Palantir’s Karp Says Businesses Are ‘Unhappy’ with Frontier AI Labs, Here’s Why

Alex Karp, the CEO of Palantir, is not known for subtlety. And in June 2026, he delivered one of his most unflinching critiques yet.

“It’s not just the man and woman on the street that is unhappy with the frontier labs,” Karp told CNBC‘s Sara Eisen. “It’s in private every single enterprise we deal with.”

He went further on a podcast appearance, comparing the way companies consume AI tokens to… well, let’s just say adult content addiction. “People are just, like, sitting there all day … enterprises are, like, ‘Okay we believe this will create value, but we cannot have people just checking the weather with it, just rearranging deck chairs on their personal Titanic,’” Karp said.

That‘s the headline. But behind the provocative language is a serious indictment: frontier AI labs, OpenAI, Anthropic, Google, are selling enterprises a dream they can’t deliver on. And the bill is coming due.

Let’s unpack what Karp actually said, why enterprises are so fed up, and what smart companies are doing differently.

---

What Did Alex Karp Actually Say?

Karp‘s critique lands on two main axes: tokenmaxxing and misaligned incentives.

The “Tokenmaxxing” Quote That Broke the Internet

Tokenmaxxing is the compulsive over‑consumption of AI tokens, you pay per prompt, per generated word, per API call, without any meaningful business outcome. Companies buy into the hype, pump tokens through frontier models, and mistake activity for productivity.

Karp’s analogy was deliberately jarring: “People are just sitting there all day like a porn addiction. Enterprises are like, okay we believe this will create value but … it is literally like porn. People are full‑on addicted.”

He even claimed Palantir has built an internal tool to wean companies off tokenmaxxing. “We have a product, eternally called something, but internally we call it the ‘demasturbatory get off masturbation’ thing.”

Funny? Yes. Absurd? Also yes. But underneath the shock value is a deadly serious observation: most enterprises are burning token budgets with no measurable return.

“The Frontier Labs Are Charismatic with Investors, Not with Enterprises”

Karp drew a sharp distinction between how frontier labs present themselves to Wall Street versus how they serve actual paying customers.

“The large language model frontier companies are super charismatic with investors,” he said. “I‘ll give you some news: they’re super not charismatic with enterprises.”

In private conversations, Karp claims, enterprise customers consistently say these labs don‘t understand their businesses and only care about one thing: tokenmaxxing — burning through tokens to signal productivity metrics that don’t translate to real value.

His advice to prospective Palantir customers is almost sarcastic: “Don‘t even call. Don’t come talk to us. There‘s a frontier company. Go spend two days with them. And if you’re lucky, after you‘re done, I’ll let you in my door. They‘re like clamoring.”

He’s confident they‘ll come back disappointed. And the data suggests he’s right.

---

Why Enterprises Are Revolting Against Frontier AI Labs

Karp identified three structural reasons businesses are so unhappy. Let‘s break each one down.

Reason #1: Tokenmaxxing Burns Budgets Without Outcomes

This is the core of the problem. Enterprises pay per token, and those tokens add up fast. Frontier labs have every financial incentive to maximise token usage, not to optimise for business value.

Think of it like a hospital that gets paid per X‑ray ordered, not per patient cured. The incentives are completely misaligned.

Cognizant CEO Ravi Kumar put it bluntly in June 2026: “There‘s been a sense of FOMO, fear mongering and that has led to token consumption without linkages to ROI and without linkages to outcomes.”

Microsoft has reportedly told employees to wind down usage of Claude Code. Uber limited spending on AI‑powered coding tools to manage ballooning costs. The pattern is everywhere: companies sign up, burn through their annual AI budget in weeks, and have nothing to show for it.

Reason #2: Frontier Models Lack Business Context

A frontier LLM knows a lot about the internet. It knows almost nothing about your business, your data models, your compliance rules, your unique workflows.

Karp‘s framing is succinct: “It is not that large language models aren‘t crucial for the world. It’s just the implementation is where the value is, certainly in the next seven years.”

In other words: the model alone is worthless. The real work is integrating it into an actual business environment with governance, security, and use‑specific logic. That‘s not what frontier labs sell. They sell tokens.

Reason #3: No Governance, No Guardrails, Just Tokens

For any organisation in a regulated industry, finance, healthcare, defence, energy, running a frontier model out of the box is a non‑starter. You need audit trails. You need role‑based access. You need to know that the model isn’t hallucinating a compliance violation.

Frontier labs are only beginning to build those enterprise features. Palantir, by contrast, built its entire platform around them. That‘s not an accident, it’s a fundamental difference in go‑to‑market strategy.

---

The Economics of Tokenmaxxing (A Quick Reality Check)

Let‘s put numbers on this, because the scale is genuinely shocking.

The bottom line: enterprises are pouring trillions into AI, more than half see no ROI, and the majority of pilots never make it to production. That‘s not a healthy market. That’s a bubble inflated by token sales.

---

Palantir‘s Answer: Implementation Over Model Worship

So what‘s the alternative?

Palantir doesn’t sell a frontier LLM. It sells an Artificial Intelligence Platform (AIP) that integrates with any model, OpenAI‘s, Anthropic’s, Google‘s, open‑source, and wraps it in what Karp calls a “no‑slop zone” of ontology, governance, and operational logic.

And the market is responding. In Q1 2026, Palantir reported:

• $1.63 billion in revenue, up 85% year‑over‑year
• U.S. commercial revenue up 133% to $595 million
• Full‑year guidance raised to ~$7.66 billion (71% growth)

Those aren‘t hype numbers. Those are enterprise customers voting with their wallets for a platform that actually solves problems, not one that just sells tokens.

Karp himself said that most of Anthropic’s public‑discussed projects are “running on Palantir.” In other words: even the frontier labs depend on Palantir‘s implementation layer to make their models usable in the real world.

---

This Isn’t Just Karp, The Industry Is Waking Up

Maybe you think Karp is biased (fair, he‘s selling a competing platform). But here‘s where it gets interesting: other industry leaders are saying the same thing.

Cognizant CEO: “Reckless Token Consumption”

Ravi Kumar S of Cognizant didn’t mince words: “There is a gap between AI capability and production value because of the reckless token consumption without linkage to outcomes. Companies need to create more efficient, more predictable and effective economics for token consumption.”

56% of Enterprises See No Significant ROI

A PwC survey of more than 4,450 CEOs across 95 countries found that 56% have yet to see any significant financial benefit from their adoption of AI.

Let that sink in. More than half of the world‘s largest companies are investing billions and getting… nothing measurable in return.

Even OpenAI’s IPO Filing Admits Enterprise Gaps

In June 2026, OpenAI confidentially filed for an IPO, a week after Anthropic did the same. But as one short‑seller analysis noted, corporate buyers are struggling to find ROI, and cheaper open‑source alternatives often perform just as well for routine tasks.

If the frontier labs‘ own customers can‘t justify the spend, that‘s not a sustainable business model. That’s a waiting game until the music stops.

What Smart Enterprises Do Differently

So, if you‘re an enterprise leader staring down a seven‑figure AI budget and wondering where the value went, here’s what the data, and Karp‘s critique, suggests:

1. Start with the problem, not the model. Don’t ask “Which LLM should we buy?” Ask “Which business process can we materially improve with AI?” The model is a means, not an end.

2. Measure outcomes, not tokens. Tie every AI initiative to a specific KPI, cost reduction, revenue lift, cycle time compression. If you can‘t draw a line from token spend to business value, you‘re tokenmaxxing.

3. Look for implementation platforms, not model providers. The frontier labs are great at building models. They are not great at integrating those models into complex enterprise environments with governance, security, and workflow logic. That‘s where platforms like Palantir’s AIP (and others) come in.

4. Treat token consumption as an architectural problem. Right‑size your models. Use smaller models for routine tasks. Cache results. Route queries intelligently. Reducing token waste is not a technical detail, it’s a business imperative.

5. Don‘t conflate activity with progress. As Karp put it, one more dashboard, one more query “can’t hurt that much”, except when you‘ve burned your entire annual budget on “testing.”

---

Frequently Asked Questions

What is tokenmaxxing?

Tokenmaxxing is the compulsive over‑consumption of AI tokens (pay‑per‑use API calls) without delivering measurable business outcomes, often driven by misaligned incentives from frontier AI labs.

Which frontier AI labs is Alex Karp criticising?

Karp’s critique broadly targets OpenAI, Anthropic, Google, and any lab that prioritises token volume over enterprise value. He specifically called out Anthropic‘s public projects running on Palantir as proof that models alone are insufficient.

Is Palantir’s AIP a replacement for frontier LLMs?

No. AIP is an integration and orchestration platform that can work with any LLM, including frontier models. The key difference is that AIP adds ontology, governance, security, and workflow logic to make AI production‑ready in complex enterprises.

What percentage of enterprises see real ROI from AI?

PwC found that 56% of surveyed CEOs reported no significant financial benefit from their AI adoption as of 2026. Meanwhile, successful deployments (e.g., code generation, customer support) show strong but narrow ROI.

How can my company avoid tokenmaxxing?

Focus on outcomes, not tokens. Right‑size models for each task, implement usage guardrails, tie AI spend to specific KPIs, and consider implementation platforms that optimise for business value rather than token volume.