Own the Means of Production

On July 1, 2026, Alex Karp went on CNBC’s Squawk Box and said the quiet part in a register most enterprise CEOs reserve for closed rooms. He was there to talk about Palantir’s new partnership with Nvidia. He spent the segment accusing the entire frontier-AI industry of running a con.

The labs, he argued, are “irresponsibly overselling” their models — charging enterprises for “tokens that create no value” while quietly absorbing the one thing that makes those enterprises defensible. He has a word for that thing, and it is a good one. He calls it your alpha: the proprietary data, processes, and hard-won edge that let a business earn more than the market average. And then he named the fix in language he almost certainly knew would travel.

A Palantir CEO borrowing Marx to sell a Nvidia deal is its own kind of theater, and we’ll get to the theater. But strip the performance away and there is a real economic claim in there — one that half his critics and most of his fans have gotten wrong in the same way. So let’s do the thing nobody on cable has time for. Let’s check whether he’s right.

He is. Halfway. And the missing half is the interesting one.

The line that made the clips was “something has gone completely wrong.” When Becky Quick observed that he sounded angry, Karp waved it off: “No. This is the voice of American business that is being channeled through me.” He claimed the CEOs he talks to are privately “twice as livid” about the direction of enterprise AI. Palantir’s stock jumped roughly 9% on the segment and the surrounding Nvidia news. The market, at least, heard something it agreed with.

Underneath the theater, his argument has a specific shape. He says the technical buyer wants control over four things — compute, models, data stack, and alpha — and that the dominant way AI is sold today, metered access to a closed frontier model through someone else’s infrastructure, quietly strips all four. Every prompt, every document, every retrieval-augmented query routes your proprietary workflow through a third party. You rent the intelligence; you hand over the context that makes it valuable; and you get back a bill for tokens.

That is the accusation: you are paying to lose the thing that made you worth more than average.

The crude reading of Karp — “OpenAI and Anthropic are training their models on your prompts” — is mostly contractually false. At the enterprise and API tier, the major labs explicitly commit not to train on business data by default. If your objection is literal data theft, you can usually point to the terms of service and win the argument. That is the half where Karp is overstating for effect, and he surely knows the audience won’t check.

For a regulated or sovereign buyer — defense, intelligence, critical infrastructure, the customers Palantir actually serves — there’s a fourth, plainer problem: routing sensitive data through a third party is a governance risk on its own, independent of any training question. So Karp is not hallucinating. He has taken a real structural tension, sharpened it into the word “alpha,” and timed it to the day of a product launch. That’s not a lie. It’s a sales pitch built on a true thing — which is the most durable kind.

But to see why the true thing is true, you have to stop talking about theft and start talking about supply and demand. Because everything Karp is describing is what happens when the cost of producing something suddenly collapses.

Here is the mental model that makes the whole argument legible. AI is not, primarily, a demand technology. It doesn’t make the world want more things. It is a supply technology: it makes producing things — code, copy, analysis, cognitive labor — radically cheaper. On a price-versus-quantity chart, that shoves the supply curve to the right. What that does to price, volume, and profit depends entirely on the shape of demand.

Start with the case Karp is really worried about — the one that describes most “AI content” today. When demand is inelastic — when the world only wants so many generic blog posts, stock images, or boilerplate translations — the new supply has nowhere to go but down in price. The equilibrium slides down a steep wall.

Read the diagram: supply jumps from S₀ to S₁, and the price craters (P₀ to P₁) while quantity barely moves (Q₀ to Q₁). The producer surplus — the profit — is destroyed. It doesn’t disappear into thin air; it leaks out to buyers as lower prices. This is a gift to consumers and a catastrophe for anyone whose business was selling the commoditized thing. It is the economic shape of “paying for tokens that create no value”: not that the tokens do nothing, but that whatever they produce is now too cheap to defend a margin around.

The same supply shock can do the opposite, and this is where the “AI is deflationary and value-destroying” doomers get it wrong. When there is a vast pool of demand sitting below the old price floor — buyers who were priced out entirely — cheaper supply doesn’t crash the price. It unlocks the buyers.

Custom software is the cleanest example. Bespoke internal tools used to cost $80,000 and six months, so a million small, idiosyncratic needs never got built — the “I wish we had a little app that just did X” backlog that never cleared. That wasn’t a lack of demand. It was demand suppressed by a supply-cost floor. Collapse the cost and the whole tail crosses the viability line at once.

With demand this elastic — flat — the supply shift is absorbed as volume: quantity explodes while price barely moves. The pie genuinely grows. This is real value creation, not a mirage.

Now play it forward. Adoption isn’t a one-time move. Everyone gets the same cheaper supply, and they get it again next year, and the year after. Each round, the supply curve shifts right again — S₀ to S₁ to S₂ — and the equilibrium marches down the demand curve toward marginal cost.

This is the Red Queen problem, straight out of Through the Looking-Glass: “it takes all the running you can do to keep in the same place.” Every firm adopts more AI just to hold position. Nobody pulls ahead, because the tool is available to all, and the margin is competed toward zero. The CEO who expected AI to mean “big dollars” instead gets “we spent a fortune to remain exactly as competitive as before.” Adoption becomes mandatory and unprofitable — table stakes, not an edge.

That gap — between the promised windfall and the treadmill reality — is the disillusionment you can hear in Karp’s voice. He is not describing a swindle. He is describing competition doing exactly what competition does to a productivity technology: hand the gains to customers.

The surplus that leaves the sellers doesn’t vanish. It flows two places: to buyers, as lower prices, and as rent to whoever owns the one scarce input everyone still needs. In a world where cognitive output is commoditized, the durable money migrates to the bottleneck — and the bottleneck is almost never the firm that merely adopted the tool.

What stays scarce? Compute (the reason Nvidia collects a toll no matter who wins downstream). Energy and datacenters. Distribution and trust. And, most durably of all, proprietary data and the physical means of making things. AI is a supply shock to bits. It does almost nothing to the cost of atoms — of a $30-billion chip fab, a certified jet engine, a cured wheel of cheese, a poured foundation. The moats that survive the AI wave best are the ones made of atoms, capital, certification, and tacit process knowledge that no model can download.

This is exactly what Karp means by “own the means of production,” and on this he is simply correct. If your edge is a scarce, hard-to-replicate asset — your own data, your own process, your own plant — then AI is a tool you point at your moat to deepen it. If your edge was just “we can produce the commoditized thing,” AI is the flood that fills your trench. The strategic question is not “should we adopt AI.” Everyone will. It’s “do we own a scarce input, or are we renting our intelligence and hoping the treadmill slows down.”

You don’t have to go to Silicon Valley to find the scarce-input economy. Drive an hour in any direction from Wausau. The stretch of central-to-northern Wisconsin that UpNorthDigital calls home is a dense cluster of mid-market manufacturers whose defensibility lives almost entirely in atoms.

The anchor is Greenheck, in Schofield — the global leader in commercial air-movement and ventilation equipment, grown from a 1947 sheet-metal shop into a campus of 2,000-plus people with thousands more across the state. Its moat is decades of application-engineering know-how, scale, and a made-to-order catalog no model can regenerate. Down the road, Kolbe builds premium windows and doors on brand and craft; County Materials and Wausau Tile win on the physics of freight — concrete is so heavy relative to its value that every plant is a near-monopoly inside its own shipping radius. You cannot import that competition. AI cannot download it.

Marathon County grows roughly 95% of all U.S. ginseng — one county, a few hundred growers, exporting most of the crop to China for something like $40 million a year. That is a terroir moat in the most literal sense: soil, climate, and generations of cultivation know-how that a competitor genuinely cannot replicate. It is also the perfect cautionary tale, because ginseng’s Achilles’ heel isn’t supply — it’s a single concentrated buyer. When trade policy with China turns, an unbeatable physical moat gets throttled on the demand side. Same lesson as the diagrams, wearing a straw hat.

And the region’s hardest chapter proves the rule from the other direction. Central Wisconsin was built on paper — and when Verso idled the Wisconsin Rapids mill in 2020, roughly 900 jobs went with it. That mill had a massive capital moat. What it didn’t have was demand: graphic paper is in secular decline, and no amount of capital intensity saves you when the market for your output evaporates. The survivors pivoted to packaging, tissue, and specialty films — same atoms, different demand curve. Meanwhile the dairy and food processors keep pouring capital into the region precisely because milk is perishable and has to be turned into cheese near where it’s made. Logistics is the moat.

So back to the theater. Karp gave that interview on the day Palantir announced a “sovereign AI” stack with Nvidia — on-prem, air-gapped, own-your-weights. He defined the disease in the exact shape of his cure. “They’re after your alpha” is a genuinely true structural argument and a genuinely self-interested product pitch, at the same time, and both facts are worth holding.

It’s worth remembering that “own the means of production” via Palantir still means routing your most sensitive data and workflows through Palantir. And Palantir runs frontier models under the hood, and aligns with Nvidia — the one company that collects a toll no matter which side wins. The real choice on offer was never “own versus rent.” It was which intermediary you depend on.

But the economics survive the sales pitch. AI reliably shoves the supply curve right. Who keeps the money is decided by the elasticity of demand and by who owns the scarce input — almost never the firm that simply adopted the tool. Karp is right that the alpha is the thing to protect. He’s selling you one way to protect it. The cheaper way is to already own something that can’t be commoditized — a dataset, a process, a plant, a wheel of cheese aging in a cave in Wood County.

Own the means of production. He’s not wrong. He just isn’t the only landlord.