The Industry Just Admitted the Racks Were Right
Local LLMs and on-premise hardware racks are the way of the future, and the tech industry and its investor class are only just now coming around to that fact. On June 29, Palantir and NVIDIA launched a sovereign AI stack built to run NVIDIA's open Nemotron models inside air-gapped and classified environments, weights and data under the customer's control. Two days later, Palantir CEO Alex Karp went on CNBC and said "something has gone completely wrong" with how the frontier labs sell AI, tore into per-token pricing bleeding American business, and dropped a nine-point AI sovereignty manifesto on anyone who'd read it. The company touting sovereign AI is now telling its customers what we've been telling ours: own your compute, your models, and your data stack.
Palantir's project and direction are admirable, and they mean more coming from an organization that doesn't need to look over its shoulder to see where next quarter's revenue comes from. When the firm whose whole business is government-scale software says sovereignty means owning the weights, the argument has left the hobbyist forum and walked into the boardroom.
Right Side of History, Light on Follow-Through
It's not just Palantir. Even the hyperscale visionaries, the folks whose business is selling metered inference by the rack-hour, are preaching open source now. Cerebras CEO Andrew Feldman sat down with All-In's Liquidity show yesterday and said the quiet parts out loud, in order: the United States doesn't have a serious homegrown open-source alternative while Chinese models like GLM and Kimi set the pace; regulated industries in finance and healthcare need on-premises, domestically controlled models their HIPAA and FINRA obligations can live with; and, in three words that cover the entire sovereignty argument, "nobody likes being dependent." That's from a man sitting on a $25 billion backlog for chips that aren't even finished being built, whose hardware runs OpenAI's own open-weight model faster than anybody. Good on him. It's the right stance and the right side of history to be on.
But the YouTube-circuit lip service only goes so far, even when the diagnosis is dead-on. For all the magnanimous feel-good confessions punctuating social media commentary these days, talk needs the bite of palpable action and enforced policy to feel real. And palpable action has a shape: release weights under licenses that allow real commercial use, publish the data recipes and training code instead of a victory-lap model card, commit to a cadence instead of a one-off, and above all, fund the work like you mean it. A truly meaningful gesture from the golden-tongued would be a generous philanthropic project that puts their money where their mouths are.
The Combination Nobody Has Funded
Here's the gap that gesture would fill. American open-weight models exist: NVIDIA's Nemotron, OpenAI's gpt-oss, Meta's Llama line, Google's Gemma. Every one of them is a side project of a giant whose real business is something else, chips, ads, or a closed frontier model the open release must never embarrass. Meanwhile the models setting the open-weight pace, DeepSeek, Qwen, MiniMax, GLM, Kimi, ship from Chinese labs on a rhythm the American releases haven't matched. And that's not our characterization; it's Feldman's, from the same All-In appearance: the strong domestic open-source alternative doesn't exist. The one genuinely independent American effort, the Allen Institute's OLMo, is the truest open source in the field, weights, data, and code all published, and it runs at a fraction of frontier scale.
Independent. North American. Frontier-following. Free of any meter. Pick any three and somebody already exists; nobody's funded all four at once. The ask is specific: a substantial, North American-originated open-source LLM company that lands within a six-month catch-up curve of frontier release rhythms and answers to no shareholder pressure to charge exorbitant token rates, because it doesn't have token rates at all. Heck, make that the credo and write it into the charter: we're committed to the cheapest, most capable open-source LLMs, built from scratch on American soil, full stop, end of story. Give it to us.
The Check Is Smaller Than You Think
DeepSeek published a compute figure of roughly $5.6 million for the pretraining run of its V3 model. Grant every caveat, that number leaves out salaries, research, failed runs, and the cluster itself, and the all-in cost of a serious fast-follower still lands in the hundreds of millions a year, not the tens of billions the closed frontier burns. For calibration: Michael Bloomberg gave $1.8 billion to Johns Hopkins in one gift. One naming-rights-sized act of philanthropy funds this lab for years, and unlike a stadium, the return compounds. Every business, clinic, school district, and tribal government in the country gets frontier-adjacent intelligence it can own outright, forever, for the cost of the hardware it runs on.
And it should be built exactly this way, not as a national champion on a state budget. The UAE funds Falcon. Beijing's labs run with state wind at their backs. Brussels subsidizes its contenders. The American version of this story is supposed to be different: independently funded, privately governed, publicly released, beholden to nobody. That's what non-state-sponsored open source looks like when it's done on purpose instead of as a marketing line item.
What It Buys the People We Build For
For the organizations we build for, the payoff is concrete. Defense contractors, government agencies, and tribal nations that run open weights today do it over a procurement objection: the strongest open models originate overseas. The math inside the weights is auditable wherever it was trained, which is why air-gapped deployments of those models are defensible right now, but procurement committees run on provenance as much as math. An American open-weight lab at frontier cadence deletes the objection outright. Europe's already written sovereignty into a four-tier legal framework; an independent American lab is what keeps the United States on the right side of that scoreboard without a subsidy in sight.
What You Don't Get
Honesty, as always. A fast follower follows: philanthropy won't out-spend the closed labs on the bleeding edge, and the hardest reasoning tasks will stay a closed-model advantage for stretches at a time. Six-month lag is a target, not a birthright, and holding it takes real engineering discipline year after year. And one more thing the speeches skip: an American open model served through somebody else's metered API is the same tax with a flag on it. Open weights deliver sovereignty only when they run on hardware you own. The lab is the missing half of the argument. The racks are the half that already ships.
Now Is the Time
So here's the call to action, addressed in order.
To the golden-tongued: you've said the words on every stage that'll have you. A few dozen people in this country could endow this lab without refinancing a thing. The mission statement's drafted above, free of charge. Take it, fund it, name it after yourself if that's what it takes, and be remembered as the reason American open source stopped being a side project. That's more than a gesture or trite symbolism. That's what independently funded, non-state-sponsored open source should look like, and it's the whole game.
To everyone else: don't wait on them. The hardware half of sovereignty ships today. A rack you own runs today's best open weights right now, and the day that American lab drops its first release, the same rack runs it that week. No meter, no permission slip, no renegotiated contract. Own the metal. The models keep getting better on their own, and history's currently taking names. Talk to the builder.