There is a version of the next ten years where every piece of intelligence your business touches is rented from four companies, at whatever price they settle on some Tuesday you were not invited to. We are closer to that than most people think. The main thing standing in the way is open source AI, and it is the most important argument anyone should be having right now.
I have been building on the internet since 1999, which means I have watched this movie before. The web started open. Anyone could put up a page, anyone could read it, and nobody needed permission. Then the useful parts got fenced. Search results filled with ads, social networks opened their doors wide, gathered everyone inside, and quietly shut the windows. Every time, the fence went up after the thing became essential, never before. The people who ended up trapped had no idea they were making a decision at the time.
Artificial intelligence is at the fence-building stage right now. Not the model releases, not the funding rounds. The part worth watching is who ends up owning the shelf that everything else is built on.
A few days ago Nvidia agreed to buy Hugging Face for almost thirteen billion dollars. I wrote about that here already, and the short version is that the company selling the chips just bought the warehouse where the open-source AI world keeps its work. In the same stretch of days, one lab locked in roughly fifteen gigawatts of computing capacity on a deal that could run past half a trillion dollars over its life. That is a power plant footprint, committed by a single company, to run models nobody outside it can inspect.
Set those two things next to each other and the shape of the next decade gets easier to see. Enormous private compute on one side. On the other, a shrinking number of places where the models are open enough that you can download the weights, run them yourself, and see what you are actually working with.
Here is the part I want you to pay attention to. When someone argues that open models are too dangerous to release, listen carefully to who is making the argument. Sometimes the concern is sincere, and there are real questions worth taking seriously about what gets published and when. But notice how often the proposed cure is a licensing regime, a compliance burden, or an approval process that only a company with a legal department and a data center can survive. The safety framing arrives with a business model attached.
Question any company working to squash open source AI. Not because their stated reasons are always wrong, but because of who happens to benefit when they win. A rule that says only the biggest labs may publish capable models is not a safety rule. It is a moat with a safety sticker on it.
None of this is theoretical for me. Most of what I do day to day runs on open models. There are high-end machines sitting around my office in Atlanta doing the work, and a good share of it never leaves the building. I did not set it up that way to make a point. I set it up that way because it works.
Part of it is the environmental math, and I want to be precise about this rather than hand-wave, because the sloppy version of this argument is easy to knock down. A big graphics card in my office is not automatically greener than a hyperscale data center. Those facilities are engineered for efficiency in ways my back room will never match. That is not the comparison that matters.
The comparison that matters is what you send where. Most business tasks are small. Rewrite this paragraph, pull the addresses out of this list, sort these emails, summarize this call. Sending every one of those to a frontier model burns far more than the job requires, and it does it in somebody else's building, every single time, forever. A right-sized open model running on hardware I already bought and already power handles the same work at a fraction of the cost, and the marginal footprint of one more task is close to nothing. The waste is not in where the computer sits. The waste is in routing trivial work to enormous models out of habit.
The second reason is less high-minded and more persuasive: they get the job done. That is the part people who have not tried recent open models tend to miss. The gap between the best closed model and a good open one closed a lot faster than the marketing suggested it would. For most of what a real business actually needs, the open model finishes the task, and it does it without a per-token meter running and without your customer data taking a trip through a service whose terms change when the quarter goes badly.
Owning your AI instead of renting it changes your position. Rented tools can be repriced, deprecated, rate-limited, or discontinued while you are mid-project, and none of those decisions will consult you. A model sitting on your own hardware keeps working on the morning a vendor changes its plans. That is not ideology, it is just not wanting your operations to depend on somebody else's roadmap.
You do not need to follow every model release to run a company. That is what people like me are for, and it is most of what our AI integration work amounts to: figuring out which jobs need the expensive model, which ones an open model handles on hardware you control, and building the thing so you are not renting your own business back from someone.
The shelf is still open today. It is worth noticing who keeps suggesting it should not be, and worth asking what they get if everyone agrees.