构没有任何稀释机制。三位创始人即使卖掉自身 99% 的经济权益，依然能牢牢掌控公司的发展方向。这种近乎中世纪君主制的权力架构，让他们在资本市场上具有极强的免责属性。这也解释了为什么在过去三个月中，内部人交易以 9.3:1 的压倒性比例狂揽 4.33 亿美元现金，其中 Peter Thiel 单日套现 2.9 亿，而 Alex Karp 亦套现 1.32 亿。这种“只要控制权、不要股权”的套现行为，实际上是在向散户转嫁风险，而创始人们依然保留着对公司的终身支配权。
Original English Source
Before I get into the earnings themselves, which again are generally strong, I need to explain something about Palunteer's corporate structure that most investors don't know and that almost no coverage of the Q2 results has mentioned. I know the numbers have been lovely, but the house is the foundations as well and not just the pretty facade. Most major tech companies use what's called a dual class share structure. Google, Meta, and Snap all have this. The idea is pretty simple. Founders get shares with extra voting power, typically 10 votes per share instead of one, so they can maintain strategic control even as they sell stock or the company issues new shares. It's relatively common in Silicon Valley, and there are reasonable arguments both for and against it. The for argument is that it lets visionary founders make long-term decisions without being pressured by short-term shareholders. The against argument is that it concentrates power in the hands of a few individuals who may or may not deserve it indefinitely. Palunteer does something very different. It has a third class of shares, class F, held exclusively by the three founders, CEO Alex Karp, co-founder Peter Thiel, and co-founder Steven Cohen. The class F mechanism automatically adjusts its voting weight to guarantee that the three founders collectively hold at least 49.999% of total votes as long as they collectively hold a minimum number of shares. So let me say that again because the implications are super significant. Most dual class structures have a natural decay. As founders sell shares, their voting power gradually decreases, right? Eventually, if they sell enough, they lose control. That's the built-in check. Palunteer's class F structure has no such decay. It is mathematically engineered to maintain founder control regardless of how much stock they sell. They could sell 99% of their economic interest in the company and still control its direction. It is a level of permanent unshakable power usually reserved for medieval monarchs, maybe North Korean dictators and people who own the physical rights to the France theme song. This is again unusual even by Silicon Valley standards. I'm aware of very few public companies with this kind of permanent self-adjusting voting guarantee. It's closer to how a family trust or a sovereign wealth fund operates than how a publicly traded company typically works. And this matters because of what the founders are actually doing with their shares, which of course connects to the Q2 story. Over the past 3 months, Palunteer Insiders have sold approximately $433 million worth of stock. The selling to buying ratio across 72 recent insider transactions was 9.3 to 1. So that's nine shares sold for every one share bought. Peter Thiel, the man who recently called the Pope a Chinese Communist agent, which I addressed at length in a previous video, having been born in the Vatican, of course, sold 2 million shares in a single day in March, pocketing an estimated $290 million without making a single purchase. That is the financial equivalent of drinking a dirty martini made entirely of olive brine and liquid cash, wiping your mouth with a tax loophole, and leaving the bill for retail investors. Karp executed 32 separate sales totaling roughly $132 million. Steven Cohen sold shares worth about $100 million. Zero open market purchases from any insider. When insiders sell at this scale, those shares flood into the market, increasing supply and putting downward pressure on the stock price, which is part of why even after a blowout quarter in Q2, the stock is still down 10% year to date. The people with the most information about the company's future are choosing cash over their own equity, and everyone else's shares get a little less valuable each time they do it. Now, the standard defense here is that these are pre-planned sales under rule 10b5-1 trading plans set up months in advance. Executives diversify their wealth. Each transaction is defensible, but taken collectively, $433 million in three months, a 9.3:1 ratio, zero purchases, the pattern is notable. These are the people closest to the company. They have the most information about its trajectory. And they're choosing liquidity now over concentration in their own stock at these prices. That is not illegal, by the way. It is not necessarily bearish either, but it is the behavior of people who are comfortable selling and the class F structure guarantees they can do so indefinitely without ever losing an atom of control. And then there's the dilution. In fiscal year 2025, Palunteer stockbased compensation, that's shares given to employees as part of their pay, total $684 million. Buybacks, which are shares the company repurchased to offset that dilution, total just $75 million. So that's a net dilution of over $600 million. In simple terms, the pie is getting bigger, but existing shareholder slices are getting thinner. The founders don't feel this because their voting power adjusts automatically, but everyone else's does.主权 AI 的崛起：地缘政治与数据主权的逻辑闭环
在公司治理的阴影之外，必须承认 Alex Karp 对数据主权（Data Sovereignty: 确保数据在物理和法律上受本国或本地实体控制）的判断是准确的。当前主流的企业级 AI 采用的是 API 云端调用模式——企业需要将自身的商业机密或敏感数据发送至 OpenAI 或 Anthropic 的第三方服务器进行处理。对于普通消费级场景这无关痛痒，但对于国防承包商、金融巨头或医疗机构而言，数据跨越物理边界意味着致命的合规与竞争风险。这种传统模式本质上无异于给一个可能在五年内倒闭的公司支付月租，并任由其使用你的专有数据来训练模型，最后再将这些模型卖给你的竞争对手。
Palantir 倡导的主权 AI（Sovereign AI: 运行于客户本地或托管环境下、不依赖外部云端的数据与模型闭环体系）恰恰解决了这一痛点。Palantir 允许客户在自己控制的算力与本地服务器中运行开源/权重开放模型（Open-weight Model: 允许本地部署并自由调整权重的 AI 模型）。今年 6 月，Palantir 宣布与 Nvidia 达成战略合作，在物理隔离的政府涉密网络中部署 Neotron 混合架构模型。该模型每处理一个 Token 仅需激活约十分之一的参数，极大地降低了本地算力的部署成本。Palantir 第二季度美国商业收入飙升 149% 的亮眼数据证明，不仅是政府，私营企业也开始为这套“数据不出户”的安全叙事买单。
Original English Source
Now, we are nothing if not balanced and nuanced. So, I want to give credit where it's due because I think intellectual honesty requires it and because what I'm about to say might surprise some of you given the tone of the previous section. Alex Karp has a genuine important and I think correct point about data sovereignty and the Q2 earnings are in part evidence that the market agrees with him. Right now, the dominant model for enterprise AI works kind of like this. You send your data to OpenAI or Anthropic through an API. Their models process it on their servers. You pay per token and your proprietary information, your business secrets, your customer data, your competitive intelligence travels through somebody else's infrastructure. For casual use, that's I guess fine. We can all decide on our individual risk tolerance. But for a defense contractor, a bank, a hospital, or any serious enterprise handling sensitive information, it is a genuine problem. From speaking with my colleagues across the finance and technology sectors, this resonates across the whole field. The biggest consideration for large organizations evaluating AI adoption is not capability. The models are good enough for many use cases. It is that the data leaves the building and once the data leaves you're trusting a company that may be losing billions of dollars a year that may not even exist in 5 years and that has every incentive to use your data to improve its own models which then get sold to your competitors. It's a bit like paying a parasite a monthly subscription fee to feed on your internal organs only to find out that it's using your own DNA to breed a stronger version of the guy trying to steal your job. Karp's alternative, what Palanteer calls sovereign AI, runs openweight models on the client's own servers inside their own network. The data never leaves the premises. The organization owns the model weights, controls the compute, and doesn't pay per token fees to a third party. In June, Palanteer announced a partnership with Nvidia to deploy Neotron open weights models inside airgapped government environments. Exactly the kind of deployment that the API model cannot service. The Neotron models use a hybrid architecture that activates only about a tenth of parameters per token, making them dramatically cheaper to run on premises than full frontier models. The US commercial revenue surging 149% is evidence that private companies, not just government agencies, are buying this argument, and I think that they should be. Data sovereignty is the direction enterprise AI needs to go and Karp has been saying this louder and more chaotically than almost anyone else in the industry. Admittedly, the chaos isn't really serving him, but Pobet isn't perfect. Still, Alex, if you're listening, being right about the problem doesn't necessarily make you the cure. Just saying.天花板与商业摩擦：驻场工程师模式的规模化困境
然而，主权 AI 的商业化并非没有天花板。随着 OpenAI 与 Anthropic 在企业级市场的定价竞争趋于白热化，Token 调用成本持续雪崩，Palantir 原本依靠本地部署带来的成本差价（Cost Arbitrage）优势正在不断收窄。与此同时，Palantir 引以为傲的“无庞大销售团队”神话在政府和大型政军部门或许有效，但在强手如林的市场竞争中却成为了掣肘。面对微软、谷歌以及亚马逊 AWS 这类拥有庞大商业网络和完备主权/本地化云服务的巨头，Palantir 仅凭其特立独行的硬核销售方式正面临巨大的增长摩擦。
更深层的瓶颈在于 Palantir 的商业交付模式。为了让系统成功嵌入，Palantir 需要向客户内部派遣由前线部署工程师（Forward Deployed Engineer: 直接进驻客户业务一线负责系统集成与定制开发的工程师团队）组成的精英小队，进行极其重度的定制化改造。这更像是一种让昂贵、高压的外部工程团队直接“住进企业客房”直至重构整个网络架构的过程。虽然这样建立的业务极其“粘性”，一旦嵌入便极难被移除，但这根本不是边际成本为零的标准软件模式，而是带有浓重咨询服务色彩的埃森哲式重资产模式。这种高人力开销的交付逻辑限制了其作为纯软件公司应有的爆发式规模化（Scalability）潜力。
Original English Source
The sovereign AI thesis has a ceiling that Palanteer's earnings call didn't really acknowledge. If OpenAI and Anthropic race to the bottom on enterprise pricing, which they're actively doing, the cost arbitrage of running your own models narrows, token costs are crashing. And Palanteer's minuscule sales team narrative, I'm talking about car bragging about having fewer salespeople than almost any company this size, works when you're selling to governments who already know you. It's a very different proposition in commercial markets when you're competing with Microsoft, Google, and AWS, all of whom have their own sovereign and on premises offerings and considerably more commercial sales infrastructure. Then there is the political risk. Of course, Palanteer's government business, which still accounts for roughly half of total revenue, depends on contracts with the Department of Defense and Intelligence agencies. The DoD and Dodge have already collectively reduced $800 million in spending. Karp says meritocracy favors him. That Palunteer's products are so good that budget cuts will hurt competitors more. But that's the kind of thing you say right until you lose the contract. No. And the political entanglement cuts both ways. Feel bankrolled JD Vance's political career. Palanteer has deep proximity to the current administration and 13 former employees publicly criticized the company's direction in early 2026. Political relationships are assets in one administration and liabilities in the next. And there is a scalability question buried in the business model here. Palunteer's approach involves sending forward deployed engineers to sit inside client operations, custom building the integration, making themselves indispensable. It's less like buying a software package and more like adopting a team of very expensive, highly stressed engineers who move into your spare bedroom and refuse to leave until you reconfigure your entire network architecture. That is very sticky business. Once Palunteer is embedded, it's very hard to rip it out. But it doesn't scale like software. It scales like Accenture. And Accenture, for all its success, doesn't trade at a bazillion times earnings.完美主义陷阱：148 倍市盈率与不容出错的估值钢丝
Palantir 在第二季度递交了堪称惊艳的财务答卷：GAAP 口径下净利润达 10.6 亿美元，利润率高达不可思议的 55%，自由现金流达到单季度 12.2 亿美元。但这近乎完美的财务表现背后，是资本市场给予的极度苛刻的“完美估值陷阱”。当前 Palantir 的滚动市盈率（P/E Ratio）高达约 148 倍。作为对比，支撑起整个生成式 AI 算力底座的 Nvidia 市盈率仅为 33 倍，而 Apple 为 30 倍。这意味着 Palantir 的估值溢价是 Nvidia 的四倍有余。
148 倍的市盈率与其说是一个严谨的财务估值指标，不如说更接近于一种宗教式的集体狂热。它要求投资者相信线性数学在企业级软件领域会失效，要求 Palantir 每一个季度都必须交出超越市场极限的爆发式业绩。高估值意味着没有任何容错空间，一旦未来某一个季度其营收增速仅是持平甚至稍有放缓，都将引发惨烈的估值戴维斯双杀。即便在公布了如此强劲的财报且单日股价暴涨 30% 之后，其股价在当年依然累计下跌了约 10%，这一市场表现深刻折射出财报落地前其估值体系承受了多么沉重的泡沫压力。
此外，未被写进财报的伦理冲突与政商隐患依然在阴影中发酵。从为移民与海关执法局（ICE）提供驱逐追踪技术，到深度渗透五角大楼的战场杀伤链，再到饱受争议的预测性警务系统，Palantir 始终处于舆论的风口浪尖。甚至有 13 名员工曾在 2026 年初罕见地发表公开信，直言不讳地指责公司在政商博弈中的走向。这类系统性风险平日里并不会直接反映在资产负债表上，但一旦越过红线，就会在人才吸纳、声誉减损和监管重锤上给这家没有任何估值容错空间的“完美科技公司”致命一击。