On July 24, thirty-five organizations put their names on an open letter to Washington arguing that the United States should not restrict open-weight AI models. Microsoft. Meta. OpenAI. Nvidia. IBM. Mozilla. Y Combinator. Andreessen Horowitz. Companies that normally spend their time competing signed the same page.

I read it twice. I think they are right.

I also noticed who was not on the list. Not one small business. Not one solo operator. Not one of the 29.8 million American businesses that have no employees at all.

That is not a scandal. Lobbying letters get written by people with lobbyists. But it is worth understanding what was actually being argued, because the outcome lands on your desk whether or not anyone asked you.

What an open-weight model is, in plain terms

An AI model has weights. The weights are the thing that makes it work, the accumulated result of the training. A closed model keeps its weights private. You rent access through an interface and the company decides the price, the terms, and how long it exists.

An open-weight model publishes them. You can download it, inspect it, change it, and run it on your own hardware. You are not renting. You have a copy.

That is the whole fight. Whether the second kind stays legal and available, or whether the only serious AI runs on machines you do not control.

Why this reaches a business with four employees

Three things follow from that distinction, and none of them are theoretical.

You stop choosing what things cost. Today most AI use is small. You draft an email, you summarize a document. The cost is invisible. That changes when AI stops being a thing you open and becomes a thing that runs. When it answers your phone, follows up on every lead, and reconciles your books nightly, the bill scales with usage and the price is set by whoever owns the model.

I run a set of AI agents that work continuously across several machines. The single biggest lever I have on what that costs is choosing which model does which job. Routine monitoring runs on something cheap. Hard reasoning runs on something expensive. If every task had to run on one company's frontier model at that company's price, the operation would not be viable. The letter puts it as matching the right model to the right job at the right cost. That is not an abstraction to me, it is the monthly bill.

You do not own what you build. Every small business owner already knows this story. You build an audience on a platform and the reach gets throttled. You build a storefront on a marketplace and the fees climb. You rank on a search engine and the algorithm moves. The asset you thought you were building turned out to be rented.

AI is that same arrangement, earlier in the cycle. If your customer intake, your scheduling, and your follow-up all depend on one vendor's model, you have handed over your pricing power and your continuity. When the terms change, and terms always change, your options are to pay or to rebuild.

Your client data goes somewhere. Closed models mean your information, and your clients' information, is processed on somebody else's infrastructure. For a lot of businesses that is fine. For some it is not.

I do work for a medical practice. There is a signed agreement governing how patient information is handled and a legal duty to safeguard it. The question of where a model physically runs stops being a technical footnote the moment you are the one responsible. Anyone handling health records, financial data, or legal files runs into the same wall.

The number that actually worries me

The US Census Bureau tracks AI use by business size. As of early 2026, firms with 250 or more employees use AI at 31 percent. Firms with 1 to 10 employees sit at 18 percent.

A gap is not surprising. Big companies adopt new tools first. The part that should get your attention is the direction.

Census asked businesses whether they expected to be using AI within six months. The largest firms said 40 percent. The smallest said 21 percent. And between December 2025 and May 2026, reported AI use rose among firms with 20 or more employees while showing no statistically significant change at all for firms under 20.

The gap is not closing. It is opening, and both sides know it.

Meanwhile the capacity to build these models is concentrating. Stanford's AI Index reports that industry produced more than 90 percent of notable AI models in 2025, and that a single company accounts for over 60 percent of the compute they run on. One firm reported more than 150 billion dollars in capital spending last year.

Put those together. The ability to make advanced AI is consolidating into a handful of organizations, and the businesses least able to absorb whatever they decide to charge are the ones falling behind fastest.

Why this matters more than it sounds like it should

There are 36.2 million small businesses in the United States. They are 99.9 percent of all American businesses, employ 45.9 percent of the private workforce, and accounted for 88.9 percent of net new jobs in the most recent year measured.

And 78.4 percent of all US businesses have no employees at all. Not small. Solo. One person, doing the work. That share has grown every year for over a decade.

This is the part the letter does not say, because it was not written by us. If AI becomes something only large organizations can afford to run at scale, the productivity gap between a company with 500 employees and a company with 5 does not stay where it is. It widens, permanently, in a country where the second kind is almost all of us.

Open weights are not a complete answer to that. They are a precondition for one. You cannot compete with a tool you are not allowed to have.

What I am not saying

I am not saying open models are safe by default. The letter's own signatories concede the risk plainly: once weights are released they cannot be recalled, and modified versions are hard to trace. That is real.

I am also not saying the big AI companies are villains. I use their products daily and they are extraordinary. The concern is structural, not personal. Concentration is a bad outcome regardless of who ends up concentrated, including the companies I like.

And I am not qualified to tell Washington how to write the rule. What I can tell you is what it feels like from a one-person shop, which is a perspective that was not in the room when those thirty-five names went on the page.

The question I keep coming back to

The letter argues that open models keep the frontier plural, that competition keeps the gains of AI broadly shared instead of concentrated in a few hands. I believe that. I also notice that the argument was made entirely by organizations that have already won, and that their interests and mine align right now but will not align forever.

So here is what I actually want to know from other people running small operations.

Are you making decisions about AI based on what happens if the price triples, or the vendor changes the terms, or the model you built on gets retired? Or are you doing what I did for the first year, which is picking whatever works today and worrying about the rest later?

I ask because I think most of us are in the second group, and I am no longer sure that is survivable.

Sources