Palantir tied up with Nvidia on a “sovereign AI” push, then dropped a nine-point manifesto telling institutions to own their data and models instead of renting from OpenAI or Anthropic. Two big US-and-Canada builders, McCarthy and Cavanagh, have already gone all in on Palantir. On the show we argued about whether that’s a real edge for construction or an expensive way to repackage risk and reporting. Here’s the honest version.
Alex Karp doesn’t do subtle. Inside a single week Palantir announced a sovereign AI deal with Nvidia, posted a nine-point manifesto telling everyone to stop renting their intelligence from the big labs, and sent its CEO on CNBC to call the whole token-pricing model broken. Boiled down, the pitch is this: whoever controls your data and your model weights controls your future, so don’t hand either to someone else.
That’s a national-security argument on its face. It’s landing in construction too. In June, McCarthy, one of the oldest big contractors in the US, signed a multi-year deal to build its entire AI operating system on Palantir. Around the same time, Canada’s Thomas Cavanagh Construction locked in an eleven-year deal and spun out a subsidiary, Cavtera, whose whole job is selling Palantir setups to other builders.
So we put it on the table on the latest episode. Is Palantir actually built for construction, or is this a very expensive way to make your reporting look clever? We didn’t fully agree, which is exactly why it’s worth writing up.
Palantir’s sell is “own your AI, don’t rent it.” Two major builders bought the pitch. Whether it moves the needle on productivity, or just reprices risk, depends entirely on the data discipline underneath it.
What Palantir actually sells, and why Karp is picking a fight
Ontology, embedded engineers, and “own your weights”
Palantir’s core product is a thing called Ontology, a software layer that sits between a raw AI model and your messy company data. It maps your people, equipment, contracts, and jobs into one structured model the AI can reason over, and it’s built to stop the model wandering off with your data or IP. The other half of the Palantir method is people: forward-deployed engineers who embed with your teams in the field, learn how decisions actually get made, and wire that straight back into the software.
The Nvidia deal, announced in late June, lets customers run Nvidia’s open Nemotron models inside their own secure walls using Palantir’s stack, so agencies and critical-infrastructure operators get powerful AI they fully control. The nine-point manifesto that followed is the marketing engine bolted around it. It tells institutions their data is treasure, that “controlling your weights is controlling your fate,” and takes direct aim at “tokenmaxxing,” Palantir’s word for burning cash on metered AI from OpenAI and Anthropic. Strip the theatrics and the point is fair: data and model weights are true sources of advantage, and Karp has built a loud brand around that anxiety.
- Ontology. A structured data layer between the AI and your business, so the model reasons over your real operations instead of guessing.
- Forward-deployed engineers. Palantir staff embed with your teams, learn the workflows, and build the software around them.
- Sovereignty. Run open models inside your own walls, keep your data and weights, and stop renting your edge from a frontier lab.
Two big builders already bought in
McCarthy’s “Pulse” and Cavanagh turning into a reseller
McCarthy’s deal is the headline. The contractor is using Palantir’s AI Platform and Ontology to connect estimating, bidding, buyout, QA/QC, logistics, and field execution through one shared model, with an AI-native operations tool called Pulse sitting on top for real-time insight and risk analysis. It’s also copying Palantir’s playbook internally, building its own embedded-engineer team rather than buying finished tools off a shelf.
Cavanagh went further down the same road. After rebuilding its core workflows in Palantir’s Foundry in about a year, replacing several separate software platforms along the way, it signed on through 2035 and launched Cavtera to sell that experience to other contractors. Read that back: a construction client became a Palantir channel. That tells you how sticky this gets once the ontology is in place, and why more tech-literate builders are choosing to build on a platform rather than buy finished tools off a shelf.
We believe we’ve invested in the premier platform on the market.Justin McFarland, Chief Digital Officer, McCarthy
The construction tech moves that matter, before everyone else clocks them
Deals, tech, and honest takes for people who build, fund, and sell into the industry. One email, no fluff.
Join 3000+ ReadersSo is it built for construction? The argument we had
Productivity, or insurance with extra steps?
Here’s where the table split. Dustin’s view: the forward-deployed-engineer model, armies of pricey people hand-mapping your business, is starting to look dated, because AI can now map a company’s ontology quickly and cheaply. If a model can learn your data structure in days, why pay a small consulting firm to live on your site for a year?
Alain pushed a different angle. Strip a big general contractor back to what it really is at the top, and it looks a lot like an insurance business: it prices risk, carries it, and reprices it as jobs move. Palantir is strong at exactly that kind of risk work. But repricing risk isn’t the same as making the actual building go up faster, and if you sell it as a productivity story, well, that’s marketing.
The counter is just as strong. Connecting estimating, bidding, and field data through one clean model is precisely the data-plumbing problem construction keeps failing at. We’ve made that case before, that most AI budgets are inverted and the real work is fixing the data underneath, and that the flashy deals change nothing until the plumbing gets sorted. If Palantir forces a firm to finally build that layer, the value can be real. The honest answer landed somewhere in the middle: it works if you’ve got the scale and data discipline to feed it, and for most firms the bottleneck sits upstream, not in the model.
Rent the model, or own the stack?
The two camps, side by side
The whole fight comes down to one choice: rent intelligence by the token, or own the stack and run models inside your own walls. Neither wins outright. Here’s how they line up for a builder.
| Dimension | Rent from the frontier labs | Own the stack with Palantir |
|---|---|---|
| Best fit | Fast, flexible use cases and smaller teams | Regulated, data-sensitive, large operations |
| Your data | Flows through someone else’s model | Stays inside your own walls |
| Cost model | Metered, per-token, climbs with use | Platform plus engineers, big upfront |
| Construction example | Copilots bolted onto existing tools | McCarthy’s Pulse, Cavanagh’s Foundry rebuild |
| The catch | Bills balloon, your edge leaks out | Heavy lift, needs data discipline and scale |
A sovereign AI partnership that lets US government agencies and critical-infrastructure operators run Nvidia’s open Nemotron models inside their own secure environments using Palantir’s platforms. The point is control: your data, your models, your walls. (Source)
A statement Palantir posted on X on June 30, 2026, urging institutions to own their data and model weights rather than renting AI from the big labs. It attacks “tokenmaxxing” and argues that controlling your weights is controlling your fate. (Source)
McCarthy, one of the oldest big US contractors, signed a multi-year deal in June to run its AI operating system on Palantir. Canada’s Thomas Cavanagh Construction signed an eleven-year deal and now resells Palantir setups through its Cavtera subsidiary. (Source)
A subsidiary Cavanagh spun out after rebuilding its own workflows in Palantir’s Foundry. It’s positioned as a Foundry service provider built specifically for construction and heavy industry, selling that setup on to other builders. (Source)
A structured map of your business, your people, equipment, contracts, and jobs, that sits between the raw AI model and your data so the model reasons over your real operations instead of guessing. It’s the layer Palantir sells hardest. (Source)
Both arguments have teeth. It can connect the messy, disconnected data construction runs on, which is a real problem. But critics on our show argued it’s better at repricing risk than speeding up the build itself, and it only pays off if your underlying data is already in shape. (Source)
Palantir staff who embed directly with your teams, on site or remotely, to learn how work really happens and build the software around it. It’s central to how Palantir deploys, and McCarthy is now copying the model with its own internal engineers. (Source)
Related Articles
Why Most ConTech AI Budgets Are Inverted
Inside OpenAI’s Construction Power Move
