Palantir Is a Plumbing Company That Learned to Sell Water

Bottom Line Up Front

On August 3, 2026, Palantir reported the fastest revenue growth in its history: $1.935 billion for the quarter, up 93% from a year earlier. U.S. commercial revenue — sales to American businesses rather than governments — grew 149%. Net income was $1.07 billion, against roughly $329 million a year ago. Profit more than tripled.

Alex Karp, the chief executive, called the quarter "otherworldly." That is not a word most public-company CEOs use, and Palantir is not most public companies.

Here's the part worth understanding, and it has almost nothing to do with the stock. Palantir spent roughly twenty years building deeply unglamorous infrastructure — the software equivalent of the pipes under a street — while a great many serious people considered it an overhyped consulting firm with a spy-agency clientele. Then large language models arrived, and it turned out the pipes were the scarce thing.

Every company now wants AI to do useful work. Almost none can, because their data sits in systems that don't know each other exist. Palantir's entire twenty-year project was solving that exact problem. The AI boom didn't create this business. It revealed it.

We think the growth is real and the mechanism behind it is durable. We also think the price assumes very little goes wrong for a very long time. Both can be true.

---

The dots nobody could connect

After September 11, 2001, a painful finding emerged: the United States government had, scattered across its agencies, much of the information needed to disrupt the plot. Names in one database. Visa records in another. A flight-school enrollment somewhere else. The failure wasn't a shortage of data. It was that no system could see across the others.

That's a strange kind of failure. Nobody was lazy and nothing was hidden — a name in one database and a name in another were, as far as any computer was concerned, unrelated strings of text.

A few years earlier, engineers at PayPal had wrestled with a smaller version of the same thing. Fraud rings were draining the company. Any single transaction looked fine; the pattern only appeared when you connected accounts, devices, addresses and timing — data living in separate systems with no reason to be joined. They built tools that made those connections visible, and the losses fell.

Palantir was incorporated on May 21, 2003, on the bet that the second problem and the first were the same problem at different scales.

Who they are

The founding group was unusual even by Silicon Valley standards: Peter Thiel, who had co-founded PayPal and put in roughly $2 million of seed money; Alex Karp, a Stanford Law classmate of Thiel's with a doctorate in social theory from Frankfurt and no technology background at all; and three engineers — Nathan Gettings, Joe Lonsdale and Stephen Cohen.

Karp became CEO and remains an odd fit for the role in ways that matter. He discusses philosophy on earnings calls and is openly combative about the company's politics rather than diplomatically quiet, which has cost Palantir some customers and won it others.

Early credibility came from In-Q-Tel, the CIA-backed venture fund, which invested and opened doors. The first customers were intelligence and defense — an origin that shaped everything after: the products, the culture, the criticism, and the strange position the company occupies now.

The problem, in a warehouse

Imagine a manufacturer with a warehouse in Ohio, a shipping partner in Rotterdam, and a finance team in spreadsheets. The warehouse calls a shipment PLT-4471. The shipper calls that same physical pallet RTM_88213. Finance knows it only as dollars. Three systems, one pallet, no shared understanding that they describe the same object in the world.

Now ask: if Rotterdam slows down, which customer orders are late and how much revenue is at risk?

A person who walks the building for a week can answer that. Software essentially cannot, because nothing knows that PLT-4471 and RTM_88213 are the same thing — or that a pallet relates to an order relates to a customer relates to revenue.

What Palantir built is a map that holds those relationships. The company calls it an ontology, a word that will glaze most eyes, so ignore it and keep the idea: a working model of what exists in your organisation and how it connects. Once that map exists, software can reason about the business itself instead of shuffling database rows.

That's the whole company. Everything else follows from it.

The four systems

Gotham is the original, built for defense, intelligence and law enforcement. It joins vast amounts of fast-moving data and puts it in front of operators in one view, so people in different roles and places see the same picture and can act together.

Foundry is the same idea aimed at ordinary companies — manufacturers, airlines, hospitals, banks. It builds the map across the dozens of disconnected systems a large business accumulates over decades.

Apollo is the least discussed and quietly explains why Palantir wins certain contracts. It's the delivery machinery: it ships the software into whatever environment the customer actually runs — someone else's cloud, their own data centre, a classified network, or a disconnected system on a ship with no internet. Most enterprise software assumes a normal cloud. Much of Palantir's world is not normal.

AIP, launched in April 2023, connects large language models to the map, so a model works with real objects and real permissions rather than loose text.

They're sold separately and worth more together. That's the business model.

What it looks like when it works

Abstractions are cheap, so here's the public record.

Airbus merged 25 data silos and more than 400 datasets across A350 production, raised output 33%, and identified an estimated $1.7 billion a year in savings. The mechanism is unglamorous and exactly the point: when a production problem appeared, they could see where it came from and what it touched, rather than convening a meeting to find out.

NHS England ran pandemic logistics through Foundry — distributing 6.9 billion items of protective equipment and supporting a vaccination programme that at peak delivered up to 771,000 doses in a day. In 2023 the NHS awarded Palantir a £330 million contract for a central data platform, which drew sustained criticism about a private, American, intelligence-linked firm sitting at the centre of national health data. That objection is real, not hysterical.

Ukraine has used Palantir software since 2022 for target detection, intelligence, demining and the logistics of moving supplies and refugees.

Notice what these share. None is "we bought an AI." All are versions of: we already had the information, in pieces, and now we can see it as one thing.

The twenty years nobody cared

The long-running criticism was not unreasonable. Palantir sold software but behaved like a consultancy, sending its own forward-deployed engineers — people who sit at the customer's site and build the thing with them rather than shipping a product and hoping. That's expensive and doesn't scale the way software should. Every deployment looked handmade.

Investors are rightly suspicious of software companies whose growth requires bodies — it's the difference between selling a machine and renting a craftsman. The numbers gave skeptics plenty. Palantir went public in September 2020 by direct listing, opening at $10 a share for roughly $16.5 billion — below its $20.4 billion private valuation from 2015 — on about $1.1 billion of revenue, still mostly government, while losing money at scale.

What the critique missed took twenty years to become visible: building the map is the hard part, and it can't be automated by a vendor who doesn't understand your business. All that expensive, handmade work was constructing exactly the asset that would later become the bottleneck for everything else.

What changed

Large language models are extraordinary at language and unreliable about facts. Ask one about your supply chain and it produces something fluent, confident and possibly invented, because it has no grounding in what's true inside your company.

Give that model the map — real objects, real relationships, real permissions about who may see what — and it becomes something else. Not a chatbot. A system that can answer which orders are at risk and be checkably right, because every claim traces to a system of record.

Palantir paired this with an unusual tactic: short, intense "bootcamps" where a prospective customer brings real data and builds something working in days. Hard to fake. Either the map gets built and the answers are right, or they aren't.

A cardinal doesn't rebuild its map of the forest each morning. It remembers where food reliably appears and ignores nearly everything else — which is why it eats while other birds are still searching. Palantir spent twenty years building the map. The AI boom mostly made everyone else notice they were still searching.

The numbers, honestly

From the company's release for the quarter ended June 30, 2026:

• Revenue: $1.935 billion, up 93% — the fastest in company history
• U.S. commercial revenue: up 149% year over year
• U.S. government revenue: up 90% year over year, 18% from the prior quarter
• Net income: $1.07 billion, or $0.41 per share, against ~$329 million and $0.13 a year earlier
• Operating income: $912 million, a 47% margin, against $269 million
• Adjusted free cash flow: $1.22 billion in the quarter
• U.S. commercial remaining deal value: $6.24 billion, more than double a year ago
• Deals closed: 220 worth $1 million or more; 98 worth at least $5 million

Two need translating. Remaining deal value is money customers have contracted to pay that isn't yet revenue — the closest thing software has to visible future income. When it more than doubles, much of next year's growth is already sold. The Rule of 40 is the industry's rough health check: growth rate plus profit margin, where 40 or better means you're growing without setting money on fire. Most good software companies live in the 40s. Palantir reported 155%.

That's genuinely unusual, and we won't quietly convert it into a thesis. A score of 155 tells you this quarter was exceptional. It doesn't tell you it repeats.

Full-year guidance is now $8.150–8.158 billion, roughly 82% growth, with adjusted free cash flow of $4.5–4.7 billion.

The case against

The valuation assumes near-perfection. Even after this quarter, the multiple requires years of compounding at rates almost no enterprise software company has sustained. You can believe the business is excellent and the stock is priced with no room for one bad year. Bloomberg framed it as a "worst-of-both-worlds bind" — valued like pure software, operated with the labour intensity of services.

Government concentration cuts both ways. Ninety percent growth is remarkable and also means much of the business rides on procurement cycles and political weather no company controls.

The services critique was outrun, not answered. At 93% growth, deployment costs look trivial. At 20%, they would not.

Competitors want this ground. Databricks, Snowflake and the major clouds all understand the map is where the value sits. None has produced a convincing substitute yet. None is ignoring it either.

And one question isn't financial. Palantir builds systems that make institutions — including military and law-enforcement agencies — dramatically better at finding patterns in data about people. In Ukraine that has meant helping locate artillery. In Britain it meant a sustained argument about who holds health records. Reasonable people land differently on this, and it doesn't belong in a footnote: anyone deciding whether to own this company is also deciding how they feel about that.

Why this matters beyond one stock

Palantir is the clearest live test of a question the whole market is guessing at: does enterprise AI produce measurable value, or mostly impressive demonstrations?

For three years the honest answer was unclear — pilots multiplied, durable deployments were scarce. These commercial numbers are the strongest public evidence yet that the value is real, with the large caveat that capturing it seems to require infrastructure most companies never built.

Sit with the consequence. If AI's usefulness depends on having a clean, connected model of your own operations, the binding constraint on enterprise AI isn't model quality. It's data plumbing. That reorders who captures the value — a very different world from the one where the best model simply wins.

What We're Watching

1. Whether U.S. commercial growth holds above 100% next quarter. This single number separates "AI made our business inevitable" from "AI gave us one exceptional year."

2. Net dollar retention among existing commercial customers. Growth from new customers is a sales story. Growth from existing customers buying more is a product story. Only the second compounds.

3. Whether a credible substitute for the map appears. We'd change our mind about durability on evidence of large customers replacing Foundry rather than adding to it.

4. Government revenue through a full appropriations cycle — worth watching through a real budget fight, not just a quarter.

5. Headcount against revenue. If revenue keeps growing much faster than engineers, the services critique dies for good. If they track each other, this was always a consultancy with excellent margins, and should be valued as one.

---

SOURCES

• Palantir Technologies, Q2 2026 results release, August 3, 2026 — investors.palantir.com
• Palantir platform documentation, AIP / Foundry / Apollo — palantir.com/docs
• CNBC, "Palantir (PLTR) earnings Q2 2026," August 3, 2026 — cnbc.com
• CNBC, "Palantir files to go public, lost about $580 million last year," August 25, 2020 — cnbc.com
• Bloomberg, "Palantir Raises Full-Year Sales, Income Forecasts After Strong Quarter," August 3, 2026 — bloomberg.com
• Bloomberg, "Palantir Is Stuck in Worst-of-Both-Worlds Bind With AI, Software," August 3, 2026 — bloomberg.com
• Palantir Technologies, company history and deployments — Wikipedia

Figures are as reported by the company for the quarter ended June 30, 2026. Where sources disagreed — two outlets gave different counts of deals above $10 million, and different after-hours share moves — we omitted the figure rather than choose between them. Deployment details for Airbus, the NHS and Ukraine come from public reporting and the company's own material; we have not independently audited the savings or production claims, and we'd treat vendor-supplied numbers with the caution they deserve. This is research, not investment advice.