Each week we find a new topic for our readers to learn about in our AI Education column.
Artificial intelligence and the financial services industry—infamous for being slow to embrace new and developing technologies—may seem like they’re very different.
But they’re actually extremely similar in one key attribute: Despite what appears to laymen as a dizzying diversity, they’re pretty centralized, with most of the power and influence being held by a relative handful of dominant organizations and companies.
Chances are, if you’ve used AI, you know these actors—pure-play AI model providers like Anthropic, OpenAI; tech giants like Microsoft, Amazon and Google. The hyperscalers and incumbent big tech companies. They’re like the Goldman Sachs, Bank of America-Merrill Lynches, Morgan Stanleys, Blackrocks and Fidelitys of AI. They always seem to be one of the turtles at the bottom of the stack holding everyone up. Try as we might, we keep running into them as we dig down into technology.
And just as the traditional, centralized finance industry has been changed, perhaps improved, by a challenge in the form of decentralized finance, the centralized artificial intelligence industry may be altered by a challenge in the form of what’s being called decentralized artificial intelligence.
And just like decentralized finance, decentralized artificial intelligence, sometimes shortened to DeAI, is built primarily on blockchain technology.
What Is DeAI?
What we’ll call traditional AI is built and trained using massive repositories of stored data and computing power, which require information infrastructure in the form of processors and storage, and physical infrastructure in the form of energy, space and cooling. In the traditional approach, most of this infrastructure—but especially the information infrastructure—is controlled by one of those dominant AI companies like the ones we named above, in the form of data centers, which each contain hundreds or thousands of powerful semiconductors. However, millions of powerful semiconductors are “out in the wild,” installed in business and personal computers. In a decentralized AI model, information infrastructure is distributed among a much larger number of actors—including organizations, businesses, and individuals.
This is accomplished using technology similar to that which powers cryptocurrencies, but has roots considerably older than today’s blockchains. DeAI as a concept actually evolved from distributed intelligence and computing systems, like SETI@home, in which the Search for Extraterrestrial Intelligence (yes, those guys listening to space for aliens) used the idle processing cycles on volunteers’ personal computers to analyze radio telescope data. That was happening over 20 years ago. Today, we can use better technology to distribute the training of an AI model to a universe of computers on a startlingly diverse amount of data. We’ll revisit that idea shortly.
DeAI Vs. Traditional AI
Blockchains manage the network supporting a decentralized artificial intelligence, offering tokens to participants, or users, who offer their local processing power, storage, and data to the network and help train models. Participants trade those tokens back in to the network to use its power, storage and data resources.
It was only by chance that we’re writing this column just after a few pro-DeAI pieces of thought leadership were published. In one, published in CoinDesk, the editors point out that DeAI providers currently have a collective enterprise value, all told, of around $12 billion, which sounds like big money until they mention that the big centralized AI providers are worth $12 trillion. The big providers control 70% of global cloud infrastructure. The world runs on their pipes.
The CoinDesk writers argue we’re entering an agentic AI age where information technology capable of reasoning is going to be shouldering some portion of the decision-making responsibilities in areas like government, healthcare and finance. Those decisions probably shouldn’t be made by models controlled by a handful of ultra-powerful organizations—there’s no guarantee that all those AI agents are actually acting autonomously if they’re all built on a centralized model. They may be more inclined to act in the interest of the AI hyperscaler du jour, or its biggest investors.
Or at least they may be perceived to be acting more in the best interests of the Amazons and Microsofts of the universe than those of doctors and financial services providers, not to mention the patients and wealth management clients in whose interests everyone is actually supposed to be acting in.
A Case for DeAI
On the other hand, hey, maybe AI is like web search, where we really only needed one big-huge search engine and all those little companies just kind of folded into each other or withered up and died… but probably not. Centralized AI presents some issues for people, the biggest one is that we need to send data to a centralized location for it to be used, recorded and processed before our model is trained, or, in the inference phase, before we get whatever result that we’re looking for back from the AI. Basically, we’ve been trading a mother lode of digital gold, data, to the big AI interests in exchange generative AI models. In DeAI systems, data stays local in the inference phase.
In fact, in decentralized AI, data stays local—and can be kept private and protected—in the training phase, too, thanks to what’s called federated learning. In federated learning, models are trained on localized edge devices, without transferring the data off the device. The model actually downloads itself onto the edge device for training, then, uploads itself back onto its distributed network. When the model is off the edge device, it only shares the insights it has learned, not the data itself, throughout the network—allowing the model to improve without sharing or centralizing information. In this way, we can train models on protected and regulated health or financial data without putting the actual information at risk.
Blockchains come back in to make sure the version of the model that is uploaded back to the global network from end users, and their edge devices, is verified and validated, basically agreed upon and made usable, by all the users of the network.
DeAI advocates argue that decentralized AI is potentially cheaper, or more energy efficient, than centralized AI. Blockchains are notoriously energy consumers. On the other hand, decentralizing AI eliminates the need for huge amounts of centralized storage, potentially cutting back on demand for data centers and energy use. Centralized AI also carries a fundamental risk—any disruption or breach to a centralized location for processing power, energy production or data storage is more likely to imperil the entire network and all of its data than a disruption or breach to an edge participant in a DeAI network. While maybe not exactly cheaper, then, DeAI is potentially a more stable solution for key industries as well.
