Palantir Is Mostly Right. But They're Also Selling Something.

Palantir Is Mostly Right..

9:31

Palantir published a nine-point manifesto on AI sovereignty this week that's been making the rounds. The highlights: sovereignty dictates your institution's future, data retention is your treasure, tokenmaxxing hijacks your value orientation, and controlling your weights is controlling your fate. CEO Alex Karp followed it up on CNBC, calling the per-token pricing model a "tax" and suggesting something has gone "completely wrong" with how AI is sold to enterprises.

It's worth taking seriously. It's also worth separating the signal from the pitch.

Because Palantir is approximately 80 to 90 percent right on the strategic direction — and the 10 to 20 percent where they're wrong is specifically the part that benefits Palantir.

Let me start with what they're getting right.

The sovereignty argument is real. The question of who controls your AI infrastructure is going to be one of the defining strategic decisions for governments and enterprises over the next decade. Not in a vague, futurist way. In a very concrete, contractual, operational way.

Sovereignty here doesn't mean owning every model yourself. It means retaining the ability to choose: where your models run, who has access to your data, whether you can switch providers, who controls upgrades and pricing, and what happens when geopolitical or business conditions change.

If your institution's decision-making is running on an AI stack you don't control, and that provider changes their pricing, changes their policies, gets acquired, gets sanctioned, or simply decides your use case doesn't fit their risk appetite anymore — you're exposed in a way that's hard to unwind quickly. The decisions made by that system are already embedded in your operations.

This is the cloud lock-in argument, but the stakes are higher. Cloud lock-in meant your data was somewhere inconvenient to move. AI lock-in means the judgment layer of your organization is somewhere inconvenient to move. Those are meaningfully different problems. The cloud parallel is instructive. During the hyperscaler land-grab of the 2010s, enterprises got lured in with free credits and then found themselves trapped by data egress fees and proprietary tooling. The "multi-cloud" movement that followed — attempting to maintain portability across AWS, Azure, and GCP simultaneously — turned out to be enormously expensive and operationally complex. Most enterprises ended up defaulting to the lowest common denominator anyway. The lesson: lock-in is easy to get into and hard to get out of, and the cost shows up years after the decision.

The AI equivalent is data gravity. Once your operational data lives inside a closed ecosystem's vector stores and agent memory — trained against, fine-tuned on, embedded in — migrating to a different model provider isn't a configuration change. It's a reconstruction project.

The data argument is the strongest point in the manifesto, and it's one I think is still underappreciated across the industry. Most companies believe their competitive advantage in AI is "using ChatGPT." It isn't. Everyone has access to similar foundation models. GPT-5 versus Claude versus Gemini is increasingly a commodity distinction for most use cases.

What your competitors don't have is your data. Your customer interactions. Your historical operational records. Your internal processes, institutional knowledge, and ten years of accumulated decisions. That's the fuel. The model is the engine, and right now engines are cheap and getting cheaper.

We're entering an era where data quality matters more than model quality for most businesses. A company with ten years of customer service transcripts, engineering notes, sales calls, contracts, and internal documentation — properly organized, properly retrieved — is going to outperform a company with better model access and worse data. Not because the model is smarter. Because the model has better material to work with.

The numbers bear this out. RAG — retrieval-augmented generation, where you inject your proprietary data into a rented foundation model at inference time — accounts for over 50 percent of active enterprise AI production deployments right now. Fine-tuning on proprietary data accounts for roughly 9 to 15 percent. Enterprises are already making this bet, mostly intuitively. The data advantage is real.

Tokenmaxxing is where it gets interesting.

Palantir coined the term — derived from internet slang like looksmaxxing and wealthmaxxing — and it's crystallized quickly into a genuine industry critique. The phenomenon it describes is real: organizations celebrating high token consumption as a proxy for AI sophistication. An agent used two million tokens on a task. Impressive. But did it create value? Those aren't the same question, and a lot of organizations aren't asking the second one.

Large context windows and massive token consumption can produce the illusion of thoroughness. The AI processed everything. It must have been comprehensive. Sometimes a thirty-second script with a single targeted API call solves the same problem better. Karp's specific critique is that the metered, per-token pricing model actively incentivizes this. If the vendor charges by the token, inefficient agentic loops that burn ten million tokens on a minor database query are the vendor's profit center. The enterprise bleeds margin. The provider counts revenue.

But here's where I'd push back: token usage isn't inherently bad. Some workloads genuinely require millions of tokens. Legal discovery. Codebase analysis across a large repository. Medical research synthesis. Enterprise document comparison at scale. For those workloads, high token consumption is the appropriate cost of doing the work properly.

The metric isn't lowest token count. The metric is cost per useful outcome. Palantir's framing slides toward implying that token consumption is the problem — which is convenient if you're selling a platform that runs on local infrastructure rather than metered APIs. The actual insight — measure value, not volume — is correct. The framing is self-serving.

Weights are the most nuanced point.

Weights represent learned behavior — the distilled output of training. Palantir's argument is that organizations should own as much of this stack as practical. Directionally correct. But the practical implications are where this gets complicated. There are layers here that often get collapsed: foundation model weights, fine-tuned weights, retrieval systems, internal knowledge bases, agent workflows, evaluation pipelines, prompting strategies. Palantir's manifesto treats this as a unified stack that organizations should control top to bottom.

Most organizations should not be training foundation models. OpenAI, Anthropic, and Google spend billions of dollars and consume gigawatts of compute to build these systems. The idea that individual enterprises should recreate this work is not strategic advice — it's a setup for very expensive failure. What organizations should own — actually own, with genuine control — is everything above the foundation layer. Their data. Their workflows. Their prompts. Their evaluation systems. Their agent orchestration logic. Keep those portable and you can swap the foundation model underneath without losing the institutional intelligence built on top of it.

Think of it like semiconductor fabs. Almost nobody owns one. But everyone who builds on silicon owns the product built on top of it. You don't need to own TSMC to have a durable hardware business. You need to own what differentiates you at the product layer. The equivalent for AI: rent the frontier intelligence, own the uniquely yours parts. Build portability into the architecture so the engine can be swapped without losing the vehicle.

Here's the irony that doesn't get enough attention.

Palantir's answer to vendor lock-in is Palantir. Their Foundry and AIP platform is proprietary, deeply sticky, and designed to sit at the center of enterprise data infrastructure. France's internal intelligence service and Germany's military have both been actively building alternatives specifically to reduce their dependence on Palantir. They are pursuing AI sovereignty from the sovereign AI platform.

The sovereign AI market — infrastructure and enablement services for organizations that want to run models inside their own environments — is projected to exceed $11.5 billion in 2026 and scale toward $177 billion over the next decade. Palantir's manifesto is, among other things, a highly calculated attempt to own that narrative and that market. This doesn't make the advice wrong. But it does mean you should read it the way you'd read any strategic framework from a vendor with a strong commercial interest in a particular conclusion.

There's something the manifesto doesn't say, and it's the most important thing. The genuine scarce resource in AI right now isn't GPUs. It isn't tokens. It isn't even data, though data matters enormously. It's attention and workflow design.

We've crossed a threshold where the models are good enough for most tasks. The bottleneck has shifted. The problem isn't getting AI to do the work. The problem is designing systems that know when to invoke AI, know what context to provide, know when to involve a human, and continuously improve from outcomes. That's where durable competitive advantage actually comes from. Not from owning the weights. From owning the design of the loop — the judgment architecture, the feedback system that makes the whole operation get smarter over time from real results.

This doesn't show up in Palantir's framework because it's hard to sell and impossible to productize in the conventional sense. It's the kind of capability that has to be built into how an organization operates, not installed from outside. Every organization has to develop it differently because every organization has different judgment requirements.

So here's how I'd restate the framework after separating signal from pitch:

Own your data rigorously. Not "store it somewhere" — actively curate, structure, and protect it as a strategic asset. The compounding value of proprietary operational data is real, and most organizations are still treating it as a byproduct rather than the asset it actually is.

Avoid lock-in on the things that are hardest to move. Your workflows, your decision logic, your evaluation systems — build those to be portable. Don't architect your core operations around a single model provider's specific capabilities in ways that make switching expensive. Measure AI by business outcomes. Not token counts, not benchmark scores, not demo impressiveness. Cost per useful outcome is the right unit. Everything else is a vanity metric.

Invest in the judgment layer. The organizations that win over the next decade won't be the ones that bought the best model. They'll be the ones that built the best systems for directing, evaluating, and improving AI-driven work. That's a design and governance challenge, not a compute procurement challenge. Palantir is making a sophisticated strategic argument and they deserve credit for it. They're also a company whose entire value proposition depends on enterprises believing that AI deployment requires specialized, proprietary infrastructure and deep vendor relationships. France and Germany figured this out. Read the framework with that in mind.

The goal isn't to own the engine. It's to ensure the engine can be swapped without losing the vehicle you've built around it. That's AI sovereignty for organizations that aren't buying ten-year, ten-billion-dollar Army contracts. Which is most organizations.

---

Ad: Use code: RICH99 for a discount.

Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital and CIO of Data Power Supply.