On 2 November 2026, Palantir reported its third-quarter results, and the Palantir earnings headline was growth. Revenue "grew 96% year-over-year and 20% quarter-over-quarter to $2.313 billion", according to the company's release. GAAP net income was $1.128 billion, a 49% margin. And for the first time, the release says, "U.S. commercial revenue surpassed U.S. government revenue". Its chief executive, Alex Karp, put it more personally: "Our U.S. commercial revenue has surpassed our U.S. government revenue for the first time in our company's history."
Searches for "pltr stock" spiked on Google Trends as the share price moved. This piece is not about the share price. It sets out what Palantir reported, what its approach actually is, where the numbers and the claims need a careful reading, and what a business deciding how to adopt AI can take from it.
Nothing here is investment advice. It is about building systems, not buying shares.
What Palantir reported for Q3 2026
The three figures most often repeated in search results all match the release. The table sets out the main numbers alongside them.
| Measure | Q3 2026 | Change |
|---|---|---|
| Total revenue | $2.313bn | up 96% year on year |
| US revenue | $1.937bn | up 119% |
| US commercial revenue | $984m | up 148% |
| US government revenue | $953m | up 96% |
| GAAP net income | $1.128bn | 49% margin |
| Adjusted operating income | $1.427bn | 62% margin |
| Total contract value closed | $2.918bn | up 6% |
| US commercial contract value closed | $1.463bn | up 11% |
| Customers (trailing twelve months) | 1,090 | 911 a year earlier |
Sources: Palantir's Q3 2026 press release and quarterly report on Form 10-Q.
The company closed 212 deals of at least $1 million and 64 of at least $10 million. It raised its full-year 2026 revenue guidance to "between $8.530 - $8.534 billion" and its US commercial revenue guidance to "in excess of $3.559 billion", which it says represents growth of at least 143%. For the fourth quarter it expects revenue of $2.649–2.653 billion.
On the earnings call, Palantir's David Glazer said "Our U.S. commercial customer count grew to 686 customers, reflecting growth of 29% year-over-year and 5% sequentially", and that "Net dollar retention was 165%", according to the Motley Fool transcript. That last figure means, broadly, that existing customers spent far more than they did a year earlier. The company's quarterly report puts the total customer count at 1,090, counting any organisation that generated revenue in the previous twelve months.
Palantir AIP, the ontology and bootcamps, explained plainly
Palantir sells software platforms, not AI models. Three ideas explain most of what it does.
Foundry is the data layer. The 10-Q calls Foundry "our foundational data operations platform". It pulls data out of a company's existing systems, such as its ERP, its maintenance logs and its spreadsheets, and joins it into one place that can be queried and governed.
The ontology gives that data meaning. In Palantir's words, "Ontology generally refers to the systematic mapping of data to meaningful context." In practice, it means that a row in a table becomes an object people recognise: a customer, an order, a machine, a shipment, with the relationships between them and the actions that can be taken on them. The filing describes it as going beyond data mapping by bringing data, logic and actions together. This is the part that matters most, and the part most companies skip.
AIP connects models to that structure. "AIP is our generative AI platform, which provides secure connectivity to third-party-provided large language models", the 10-Q says, along with tools for building agents. A model working through the ontology is not reading loose documents. It is working with defined objects, permissions and permitted actions, which is what makes it possible to let it do something in a live process rather than only answer questions.
Bootcamps are how it sells. Instead of a long pilot, Palantir runs short, intensive sessions in which a customer's own data and a real problem are put into the platform. The 10-Q says the company runs "pilots and bootcamps with customers, generally at our own expense and without a guarantee of future returns". On the call, Karp said "the time line from interaction to them buying pilots or going into production is completely shifted", and described first contracts that start, "instead of being, like, $250,000, it's like $25 million to start."
The executives also framed where value comes from. "The fundamental unit of the enterprise is the workflow," said Shyam Sankar, another Palantir executive on the call. His argument was that making individual staff more productive does not by itself produce a return; changing a whole process end to end does. The company is now pushing what it calls sovereign AI: fine-tuning open-weight models on a customer's own operational data, which Sankar described as "better than frontier performance on weights you control".
Enterprise AI adoption: from pilots to production
The US commercial figures are the clearest signal in this report. A 148% rise in revenue from US businesses, with existing customers expanding sharply, suggests that a meaningful number of companies have moved AI past experiments and into processes they pay to run every day.
The reasons Palantir gives are worth noting, because they are not about better models. They are about structure: data joined and defined first, a model allowed to act only through defined objects and permissions, and success measured against a specific workflow rather than general productivity. That matches a common pattern. Many AI pilots stall not because the model is weak but because the data underneath is scattered, inconsistent or owned by nobody. Palantir's approach starts there, which is a large part of why it can move a customer into production quickly.
Where to be sceptical about the Palantir earnings story
The figures are real. Several readings of them deserve care.
Bookings slowed sharply. Total contract value closed rose only 6%, and US commercial contract value 11%. CNBC reported that analysts pointed to the slowdown in contract value and other "forward indicators", noting that the US commercial segment closed a record $2.1 billion in the second quarter. Glazer said "We signed multiple 8-figure deals in the quarter that were shorter in duration than usual," and that "over time, we expect both the duration and size of these contracts to increase." That is a plausible explanation. It is also a forecast.
The valuation assumes a great deal. Shares fell about 3% in after-hours trading on Monday, TipRanks reported, and CNBC reported a further fall in premarket trading on Tuesday. TipRanks put the trailing non-GAAP price-to-earnings ratio at 153.43 against a sector median of 23.26. A price like that leaves little room for growth to slow.
Revenue is concentrated. About half of revenue still comes from governments: "51% of our revenue came from government customers" in the first nine months of 2026, the 10-Q says. Most of it comes from one country, the United States. And the average revenue from its top twenty customers over the last twelve months was $146 million each. No single customer exceeds 10% of revenue, but a small group matters a lot.
Profit depends on the measure. The 62% adjusted operating margin excludes stock-based compensation, which the 10-Q puts at $318.9 million in the quarter, up from $172.3 million a year earlier. GAAP net income of $1.128 billion is the more conservative figure.
The executives are selling. Karp's release frames the moment as "the most dramatic dichotomy between haves and have nots that I have witnessed in 30 years of business", with every institution "rapidly being forced, voluntarily or not" into one category or the other. That is a sales message from the person whose company sells the solution. Customer stories told on an earnings call, such as one customer that piloted "a nearly $3 million engagement" before signing a $29 million deal, are chosen to impress and cannot be checked.
Lock-in is the price of depth. The ontology is valuable because it becomes the shared model of how a business works. That is also what makes it hard to leave. If your processes, permissions, agent logic and fine-tuned models all live inside one vendor's platform, moving away later means rebuilding them.
What it means for a business adopting AI
Most businesses will never buy Palantir. The lessons apply anyway.
Get the data foundations right before choosing a model. Decide which business objects matter, such as customers, orders, assets and cases, where the trusted version of each lives, and who owns it. A model on top of messy, contradictory data produces confident, contradictory answers. This is unglamorous work, and it is where most of the value is.
Start from one workflow, not from the technology. Pick a process with a clear cost or delay, measure it now, and define what better looks like. If you cannot name the workflow, it is worth asking first whether your business really needs AI for that problem.
Evaluate a platform against building your own, honestly. A platform buys speed, security controls and a team that has solved the same problem before. Building your own, from a data warehouse, an orchestration layer and model APIs, buys control and lower long-term licence cost, at the price of more engineering and slower first results. The useful questions are how many workflows you will run, whether you have people to maintain what you build, and how far your processes differ from what a platform assumes.
Ask the lock-in questions before signing. Can you export your data, your ontology definitions and your business logic in a usable form? Who owns models fine-tuned on your data, and can you take the weights with you? Can you swap the underlying model provider? What happens to pricing at renewal, once your processes depend on it? Our piece on Big Tech AI spending after Q3 2026 covers the same questions for cloud and model contracts.
Insist on a time-boxed proof with your own data. The bootcamp idea is a good one whichever vendor you use: a short, fixed period, a real problem, your real data, and an agreed measure of success. A vendor that cannot show results that way is unlikely to show them later.
The short version
Palantir's third quarter shows that a growing number of companies are paying to run AI in production, not just test it, and that the ones getting there have done the data work first. It also shows slowing bookings, heavy concentration and a valuation that assumes the growth continues. For a business, the practical lesson is less about Palantir than about order: define your data, choose one workflow, prove it with your own information in a fixed time, and know how you would leave before you commit.