Industry Research · AI Applications Series, Part 2. In the previous installment, we argued that most AI application layers are thin wrappers with no moat — real moats lie in data, workflows, and distribution. This piece examines a case that takes "workflow lock-in" to its extreme: Palantir, and the category it represents, the AI Operating System.
I. Palantir Is Doing the Exact Opposite of Wrapping
In the last piece, I laid out the wrapper trap — most AI applications are thin shells wrapped around a large model, easily copied and swallowed. What makes Palantir interesting is that it's doing almost the exact opposite.
A wrapper app is "thinly layering AI capability on top for the user." Palantir "thickly packs the client's entire operational reality into its own system, then lets AI work on top of it."
Concretely, Palantir's core isn't some AI feature — it's something called Ontology. The term sounds esoteric, but its meaning can be stated concretely — Ontology is a digital mapping of everything in a company's real world (factories, equipment, people, orders, processes, decisions) into an interconnected digital model, effectively creating a "digital twin" of the company's operations that software and AI can act on.
Example: in a manufacturing company, every machine, batch of raw materials, order, worker, and production line rule becomes a digital object with attributes and relationships inside the Ontology. So AI is no longer a "chatbot that answers questions" — it can make decisions, issue commands, and trigger actions directly on this digital twin — like detecting an impending line failure, automatically adjusting production scheduling, notifying personnel, and rerouting materials.
That's what "AI Operating System" means: AI stops being a sidecar analytics tool and becomes an "operational layer" embedded in the core of the enterprise, capable of actually running business operations. This is the qualitative shift from AI as a "feature" to AI as an "operating layer" — and it's the position Palantir wants to own.
II. Why Ontology Might Be the Deepest Moat at the Application Layer
Once you understand Ontology, you see why it's an extremely deep moat — it takes what I called in the previous piece the hardest of the three moats (proprietary data, lock-in, distribution) — workflow lock-in — to its extreme.
Think about it: once a company has modeled its entire operational reality into Palantir's Ontology — all its data, processes, and decision logic running on top — can it still switch?
Almost impossible. Switching means tearing down and rebuilding the entire digital twin of the business, rewriting all embedded processes and decision logic, and taking on the massive risk of operational disruption. This isn't "switching software" — it's "replacing the chassis of a car going 100 mph." When you become the "operating system" of a client's operations, you achieve the strongest lock-in software can offer: switching costs so high the client never even thinks about it.
This is exactly the highest level of moat I described in the moat piece: the ultimate form of moat isn't how good the product is, but how deeply you are woven into the client's fabric, becoming a part it cannot do without. Palantir's Ontology is an extreme example of this "embedded moat" in the AI era.
What's more, this moat has a self-reinforcing property: the longer a client runs on Ontology, the more data and processes accumulate, the deeper the lock-in, and the greater the value. Time is on Palantir's side — it's a "stickier the more you use it" flywheel. Palantir started in military/intelligence scenarios (modeling the chaotic reality of battlefields into actionable situational awareness), then replicated the same capability (AIP, the AI Platform) into the commercial domain — the logic is the same.
Financially, the "quality" of this moat shows: the company is rare in simultaneously achieving high growth + high profitability (I've written about its "Rule of 120" — growth rate plus profit margin at an extremely high level, far exceeding the software industry's conventional Rule of 40). A company with high growth, high profitability, and extremely deep lock-in has fundamentally impeccable basics.
III. But "Impeccable Company" ≠ "Worthwhile Stock"
If I stopped here praising Palantir, I'd be betraying my own discipline. So I must pivot to the more important question — as an investment, what's the problem?
Palantir is the most extreme example I've seen of the tension between "great company" and "possibly terrible investment." Its fundamentals are nearly flawless, but its valuation is likely one of the most expensive in all of software.
This brings us back to the iron law I've repeated — great companies and great investments are two different things (The Most Important Thing, Mr. Market). The return on a stock doesn't depend on "how good the company is" — it depends on "what price you paid for that goodness." When a stock's price already embeds the perfect assumptions of "moat never eroded, growth always high, margins always expanding," any hint of imperfection is enough to trigger a sharp correction — even if the company itself remains excellent.
Several real risks Palantir faces:
First, the valuation prices in too much perfection. Its price-to-sales and P/E ratios are at extreme highs for software. That means market expectations are fully stretched, leaving almost no margin for error. It doesn't need to get worse — it only needs to fall short of "as perfect as the narrative" to see a violent drawdown.
Second, reliance on government business and its uncertainty. A substantial part of Palantir's revenue comes from government/defense. This business has its moat (deep ) but also its uncertainties (budget cycles, political winds, project concentration). Commercialization (AIP's expansion into enterprises) is its second growth curve, but whether that curve can support its high valuation needs time to prove.
Third, will the "AI Operating System" category be diluted by incumbents and open-source? The Ontology moat is extremely deep for locked-in clients, but for winning new clients, Palantir competes against Microsoft, Amazon, and a growing array of new tools. As the practice of "modeling enterprise data into AI-actionable objects" becomes achievable through more (even open-source) approaches, whether Palantir can maintain its "category exclusivity" premium is an open question. The moat protects the installed base; high valuation demands sustained high-speed incremental growth — there's tension between the two.
Fourth, it carries a strong "narrative premium." Part of Palantir's stock price is premium for the sexy narrative of "AI Operating System," not pure cash flow. Narrative premiums act as a tailwind in good times, but they are the first to evaporate when sentiment turns.
IV. My Take: Separate the Category from the Price
So how should we view Palantir, and the "AI Operating System" category it represents? My judgment has two layers:
On the category, I'm convinced. "Modeling enterprise operational reality into an AI-actionable operating layer" — I see this as one of the core paths for AI to create real value on the enterprise side, far more promising than wrapper apps. It represents the right direction for AI to move from "chat tool" to "operating system." This category will produce real, deeply moated winners, and Palantir is among the most advanced.
On the price, I'm cautious. Acknowledging the category and the company does not mean endorsing today's price. Palantir is a "great company + extremely high valuation" combination — and the historical returns of such combinations depend heavily on your entry point. Buying a great but extremely expensive company at the peak of sentiment, you may have to endure a long drawdown and consolidation while the fundamentals slowly "grow into" the valuation — much like buying Amazon in 2000 and enduring a 90% drawdown before eventually realizing its greatness.
So my stance on names like Palantir is consistent with my stance on Nvidia (remember the last piece?): Will it fail? Probably not — its moat is real. But has today's price paid too high a premium for its goodness? Probably yes. My discipline therefore is — deeply understand the category and the company (it truly represents the right form of moat at the AI application layer), but exercise extreme restraint on the purchase price. Strictly separate the question "is this a good company?" from "is this a good price?"
V. Final Thoughts
Palantir is a mirror, showing both the deepest moat possible at the AI application layer and the biggest trap in investing in such companies.
Its moat — Ontology, AI Operating System, extreme workflow lock-in — is real, profound, and the most powerful rebuttal to the "wrappers have no moat" dilemma. It proves that in the AI application layer, the real winner is not the one wrapping a thin shell around a model, but the one that becomes the operating system the client can't do without. I believe this direction without reservation.
But it also reminds me: the deeper the moat and the sexier the story, the more easily the market will give it a "perfect" price — and a perfect price leaves no room for error. Palantir's fundamentals may deserve "great," but whether its stock price deserves "buy now" is a separate question that must be answered with cold clarity.
That's how I view the "AI Operating System" category: get excited about the direction, applaud the company, but hit the brakes on the price. Recognizing a great category and a great company, while refusing to pay a price that mortgages the future — in my book, these are never contradictory; they are two sides of the same rational coin.
If I leave only one sentence —
Palantir represents the rightest direction at the AI application layer — become the client's operating system, not the model's thin shell. Its moat is real. But deep moat does not equal reasonable price; a great company never automatically equals a great investment. Separating category from price is the discipline to hold when facing every "perfect story."
The AI applications series concludes here. Next, we step away from the compute and software mainlines to look at a few smaller, more patient-requiring tracks I've written about before — space, optical communications, innovative drugs, and robotics. They are farther from the center of AI, but each holds a distinct, long-term industry curve worth tracking.
——
Risk disclaimer: This article is an industry research piece. The companies mentioned are for analytical illustration only and do not constitute investment advice. Markets are risky; invest cautiously.
On investment analysis, market insight, and asset allocation — not chasing short-term noise, but understanding what truly drives long-term returns.
