Essay

Same model, same customer, same timeline: the frontier lab still lost

A frontier lab and Palantir ran the identical assignment for the identical customer, on the same models, and the frontier lab still lost. The story is Palantir CTO Shyam Sankar’s, retold on this year’s Q2 earnings call. A “major Silicon Valley tech company” ran a bake-off: a frontier lab and its deployment team against Palantir’s AIP platform and its Forward Deployed Engineers. The lab picked a ticketing-automation problem and didn’t ship value. Palantir’s team built agent swarms aimed at revenue and utilization, and converted the pilot into a $10 million ACV contract. Sankar’s line for the record: “same customer, same timeline, same models.”

That is a company telling a story about itself, on a call built to move a stock price, with no customer named and no way for anyone outside the room to check it. Take it as color, not evidence.

The half you can check went to the SEC. Revenue for the quarter, per the 8-K Palantir filed, came in at $1.935 billion, up 93 percent year over year. US commercial alone grew 149 percent to $764 million. The company closed 220 deals worth at least $1 million apiece and raised full-year guidance to $8.150 to $8.158 billion. The range sits in the release. The superlatives came on the call, where CFO David Glazer called it an eleven-point increase over last quarter’s guidance and the largest full-year revenue guidance raise in Palantir’s history. Sankar didn’t write those figures.

A CEO who misstates them answers to the SEC.

Isn’t this just Palantir’s sales pitch about itself? Fair question, and the honest answer is: more than half of it, yes. The release proves the growth; only the call supplies the cause. Nobody on that call credited the quarter to Forward Deployed Engineers. The drivers the executives actually named were the US business, sovereign demand for AI, and AIP itself. The FDE credit lives inside the bake-off story I just told you to discount. And a fat slice of any 93 in this market is tide: enterprise AI is the largest spending wave software has seen, and every vendor in the blast radius printed a good quarter.

Held at arm’s length, though, the release still proves one thing, the thing skeptics of high-touch always said was impossible: scale at software economics. The knock on the embedded engineer was never quality; it was that a business built on him is a consultancy wearing a software multiple, and consultancies don’t compound. A two-billion-dollar quarter growing 93 percent at a 62 percent adjusted operating margin retires the knock twice: the growth says it compounds, the margin says the engineers aren’t eating it. What the numbers cannot do is name the layer that produced them. No filing can. For cause, the bake-off controls for the tide better than any filing could, and it is Sankar’s word from end to end. The better evidence arrived three months earlier, from the other side of the fight. “Only Palantir has FDEs,” Sankar told the same call. “Everyone else has sparkling sales engineers.” That line had been false since May, when OpenAI, which had been hiring forward-deployed engineers since late 2024, stopped hiring and started buying: the OpenAI Deployment Company, majority-owned, launched with more than $4 billion of initial investment, seeded with roughly 150 forward-deployed engineers acquired in a block from Tomoro, with McKinsey, Bain, and Capgemini attached and TPG and Goldman Sachs among the backers. There is a deflationary way to read that, and half of it lands. A layer your competitor can replicate with one acquisition and a consulting alliance stops looking like a moat and starts looking like scarce labor getting bid up on its way to systems-integrator margins. The rest dies on who the buyer is: the company that builds the frontier model, holds the weights, and sells intelligence through an API decided it could not close the enterprise without engineers in the customer’s building.

When the model layer buys the human layer, the confession is in the wire.

And the lab in Sankar’s bake-off had a deployment team of its own. That detail does more work than he meant it to: both sides had humans embedded with the customer, so warm bodies were not the variable. His full attribution, the half that never makes the quote cards, is “The only difference was AIP and Palantir’s unique FDE tradecraft, it was determinative.” Platform and practice, jointly. On the anecdote’s own terms the platform is the copyable half: AIP is software, and the models under it were the same by construction. That leaves tradecraft as the residual, and residuals have a problem: nobody outside the room can falsify one. So watch what the buyers did instead of what the sellers said. When OpenAI went shopping for the layer, the asset it actually acquired was people, the Tomoro block, engineers who had already done the deployments, because tradecraft lives nowhere else. A decade of sitting inside customers’ ticketing queues compounds in a person. It does not transfer by API.

And here’s the part I have a stake in, so weigh it accordingly: I bet a company on the belief that revenue stays high-touch as AI eats the codebase, and Closin exists because I think the go-to-market function doesn’t hollow out the way the model layer does. A frontier lab losing a bake-off on the same models is a convenient story for that bet, so discount me the way you’d discount Sankar. What survives the haircut is the size: the revenue, the deal count, the raise, all checkable. The shape of the loss is Sankar’s word, but I didn’t get the shape from Sankar. My first company died on it: I wrote working code and made no money, because writing the code was never the part that was hard. I have held that prior since 2015, and a convenient anecdote doesn’t upgrade a prior into a receipt. The numbers carry the weight. The story just draws the picture.

The model gets cheaper every quarter. The seat next to the customer is going the other way: Palantir grew 93 percent with that seat at the center of its operating model, and OpenAI just paid four billion dollars to build its own. The engineer who sits with the customer until the deal closes didn’t get automated by this model release. He got acquired.

Sources: