Over the past couple of years, the word "AI" has become like a broken record, heard at least once almost every day, often followed by a wave of anxiety.
What has happened amid all the FOMO and paranoia is that users have begun sharing virtually everything deemed "confidential" under the sun in search of answers.
Businesses pay for intelligence, but for that to be useful, you need to present the AI model companies with proprietary data, workflows, and corrections that give them a competitive edge.
The buyer is essentially giving up their knowledge simply to make use of what they have purchased.
Nadella's concern is that companies ultimately pay twice, once in cash and again with institutional know-how over time.
Satya Nadella says companies may be paying for AI twice
Microsoft CEO Satya Nadella argued that the visible cost of AI might just be the beginning.
"You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful," Nadella wrote in a recent blog post.
For AI systems to perform better, there needs to be higher-quality internal context, which likely includes employee prompts, operational procedures, agentic activity, and corrections.
"Models learn 'from exhaust,' the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong," Nadella said. "Every correction is distilled into institutional know-how."
"I am paying for tokens that create no value," Karp said in his most recent appearance on CNBC's "Squawk Box," describing the frustration he hears from enterprise customers. "These people are stealing the weights and alpha of my business."
Additionally, Karp also challenged the industry's basic pricing model: "If I can make you $1 billion tomorrow, wouldn't I say I'll make you $1 billion, and I want 30%? Why are they charging for tokens if it's so valuable?"
Nadella's version feels a lot less confrontational, but far more coherent, than Karp's. Still, the underlying warning remains the same.
Businesses are effectively renting models while donating the knowledge that makes them much more capable.
"In consuming intelligence, you are creating intelligence, and what you create should belong to you," as Nadella puts it.
Nadella's warning strengthens Palantir's core AI pitch
For Palantir (PLTR) stock investors, Nadella's warning is important and may have indirectly validated the problem Karp says Palantir was built to solve.
The CEO of the controversial tech firm Karp argued that enterprises should not expose their proprietary data, workflows, and operational knowledge directly to large language models outside their organizations.
Palantir's answer is Ontology, an application layer that connects models to company operations while controlling what models can access and retain.
Karp said Ontology makes AI "safe and useful and precise," preventing models from caching customer data, replicating the business, or transferring sensitive intellectual property.
He went a step further in his interview with podcaster Mathias Döpfner, saying businesses need an application layer that "protects your data from being essentially abused by large language model providers."
If customers become more wary of the data they give up, Palantir could be in line for a massive long-term windfall, but it could also create valuation risks elsewhere in the AI sector.
Palantir needs to prove Ontology can turn that strategic concern into durable contracts, expanding margins, and measurable customer returns.
It's worth mentioning that the stock is down 27% in the past six months and more than 26% year-to-date, according to Seeking Alpha data. Still, Palantir stock is changing hands at 88 times non-GAAP forward earnings, a steep premium, to say the least, compared to the sector median of around 25 times.
Nadella's warning raises the stakes for the AI trade
The interesting part is that the broader AI trade is already up against the uncomfortable question that Wall Street hasn't answered: Who will earn enough money to justify the extraordinary spending?
For perspective, Amazon, Microsoft, Alphabet, and Meta are projected to spend about $630 billion on data centers and AI chips in 2026 alone, according to Reuters, more than 4 times their 2023 guidance.
However, with recent developments, it seems the chickens are finally coming home to roost as the AI trade undergoes a shakeout.
Bank of America's latest survey found that 45% of fund managers view an AI bubble as the market's biggest tail risk, Reuters also reported. Yet investors remain heavily committed to the chip stock trade.
Moreover, several of Wall Street's most popular personalities have sounded alarms.
Ray Dalio says AI is "now in the early stages of a bubble," while Jeremy Grantham warns that "sooner or later, the bubble will burst."
"Big Short" investor Michael Burry has long been skeptical of the AI boom, calling semiconductor valuations "a pure form of overvaluation" and warning that the end may be near.
Nadella's argument adds to those vulnerabilities.
The reverse information paradox may lead customers to redirect spending toward private, model-agnostic systems, weighing on the biggest names in AI and calling their nosebleed valuations into question.
Facts Only
* Users share information in search of answers amid FOMO and paranoia.
* Businesses must provide proprietary data, workflows, and corrections for AI intelligence to be useful.
* Satya Nadella argues companies pay for intelligence twice: once in cash and again with institutional know-how.
* AI models learn from prompts, agent activity, and corrections made by users.
* Karp expressed frustration over paying for tokens when the value seems tied to stolen business alpha.
* Palantir's Ontology aims to connect models to operations while controlling access and retention of data.
* Ontology is designed to prevent models from caching customer data or replicating the business.
* Bank of America survey found 45% of fund managers view an AI bubble as the market's biggest tail risk.
* Ray Dalio suggests AI is in the early stages of a bubble, and Jeremy Grantham warns of a future burst.
* Amazon, Microsoft, Alphabet, and Meta are projected to spend approximately $630 billion on data centers and AI chips in 2026.
Executive Summary
The discourse surrounding AI adoption highlights a tension between the perceived value of AI intelligence and the cost associated with acquiring it. Leaders like Satya Nadella suggest that companies are paying for AI twice: once in monetary terms and again through the proprietary knowledge they must provide to make the models useful. This knowledge includes employee prompts, operational procedures, agentic activities, and necessary corrections, which serve as the "exhaust" from which models learn. This dynamic leads to concerns among some actors, like Karp, who feel that users are essentially forfeiting their proprietary knowledge in exchange for access to the purchased intelligence.
The argument extends to the pricing structure, with frustration expressed over charging for tokens when the value seems tied to the resulting output. Palantir's approach, exemplified by Ontology, attempts to address this by creating an application layer designed to secure and control proprietary data, aiming to prevent models from caching sensitive information or replicating business structures. This suggests a focus is shifting toward building protective frameworks around AI consumption to manage risk and ensure that generated intelligence remains an asset of the owner.
Furthermore, the broader market context involves significant expenditure on AI infrastructure, with major tech companies projecting massive spending on data centers and chips. While some experts express skepticism regarding the sustainability of current valuations—citing concerns about a potential bubble—investors remain heavily committed to the semiconductor trade. This juxtaposition suggests that while high-level economic activity continues, underlying vulnerabilities related to value justification are being publicly debated by various stakeholders.
Full Take
The narrative establishes a core tension between the commodification of intelligence via large language models and the retention of proprietary organizational knowledge. The mechanism described—where user interactions become the training ground for institutional know-how—suggests an asymmetric value transfer: users provide the necessary input, which generates superior outputs, and this resulting knowledge accrues to the system operators. This mirrors historical patterns where access to information is bundled with control over its use.
The reaction from figures like Nadella shifts the focus from simple cost accounting to intellectual property ownership within the AI ecosystem. The need for an application layer, like Palantir's Ontology, reflects a systemic distrust in the default architecture of LLM providers, signaling a necessary pivot toward decentralized control mechanisms over data governance rather than mere access provision. This transition is not just about security; it challenges the fundamental assumption that external models can be safely utilized without internal contextualization being extracted as capital.
The broader market sentiment, characterized by skepticism from established investors and warnings from prominent figures, suggests an underlying pattern of recognizing speculative risk in rapidly expanding technological sectors. The juxtaposition of massive capital deployment in hardware versus concerns over the sustainability of the investment thesis points toward a potential systemic misalignment where hype-driven valuations may obscure tangible returns or long-term structural risks related to data ownership and intellectual control. What is unstated is the trajectory: whether this shift toward private, model-agnostic systems will ultimately lead to more robust economic models or simply create new, fragmented arenas of competitive risk.
Bridge Questions: How can valuation models be adjusted to accurately account for the extraction and monetization of institutional know-how from AI training data? What specific regulatory frameworks are necessary to enforce the separation between purchased intelligence tokens and proprietary knowledge embedded within them? If customers migrate toward private systems, what is the long-term competitive viability for large model providers whose primary value proposition relies on generalized access?
Sentinel — Human
The article builds an argument by framing the AI investment debate around proprietary knowledge, using executive commentary to pivot toward specific business solutions and broader market instability.