Code got cheap. Deployment became the alpha. Karp gives away the expensive part and licenses you the part that should be yours.
The franchise
Alex Karp has spent 2026 as enterprise AI’s loudest critic of rent. Usage pricing, he told CNBC in July, is a wealth tax on American business, and customers deserve to own the means of production [2]. On the Financial pillar, nobody has said it better. It may be the most persuasive case for owning your stack ever made by a man who licenses his.
Then, in the space of twelve days, he built a distribution network for the one layer customers cannot own.
On October 1, Palantir named Armada its first Certified Modular Data Center Partner. Ruggedized Galleon modules, built in the US and allied nations, running Palantir’s Sovereign AI OS, with the release promising a stack customers own end to end. Armada’s tagline: “Sovereignty is not something you rent.” [3] Correct. You license it, annually. On October 12, Ondas and Palantir launched Regna AI, a standalone company that fine-tunes open-weight models on mission data and ships them inside customer environments on Palantir’s AIP, Foundry and Apollo [4]. Karp’s line: sovereign AI trained on tradecraft beats the frontier labs “while removing dependence on them.” [5] The announcement skipped one detail. How Palantir gets paid [5]. Presumably not per token.
Look at the shape. Hardware: a certified partner, and Territorial passes. Models: open and swappable, and the model layer passes Technological. Mission data: yours. And in the middle of all of it, the OS: Palantir’s, licensed, released by Palantir, and now carried into new markets by companies Palantir does not own. That is not a product. That is a franchise. Call it the McDonald’s model of sovereignty: you own the building, the fryers and the staff, and Palantir owns the recipe, the menu and the right to approve every change in your kitchen. Every Galleon module and every Regna fine-tune opens another storefront for the one layer nobody else can fork.
In July, Dave Vellante and I gave this stack the best grade any major vendor has earned, with one asterisk: the model is open, the OS is not [2]. A franchise is what happens when a vendor turns its asterisk into its business model. Like any good horse trade, the terms look generous right up until you check who is holding the reins.
Code got cheap. Deployment didn’t.
Something changed in the value of software this year, and it changes the sovereignty math.
Writing code stopped being scarce. Stripe’s internal agents now merge more than 1,300 pull requests a week that contain no human-written code, every one human-reviewed, on systems that move over a trillion dollars a year [19]. When the next line of code costs a few cents of compute, the line itself is not where the value sits.
Reading code stopped being scarce too. REA, an MIT-licensed project with roughly 208,000 GitHub stars and 50,000 forks, lets an agent inspect an application without its source, explain how a feature works, show the evidence and build a version for your own project [20]. Its own disclaimer limits it to software you own or are authorized to examine, which is exactly the point. Closed source used to be a vault. In 2026 it is a curtain, and the audience brought night vision. What keeps a proprietary stack proprietary now is not secrecy. It is the license. The lock moved from the code to the contract, which means it moved from the Technological pillar to the Legal one.
So where did the value go? Into deployment. Getting software to actually run inside your environment, against your data, under your constraints and classification level, and keeping it running. That is the scarce thing now, and the industry knows it: OpenAI, Anthropic, Google and IBM have all stood up forward-deployed teams of their own this year [21][22].
Credit where it is due. Karp understood this a decade before everyone else. Palantir’s genius was never the code. It was the forward-deployed engineer and Apollo: working software in places nobody else could reach. On where the value lives, he is right.
Where he goes wrong is how he collects it. Palantir does not charge you for deployment and hand you what it deployed. It charges you for deployment by licensing you the code, indefinitely, and keeping the source. When code was expensive and secret, that was at least defensible as IP protection. Now that code is cheap and readable, withholding the source protects nothing but the meter. That is a Financial pillar decision dressed up as a Technological one.
The sovereign model separates the two. Pay for deployment: the engineers, the know-how, the speed. Own what gets deployed: the code, the ontology, the pipeline. Deployment becomes a service you can fire. The code becomes an asset you keep. That is the difference between a partner and a landlord.
Why the OS is the alpha
Weights are becoming a commodity. This month alone, a trillion-parameter European model and a 501B American one were announced with open weights promised within weeks [16][17]. Whatever model you fine-tune today, you will swap within a year, and you should.
What you cannot swap is the layer that encodes how your organization actually works. Your objects and the relationships between them. Your workflows. Which agent may take which action, on whose authority. Your evals. Your audit trail. Your release pipeline. That is the agentic OS, and it is where tradecraft becomes software.
Palantir agrees. On its Q2 call, the company’s CRO described what customers are buying under the banner of AI sovereignty as control over their data, their logic, their workflows, their security and their model outputs [6]. Correct. Now ask who holds the source code for the logic and the workflows.
Karp, to his credit, does not hide it. His sovereign stack, he says, offers “alpha protection qualities unavailable otherwise.” [7] Read that through the Technological pillar. The alpha is protected, but by a moat you sit outside of. Unavailable otherwise is not a feature. It is the lock. As TNW’s editor noted, an ontology is not something you swap out later [7].
The weights are the engine. The OS is the driver’s license, the road map and the insurance policy. Palantir lets you keep the engine.
The litmus test, run at the OS layer
Same five pillars, same standard as July. The difference: this time we grade the layer that holds the alpha, not the layer that holds the data.
01 Territorial
The question. Where do the data and compute physically reside?
The story. Air-gapped, on customer hardware, now in modular data centers manufactured in the US and allied nations [3]. Nobody in the market does this better.
Under pressure. The data lives in your building. The software that touches it does not originate there. Palantir’s own documentation describes air-gapped Apollo hubs running off a copy of a product catalog generated on Palantir’s SaaS hub, carrying delivery paths and recall instructions across the gap [8]. Your crown jewels sit on sovereign soil. The code that governs them is an immigrant with a standing return ticket.
Verdict. Demonstrated for data and compute. The OS is a resident, not a citizen.
02 Operational
The question. Who runs it, who approves changes, who takes the 3 a.m. page?
The story. Customer-held keys, enforced isolation, full auditability. In July this earned a clean check [2].
Under pressure. Apollo is the change authority. In Palantir’s published IL6 model, changes go to Palantir’s centralized operations team, Baseline, for approval before Apollo applies them, and vulnerability findings route back to the Palantir team that owns the product [9]. That is excellent security engineering. It is also a vendor-held release train running inside your perimeter. Operating a box is not the same as governing what runs on it. You hold the keys to the house. Palantir holds the keys to the renovations.
Verdict. Partial. You run the hardware. Palantir runs the release train.
03 Technological
The question. Can you audit, fork and self-host the stack without the vendor?
The story. Nemotron is open. You keep the fine-tuned weights [2]. Regna’s whole premise is that customers own the specialized model [4].
Under pressure. AIP, Ontology, Foundry and Apollo are proprietary [1][2]. Your ontology lives in Palantir’s schema. Your agent logic lives in AIP. Your deployment lives in Apollo. Fork the model? Yes. Fork the thing that makes the model useful to your organization? No. And every Regna fine-tune is trained on mission data organized by an ontology you cannot take with you. The weights you own were shaped by a structure you rent.
Verdict. Not Demonstrated at the layer that holds the alpha. Open weights in a closed OS is a nice apartment in someone else’s building. Karp wants you to own the means of production. He means the GPUs.
04 Legal
The question. Which jurisdiction governs access, and which contract governs the code?
The story. For a US agency on air-gapped US-built hardware, this is as contained as it gets [2].
Under pressure. For everyone else, the license is the jurisdiction. France’s DGSI renewed its Palantir contract in December 2025 and announced its exit to ChapsVision six months later [11], with the Prime Minister saying France can no longer depend on partners capable of “turning off the access tap.” [10] Palantir’s response: the contract remains in force [11]. That is what the Legal pillar sounds like at 3 a.m. Germany’s domestic intelligence agency had already picked ChapsVision [12], and Berlin ruled Palantir out for military decision support in April [7]. In London, Palantir is challenging City Hall in court for letting concerns about a vendor’s values influence a procurement decision [13]. Read that one twice. The vendor’s position is that its values are none of your business. A bold stance for a company whose CEO wrote a book about values.
Verdict. Partial for US buyers. Not Demonstrated for buyers outside the US, where the OS license is the controlling legal instrument.
05 Financial
The question. Is the cost predictable, and can you leave?
The story. In July, this earned a check. Open weights on owned GPUs kill the per-token meter, and Karp’s “wealth tax” rant against usage pricing is the Financial pillar at full volume [2].
Under pressure. The meter did not disappear. It moved up the stack. Palantir’s Q2 net dollar retention was 157%, meaning existing customers spent 57% more than a year earlier, and US commercial revenue grew 149% while the customer count grew 35% [6]. That is financial alpha. It is just flowing to Denver. The US Army consolidated 75 Palantir contracts into one ten-year enterprise agreement with a $10 billion ceiling [14]. A smart procurement move, and a decade-long commitment to one OS. On the way out, the bill gets worse. The NHS is reportedly studying how to trigger the break clause on its £330 million Federated Data Platform contract without inviting a legal challenge, and the precedent haunting Whitehall is Fujitsu: a 2008 termination, a decade of litigation and a £700 million payout to Fujitsu [15].
Verdict. Not Demonstrated. You canceled the token tax and signed a platform tax. Karp calls usage pricing a wealth tax. A 157% retention rate is what one looks like from the customer’s side of the invoice. On the same Q2 call, Palantir described its model as non-parasitic [18]. The Financial pillar measures who keeps the alpha, and the retention rate answers that.
The score, and why it is worse than it looks
One Demonstrated, two Partial, two Not Demonstrated. On a report card that is a C minus, the kind that comes home with a polite note. In practice it is worse, because every pillar routes through the same layer. Territorial covers where the data sits, but the OS decides what runs on it. Operational covers who holds the keys, but the OS decides what changes. Financial covers the meter, and the OS is the meter. Lose the OS and you have not lost one pillar out of five. You have lost the load path all five stand on.
The standard: the customer must own the OS
This is not a Palantir problem. It is a category problem, and Palantir is simply the best-built example of it. So here is the test, and it applies to every vendor selling an agentic or sovereign OS, including the company I run.
- Source in hand. Do you hold the source code of the orchestration layer under a license that survives the end of the vendor relationship?
- Run without us. Can your own team build, patch and release it if the vendor disappears tomorrow?
- Logic that travels. Do your ontology, workflows, agent policies, evals and audit trail export in an open format that another runtime can execute, not merely read?
Three yeses and you own the alpha. Anything less and you are leasing it, however good the hardware looks. Notice what the test does not ask: whether the vendor’s engineers stay. Keep them as long as they earn it. Deployment is worth paying for. Captivity is not.
The executive TCO
- Inference. Owned compute on open weights. A genuine win, and Palantir earns credit for making it the default.
- Platform. A license that compounds with your success. The more of your organization you encode in it, the more it costs to leave and the more it costs to stay.
- Exit. The unpriced liability. France needed a decade and a domestic champion. The NHS is weighing legal risk against a nine-figure precedent. If it is not in your business case, it is in your vendor’s.
The cheapest year of a closed OS is year one. Year two is when you meet the renewals team.
Monday
- Map where your logic lives. List every object model, workflow, agent permission and eval, and the system that holds each one.
- Send your OS vendor the three questions above, in writing.
- Price your exit now. Put a number on migrating your logic in twelve months and add it to the business case. It is cheaper than hearing the number from a judge in 2036.
- Split the layers in procurement. Buy models and hardware as swappable parts. Buy the OS as an asset you own.
- Keep what works. Just stop calling a licensed OS sovereign.
The close
On any other day, this is where I would call Alex Karp the arsonist who moonlights as the fire department. I have run that headline before: on Broadcom, the arsonist now selling smoke detectors [24], and on the frontier labs, the arsonists now selling fire trucks [25]. Set the fire, sell the hose, bill for the water.
Today I am going to surprise every reader of this newsletter, and possibly my editor, and give him credit.
Karp is not masking anything. Every symptom he names maps to a pillar, and every one is real: the token meter on Financial, the rented model on Technological, the alpha walking out the door to the frontier labs. And look at the trade he is actually offering. He gives away, or close to it, the one thing that genuinely costs him money: his engineers’ time. The AIP bootcamp puts forward-deployed engineers on your own data for five days at a price buyer-side advisors describe as cheap or free, and they note why: the ontology you build that week encodes your business logic into Palantir’s data model before anyone has agreed a production price [23]. In exchange, he takes a license on the one thing that should be free to you as the customer: the code that encodes your own business. The expensive thing goes out the front door at cost. The cheap thing comes back through the side door at a margin, indefinitely. A 47% GAAP operating margin tells you how the side door is doing [6].
That is not a con. It is a trade, made in plain sight. Karp is not the arsonist in this story. He is the horse trader, and he is better at it than you are.
Rule one of horse trading: when the saddle is free, ask who keeps the horse.
So here is the trade, pillar by pillar:
- What you get. Deployment speed, engineers on site, a working use case in a week. That is real Operational value, and it is worth paying for.
- What you give. The source of your logic (Technological), the release train (Operational), a license that becomes your jurisdiction (Legal), and a meter that grows with your success (Financial).
- What a good horse trader asks for. Pay for the engineers, keep the code. Source in escrow at minimum, an open export of the ontology, a license that survives the relationship, and an exit price written into the contract before the bootcamp, not after.
The rule has always been simple. Own the core. Stay movable at the edges. Karp will happily sell you the edges and rent you the core. Your job at the table is to make him sell you the core too. He respects a good trader. He is one.
If your OS vendor answers the three questions, send me the answers. I will grade them on the record, with the same standard.
Amit
Amit Eyal Govrin is lead analyst for sovereign AI at theCUBE Research and CEO of Agentcy Labs, which designs and builds sovereign AI platforms for enterprises and governments. His research for theCUBE Research is conducted under theCUBE Research’s editorial standards and is not sponsored or directed by Agentcy Labs. Agentcy Labs may hold commercial relationships with vendors covered in this research.
References
[1] Palantir / NVIDIA, “Palantir and NVIDIA Team to Deliver Sovereign AI Operating System Reference Architecture,” BusinessWire, Mar 12, 2026. https://www.businesswire.com/news/home/20260312795208/en/
[2] Govrin and Vellante, “Open Model, Closed Environment: pushing the Sovereignty Litmus test to the limit,” theCUBE Research, Jul 15, 2026. https://thecuberesearch.com/open-model-closed-environment-pushing-the-sovereignty-litmus-test-to-the-limit/
[3] Palantir / Armada, “Palantir and Armada Partner to Accelerate Sovereign AI Infrastructure,” Oct 1, 2026, via HPCwire (https://www.hpcwire.com/bigdatawire/this-just-in/palantir-and-armada-partner-to-accelerate-sovereign-ai-infrastructure/) and StorageNewsletter (https://www.storagenewsletter.com/?p=305751)
[4] Ondas, “Ondas Announces Regna AI, Powered by Palantir, to Advance Sovereign AI,” Oct 12, 2026. https://ir.ondas.com/press-releases/detail/334/ondas-announces-regna-ai-powered-by-palantir-to-advance
[5] Benzinga, “Ondas Wants to Run Pentagon AI with Palantir-Powered Regna AI,” Oct 12, 2026. https://www.benzinga.com/markets/tech/26/10/62305046/ondas-wants-to-run-pentagon-ai-with-palantir-powered-regna-ai
[6] Nasdaq, “Palantir Technologies Q2 Earnings Call Highlights,” Aug 2026. https://www.nasdaq.com/articles/palantir-technologies-q2-earnings-call-highlights
[7] Ana-Maria Stanciuc, “Nvidia and Palantir are selling a sovereign AI stack, starting with Nvidia’s own supply chain,” TNW, Sep 10, 2026. https://thenextweb.com/news/nvidia-palantir-sovereign-ai-supply-chains
[8] Palantir, “Deploying Across Security Domains,” Palantir Blog. https://blog.palantir.com/deploying-across-security-domains-449c786d92c0
[9] Palantir, “How Palantir Meets IL6 Security Requirements with Apollo,” via Carahsoft. https://www.carahsoft.com/article/palantir-meeting-il6-security-requirements-with-apollo-blog-2023
[10] AFP, “French spies drop AI giant Palantir over US overreliance fears,” Jun 16, 2026. https://zeta.com.pa/ingles/75525-french-spies-drop-ai-giant-palantir-over-us-overreliance-fears.html
[11] Kurrant, “After The Fable 5 Cutoff, DGSI Drops Palantir For France’s ChapsVision,” Jun 2026. https://kurrant.com/kurrantly-news/after-the-fable-5-cutoff-dgsi-drops-palantir-for-frances-chapsvision/
[12] Sifted, “ChapsVision to replace Palantir in major contract with French intelligence agency,” Jun 16, 2026. https://sifted.eu/articles/chapsvision-to-replace-palantir-in-major-contract-with-french-intelligence-agency
[13] Computing, “MPs urge government to cancel Palantir/NHS contract,” 2026. https://www.computing.co.uk/news/2026/government/mps-urge-government-cancel-palantir-nhs-contract
[14] Breaking Defense, “Army consolidates dozens of Palantir software contracts into one deal worth up to $10 billion,” Jul 31, 2025. https://breakingdefense.com/?p=46928
[15] Digital Health News, “Pressure mounts from MPs on Palantir’s role in the NHS,” 2026. https://www.digitalhealth.net/?p=308053
[16] BetaNews, “Mistral opens trillion-parameter Large 4 preview, weights due in October,” Oct 8, 2026. https://betanews.com/article/mistral-large-4-preview-open-weights/
[17] Reflection, “Introducing Beam: Reflection’s 501B open-weight model,” Oct 5, 2026. https://reflection.ai/blog/introducing-beam
[18] GuruFocus, “Palantir Technologies Inc (PLTR) (Q2 2026) Earnings Call Highlights,” Aug 2026. https://www.gurufocus.com/news/9000347/
[19] InfoQ, “Stripe Engineers Deploy Minions, Autonomous Agents Producing Thousands of Pull Requests Weekly,” Mar 2026. https://infoq.com/news/2026/03/stripe-autonomous-coding-agents/
[20] morluto, “REA: Reverse Engineer Anything,” GitHub, MIT license, accessed Oct 14, 2026. https://github.com/morluto/rea
[21] Context Studios, “Forward Deployed Engineers: Why the Palantir Model is Becoming the Operating System of Enterprise AI,” 2026. https://www.contextstudios.ai/blog/forward-deployed-engineers-palantir-model
[22] andrew.ooo, “IBM FDU vs OpenAI Deployment Company vs Palantir FDE,” May 2026. https://andrew.ooo/answers/ibm-fdu-vs-openai-deployment-company-vs-palantir-fde-may-2026
[23] Redress Compliance, “Palantir AIP and Foundry negotiation,” 2026. https://redresscompliance.com/palantir-aip-foundry-negotiation
[24] Govrin, “The Arsonist Now Sells Smoke Detectors,” theCUBE Research, Sep 7, 2026. https://thecuberesearch.com/the-arsonist-now-sells-smoke-detectors/
[25] Govrin and Vellante, “EMERGENCY EDITION: The Arsonists Are Now Selling Fire Trucks,” theCUBE Research, Sep 15, 2026. https://thecuberesearch.com/emergency-edition-the-arsonists-are-now-selling-fire-trucks/
