On Tuesday, Palantir posted a nine-point manifesto on X titled “thoughts on the importance of AI sovereignty,” first flagged by Business Insider. The document is a direct assault on the prevailing business model of the AI industry, which relies on metered per-token pricing. Palantir argues that whoever controls your data and your models controls your future, so institutions should never hand either to a third party.
The first point declares that “your AI sovereignty dictates your institution’s future.” The second warns that “data retention is your treasure.” Point four is blunter: “Controlling your weights is controlling your fate.” Weights are the numerical values that encode what a model has learned. Palantir casts them as an organisation’s crown jewels.
The sharpest barb targets the frontier labs. Palantir attacks “tokenmaxxing,” the habit of spending as much as possible on AI. It calls the practice “the addictive feeling of false progress.” Heavy token use, it argues, rewards throwaway scripts over solid software. Then comes the tell: “There is a reason why those selling tokens refuse to charge based on value.”
That is a direct shot at OpenAI, Anthropic, and similar companies. Their businesses run on metered, per-token pricing. Chief executive Alex Karp put it more crudely in a follow-up. “Why are they charging for tokens?” he asked, if the value is really there. The complaint has teeth. Enterprise AI bills keep climbing even as token prices fall. Many organizations are locked into contracts where cost scales with usage, often with no cap, creating unpredictable expenses. Palantir positions its own model as a fixed-cost alternative: customers pay for the software platform and run models on their own infrastructure, decoupling cost from the volume of queries.
It is worth remembering what Palantir sells. Its business is deploying software, and increasingly AI models, inside a customer’s own walls. That includes air-gapped systems cut off from the outside world. This week it also said it had won accreditation for NATO’s classified network. Palantir is packaging open models from Nvidia for government use under a “sovereign” banner as well. A manifesto that tells institutions to own their stack is, conveniently, a manifesto for buying Palantir.
Karp has never hidden his contempt for rivals. He recently told CNBC that AI companies “don’t understand how unlikeable they are.” Their products, he added, “don’t actually work the way” customers expect. He has also predicted the nationalisation of AI firms. The sovereignty manifesto fits that worldview. It is combative, absolutist, and pitched at states and armies rather than startups.
The irony is that “sovereignty” is the same argument now used against Palantir. Europe has turned the word into policy, and the debate over who really controls the continent’s AI is raging in Brussels and national capitals. France’s foreign intelligence service dropped Palantir for a homegrown rival, and Germany’s military has kept it at arm’s length. For those governments, true sovereignty means not depending on a single American vendor either, however loudly it preaches independence.
Strip away the grandeur and Palantir’s core claim is reasonable. Data and model weights are real sources of advantage, and handing them to a third party carries real risk. But a nine-point scripture about “alpha” and “the means of production” is also brand-building. It courts a moment when every Western institution suddenly worries about where its AI runs. Palantir did not invent that anxiety. It is just very good at selling the cure.
To understand Palantir’s strategy, one must examine the evolution of the AI market. Over the past two years, generative AI has exploded in popularity, driven by large language models like GPT-4, Claude, and Gemini. These models are typically accessed via APIs, where customers pay per token — roughly per word or sub-word processed. The token billing model was inherited from earlier cloud services, but it has become a lightning rod for cost-conscious enterprises. As token prices drop, usage typically rises even faster, meaning total spend often increases. Critics argue that this model incentivizes model providers to maximize token consumption rather than delivering efficient, reliable solutions.
Palantir’s counter-proposal is not new. The company has long championed on-premise data integration and analysis, often in government and defense sectors. Its Foundry platform allows organizations to fuse disparate data sources and build operational AI applications. By running models locally, customers avoid sending sensitive data to external servers, reduce latency, and maintain full control over their intellectual property. This approach aligns with the growing push for “AI sovereignty” — the idea that nations and institutions should retain governance over their AI systems and data.
The sovereignty concept has gained traction in Europe, where the General Data Protection Regulation (GDPR) and digital sovereignty initiatives encourage homegrown technology. The European Commission has funded projects to develop European AI models and cloud infrastructure, partly to reduce reliance on US tech giants. Palantir’s manifesto taps into this mood, but its own position is complicated. The company is American, based in Denver, Colorado, with a controversial history of contracts with US immigration and law enforcement agencies. European critics argue that true sovereignty requires European-owned platforms, not just on-premise deployment by a foreign vendor.
Karp’s combative tone also reflects a deeper schism in the AI industry. On one side are the “tokenmaxxers” — the large labs that see exponential scaling as the path to artificial general intelligence. They burn billions of dollars on compute and struggle to make their products consistently reliable. On the other side are “vertical AI” companies that embed models into specific workflows and charge by value, not by usage. Palantir is trying to straddle both camps: it uses open source or third-party models but wraps them in its own platform and charges a subscription fee based on the value delivered to the customer.
The manifesto’s military language is deliberate. Palantir’s core customer base remains the US Department of Defense and allied intelligence agencies. It has built a reputation as the go-to provider for battlefield AI, data fusion, and targeting systems. The company believes that future warfare will be decided by AI that operates without external connectivity, processing captured data in real-time to deliver tactical advantage. For such applications, token-based pricing is absurd — a soldier cannot afford to pause a mission to pay for each image analyzed. Palantir’s “sovereign AI” is thus a sales pitch for defense modernization, wrapped in philosophical principles.
But not all customers are convinced. In the private sector, many enterprises are comfortable using cloud AI services for routine tasks like customer support, content generation, or analytics. They value the scalability and continuous improvement that comes from shared models. Palantir’s arguments resonate most with organizations handling classified, proprietary, or heavily regulated data. The healthcare industry, for example, may fear sending patient records to third-party AI providers, even under strict privacy agreements. Financial firms worry about trading algorithms being reverse-engineered via token output analysis.
Palantir’s attack on token pricing also overlooks a key benefit: the pay-as-you-go model lowers the barrier to entry for small and medium businesses. Without tokens, they would need sizeable upfront investment in infrastructure and expertise. Palantir’s platform itself is expensive, often costing millions of dollars annually, placing it out of reach for many potential users. The sovereignty manifesto thus targets a wealthy, risk-averse elite — governments, large corporations, and defense contractors — rather than the broader market.
Nevertheless, the manifesto has stirred debate. It has been shared widely on social media, praised by some privacy advocates and criticized by AI developers who see it as fear-mongering. The timing is notable: Palantir’s stock has risen significantly in 2024 as the company posted better-than-expected earnings and secured new government contracts. The manifesto helps reinforce its brand as the sober, security-conscious alternative to flashy Silicon Valley AI startups.
In the end, Palantir’s war on tokenmaxxing is a war for customers’ minds and wallets. It offers a stark choice: either control your data and models with Palantir, or surrender them to the token sellers. It is a powerful narrative, especially at a time when AI safety, data breaches, and regulatory uncertainty dominate headlines. Whether that narrative translates into lasting market dominance remains to be seen, especially as competitors like Microsoft and AWS push their own sovereignty solutions with broader ecosystems.
One thing is certain: the debate over who controls the future of AI will intensify in the coming years, and manifestos like Palantir’s will be used to frame the terms of that battle.