51³Ô¹Ï

In October 2024, I we were about to enter a decentralized AI training race. Eighteen months later, that race is already happening. In June, the decentralized AI training protocol Pluralis ran a new kind of distributed training of an 8-billion-parameter model across thousands of heterogeneous graphics processing units (GPUs) — the specialized chips used to train AI models — ranging from high-end machines down to consumer gaming devices. Earlier this year, a 72-billion-parameter AI was trained on a decentralized network of independently owned computers, setting a new headline record for parameter count. What’s significant is that both runs took place entirely outside a corporate data center.

The implication is bigger than it sounds: For the first time, the most resource-intensive part of building AI — the training — can happen without a billion-dollar facility. Every major objection to that approach has been , one by one, including feasibility, cost and scalability of the approach.

The AI arms race is real. The geopolitical stakes are as high as advertised. But the assumption that only governments and trillion-dollar companies have the investment capability and thus are the only players who can shape AI’s future is exactly the kind of incumbent logic that gets disrupted. We have seen this before. IBM once believed serious computing power had to live in a centralized mainframe, managed by institutions with the resources to operate it. Then the personal computer arrived, and that assumption collapsed. The internet followed the same pattern, and so did open-source software. In each case, the distributed challenger looked marginal until it didn’t. The incumbents were not beaten by a bigger version of themselves. They were beaten by an entirely different architecture.

There is also a geographic advantage that rarely gets discussed. Centralized data centers are prisoners of infrastructure, requiring access to power grids capable of sustaining the electricity demand of a small city, which limits where they can be built and by whom. Around half of a data center’s total cost of ownership is facilities maintenance and device cooling, all costs that are passed down to owners and renters. A distributed network has no such constraint. It sources compute wherever people live, which is also where cheap, renewable or stranded energy already exists, that is, on devices that have already been paid for and incur no cooling costs. That is a structural moat that a data center model cannot easily replicate.

The standard story of the AI race (for example, between the US and China) is a story about resources: who has the most advanced chips, who can build the biggest facilities, who can sustain tens of billions in annual capital expenditure. But it is missing something important. The race is also being run from the bottom up by a distributed network of independent contributors who, collectively, may prove harder to beat than any nation-state.

The trap inside China’s AI strategy

China’s open-source AI ecosystem has quietly become a dominant force. Models from Chinese labs now top the download charts on the platforms developers use to find and share AI tools. By some estimates, roughly of US AI startups are building on Chinese open-source models. The appeal is straightforward: These models can be up to 90% cheaper for companies to use than their closed-source alternatives, at surprisingly competitive quality. They are free to inspect, modify and build on. Most recently, Moonshot’s open Kimi K3 has Fable 5 and GPT-5.6 on key benchmarks, showing that ingenious optimizations can beat pure hardware power.

But this open-source AI has a fundamental economic problem. It costs enormous sums to build and almost nothing to copy. The organizations producing these models are, in effect, subsidizing the entire industry. We are already seeing the early signs: Alibaba, one of China’s most prominent AI labs, has now three consecutive closed-source models, signaling a strategic pivot toward monetization over openness. Meanwhile, access to Chinese models on major Western AI platforms is narrowing. The result is a potential gap between the open and closed frontiers — one that decentralized training networks may be uniquely positioned to fill. Meta, the largest US producer of open-source AI, is hedging toward a hybrid approach.

The open-source movement that drove China’s rise could be its first casualty if that rise succeeds. Without a way to make open AI economically sustainable, the gravitational pull will always be back toward consolidation among the handful of companies wealthy enough to absorb the losses indefinitely.

There is, however, an alternative, and it is one that most AI insiders dismissed until recently: training AI models not inside a single facility, but distributed across thousands of computers owned by different people in different places, following the same basic principle that made the early internet, Wikipedia and open-source software possible.

Two and a half years ago, this was considered technically impossible. Then it was done, but dismissed as too expensive. Then it was to be cheaper per training run than conventional data center economics. The objection then became scale: This was fine for small experiments, but never for serious models. But when a major decentralized training run was launched and worked, that put that argument to rest, too. As I write this article, Google DeepMind is in-house decentralization technologies.

Each time a new objection falls, the incumbents’ underlying logic gets a little harder to defend. Jack Clark, co-founder of Anthropic, recently acknowledged the progress while noting that decentralized training remains behind the frontier. He is not wrong about the current moment. But according to from watchdog Epoch AI, the distributed computing approach is closing the gap with conventional data center training at roughly four times the rate that data centers themselves are scaling, according to researchers tracking both trajectories. The question is no longer whether the gap will close, but when. It is already clear, though, that decentralized training is now a legitimate competitive vector in the AI race.

Why distributed AI will win in the long run

The most important shift happening in AI right now is not about bigger models, but about specialization. Fields like biology, materials science and mathematics are not waiting for a single general-purpose AI to solve their problems. They are actively seeking AI systems trained specifically for their domain. Those models do not require the resources of a frontier lab to build. They require expertise, data and enough computing power to train something useful. That is a competition open communities can win.

Consumer hardware is also getting meaningfully more powerful. Apple’s M5 chips, running on Apple’s open-source MLX framework, now run and fine-tune serious AI models entirely on a laptop, with no cloud account, no data center and no permission required. The device on your desk is already a viable node in the AI race.

Alongside optimized consumer hardware, AI agent tools are putting sophisticated capabilities into the hands of individual developers. The infrastructure for millions of people to participate in AI development, not just consume it, is being assembled in plain sight. The result, over time, is a network effect that no single institution can replicate: a swarm of participants whose collective compute and collective intelligence exceeds what any data center, however large, can concentrate in one place. A data center can be outspent, sanctioned, or seized. A network of millions of laptops, phones and servers spread across the world cannot.

This is where the economic model comes in and where decentralized AI offers something neither the US hyperscalers nor the Chinese open-source labs have figured out. If the people who contribute computing power to a training run, help fine-tune a model or deploy AI agents on a distributed network can receive a direct financial return for their contributions, the economics of open-source AI change entirely. You no longer need a trillion-dollar company willing to absorb losses. You need a sufficiently large network of participants, each making a reasonable return.

This is not a distant theoretical possibility. The broader financial industry is already moving toward tokenized assets at scale and AI model economics will follow. The coordination layer that makes this possible, the infrastructure for millions of independent participants to transact and collaborate without a central authority, is what crypto was built to provide.

Decentralized AI will continue to grow and, indeed, scale beyond what either nation-state can provide. Those doubting this, or threatened by it, will continue trying to move the goalposts. But the direction of travel is not in question. Open, distributed, economically self-sustaining AI is coming. And the only viable long-term competitor to a facility the size of a power plant is something that does not need to be built at all. It is the collective computational power of everyone who wants to participate. Best of all, anyone can use a network, not just those paying OpenAI or Anthropic.

[Jake Brukhman is the founder and CEO of CoinFund, which holds investments in decentralized AI infrastructure companies.]

[ edited this piece.]

The views expressed in this article are the author’s own and do not necessarily reflect 51³Ô¹Ï’s editorial policy.