DeepSeek and Open-Source AI: Does It Really Give AI Back to Us?

When a small Chinese lab shipped a model that matched the giants for a reported fraction of the cost and released it for anyone to download, the story wrote itself: DeepSeek open source AI had cracked the fortress. The share prices of the biggest AI names wobbled, a lot of value was wiped off AI stocks in a single jittery stretch, and a wave of commentary declared the monopoly over. For once, the argument went, the technology had slipped out of the hands of a few American labs and into everyone else’s. I want to take that claim seriously, because part of it is true and part of it is a comforting story we tell ourselves. The honest answer is that something real cracked open — and that a crack is not the same thing as giving AI back to us.

What actually happened

Strip away the market drama and the event is simple. A model arrived that performed close to the frontier on the tasks people care about, was reported to cost far less to train than the prevailing assumption, and was released openly so that anyone could pull the weights down and run them. That combination is what rattled people. The prevailing bet in the industry was that frontier AI required a near-bottomless budget for chips and electricity, and that this cost was itself the moat — the thing that kept the technology safely inside a handful of well-capitalised firms. A far cheaper model, given away rather than rented, suggested the moat might be shallower than advertised.

The market reaction told you what investors had actually been pricing in. It was not the intelligence of any one model; it was the assumption of scarcity. If frontier-class AI can be built for less and handed out for free, the premium attached to owning the only door starts to look fragile. That fragility is worth understanding clearly, because it is the genuine opening here. But the word doing quiet work in every headline was “open” — and that word means far less than most readers assume.

“Open weights” is not “open source,” and neither is “ours”

Here is the distinction that most of the coverage blurred. When a lab releases a model like this, what it almost always releases are the weights — the enormous grid of numbers that the training process produced. You can download them, run them on your own hardware, fine-tune them, and build products on top without asking anyone’s permission or paying a per-call fee. That is real, and it is genuinely useful. It is why people could take the model and start tinkering within days.

But open weights are not the same as open source. In ordinary software, “open source” means you get the recipe: the source code, the ability to see exactly how the thing was made and to reproduce it yourself. With an open-weights model you get the finished cake, not the recipe. You typically do not get the full training dataset, the exact data-cleaning and filtering pipeline, or the complete training code needed to rebuild the model from scratch. You can use what was produced; you usually cannot independently reproduce how it was produced, or verify what went into it.

Open weights hand you the finished cake, not the recipe. You can eat it, sell slices, even reshape it — but you cannot bake your own from what you were given.

That gap matters for three reasons. First, transparency: if you cannot see the training data, you cannot fully audit the model for what it absorbed, what it was optimised to say, or what it quietly refuses to discuss. Second, reproducibility: without the recipe, the ability to make the next model still lives with whoever holds the data and the compute. And third, control: a licence that lets you use the weights today can carry restrictions, and can be narrowed for the next release. “You can download it” is a real gift. It is not the same as “you own the means of making it,” and it is nowhere near “this technology now belongs to us.” Keeping those three things separate — usable, reproducible, ours — is the whole game.

The crack is real

I do not want to be sour about this, because the opening is genuine and worth defending. An openly released, capable model does real work in the world. A startup, a university lab, or a government in a country that will never build a frontier model of its own can now run something powerful on hardware it controls, without wiring every query through a foreign company’s billing system. A researcher can inspect behaviour that a closed API hides behind a curtain. A business can escape the quiet dependency of building everything on top of one vendor’s pricing and one vendor’s terms of service.

That is a meaningful redistribution of capability, and it puts downward pressure on the people who were counting on renting intelligence to everyone forever. It is harder to charge monopoly prices for a capability that a competitor is giving away. It is harder to claim that only you can be trusted with the technology when a downloadable model is doing much the same job on someone else’s laptop. Every open release chips at the story that frontier AI must, by its nature, be centralised. That story is doing a lot of political work, and it deserves chipping at. If you want the fuller version of that argument, I’ve written separately about who owns AI and why the ownership question sits underneath every other debate about it.

Why the enclosure survives

And yet. Follow the actual dependencies and the celebration cools. A model’s weights may be free, but the things required to make one, and to run one at scale, are not — and they remain concentrated in remarkably few hands.

Start with compute. Training and serving these models depends on specialised chips produced by essentially one designer, manufactured by essentially one foundry, in volumes that are allocated to the largest buyers first. When a model is cheaper to train, that lowers one bill; it does not change who controls the supply of the hardware, or who can afford it at the scale needed to serve millions of users reliably. Cheap-to-train and cheap-to-serve-at-planetary-scale are not the same thing, and the second is where the concentration lives.

Then data. The weights you download are the residue of a training corpus you never see — scraped text, images, code, human feedback — assembled at a scale only large organisations can gather, clean, and label. Whoever controls that pipeline controls the ability to make the next model, and the one after that. Handing out this generation’s weights does not hand out the data machine that produced them.

And capital. The lab that surprised everyone did not appear from nothing; it grew out of a well-resourced operation with the money, the talent, and the hardware to attempt a frontier run in the first place. “Cheaper” here means cheaper than an eye-watering benchmark. It does not mean cheap enough for the rest of us. The barrier to entry moved; it did not disappear.

An open model can loosen one company’s grip while leaving the underlying structure — the chips, the data, the capital — exactly where it was.

This is the pattern I keep returning to across every technology, because it is the same move dressed in new clothes each time: the same move, a new machine, every time. A tool arrives, it briefly looks like liberation, and then the layer underneath it quietly re-concentrates. The web was going to flatten publishing; a few platforms became the new gatekeepers. Open protocols were going to keep communication free; they were folded into a handful of walled apps. I’ve traced this arc in detail in how technology gets captured — the recurring question of who takes, who pays, and who fights back. Open weights can be the thing that fights back and, if we are not careful, the thing that gets captured. One enclosure can quietly replace another: a business that was renting you access to a closed model can pivot to renting you the compute to run the open one, and you are still paying rent to roughly the same layer of the economy.

That is the shape of what I’ve elsewhere called technofeudalism — not ownership of a product you buy once, but perpetual tenancy on infrastructure someone else controls. An open model does not escape that arrangement on its own. If the chips, the data, and the capital stay enclosed, then “open” can become the friendly face of the same dependency, the moat rebuilt one layer down where nobody is looking.

What would actually give AI back to us

So what would count as giving AI back to us, as opposed to swapping one landlord for a slightly nicer one? I think it comes down to three things, and open weights are only the first of them.

Real openness, not just free weights. The recipe matters as much as the cake. Genuinely open models would release the training code, document the data, and make the whole thing reproducible and auditable — so that “open” means you could rebuild and inspect it, not merely run it. That is the difference between a gift you are allowed to use and a commons you are allowed to build on.

Public and shared compute. If the chips and data centres remain a private toll road, openness at the model layer only goes so far. Public research compute — funded like other public infrastructure, available to universities, startups, and public bodies — would put the means of running and training these systems within reach of people who are not among the handful of firms that can currently afford it. Roads and power grids were not left entirely to whoever could privately build them; there is no reason the substrate of AI must be.

Ownership and governance, not just access. Access is being allowed to use something on terms someone else sets. Ownership is having a say in the terms. That means data used with consent and, where possible, held in common; it means public and cooperative institutions in the mix, not only venture-funded labs; it means the people whose work trained these systems having some claim on what the systems do. Giving AI back to us is ultimately a question about power and ownership, not about download links.

None of that happens automatically because one impressive model was released for free. It happens through deliberate choices — funding, law, licensing, and the plain refusal to accept that the current concentration is natural or permanent.

Hold both thoughts at once

So did DeepSeek and open-source AI give AI back to us? My honest answer is: it opened a door that the incumbents wanted kept shut, and that is worth celebrating without pretending the room beyond it is already ours. An openly released, capable, far cheaper model is a real crack in a real monopoly. It proves the moat was narrower than we were told, and it hands genuine capability to people who had none. That is not nothing. It is a lot.

But it is a crack, not a demolition. The compute, the data, and the capital that decide who gets to build the next generation are still held by very few, and an open model can just as easily become the new front door to the same old dependency as it can become a genuine commons. Which of those it turns into is not a technical question that the model answers for us. It is a political and economic question about who owns the layer underneath — and that question is still wide open. The task is to press on the crack, not to mistake it for the daylight on the other side.