Something has gone completely wrong in enterprise AI

Something has gone completely wrong

There is a moment in every boom when the music is still playing, the lights are still flashing, and someone finally notices that the bar tab has been left on the table.

That is roughly where enterprise AI feels today.

Palantir’s Alex Karp may have delivered the message in his usual “ effing” blunt-force style, but underneath the fireworks sits a serious concern. The frontier labs have built extraordinary machines. No one sensible disputes that. The issue is whether the people renting those machines are actually building a business advantage, or simply feeding coins into someone else’s meter.

Karp’s complaint is not that AI does not work. It is that the economics increasingly look like a casino where the house owns the chips, the tables, the cameras and perhaps a copy of every card you have ever played.

Companies are being told to embrace token usage, embed models into every workflow, build agents, automate decisions, scale experimentation and move faster than the competition. The promise is seductive. The more AI you use, the more productive you become. The more productive you become, the more valuable the business becomes.

But there is a catch.

Every time a company pushes proprietary data, internal process knowledge, customer information and years of accumulated institutional judgement through an external model, it is not merely consuming AI. It may be exporting part of its operating memory.

That is why Karp’s language around “AI sovereignty” matters more than the headline theatrics. He is asking a question that many boards have not yet properly confronted: when you build on someone else’s model, someone else’s cloud, someone else’s weights and someone else’s pricing system, how much of the future business do you really own?

The token model sits at the heart of that tension.

On paper, it looks elegant. You pay for what you use. A few tokens here, a few million there. It feels like turning on a utility. But utilities normally get cheaper as they scale. AI can feel like the opposite. The more deeply a company embeds it into research, coding, customer service, compliance, trading, legal workflows, logistics and internal decision-making, the more the meter starts spinning.

The demo may cost pennies. Production is where the bill arrives.

An AI agent running across a corporate system is not one prompt and one answer. It can call multiple models, retrieve documents, scan data, invoke tools, write code, check code, run another model, audit the output and then start the whole process again. Multiply that across an enterprise and the token jar begins to look less like a subscription service and more like a taxi with the meter running in heavy traffic.

Karp’s point is that companies may be confusing activity with ownership.

The dashboards show rising AI usage. Token consumption climbs. Internal teams report more pilots, more prompts, more automation and more experimentation. It looks like progress because everything is moving.

But a hamster wheel moves too.

The question is whether all that motion compounds into proprietary intelligence, or whether it simply compounds into a larger invoice for the model provider.

That is where the data issue becomes central. Data is not just fuel. It is the memory of the institution. It is the record of what worked, what failed, how customers behave, where risk lives, what pricing decisions produced good outcomes and what patterns only become visible after years of repetition.

A company’s edge is rarely one giant secret sitting in a vault. More often it is thousands of small decisions, tiny operational habits, customer relationships, historic exceptions and accumulated scar tissue. Put enough of that through an external system and the danger is not that someone steals the whole vault overnight. The danger is that the moat slowly turns into a public road.

Karp’s line that “controlling your weights is controlling your fate” is intentionally dramatic, but not entirely wrong.

Weights are where the learning lives. They are the compressed residue of data, training, fine-tuning and repeated interaction. If a company gives up control of the intelligence layer, it may eventually find itself renting back part of the competitive advantage it helped create.

That is a difficult proposition for any serious enterprise. It is even more difficult for governments, defence organisations and critical infrastructure operators.

You would not outsource the command room of a battleship to whichever vendor has the most polished sales deck that quarter. You would not let a third party own the map, the radar, the radio and the operating manual, then charge you by the message every time a storm appeared on the horizon.

Yet that is not far from the question Karp is raising around national security. If AI becomes embedded in intelligence, logistics, battlefield decisions, cyber defence and critical systems, then control over the data, models and deployment architecture is not a procurement detail. It becomes part of national capacity.

This is also why the growing interest in Chinese open-weight models is more than a curiosity.

The shift is not necessarily a declaration that Chinese models are better across the board. The frontier US labs still lead in many areas, particularly at the cutting edge of reasoning, coding and multimodal capability. But enterprises are beginning to behave like rational buyers. They are comparing performance, cost, reliability, deployment flexibility and the ability to keep the system close to home.

For some use cases, the most advanced model in the world is not the most useful model in the building.

A cheaper open-weight model that can be hosted internally, tuned around proprietary data and controlled by the enterprise may deliver a better economic outcome than a brilliant frontier model accessed through an expensive metered pipe. That does not mean the premium model loses. It means the market starts asking the question it always asks eventually: what am I getting for the price?

That is where the AI boom is beginning to change character.

The first phase was awe. Look what these models can do.

The second phase was fear. What happens if we fall behind.

The next phase is audit. Who owns the system, who owns the data, who owns the weights, who owns the customer relationship and who captures the margin.

This is the part of the cycle where the slogans get tested against the spreadsheets.

Palantir’s response is to push a sovereign deployment model with Nvidia, where the customer retains control over compute, models, data and weights rather than simply renting intelligence through a frontier API. That is not merely a technical architecture. It is a different answer to the value-capture question.

The frontier labs want to become the intelligence layer of the global economy. Palantir is arguing that no serious institution should hand over the keys quite so easily.

Both sides have a point.

The frontier labs have created products with genuine, extraordinary capability. They are not selling smoke. But capability alone does not settle the economics. A model can be brilliant and still be too expensive. It can be powerful and still be too externally controlled. It can save time for a department while quietly transferring long-term value away from the enterprise.

That is the uncomfortable part of the story.

The real AI race may not be between OpenAI, Anthropic, Google, Meta, DeepSeek and the rest. It may be between companies that use AI to compound their own institutional intelligence and companies that use AI to become more dependent on someone else’s.

The difference may not show up in the first quarter.

It may only become obvious years later, when one company owns the factory and the other is still feeding coins into the machine.

Karp had warned against underestimating China’s progress; these examples illustrate the trend in real time.