Finally, a company that uses AI found its business was taken over by large model companies.

Author: Yuhangyuan; Source: Geek Park

On July 1, Palantir CEO Alex Karp walked into the CNBC studio and hurled a bombshell in a near-uncontrolled tone.

He said the AI industry is “effing insane” (crazy). He said U.S. corporate CEOs are “livid” (furious) about OpenAI and Anthropic. He said companies are doing something absurd—paying wildly for tokens while handing their most core operational data over to model vendors. Yet the commercial value returned is almost impossible to measure.

The host asked whether this was “passing the buck.” Karp replied, “No, I’m just stating the facts.”

Palantir’s stock price rose 9% that day. The number itself is a kind of vote—the market believes he said a lot of what many people want to say but haven’t been able to say.

This isn’t just one person venting. When the leader of a company with a market cap of more than one trillion dollars fires at the entire large-model industry on live national television—and the market responds with real money—it means a collective emotional state has reached a critical point.

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

01From “getting carried away” to “not being naïve”

Looking back to early 2024, corporate attitudes toward large models could be summarized in four Chinese characters: “use it first.”

No matter the ROI or not, no matter where the data flows—just don’t fall behind. At that time, the mainstream narrative was “the AI revolution is here; if you don’t embrace it, you’ll be eliminated.” CIOs and CTOs across industries were under enormous pressure, shoving AI into every business process they could shove it into. This is a typical decision driven by tech panic.

By 2025, “full rollout” became the keyword. Companies began seriously embedding large models into core business workflows, no longer just doing demos or running internal hackathons. From customer service to code generation, from market analysis to product design, the depth and breadth of AI penetration expanded exponentially.

But entering 2026, a subtle shift in sentiment is taking place.

Salesforce’s research shows that only half of IT leaders are confident that their company’s data infrastructure can support successful AI deployment. A research report released by NTT DATA in May this year directly used the term “hitting a wall”—enterprise AI is encountering architectural bottlenecks caused by data privacy and sovereignty requirements. Gartner predicts that by 2027, 35% of countries will rely on regionalized AI platforms, while that figure is only 5% today.

Karp put this shift even more plainly. He said enterprises are moving away from “tokenmaxxing”—mindlessly consuming tokens—and toward truly asking about return on investment. “The basic viewpoint is, don’t waste time on tokens anymore.”

This isn’t denying large models; it’s that the entire industry is moving from “getting carried away” to “not being naïve.” After the fever passes, companies start looking at a fundamental question with a calmer perspective: What I give away, and what I get back—does the accounting actually add up?

02When partners become competitors

Karp’s criticism still stays at the business model level. But what truly chills people is another, more direct threat: your AI service provider may be using the data and scenario understanding you contribute to build a product that replaces yours.

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,” integrating Claude-generated code seamlessly into Figma’s design workflow. The two companies looked like close partners.

On April 14, Anthropic Chief Product Officer Mike Krieger quietly stepped down from Figma’s board of directors.

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

That day, Figma’s stock price fell by nearly 8%.

A later report by Fast Company included a telling detail: Figma and companies like Adobe and Canva have had multi-year partnership relationships with Anthropic, but before Claude Design was released, no one was informed. Everyone realized—caught off guard—that their AI partner, right under their noses, had turned into a competitor.

This story is worth deep thought because it exposes a structural problem in the large-model era—more dangerous than ever before: when you deeply collaborate with an AI company, you don’t just hand over a market entry point; you also hand over your core scenario understanding and user needs data.

Anthropic’s ability to build Claude Design is largely due to how, through its partnerships with design-tool companies, it has deeply understood designers’ workflows and pain points.

But if you widen the lens, this isn’t a new script in the history of technology.

Amazon built its own brands from its e-commerce platform—using platform data to precisely identify the most profitable categories, then launching its own products to eat into third-party sellers. Microsoft started from the operating system, absorbing browsers, office software, and communications tools one by one—Netscape was killed, and Slack was forced to sell itself. Google extended from search engines, answering user questions directly on search-results pages—pushing Yelp and many vertical information service providers to the margins.

The iron law of the tech industry has never changed: once a platform has enough data and user understanding, it will erode upstream.

In the large-model era, this law becomes even more aggressive. Traditional platform erosion still needs time to accumulate understanding, but large models are naturally an “understanding accelerator.” Every API call and every input of business data helps the model vendor understand your territory faster and deeper.

03The “Roche limit” of the AI era

In astronomy, there’s a concept called the Roche limit: when a celestial body gets too close to a massive star, tidal forces exceed its own gravitational pull, and the celestial body will be torn apart.

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

Large models are that massive star. Every company wants to borrow its gravitational pull to accelerate efficiency, reduce costs, and innovate. But the problem is that when you get close enough, your “material” starts getting stripped away. Your data, know-how, and understanding of user needs all flow to the gravitational center during the collaboration process.

So where is the boundary for how a company can “dance with AI” without being ultimately consumed?

This question has already been put on the table in the United States. But if you think it’s still far away for Chinese enterprises, that may be an illusion.

There are differences between China and the U.S. in the pace of AI application. U.S. companies have already entered large-scale, deeply embedded AI deployment, while Chinese companies overall are still transitioning from pilots to scaling. Research published in March this year by Lenovo together with IDC shows that 72% of domestic enterprises have already completed agent pilots and put them into formal use, deploying AI across an average of 3.5 scenarios. But the focus of challenges has also shifted—from “lack of computing power, lack of data” to “application effects not meeting expectations” and “ROI is unclear.”

In other words, Chinese enterprises are entering a “AI sobriety period” similar to that of U.S. enterprises.

Geek Park recently spoke with a number of startups and companies with traditional businesses and found an interesting pattern: when people think about these issues, often it doesn’t come from a direct sense of crisis like “I’m worried the model company will steal my business.” Instead, after AI is truly embedded into operations, they naturally begin redefining “in the AI era, what is my core value?”

This redefinition ultimately comes down to two key capabilities.

04Who controls the “AI foundation”?

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

Karp repeatedly emphasized this on CNBC. A company’s most sensitive operational data should not flow into a third-party model vendor’s black box. He positions Palantir as an application layer for “sovereign AI”—models can be someone else’s, but data must stay within your own perimeter, and deployment must run on infrastructure you can control.

This isn’t paranoia. In Chinese companies’ lived experience, it’s completely aligned. Huang Weijie, head of R&D and product at Kingsoft’s WPS 365, recently said a very apt line: “What enterprises lack today isn’t hardware and models, but a secure AI application layer.”

IDC data also corroborates this trend. In enterprise AI compute deployments, the share of public cloud is declining; the combined share of private cloud and on-prem deployment has risen from 54% to 69%. “Data must not leave the domain” is moving from being a compliance slogan to becoming the first screening condition when CTOs choose solutions.

Karp calls this “commodity cognition.” His judgment is that the quality of the models themselves is converging; the real differentiated value isn’t in the model layer, but in the application layer that binds model capabilities to enterprise-specific scenarios. Palantir’s “sovereign AI engine,” developed in partnership with NVIDIA, is a productization of this logic: using open-source models together with Palantir’s own ontology layer and governance framework, enabling enterprises to run AI in a fully controllable environment, with not a single byte of data leaving. Palantir’s revenue in the first quarter of 2026 was $1.63 billion, up 85% year over year—somewhat like the market casting a vote for this path.

There’s a signal here worth watching: in the future, companies and solutions that help enterprises run AI on “their own foundation” will be even more in demand. In China, “AI privatization brains” has already become a real track. Many startups are building products around this direction. This isn’t a tech-purity fetish; it’s a rational choice companies make after thinking it through.

05Don’t turn the organization into a “repeater”

The second capability is harder to quantify, but Geek Park increasingly feels it when talking with enterprises: when AI can replace more and more execution steps, what kind of “people” does an organization actually need?

Some faster-moving companies have already fallen into this trap.

When AI’s efficiency clearly exceeds that of humans in certain steps, the natural thought is to “cut people.” But after the organization becomes thinner, a hidden problem emerges: what AI runs is essentially the “best practices” condensed by those people in the old environment. When the environment changes, the market changes, and users change, AI continues faithfully executing the old logic—while the organization no longer has enough people to sense these changes and push the business forward to evolve.

Put simply, an organization filled with AI but emptied by people may only be efficiently repeating the past.

This doesn’t mean you shouldn’t use AI to replace execution. It means that as AI takes over more execution layers, enterprises actually need another kind of person: not the people who execute specific tasks in the traditional sense, but those who can “command” AI. This role requires understanding the business globally, being able to judge whether what AI outputs still applies to the changing reality, and being able to see new possibilities beyond the “optimal solutions” AI offers.

Some leading companies have already started taking this seriously. They find that with AI, real competitiveness isn’t about “how many people you replaced with AI,” but whether “your people can harness AI to do things that were previously impossible.” If you merely keep letting AI automate continuously from historical data and keep looping, then fundamentally you’re locked into a snapshot of the past.

The importance of this cognitive flip may be no less than data sovereignty. When AI flattens technical barriers, “human judgment” and “organizational evolution capability” become the hardest things to replicate. Some companies have already realized this; some haven’t. But this tipping point may become very clear in the next one or two years.

06The industry needs “new AI companies”

Over the past two years, an implicit assumption has dominated the entire industry: in the AI era, value will ultimately concentrate in model companies. The closer you are to the model, the higher the value.

That assumption is being shaken.

Karp actually points out something on CNBC: the model itself is becoming commoditized cognition. As the capability gaps among different large models continue to narrow, real differentiation won’t be in the model layer anymore. An industry structure where only model companies dominate isn’t just unhealthy for enterprises—it also constrains the overall pace of development across the AI industry.

Enterprises have never needed a stronger model. What they need is an entire ecosystem—one that can respond to anxiety about data sovereignty, protect competitive moats from being “siphoned,” and allow AI to truly embed into the business without going out of control. This demand is creating a market far more complex than just “selling tokens.”

Several directions already show clear signals.

“Sovereign AI infrastructure” is becoming a real, well-funded track. This isn’t a concept. In just the first half of 2026, Europe had three companies building sovereign AI infrastructure (Nebius, nScale, AtlasEsge), which together raised more than $11.8 billion. Just a few days ago, Valarian in London just secured a $50 million Series A to do something very concrete—adding a “sovereign control layer” between AI systems and sensitive data, determining which AI can access which data and under what conditions. Two years ago, this kind of thing didn’t exist as a demand at all; now governments and large enterprises are lining up to get it.

“AI gateways” and orchestration middle layers are becoming an indispensable part of enterprise AI architecture. When an enterprise uses OpenAI, Anthropic, open-source models, and also its own fine-tuned dedicated models, who provides unified routing, cost control, permission governance, and auditing? In the traditional software era, this position is called middleware; in the AI era, it’s called a gateway or orchestration layer. It’s not glamorous, but it’s the key infrastructure for enterprises to move from “using AI” to “governing AI.” Palantir essentially does this layer—just in the heaviest version. More lightweight solutions for enterprises of different sizes have huge room.

At the application layer, AI solutions for vertical industries are also moving from “wrapping shells” to “going deeper.” In the past, many so-called AI applications were, in essence, just a shell wrapped around GPT. But now, what can truly stand on its own are products that deeply understand specific industry know-how and tightly bind AI capabilities with industry logic. The value anchor of these companies isn’t the model—it’s industry understanding, which is exactly the kind of thing large-model companies find difficult to obtain through training.

Even at the “people” level, a new services market is emerging. As more enterprises realize they don’t need more AI tools, but people who can “command AI” and the supporting organizational methodology, demand for organizational transformation consulting in the AI era, talent development, and process redesign is also rapidly rising.

At the end of the day, an industry with only a “model layer” is fragile. What truly allows the AI industry to run faster and healthier is a more three-dimensional ecosystem. In this ecosystem, some build models, some build sovereign infrastructure, some build gateways and governance, some build deep vertical applications, and some help enterprises reshape organizational capabilities. Each layer addresses real needs as enterprises move from “embracing” to “harnessing” AI.

Over the past year, these needs have become clearer and clearer from vague. Next, the next generation of solutions, service providers, and products that emerge around these needs may see a clear wave of breakthroughs.

Returning to the Roche limit metaphor: finding that safe trajectory is never something one enterprise can do alone. When the entire ecosystem starts growing forces beyond models, enterprises truly gain the confidence not to be torn apart.