On July 1, Alex Karp went on CNBC’s Squawk Box to talk about a Palantir–Nvidia partnership. Within two minutes he’d torched the entire frontier-AI business model instead.
His accusation: the leading labs are running a double-dip. They charge enterprises for tokens, and then they absorb those same enterprises’ proprietary data to improve the models they’ll sell to everyone else — including the customer’s own competitors.
Karp claimed frontier AI companies like OpenAI and Anthropic siphon customers’ valuable information while delivering questionable value. His word for it was blunt: they are “stealing the weights and alpha” of his clients’ businesses.
The mechanism is worth sitting with, because it’s structural, not accidental. Every enterprise that runs its confidential documents, customer conversations, and financial models through a frontier API is potentially teaching that model how to do its job.
The vendor collects the token fee and the compounding intelligence about how that business actually works. As Karp framed the private mood of American executives: they are “livid,” as “they are paying for tokens that create no value.”
The market took it seriously. Karp declared “the jig is up” in a CNBC appearance that sent Palantir shares up 9% — while several other AI names dipped the same day.
Now, the obvious caveat: Karp sells the alternative, so he’s talking his book. But note who agreed. Moor Insights CEO Patrick Moorhead told every CEO to watch the clip, saying Karp was voicing risks “few will say” out loud.
The reason the argument landed is that CIOs have been muttering it privately for a year. Karp just said it into a microphone.
Here’s where it gets interesting for anyone thinking about where the money goes next.
The partnership Karp was supposed to be discussing is Palantir integrating Nvidia’s Nemotron open models into its Sovereign AI platform — letting government agencies and enterprises run AI on their own infrastructure, train on their own data, and keep the resulting model weights. Data never leaves the building.
And Nemotron isn’t a watered-down “open-weights, secret-sauce-withheld” release. Nvidia gave away the whole kit. They gave away a family of frontier-class AI models, complete with the training data, the training recipes, plus 10 trillion tokens of open datasets. No API paywall. No waitlist. Jensen Huang’s framing: “Open innovation is the foundation of AI progress.”
The strategic logic is the tell. Nvidia doesn’t need to own the model layer — it needs the model layer to be cheap, abundant, and everywhere, because every open model running anywhere runs on Nvidia silicon. A world of a thousand specialized open models sells more GPUs than a world of three closed APIs.
So Nvidia is deliberately commoditizing the thing OpenAI and Anthropic need to keep scarce. A $183B+ hardware giant is subsidizing the open-source counterattack against the closed labs — and it’s rational for them to do it.
The early-adopter list already reads like an enterprise roster: Accenture, CrowdStrike, Cursor, Deloitte, EY, Oracle Cloud Infrastructure, Palantir, Perplexity, ServiceNow, Siemens and Zoom are integrating models from the Nemotron family.
The old mental model — closed models are the Ferrari, open models are the reliable sedan for people on a budget — is out of date, and the numbers say so.
Late 2023: the best closed model scored around 88% on MMLU while the best open alternative managed roughly 70.5%, a gap of 17.5 percentage points.
Early 2026: that gap is effectively zero on knowledge benchmarks, and single digits on most reasoning tasks. On the Chatbot Arena human-preference leaderboard, one analysis puts the gap at 1.7% — statistical noise for most use cases.
Open models now match or beat closed models on knowledge (MMLU), math, and even graduate-level science. Closed models maintain a lead on production coding, overall human preference, and complex agentic tasks.
And then the part that actually moves budgets — the economics have inverted. Open models close 70-90% of the capability gap at 5-10× lower per-token inference cost. One worked example: a RAG pipeline that costs roughly $2,275 per month on a frontier closed API runs for about $168 per month on an optimized open-weight model — a 93% cost reduction with minimal performance trade-off at scale.
Nearly identical output. A fraction of the cost. Your data stays yours. For a CFO, that’s not a close call.
Here’s the piece the market is really under-pricing, and it’s why this is a theme and not just a trade.
Sitting underneath the open-vs-closed fight is an unresolved, enormous question: who actually owns the intelligence inside these models? The closed labs built their moats partly on data whose ownership was never settled — and that bill is now coming due.
In 2025, Anthropic agreed to pay authors $1.5 billion to settle claims over books used to train Claude — the largest copyright recovery ever, and the first of its kind in the AI era. The exposure that forced the settlement was staggering: with statutory damages potentially reaching $150,000 per work, Anthropic was staring down theoretical liability exceeding $70 billion. And that door isn’t closed — six authors have filed fresh individual suits against Anthropic, OpenAI, Google, Meta, xAI, and Perplexity, seeking $150,000 per title from each defendant.
Now connect the two threads. Karp says the labs absorb enterprise IP through the token pipeline. Authors and publishers say they absorbed creators’ IP through the training pipeline. Same structural problem, two different victims: the closed model’s value was built on other people’s property, and the ownership question was deferred, not answered.
Open, sovereign models are the architectural answer to both. When you run an open model on your own hardware, train it on your own data, and own the resulting weights, there’s no siphoning and no IP ambiguity.
The intelligence — and the liability, and the upside — stays with whoever created it. That’s a far bigger addressable shift than “we saved 90% on inference.” It’s a re-drawing of who captures the value of AI itself.
One last thing, and it’s the crux of the premium thesis.
OpenAI was founded on the open thesis. It began as a nonprofit devoted to developing open-source AI technology for the betterment of humankind. The name was not ironic. Then came the pivot to a capped-profit arm, the Microsoft deal, and finally, in October 2025, the completed metamorphosis into the closed-source, profit-seeking juggernaut it is today, with a half-trillion-dollar valuation. A co-founder’s own retrospective on the early open-sharing approach: “We were wrong.”
Here’s our contrarian read: they abandoned the right thesis at the wrong time. They walked away from “open” precisely as the evidence began stacking up that open would win — as the benchmark gap collapsed to noise, as the cost advantage blew out to 10×, as the IP time-bombs under the closed labs started detonating, and as the single most important hardware company on earth started giving frontier-class open models away to sell more chips.
We think open-source is on track to win the enterprise — not everywhere, not overnight, but structurally, where it matters most. And the ways to position for that are not obvious, not priced in, and not limited to the one chipmaker everyone already owns.
The full thesis, and exactly how we’re positioning for it:
The complete bull case for why open beats closed in the enterprise — the four structural forces, and the one scenario that breaks the thesis
The picks-and-shovels map: who captures value when models become a commodity — across hardware, the sovereignty/authorization layer, inference, and specialized fine-tuning
The IP wildcard — how the copyright reckoning could re-rate the closed labs, and which names are most and least exposed
What we own, what we’re watching, and our entry levels — the concrete portfolio, not just the narrative
The contrarian short side — where the closed-model premium is most vulnerable to being repriced
The gap has closed. The question is who profits from what comes next. → Unlock the full playbook…
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.