Companies that finally adopted AI found their businesses snatched by large model companies.

On July 1, Palantir CEO Alex Karp walked into the CNBC studio and dropped a bombshell in a nearly out-of-control tone.

He said the AI industry is "effing insane". He said American corporate CEOs are "livid" at OpenAI and Anthropic. He said companies are doing something absurd — frantically paying for tokens while handing over their most critical operational data to model vendors. And the commercial value in return is almost immeasurable.

The host asked him if he was "passing the buck". Karp replied: "No, I'm just stating the facts."

Palantir's share price rose 9% that day. This number itself is a vote — the market believes he said what many people wanted to say but never dared to say.

This is not a personal emotional outburst. When the head of a company with a market value exceeding 100 billion USD fires at the entire large model industry on a national live broadcast, and the market gives positive feedback with real money, it means a collective emotion has reached a critical point.

Over the past two years, everyone has been talking about how to embrace large models. But now, a new problem is emerging — if a company gets too close to large models, will it be torn apart?

01 From "Infatuation" to "Disillusionment"

Looking back at the beginning of 2024, enterprises' attitude toward large models can be summed up in four words — "Use it first".

Never mind the ROI, never mind where the data flows, the bottom line is not to fall behind. The mainstream narrative at that time was "The AI revolution is coming, and you will be eliminated if you don't embrace it." CIOs and CTOs from all walks of life, under enormous pressure, integrated AI into every possible business link. This is a typical decision driven by technology panic.

By 2025, "full-scale rollout" has become the keyword. Enterprises began to seriously embed large models into core business processes, no longer just making demos or holding internal hackathons. From customer service to code generation, from market analysis to product design, the depth and breadth of AI penetration are expanding exponentially.

But entering 2026, a subtle emotional shift is taking place.

Survey data from Salesforce shows that only half of IT leaders are confident that their company's data infrastructure can support the successful implementation of AI. A research report released by NTT DATA in May this year directly used the term "hitting a wall" — enterprise AI is encountering structural bottlenecks brought about by data privacy and sovereignty requirements. Gartner predicts that by 2027, 35% of countries will rely on regionalized AI platforms, compared with only 5% today.

Karp put it more bluntly. He said enterprises are shifting from mindless token consumption "tokenmaxxing" to truly questioning return on investment. "The basic point is, stop wasting time on tokens."

This is not to deny large models, but the entire industry is moving from "infatuation" to "disillusionment". After the frenzy period, enterprises began to examine a fundamental issue with a more calm perspective — can I balance what I give away and what I get back?

02 When Partners Become Competitors

Karp's criticism remains at the business model level. But what really sends chills down the spine is another more direct threat — your AI service provider may be using the data and scenario understanding you contribute to build products that replace you.

What happened in April 2026 turned this concern from theory into reality.

In February this year, Figma and Anthropic were still jointly developing a feature called "Code to Canvas", which seamlessly integrates Claude-generated code into Figma's design workflow. The two companies seemed to be close partners.

On April 14, Anthropic's Chief Product Officer Mike Krieger quietly resigned from Figma's board seat.

Three days later, Anthropic released Claude Design — an AI design tool that can directly generate interactive prototypes, PPTs and marketing materials using natural language, precisely targeting Figma's core business.

Figma's share price fell nearly 8% that day.

A detail in Fast Company's subsequent report is thought-provoking — Figma, Adobe, Canva and other companies have had multi-year partnerships with Anthropic, but no one was notified before Claude Design was released. Everyone realized in astonishment that their AI partner had turned into a competitor right under their noses.

This story is worth pondering because it exposes a structural problem in the large model era that is more dangerous than ever — when you cooperate deeply with an AI company, you not only hand over the market entry, but also your core scenario understanding and user demand data.

The reason Anthropic can build Claude Design is largely because it gained deep insight into designers' workflows and pain points through cooperation with design tool companies.

But if we broaden our horizons, this is not a new plot in the history of technology.

Amazon developed its own private labels on its e-commerce platform, using platform data to accurately identify the most profitable categories, and then launched its own products to erode third-party sellers' market share. Starting from its operating system, Microsoft took over browsers, office software, and communication tools one by one — Netscape was killed, Slack was forced to sell itself. Google extended from its search engine, using search results pages to directly answer users' questions, marginalizing Yelp and a large number of vertical information service providers.

The iron law of the technology industry has never changed — once a platform has sufficient data and user understanding, it will erode upstream.

In the large model era, this iron law has become more ferocious, because traditional platform erosion still takes time to accumulate understanding, while large models are naturally an "understanding accelerator". Every API call you make and every input of business data helps model vendors understand your territory faster and more deeply.

03 The "Roche Limit" in the AI Era

In astronomy, there is a concept called the "Roche Limit" — when a celestial body gets too close to a massive star, the tidal force will exceed its own gravity, and the celestial body will be torn apart.

This metaphor describes the relationship between enterprises and large models today with unsettling precision.

The large model is that massive star. Every enterprise wants to use its gravity to accelerate — improve efficiency, reduce costs, and drive innovation. But the problem is, when you get close enough, your "matter" will start to be stripped away. Your data, know-how, and understanding of user needs will all flow to the gravitational center during the cooperation process.

Where is the boundary where a company can "dance with AI" without being eventually devoured?

This issue has been put on the table in the United States. But if you think it is still far from Chinese enterprises, that may be an illusion.

There are differences in the pace of AI application between Chinese and American enterprises. American enterprises have entered the stage of large-scale and in-depth AI deployment in business, while Chinese enterprises as a whole are still moving from pilot to large-scale implementation. A survey jointly released by Lenovo and IDC in March this year shows that 72% of domestic enterprises have completed agent pilots and put them into formal use, deploying AI in an average of 3.5 scenarios. But the focus of challenges has shifted from "lack of computing power and data" to "unmet application effect" and "unclear ROI".

In other words, Chinese enterprises are entering an "AI sober period" similar to that of American enterprises.

GeekPark recently communicated with many entrepreneurs and traditional business enterprises, and found an interesting phenomenon — people's thinking on these issues often does not come from a direct sense of crisis like "worrying that model companies will steal my business", but after truly integrating AI into their business, they naturally begin to redefine " what is my core value in the AI era".

This redefinition will eventually fall on two key capabilities.

04 Who Controls the "AI Foundation"?

The first and most realistic one is highly consistent with what Karp said — on whose foundation are your data and business logic actually running?

This is the core argument Karp repeatedly emphasized on CNBC. The most sensitive operational data of enterprises should not flow into the black box of third-party model vendors. He positions Palantir as an application layer that provides "sovereign AI" — the model can be from others, but the data must stay within one's own walls, and deployment must be on one's own controllable infrastructure.

This is not paranoia, and Chinese enterprises have exactly the same feelings. Huang Weijie, head of product and R&D of Kingsoft Office WPS 365, said a very insightful sentence not long ago — "What enterprises lack today is not hardware and models, but a secure AI application layer."

IDC data also confirms this trend: among enterprise AI computing power deployments, the proportion of public clouds is declining, and the total proportion of private clouds and on-premises deployments has increased from 54% to 69%. "Data does not leave the domain" is evolving from a compliance slogan to the first screening criterion for CTOs when making selection decisions.

Karp calls this "commodity cognition". His judgment is that the quality of models themselves is converging, and the real differentiated value does not lie in the model layer, but in the application layer that binds model capabilities to the specific scenarios of enterprises. Palantir's "Sovereign AI Engine" launched in cooperation with NVIDIA is the productization of this logic — using open-source models plus Palantir's own ontology layer and governance framework, allowing enterprises to run AI in a fully controllable environment, with no data leaving in any byte. Palantir's revenue in the first quarter of 2026 was 1.63 billion USD, a year-on-year increase of 85%, which to some extent represents the market's vote of confidence in this path.

There is a notable signal here — in the future, companies and solutions that help enterprises run AI "on their own foundations" will become more popular. In China, the "privatized AI brain" has become a real track, and many startups are developing products around this direction. This is not technical paranoia, but a rational choice made by enterprises after thorough consideration.

05 Don't Turn the Organization into a "Repeat Machine"

The second capability is more difficult to quantify, but GeekPark feels it more and more strongly in exchanges with enterprises — when AI can replace more and more execution links, what kind of "people" does the organization still need?

Some fast-moving enterprises have already stepped into this pit.

When AI's efficiency in certain links significantly exceeds that of humans, a natural idea is to "cut headcount". But after the organization becomes leaner, a hidden problem begins to emerge — the set of rules that AI runs is essentially the "best practices" condensed by these people in the old environment in the past. When the environment, market and users change, AI is still faithfully executing the old logic, and there are not enough people in the organization to perceive these changes and promote the evolution of the business.

To put it bluntly, an organization filled with AI but emptied of people is likely to just efficiently repeat the past.

This does not mean that we should not use AI to replace execution work. Instead, when AI takes over more and more execution layers, enterprises actually need another type of people — not those who perform specific tasks in the traditional sense, but those who can "command" AI. This role requires understanding the overall business, being able to judge whether AI's output is still applicable to the changing reality, and seeing new possibilities beyond the "optimal solution" given by AI.

Some leading enterprises have begun to seriously consider this issue. They found that after having AI, the real competitiveness is not "how many people have been replaced by AI", but "whether your people can leverage AI to do things that were impossible before". If you just let AI continue to automate and cycle in historical data, you are essentially locked in a snapshot of the past.

The importance of this cognitive shift may be no less than data sovereignty. When AI levels the playing field of technical barriers, "human judgment" and "organizational evolution capability" have instead become the most difficult things to replicate. Some companies have realized this, while others have not. But this watershed will likely become very clear in the next year or two.

06 The Industry Needs "New AI Companies"

Over the past two years, an implicit assumption has dominated the entire industry — the value of the AI era will eventually be concentrated in the hands of model companies. The closer you are to the model, the higher the value.

This assumption is being shaken.

Karp actually pointed out one thing on CNBC — models themselves are becoming commodity cognition. When the capability gap between different large models gets smaller and smaller, the real differentiation no longer lies in the model layer. An industrial structure where only model companies dominate is not only unhealthy for enterprises, but also a constraint on the development speed of the entire AI industry.

What enterprises need is never a stronger model. What they need is a complete ecosystem — one that can address the anxiety of data sovereignty, protect competitive barriers from being "siphoned off", and allow AI to truly embed into business without losing control. This demand is spawning a far more complex market than "selling tokens".

Several directions have shown clear signals.

"Sovereign AI Infrastructure" is becoming a real, well-funded track. This is not a concept. In the first half of 2026 alone, three European companies working on sovereign AI infrastructure (Nebius, nScale, AtlasEsge) raised a total of more than 11.8 billion USD. Just a few days ago, London-based Valarian secured 50 million USD in Series A funding, with a very specific mission — adding a "sovereign control layer" between AI systems and sensitive data to determine which AI can access which data and under what conditions. Such a demand did not exist two years ago, but now governments and large enterprises are queuing up for it.

"AI Gateway" and orchestration middleware are becoming an indispensable part of enterprise AI architecture. When an enterprise is using OpenAI, Anthropic, open-source models, and its own fine-tuned dedicated models at the same time, who will handle unified routing, cost control, permission governance and auditing? This position was called middleware in the traditional software era, and gateway or orchestration layer in the AI era. It is not glamorous, but it is the key infrastructure for enterprises to move from "using AI" to "managing AI well". Essentially, Palantir is building this layer, but it has developed the most heavyweight version. There is huge space for lighter-weight solutions tailored for enterprises of different scales.

At the application layer, AI solutions for vertical industries are also evolving from "shallow wrapping" to "deep integration". In the past, many so-called AI applications were essentially just a shell wrapped around GPT. But now, the products that can truly gain a foothold are those that deeply understand the specific industry know-how and tightly bind AI capabilities to industry logic. The value anchor of such companies does not lie in the model, but in industry insight — which is precisely what large model companies can hardly obtain through training.

Even at the "human" level, a new service market is emerging. As more and more enterprises realize that what they need is not more AI tools, but people who can "command AI" and supporting organizational methodologies, demands for organizational transformation consulting, talent training and process redesign in the AI era are also rapidly emerging.

In the final analysis, an industry with only a "model layer" is fragile. What can truly make the AI industry develop faster and healthier is a more three-dimensional ecosystem. In this ecosystem, some build models, some build sovereign infrastructure, some build gateways and governance tools, some develop deep applications for vertical industries, and some help enterprises reshape organizational capabilities. Each layer responds to the real needs of enterprises in the process from "embracing" to "mastering" AI.

These demands have become increasingly clear from ambiguity over the past year. Next, a new generation of solutions, service providers and products born around these demands may usher in a clear boom period.

Going back to the metaphor of the Roche Limit. Finding a safe orbit is never the sole responsibility of an individual enterprise. When the entire ecosystem begins to grow capabilities beyond models, enterprises will truly have the confidence not to be torn apart.

This article is from the WeChat public account "GeekPark" (ID: geekpark), written by Yuhangyuan, authorized by