Permissionless Intelligence

Yesterday I wrote about the gating of intelligence — eighteen days in June when one of the world’s most capable models went dark on government instruction, and a frontier release that now requires customer-by-customer approval. I ended that essay by saying the answer is to build alternatives. This one is about the alternative I’ve been building, and about who it’s actually for.

Two things became true this year at the same time, and their combination changes the picture.

First, open models stopped being the compromise option. GLM 5.2, released with open weights in June, is discussed in the same sentence as Claude now — within a point of it on some of the hardest coding benchmarks. DeepSeek’s V4 release in April matched the strongest closed models on real software-engineering evaluations, open-weight under MIT. And on OpenRouter — the closest thing to a public scoreboard of what people actually run — models from Chinese open labs grew from a rounding error to roughly half of all traffic in a year. Whatever your benchmark of choice says this month, the honest summary is: for a large share of real work, open models are no longer clearly worse. They’re just less convenient to run.

Second, the convenience of the alternative — a handful of gated APIs — revealed its price. Access that can be revoked overnight, by an authority you can’t appeal to, is not infrastructure. June made that concrete.

So the capable models are free to run, and the case for running them yourself has stopped being ideological. What’s missing is the layer underneath: somewhere to run them that isn’t a hyperscale data center. That’s a network problem, not a model problem, and it’s the problem I’ve spent the last two years on.

A network anyone can join

Tenzro is an open protocol and network for decentralized AI and distributed computing. The protocol and network are open source under Apache 2.0 — the code is at github.com/tenzro — and the premise is simple to state: anyone with a computing system should be able to join the network, run a node, serve models from the registry, and earn for the compute they provide. A few clicks, not a procurement process.

That includes the machines the AI economy currently ignores: workstations, small clusters, regional data centers, edge hardware. Individually, none of them can host a frontier model. That was the design constraint — and working within it turned out to be the most interesting engineering problem I’ve taken on.

Intelligence that doesn’t need one building

The reason AI concentrated into a few giant facilities isn’t conspiracy; it’s physics and architecture. Large models wanted to live in one place, on tightly coupled hardware. The most useful thing I can report from building Tenzro is that this constraint is loosening — and modern model architectures are what loosen it.

Mixture-of-experts models — the architecture behind most serious open models today — only activate a small fraction of themselves for any given request. An MoE model is less like one giant brain and more like a panel of specialists, a few of whom are consulted at a time. Specialists don’t need to sit in the same rack. Tenzro runs MoE models across multiple nodes — locally, within a city or a facility, and globally when the request tolerates it — routing each request to the experts it actually needs. The network also supports multi-token prediction models, which generate several tokens per step; the fewer round trips a model needs, the friendlier it is to a distributed setting, so the architectures and the network are converging from both directions.

Training tells the same story. Approaches in the DiLoCo family let nodes train independently and synchronize rarely, instead of demanding the constant lockstep communication that only a single building can provide. Tenzro supports Decoupled DiLoCo — resilient, distributed training where nodes can join, drop, and rejoin without bringing the run down. Resilience is the whole point: a network where participation fluctuates and the work persists is a network with no switch to pull.

None of this required inventing exotic science. It required taking a decade of ML research seriously as infrastructure — asking, for each technique, what it means for where computation can physically live. The conclusion I keep reaching is that the centralized data center is an artifact of one era of model architecture, not a law of nature.

A practical option for small nations

That conclusion matters most for the players the current model was never going to include.

Most countries cannot build a gigawatt AI campus. The costs, the power, the chips, the expertise — the whole stack assumes a scale that a few dozen nations will ever reach. UNCTAD said as much in its 2025 technology report: the United States and China account for roughly two-thirds of the world’s AI capacity and patents, and the “AI divide” between them and everyone else is widening, not closing. For most of the world, the current architecture of AI offers exactly one role — customer — and one posture toward it: dependence.

I understand that problem from the inside. I grew up in the Maldives, a country of half a million people spread across a thousand islands, that will never host a hyperscale data center and shouldn’t have to in order to have a say in the intelligence its economy runs on. When I look at the AI map from there, the question isn’t how to win a compute race no small nation can win. It’s whether there’s a way to participate at all without renting your entire intelligence layer from someone else’s territory, on terms you don’t set.

A distributed network is that way. A country doesn’t need a national megaproject to put its universities, its businesses, and its existing data centers to work on the network — providing compute, running models and agents, keeping its own data under its own governance, and earning rather than only paying. Sovereignty stops meaning “own a campus you can’t afford” and starts meaning “participate in a network on your own terms.” The same door is open to the independent operator: the individual, the small business, the regional data center with spare capacity. Bottom-up participation and national sovereignty turn out to be the same mechanism at different scales.

This is also where the policy conversation is finally heading. Last August the UN General Assembly established a Global Dialogue on AI Governance and an independent scientific panel; the first Dialogue convenes in Geneva this month. At Davos in January, one of the sessions was titled “Digital Embassies for Sovereign AI” — how nations might extend their digital sovereignty beyond their own borders. These forums are circling the right question: how the large and the small share the same technology without the small becoming permanent tenants of the large. I think the answer is less a treaty than an architecture. You don’t legislate your way to participation; you build a network that makes participation the default, and then the governance has something real to govern.

The footprint, and who gets to participate

Two consequences of distribution matter to me beyond resilience.

The first is environmental. The AI buildout is currently measured in gigawatts — new campuses, new plants, new transmission lines, much of it duplicating capacity that already exists in the world as underused hardware. A network that puts existing machines to work, closer to where demand actually is, grows AI’s capability without growing its footprint at the same rate. Distributed computing won’t solve AI’s energy problem by itself, but concentration is quietly making it worse, and distribution is the lever pointing the other way.

The second is economic. The AI economy is becoming one of the largest in the world, and right now the only way to participate on the supply side is to own a gigawatt campus or shares in one. A network where a workstation in Malé, a render farm in Lagos, or a small data center in Jakarta can serve models and earn from day one distributes that economy the way the web once distributed publishing. Participation is the antidote to dependence.

The quiet bet

Tenzro is a bet that the future of AI looks less like a few cathedrals and more like a grid — capable open models, running on a resilient network of hardware nobody fully controls and anybody can join, governed by rules that are legible because the code is open.

June was a preview of the cathedral model under stress. The grid is the alternative, and it’s no longer hypothetical: the protocol is open, the network runs, and the door is a few clicks wide. Whether you’re a nation that will never build a campus or a person with one good machine, if you have compute, you can be part of this.