Palantir and the Layer an Agent Has to Call

The tape spent the first half of 2026 pricing Palantir as a wrapper. Frontier models would eat the application layer. AIP was a prompt on someone else's weights. The multiple was a bet that a lab's deployment team could do the work. That is the substitution thesis that de-rated enterprise software. Palantir was the cleanest ticker in it, because Palantir's product is the layer the thesis says you no longer need.

Q2 2026 is the first quarter where that test ran on the same customer, with the same models, on the same timeline. One bake-off does not close a category. It does move the burden of proof. The market's remaining sorts — Palantir as AI software, Palantir as a defense name — still mix the ontology with warehouses, copilots, and primes.

Palantir reported the second quarter of 2026 on August 3. The 10-Q, filed August 4, books the same government and commercial split.

Shyam Sankar put the mechanism on the call, not in the press release. A major Silicon Valley technology company ran a bake-off: a frontier lab and its deployment team against AIP and Palantir's forward deployed engineers. Same customer, same timeline, same models. The lab picked a ticketing automation problem and failed to deliver anything of value. Palantir built agent swarms that recommended marketing, packaging, and pricing changes, and converted the work to a $10 million ACV contract. The model was not the variable. The ontology and the people who wire it to operations were.

That is one company-disclosed anecdote. The stock page holds talent scarcity at intact for exactly that reason — a pattern needs a second and a third disclosed displacement. Net dollar retention jumping 700 basis points to 157% is the number that is not an anecdote: existing customers deepened the coupling in the same quarter the lab lost the bake-off. A multinational that started at one operating company in Q4 2025 converted to a three-year, nearly $370 million deal spanning its full portfolio. Kirkland & Ellis, on the call, said the work would be impossible without the Ontology.

If the wrapper thesis were the risk factor, the market's sorting would be simple: names that sit on models carry substitution risk, names that sell infrastructure do not. Coverage contains a natural control group. It does not sort that way. Four names sell an operating layer an agent has to call. Two sell a data plane an agent can query and leave. Microsoft is the bundle that tries to make the call optional. Lockheed is the prime the government book gets Palantir compared with.

The scorecard does not sort on who said AI this quarter. Palantir, ServiceNow, Veeva and Atlassian are the operating-layer names. Snowflake and MongoDB printed AI attach and still sit in a thinner-moat band the framework has been willing to rank up on growth. Microsoft is not a weak software company. It is the company that can price orchestration into a signed Azure agreement. Lockheed Martin is the comparison you make if Palantir is a contractor. The composite will move. The grouping is the argument.

Palantir and ServiceNow light up vendor-owned logic, the record, and the transaction path. Veeva and Atlassian lock on the record and the workflow — Vault inside GxP, Jira as the work item — and stop short of vendor-owned logic. That is the tell that a vertical SoR is not the same instrument as an ontology the vendor owns. Snowflake and MongoDB are intact across this grid: data gravity is real, and it is not encoded operational logic. Microsoft is strong on bundling, the record, the transaction path, and regulatory lock-in, and only intact on vendor-owned logic — which is why Fabric is a skip into a signed Azure agreement, not a Palantir peer. Lockheed is a regulatory and embedding fortress; its software is sustainment on an airframe, not an ontology an enterprise agent has to call. The matrix is doing the sort the AI-software label refuses to do.

A frontier model is a general reasoner. It does not know this customer's objects, decisions, or permissions. Palantir's Ontology is that map — the objects, the links, the actions — written into Foundry so an agent can act inside the operation rather than describe it. The NVIDIA stack, now in production, runs Nemotron inside the customer's security boundary and post-trains into weights the customer owns. Sankar's line on the call: tokens are the new coal, AIP is the train, and customers can now build their own locomotives. The pitch is sovereignty, not a nicer chat box.

• Business logic. Customer-specific workflows and decision trees sit in the platform. An LLM that needs those rules to reason cannot take them to a competitor by switching the model. The bake-off is the first clean test of that claim against a lab's own deployment team.
• System of record. For defense and intelligence, Palantir is the operational-AI record at clearance. In Q2 the first government program of record chose to run inside Maven, inheriting its ontology, open data standards, and developer tooling. Maven carries more than 25,000 builders. Department of War trailing-twelve-month revenue is still under 25 basis points of the Pentagon budget.
• Transaction embedding. Logistics, targeting, and command-and-control actions run through the platform. A program of record hosting on Maven is not an analysis layer you query and leave. Removal is operational risk rather than a migration project.
• Regulatory lock-in. FedRAMP, IL4/IL5, and air-gapped accreditation take years to copy. The NVIDIA sovereign engine is accredited to run open models inside those environments. A competitor cannot buy its way onto that surface in a quarter.
• The people are not a moat yet. Forward deployed engineers are the scarce implementation model. The bake-off is one disclosed win. Intact, not strong, until a second and a third lab team lose in public.

The tape is not confused about Palantir's growth. It is confused about what instrument it is holding. Three honest objections survive the bake-off. They are not the same objection.

Snowflake is what Palantir would look like if it were the AI data platform the market says it is. Q2 FY2027 product revenue was $1.49 billion, up 37% year over year, the third straight quarter of acceleration, reported September 2. CoCo surpassed 9,100 accounts. Management said AI products contributed roughly half of the acceleration. Net revenue retention was 126%. That is a real print. The moat is still intact rather than strong, and Databricks remains the residual in head-to-head AI evaluations. An agent can query Snowflake and leave. Cortex brings the model to the data; it does not encode the customer's operating logic. If Palantir were a warehouse with an AI attach, it would live in this band: accelerating consumption, thinner durability, a second source that actually exists. Ranking Snowflake up on growth while ranking Palantir as a wrapper is the mix-up. They are not the same residual.

Microsoft is the skip path that can still work. Azure printed 43% growth in Q4 FY2026. Copilot is past 30 million seats. Commercial remaining performance obligations are $678 billion. Fabric can put AI orchestration inside a signed Azure agreement at a price Palantir will not match at the cost-sensitive tier of commercial deals. The matrix shows why: Microsoft is already strong on bundling, the record, the transaction path, and regulatory lock-in. Palantir's own bear case names this. U.S. commercial TCV of $2.132 billion, up 153%, is the evidence it has not happened at the accounts Palantir is winning this year. It is not evidence it cannot happen at the next hundred. A disclosed Fortune 500 loss to a bundled hyperscaler suite, or net dollar retention falling back below 130%, is how that objection stops being theoretical. One bake-off against a lab is not a bake-off against Azure pricing.

Lockheed is the comparison the government book invites, and Palantir is not that company. U.S. government revenue was $809 million in Q2, up 90%, 42% of the total. If the commercial AIP story is a multiple painted on a defense contractor, Lockheed is the read: $20.1 billion of Q2 sales, up 11%, $230 billion of backlog, the F-35 as infrastructure. Palantir's Department of War trailing-twelve-month revenue is still under 25 basis points of the Pentagon budget. Maven's 25,000 builders are a software platform, not a sustainment flywheel on an airframe. Treating Palantir as Lockheed with extra steps understates the commercial channel, which is faster, and overstates how much Pentagon budget the company has actually captured.

The version of the contractor objection that holds is concentration, not identity. International commercial revenue grew 26% year over year and 2% sequentially, to $182 million. International government was $181 million, up 42%. Eighty-one percent of revenue is the United States. Effectively all of the growth is a single-geography book. A federal budget disruption, an extended continuing resolution, or another enterprise-software rotation hits that book undiversified. Q2 2027 laps +93% and a record $2.132 billion U.S. commercial TCV quarter. The Q3 guide grows sequentially from this print, not at this print's year-over-year rate. The wrapper thesis failed its first clean test. The geography thesis has not been tested at all. A reader who holds the bear for that reason is not making a category error. They are making a book-composition argument the print does not rebut.

The live scorecard is the positioning. Palantir sits at the top of the software label because the ontology is the product, not a threat to it. ServiceNow is the same instrument in enterprise workflow. Veeva and Atlassian are the vertical and work-graph versions, with less vendor-owned logic and more record lock. Snowflake and MongoDB are the control group: AI attach on a data plane you can leave. Microsoft is the bundle you actually have to beat. Lockheed is the name you do not own Palantir through.

What the quarter did not buy is cheapness. Remaining deal value of $13.1 billion, up 83%, and RPO of $4.9 billion underwrite the near-term guide in a way that was not true in July. The valuation pillar is the expensive half of the same name, and it recomputes against the live price. Charging a beat to both the growth band and the multiple is how a print gets counted twice. The growth estimate on the stock page was held through an 11-point guidance raise for that reason. The next print that matters is Q3 against $2.160–$2.164 billion, and then the Q2 2027 comparison that laps this quarter's rate.

When the tape files Palantir with the AI-software basket, the job is to find whether an agent still has to call the ontology to act. On this cohort the wrapper was the wrong read on Palantir, a plausible read on Snowflake, and not a read on Lockheed at all. Microsoft remains the name that can make the call optional. Geography remains the fact the bake-off does not touch.