← IndexAlex Karp spent twenty minutes on CNBC two weeks ago being angry on behalf of American business. The anchor interrupted to tell him he sounded angry, and he denied it in the most Karp way possible: “This is the voice of American business that is being channeled through me.” Somewhere in the middle of the theatrics he said the thing that mattered: at every enterprise he deals with, “these people are livid, they’re like, ‘I am paying for tokens that create no value.’”
He’s right about the anger. I hear it in the same rooms he does. The executives I talk to are terrified three ways at once: about what they’re spending, about what they’re giving up, and about what happens if they shut it all down and fall behind. Nobody — CFO, CIO, CTO — wants to be remembered as the one who turned off the token fountain and killed the company’s progress. So the spend keeps flowing. So does the fear.
But their problem sits one level below budgeting. Every one of them would happily pour unlimited tokens into the things that compound value, and next to none into helping people plan their weekends, pick a restaurant, or rebuild the CRM because somebody didn’t like the user experience. What they can’t do is tell one from the other. Nobody can disambiguate value inside a token stream. That’s the real complaint hiding under Karp’s quote.
And in my last essay I gave the anger a diagnosis: a token that creates no value is a loop cycle with zero retention. You paid for tempo and kept nothing. So when Palantir followed the TV appearance with a white paper — Institutional Sovereignty in the Age of AI, fifteen steps for governments and companies — I read it expecting to disagree with all of it.
I didn’t. That’s what makes it worth writing about.
Where Karp and I agree
Buried in Palantir’s fifteen steps is a sentence that could have come straight out of my last post: “the institution that generates signal must also capture it.” Step VI tells you to own the model flywheel. Step XV tells you to own the context flywheel. The paper’s whole economic frame — “sovereignty is your alpha… your ownership of the value you create” — is orientation capital wearing a different suit. Karp, Nadella, and a dead Air Force colonel named John Boyd are all circling the same animal: the compounding loop, and the question of who owns what it learns.
Even the vocabulary converged. Palantir’s core message, in one reading making the rounds: models are rented; your data, infrastructure, and ontology must be owned. I’ve been saying a shorter version. The model is rented. The loop is owned.
The convergence runs deeper than the white paper. Palantir’s companion nine-point manifesto coined a word for the disease — tokenmaxxing, “the addictive feeling of false progress,” spend that rewards disposable scripts over solid engineering. Read that through Boyd and it’s a zero-retention loop described from the inside: cycles that burn tokens and deposit nothing. They can see the pathology clearly. Hold that thought for when we get to their cure.
So the diagnosis is shared, and the destination is shared. The disagreement is the route. And the route is most of the paper.
Count the steps
Read the fifteen steps and sort them by posture. Zero data retention agreements. Assurance-based hardware decisions. Air-gapped deployments for sensitive workflows. Verified compute. Granular permissions. Audit and logging. Adaptive cybersecurity. Attestation for the compute you rent. “Read-only AI until the control layer passes.” A decision tree that ends, over and over, in hold: human-in-the-loop only.
By my count, thirteen of the fifteen steps are about containment — keeping what you have from leaking out. Two are about compounding. The paper diagnoses a loop problem and prescribes a vault.
I understand why. Palantir’s fear catalog is real: per-token rent-seeking on your highest-margin workflows, labs entering your vertical and owning your customers, your operational knowhow resold to competitors as generic capability. Those risks have names and case studies now — ask Figma, whose CEO says Anthropic was “not consistently candid” about the scope of the competing design tool it built while the two collaborated — and David Sacks has been documenting the pattern in public. If your mental model of AI is an adversary wearing a vendor’s badge, fortification is the rational response.
But go back to Boyd. A firm’s advantage is tempo times retention — how fast the loop runs, multiplied by how much each cycle deposits. The vault strategy maximizes retention of the orientation you already have while strangling the tempo that creates any more of it. Air gaps don’t run cycles. Read-only AI deposits nothing. And a mandatory human approval on every consequential action rebuilds, as policy, the exact bottleneck Boyd said maturing orientation is supposed to remove. His term was implicit guidance: when the internal model is rich enough, you stop stopping to decide. Palantir’s decision tree makes stopping to decide a permanent architectural feature.
A vault protects. A loop compounds. Protection without compounding is a museum: everything preserved, nothing growing, while competitors who kept their loops spinning ride an exponential you opted out of. Boyd would have recognized the fortress strategy instantly, because he spent a career explaining why it loses to the faster loop — you don’t have to breach the walls if you can out-cycle whoever is inside them.
The anger, measured
Here’s the part Karp gets closest to and still misses. “Tokens that create no value” is a measurement claim. It has a denominator. Which tokens? Which workflows? Creating no value for whom, against what baseline?
Companies are staring at an AI bill precisely as informative as a power bill. You bought ten thousand kilowatt-hours last month. Great. What did the electricity serve? Nobody knows. (Say it in your best Nate Bargatze voice — it’s the only register that does the absurdity justice.) Tokens work the same way. You spent millions on millions of tokens… to do what? For whom? The ROI question everyone asks comes second. First you have to reconstruct what was actually bought and what was actually achieved, and today’s bill answers neither. Somehow the industry took an itemized cloud bill that was already too complex to reason with and replaced it with a single number nobody understands.
Let me be precise about the target, because Karp goes one step further and asks why labs charge for tokens at all. Wrong target. Electricity is metered per unit too, and nobody demands the utility price it per delighted customer. Meters are honest. Meters with no circuits labeled are useless — and that’s the difference between resenting your bill and reading it.
The scale of the blindness is measurable: enterprises report allocating 30–36% of their cloud budgets to AI workloads, while the AI spend they can actually trace on their bills sits closer to 2.5% (Cloud Economics Pulse). You cannot be sovereign over value you can’t even locate. And anger without attribution just becomes fortification — you can’t fix what you can’t see, so you wall it off instead. The enterprises that get past livid will be the ones that treat the AI bill as what it secretly is: a report on which cycles deposited orientation and which ones evaporated. Kill the zero-retention workflows. Feed the compounding ones. Give the CFO the context to say yes with confidence and no without fear. That’s the fix for Karp’s complaint, and no ZDR clause, air gap, or attestation ceremony delivers it, because the lab was never stealing your value. You were failing to capture it.
Whose vault is it, anyway
One more thing has to be said, and Karp said it about himself first: he’s talking his book. He admitted as much on air. What he didn’t say is how precisely the book maps to the prescription. The paper’s remedy — a sovereign control layer, an owned ontology, the connective tissue between your data and whatever model you rent — has a name. The name is Foundry, with AIP sitting on top of it. Palantir wrote fifteen steps for escaping dependency on AI vendors, and the fifteenth step ends at Palantir’s front door.
Notice the economics of the swap. Karp rails against labs charging per-token rent on your highest-margin workflows; his paper literally illustrates the fear as “rent-seeking on workflows.” Foundry charges platform rent on those same workflows. The anger is about the meter. The prescription changes whose meter is running. And the timing removes any doubt about the genre: two days before the CNBC appearance, Palantir announced an expanded NVIDIA partnership packaging open Nemotron models for “sovereign environments.” The rant shipped with a product launch attached.
Europe, meanwhile, is already running the sovereignty argument back at its author. France dropped Palantir for homegrown alternatives, Germany’s military keeps its distance, and Switzerland’s army rejected the company outright — over, of all things, sovereignty risk. By the paper’s own definition, the sovereignty vendor is somebody’s sovereignty problem. That’s the trouble with locating sovereignty in a vendor’s stack: it’s always somebody’s stack.
Self-interest doesn’t make the argument wrong. It makes the argument incomplete in a specific, predictable direction. If sovereignty means owning the stack — the weights, the GPUs, the ontology platform — then sovereignty is something you buy, and the white paper is a shopping list where aisle one has a moat around it. Swapping dependency on a frontier lab for dependency on a control-layer vendor changes your landlord. It doesn’t make you a homeowner.
Boyd puts sovereignty somewhere no vendor can reach: in orientation. The synthesized model of how your business actually works — your measures, your evals, your institutional memory, the judgment encoded in your loop. You can own every GPU in the building and still think with someone else’s model of the world. You can rent every model you touch and stay sovereign, because what compounds is yours. The stack is where orientation lives, and owning it can be worth every penny. Just don’t confuse the house for the family.
The trade Karp forgets
Karp makes a strong case that every company must own its alpha. He skips the part where every company will give up alpha to get more alpha. That trade is already running on every desk in your building: each prompt that carries a slice of your operational context into a rented model is a small export of orientation, exchanged for tempo you couldn’t generate alone. Sealed borders stop the leak and the learning in the same stroke.
So the real discipline is pricing the trade, and the price has a name from the last essay: retention. Alpha surrendered against alpha compounded. A company that measures its retention rate can run the exchange aggressively — export context where the cycle deposits more than it leaks, cut the workflows that bleed orientation for nothing. A company that can’t measure the exchange gets to pick between Karp’s vault and blind faith. Most are picking blind faith with a livid expression.
Keep five steps, add the loop
If a company asked me what to actually do with Palantir’s paper, I’d say: keep the steps that serve the loop, skip the ones that only serve the fear.
Keep model liquidity (step V) — swappable models are how you survive the leaderboard churn, and Nadella calls that switch “the key test of your control and sovereignty in the era ahead.” Keep the misaligned-incentives audit (IV). Keep granular permissions and logging (XI, XII) — loops need brakes engineered in, not bolted on. And absolutely keep the two flywheel steps (VI, XV); they’re the best pages in the paper.
Then add what the paper leaves out:
- Measure retention rate. Instrument every AI workflow to answer one question: what did this cycle deposit that we own? Tokens against outcomes, not tokens against dread.
- Run tempo deliberately. Sovereignty grows with cycles. The read-only, approval-gated posture forfeits the compounding that was the whole point of owning the flywheel.
- Retire approvals as orientation deepens. Implicit guidance is the maturity signal — fewer humans in the loop per outcome, on purpose, with the audit trail to prove it’s safe.
There’s a clock on this, and it’s faster than the paper assumes. “Read-only AI until the control layer passes,” with a human approval on every consequential action, is a bet that the fully autonomous cloud stays comfortably far away — that systems which write, deploy, operate, and buy without a human in the critical path remain someone else’s decade. I’ve published my bet: as soon as 2028. A sovereignty program built around a permanent human veto is a cautious-line strategy, and the cautious line doesn’t end in a managed decline. It ends in a cliff you don’t see coming.
Alpha was never a thing
Here’s where I stop borrowing Karp’s vocabulary and start correcting it. He took alpha from finance and quietly turned it into a possession — something you inventory, something you vault. The quants who coined the term would wince. Alpha is a rate: excess return over the market, real only in motion, measurable only in flow. And every quant knows the corollary. Alpha decays. The moment a strategy diffuses, it degrades into beta, the return everybody gets. Palantir’s nightmare scenario — your knowhow turned into “generally available intelligence” — is alpha decay wearing a trench coat. The paper describes the decay perfectly, then prescribes a vault, as if decay respects walls. It doesn’t. A vault full of yesterday’s knowhow is beta with a security budget.
Run the definition forward and the furniture rearranges itself. The data, the ontology, the tribal knowledge — residue, the exhaust a moving loop leaves behind. Your alpha is the loop itself, the one at the core of your business, and its worth is set by a single number: the rate at which it compounds. The higher the rate, the higher the value. This is the seismic shift hiding under the whole debate. AI rewrote the definition of value, and the new definition is a compounding rate. Judge every wall you’re tempted to build against it, because a wall that slows the loop destroys the thing it was poured to protect.
Karp’s paper is written for institutions standing at the harbor, deciding how high to build the seawall. The contest has already moved to open water, where it’s decided by seamanship — who runs the richer loop, faster, and keeps what every cycle teaches. Guard your data. Negotiate the ZDR. But a decade from now, nobody will ask how thick your walls were. They’ll ask what your loop learned while everyone else was pouring concrete.
Vaults hold beta. Loops make alpha.