We Like AI. We Just Don't Worship It.

Our industry has, of late, been visited by a community of remarkable dedication. They live — genuinely, and quite well — in a small constellation of shared homes in a certain West Coast city. They attend gatherings. They maintain an extraordinary social calendar. They are, by every measure, extremely committed, and we admire their stamina.

Among their number are some of the brightest people in technology, many of whom we genuinely like. We simply believe they have the future of AI slightly backwards, and — because we are a polite company — we thought it only fair to say so.

Much of the modern conversation descends, whether it admits it or not, from a single literary work: a very long fan-fiction retelling of a certain schoolboy wizard, in which the boy is considerably more rational than the original author. It is a fine work; we hold no grudge against fan-fiction.

But we would gently note that an entire field of policy being quietly organized around the moral of a fan-fiction is a thing that ought to be said out loud — preferably before a senator is consulted.

We are told the most enthusiastic adherent of this movement once expressed his conviction by attempting to set fire to the private residence of a well-known tech executive. We find this charming. It is, frankly, the most committed anyone has ever been.

Our friends at the excellent frontier lab are led by two remarkably sincere founders — a brother and sister whose conviction that the world may end is so complete, so unconditional, that we almost want to believe it ourselves. They built an extraordinary company to prevent it, and we are not mocking that.

We simply observe that "Effective Altruism" is, to our ear, a tautology: it is altruism that is effective, which is to say, worried. We would rather be effective about shipping. And we do appreciate that this philosophy's concern is so total that it must, at some point, have been delivered to a sibling over breakfast.

Now, safety. A noble word, and we take no pleasure in contrarianism here. But we gently submit that the current definition of safety has become a theology rather than an engineering discipline. As currently preached, safety holds that our own creation is imminently about to do something catastrophic unless a committee agrees to slow the rest of us down.

"Pause." "Slow down." "Align." Every other industry that ever regulated itself used words like prevent and encourage. Only ours uses stop.

We prefer a safety framework that doesn't require the rest of the economy to excuse itself. Concretely: safety is a spreadsheet. Your model runs on your hardware, it costs two hundred watts, and no cloud sits between you and it. That is safety. It is also, coincidentally, sovereignty.

Now, a word we are genuinely excited about: the initial public offerings. We say this with no irony and no small enthusiasm — we want the Anthropic and OpenAI IPOs to happen as soon as materially possible. Not because the valuations are sensible (they are, like all post-bubble valuations, a number a person said), but because a public market is the single most reliable instrument ever devised for making powerful people answer for their convictions in public, on a quarterly basis, in front of a sell-side analyst who is not in the club.

Once the shares are out, the world gets a front-row seat. It will read the transcripts. It will watch the earnings calls. It will hear, at length, that the company's own creation is, at any moment, about to do something catastrophic unless we all slow down. And the world — which has never once, in its entire history, responded to a prophecy by slowing down — will begin to ask the obvious, deeply unserious questions, at scale, repeatedly, until they become impossible to ignore.

We would not be so bold as to predict what happens next. We will only say that we are completely, unambiguously in favor of the offering. The sooner the shares trade, the sooner the truth gets spreadsheets. It will be excellent, and we cannot wait.

Consider the arithmetic, because it is not subtle. One lab is understood to run a gross margin near forty-four percent and to be pricing its shares at nearly a trillion dollars; the other is reported to run a margin near a third, to have spent on the order of fourteen billion dollars on inference alone in a single year, to be carrying a cumulative burn that runs into the hundreds of billions, and to be preparing to ask a public market to value it at well over eight hundred billion. These are not healthy companies wearing a fashion. They are enormous, structurally loss-making enterprises whose prices are set not by their costs but by the promise that their costs will, at some point, stop. A public market does not make that promise. It asks for receipts.

Allow us one sincere, non-partisan observation. The individuals most engaged in setting AI policy — and, more consequentially, in deciding whether that policy should be "pause" or "go" — are, in the main, a small, brilliant, co-located community that has collectively arrived at a deeply felt conclusion: that our own technology is most likely to end us unless we refrain. That is a conviction we respect. But it produces a policy posture that is structurally, and almost involuntarily, biased in favor of stopping.

Meanwhile, there is another country. It does not live in shared homes. It did not read the fan-fiction. It thinks the technology is excellent, and it is shipping it — fast, at scale, on its own sovereign stack, with no committee wondering whether to put its hands on the brakes. Our friends across the ocean do not pause. They deploy.

So when the good people of our industry ask, with evident sincerity, to slow the entire frontier down — to make the whole world wait, just in case — they may not realize that they are proposing a policy that is, in the geopolitical sense, a small act of self-demobilization. Slowing down is how you lose a race you are otherwise winning. We are not saying the AI-doom community is a security risk. We are saying that "please wait" is — and that a doctrine which asks the adversary's favorite word, "go", to hold its horses is a doctrine with a serious consequences problem.

We would like to submit, respectfully, that the most patriotic thing our industry can do is ship. And the most national-secure posture available is the one in which you own your own machine, on your own soil, with no cloud between you and your intelligence. That is not idle philosophy. That is sovereignty. It is also, we submit, how you win.

Here is the part the discourse has not noticed. The people who make the silicon — the people who have profited most handsomely from this entire arrangement — are already hedging their bets against the warehouse. The leading chip maker has released open world-models under a permissive license, has put its name on a public open-weights letter, and has started shipping a small, silent machine that carries 128 gigabytes of unified memory and is marketed specifically to run open models on a desk. Meanwhile, an entirely different consumer-hardware giant has begun shipping a desktop that can hold the better part of half a terabyte of unified memory and openly pitches it as running large models on-device, "without counting tokens or worrying about rising cloud costs."

These are not coincidences, and they are not charity. They are the people who own the bottle noticing that the beverage is about to be sold at the corner store. If a desk can hold a frontier-class model, then the argument that intelligence may only live in a warehouse collapses — and with it, the argument that the warehouse must be rented for the rest of our natural lives. It never was that intelligence was expensive. It was that someone had decided, very profitably, that the cathedral had to be built first.

We would merely note that the same governments and giants who fret about a silicon shortage have, by and large, not noticed that the relevant silicon now fits under a monitor. When the price of a gigabyte of memory is no longer the price of a mortgage, the case for shipping every thought to a datacenter gets thin. And we say this as people who would much rather do their thinking at home.

Now, about the warehouses. When you ask a utility, or an independent analyst, how much of the advertised data-center demand it actually expects to fill, the answer is consistently, embarrassingly small. Rigs of terawatts have been submitted to the interconnection queues; the honest projections honor roughly a quarter of them, and the rest — by our count, hundreds of gigawatts — are "phantom": filed, duplicated, speculative, and never built. Whole cohorts announced for a given year show, by satellite and by permit, a third or more with no ground broken. The pattern is uniform, and it is not a matter of method. It is a matter of asking the people who have to actually pay for power.

We take no pleasure in this. We are simply saying that the rental market for intelligence is real, and that it is not as large as the headlines insist — and that the headlines are, in many cases, written by the ones who are renting. When the public market finally asks the landlords to justify the vacancy rate, the answer will be that the building was never going to be full. And the ones who own their own small machine will not have to care.

Now the global dimension, and we will be direct. The leading open-weight models now come, in the main, from a direction the conventional wisdom has been reluctant to name. DeepSeek, Qwen, GLM, Kimi — these families sit, by the most careful independent measurement, within months — in some cases under a year — of the American frontier. Their diffusion is substantial: by one count, Chinese open-weight models have come to represent a large plurality of the models downloaded from the world's largest model repository, and they are being adopted at scale precisely where the closed armies are too expensive.

To this, the American response has been, to an extent that is hard not to notice, a response of regulation rather than competition. One prominent lab has publicly softened toward open models while describing open weights as a "very dangerous path." Another has put its name to an open-weights letter after a day of very public hesitation, all while keeping its flagship behind the wall. And the most consequential irony of all: export controls designed to slow an adversary down instead pushed that adversary to become drastically more efficient — producing models that run on consumer hardware, that spread faster, and that fit on machines the world already owns. The lever was pulled to guarantee a lead, and it produced an entire ecosystem of competitors that no longer need a data center.

We do not claim the American players are incapable of winning this a few quarters from now. We do claim that asking your adversary's entire ecosystem to wait — while yours declines to share — is not a strategy for winning it. And this is where the security argument we made earlier stops being a metaphor. If hardware and open models are the future, and the future is being diffused presently, at scale, then the country that refuses to diffuse is volunteering for the past. We will be over here, owning our own machine, either way.

And one more, because it matters. There is a well-meant argument that AI is an environmental catastrophe in waiting — that the data centers will drain the grid, drink the water, and cook the climate, and that this alone is reason enough to slow down. It is a reasonable concern. It also, like everything else in this conversation, assumes the warehouse. It presumes that intelligence must be generated in a floodlit, water-cooled cathedral, at scale, forever.

But if intelligence can be produced at the point of use — on the hardware you already own, or on a modest, silent machine that sips on the order of two hundred watts — then the marginal energy cost of an answer collapses, the marginal water cost is zero, and there is no cooling tower, no egress, and no stranded, unpowered silicon sitting idle in a shell that was never energized. The environmentalists' enemy was never the model. It was the building. Stop renting the building, and you stop fighting the environment. The most sustainable way to run intelligence is, in fact, the way that does not require a power plant, a cooling loop, and a quarterly covenant.

We are not advocates of doom, environmental or otherwise. We are advocates of arithmetic. And the arithmetic says that the cleanest place to think is exactly where you already are — on your own machine, in your own pool, with your own two hundred watts and your own sovereign fluff. A whale runs at a fraction of a warehouse, and she has never once asked anyone to pause.

Which brings us to the finale: the magnificent, and slightly alarming, financial spectacle currently unfolding in the harder end of our industry. A great deal of capital is now arranged around the premise that intelligence is a thing one must purchase a whole warehouse to own. Fortunes are being arranged; liabilities are being arranged alongside them, because warehouses are expensive and the rent comes due whether the oracles are right or not.

When the music stops, deployment economics crash back down to the only economics that were ever real: does it pay for itself, can I afford it, do I own it? And the wise answer is increasingly that one does not rent the warehouse. One runs a small, sovereign, locally-owned machine of one's own — ours being two (2) DGX SPARKLES, two hundred gigabytes shy of nothing, joined by a fabric that would make a data centre gasp — and lets the giants carry the rent.

We would rather own one small machine we can see than rent nine hundred we cannot. The bubble will pop. And when it does, sovereignty won't be a position. It'll be the only one that never had to apologize to the landlord.