Web3 did not begin as a technical project. It began as a political one.
The consolidation of Web 2.0 had produced something nobody designed and few had anticipated: a handful of platform monopolies with unprecedented power over information, commerce, identity, and public discourse. The case for an alternative was not nostalgic or ideological. It was a response to observable harm - harm that was well-documented, accelerating, and concentrated in the hands of a very small number of actors.
The promise of Web3 was sovereignty. Your data, your identity, your assets, your protocols - held by you, governed by rules you could inspect, not subject to the permission or goodwill of any intermediary. This was worth building.
Ten years into serious development, the promise remains largely unfulfilled. And now artificial intelligence has arrived - the most transformative technology since the web itself. The technology most likely to complete what Web3 started is also, structurally and economically, the technology most opposed to everything Web3 stands for. This essay explores that paradox - and asks whether history offers a way through it.
Decentralisation is the property most loudly associated with Web3, but treating it as the whole story obscures the real architecture of what was promised. There are five distinct pillars - each with its own logic, its own partial achievements, and its own characteristic failures.
Decentralisation is the structural claim: no single entity should control the infrastructure on which others depend. Most loudly proclaimed. Most frequently compromised.
Trustlessness is the epistemic claim: participants should not need to extend personal trust to counterparties or intermediaries. Trust is replaced by cryptographic proof - deception would require breaking mathematics, not merely corrupting people.
Self-sovereign identity is the claim about personhood in digital space: your credentials and verified attributes should be yours to carry and disclose selectively, not owned by the databases of platforms you happen to use.
Programmable ownership is the claim about property and contract: smart contracts encode agreements in executable code, making the terms of ownership legible to machines as well as people. The asset and the rule governing it travel together.
Interoperability is the claim about composability: value, identity, and contracts should flow across chains and protocols without friction or permission - a decentralised web is not a collection of siloed blockchains any more than the original web was a collection of isolated intranets.
These pillars are in very different states of maturity. Programmable ownership has seen the most real-world implementation, despite its failures. Trustlessness remains more aspirational than actual - oracles, bridges, and governance forums all reintroduce human trust at critical junctions. Interoperability is the least developed of all. Understanding this uneven landscape is essential to understanding what AI does to it.
The relationship between AI and these five pillars is not straightforwardly hostile. AI both advances and threatens each pillar, often simultaneously.
AI agents and semantic search can reduce dependence on centralised discovery platforms. AI can dramatically improve smart contract auditing - but then who audits the AI that audits the contract? Generative AI has made synthetic impersonation trivially easy at scale, attacking self-sovereign identity at its foundations - while AI-powered zero-knowledge systems may eventually provide the cryptographic substrate to defend it. AI agents as autonomous signatories expand what programmable ownership can mean. AI already serves as a semantic bridge across incompatible protocols, an underappreciated contribution to interoperability.
In each case, the critical variable is not capability but control. And on the question of control, the paradox becomes undeniable.
Here is the thing that should give us pause.
Training a frontier AI model requires a quantity of computation that only a handful of organisations on earth can afford. This is not a temporary fact that will dissolve as technology matures. It is a structural consequence of how transformer-based scaling works. The economic floor for playing at the frontier is measured in hundreds of millions of dollars per training run.
Three to five organisations - OpenAI, Anthropic, Google DeepMind, Meta AI, and perhaps one or two others - account for the models underpinning the vast majority of AI-powered applications. This is not a diverse ecosystem. It is an oligopoly.
To build a decentralised web on a centralised AI layer is to repeat, at a deeper architectural level, the exact mistake Web3 was designed to correct. The new landlord would be more powerful than the old ones - because the AI layer would sit beneath, and coordinate, all the others.
Before accepting this as a permanent condition, it is worth asking a different question: have we been here before?
The answer is yes - more than once.
In the earliest computers, the program was the machine. Logic was hardwired into physical circuits. To change what the computer did, you rewired it. Computation was entirely centralised in hardware - expensive, inflexible, and accessible only to institutions with the resources to build and operate these machines.
The stored-program concept, formalised in the Von Neumann architecture of 1945, broke this open. The insight was deceptively simple: program and data could both live in memory, separable from the physical substrate. What the machine does need not be fixed in what the machine is. Overnight - in historical terms - computation decoupled from dedicated hardware. The same physical machine could run a payroll, simulate physics, or sort a mailing list. The program became portable, replicable, shareable. The stranglehold of centralised hardware was broken not by distributing the hardware, but by abstracting above it.
A generation later, the mainframe imposed a new centralisation. The IBM 360 and its contemporaries were monolithic machines running concurrent programs for many users - powerful, but requiring institutional ownership and operation. Access meant proximity to the institution. The personal computer broke this open again - not by building a better mainframe, but by changing the unit of ownership entirely. Intelligence moved to the edge.
Then came a problem that individual PCs could not solve alone: computations of such complexity that even a mainframe would take years. The response was distributed computing - networks of personal machines, each contributing a fragment of a larger calculation. The SETI@home project is the canonical example: millions of home computers processing radio telescope data during idle time, their screen savers doing something genuinely useful instead of drawing geometric patterns. No single machine held the answer. The answer emerged from the aggregate.
I explored a related rupture in an earlier essay - the parallel between the stored-program concept and the smart contract. Both decoupled logic from substrate. The contract, like the program, became separable from the institution that would previously have been required to enforce it.
The question worth asking now is this: is current AI more like the hardwired circuit of 1940 or the stored program of 1945? Are we at the moment just before the conceptual rupture that changes everything - or are we already past it?
The honest answer is that we do not know. But the pattern is suggestive. What broke hardware centralisation was not better hardware - it was a new abstraction layer. What broke mainframe centralisation was not a bigger mainframe - it was a different unit of ownership. What may eventually break AI centralisation is unlikely to be a cheaper way to train the same kind of model. It is more likely to be a different conception of what intelligence in a machine requires.
Edge inference is a small signal of this. The fact that a model small enough to run on a phone can now perform tasks that required a data centre five years ago is not simply an efficiency improvement. It is a change in the unit of ownership - the same conceptual move the PC made against the mainframe. Distributed training networks are a small signal of the same thing in a different direction - the SETI@home intuition applied to intelligence rather than signal processing.
Neither is sufficient yet. But both point toward the same possibility: that the Von Neumann moment for AI - the conceptual rupture that decouples intelligence from the massive substrate currently required to produce it - may be closer than the current landscape suggests.
Web 1.0 was decentralised by nature - a network of peers, open protocols, no dominant platforms. Web 2.0 was a centralising response. Not a betrayal, but a necessary development: the participatory web that billions of people actually adopted required the centralised infrastructure of platforms and data centres. At that stage of maturity, decentralisation alone could not have delivered it.
Web3 emerged from within Web 2.0 as a reaction to its excesses - but equipped with tools that only the centralised phase had made possible. The blockchain, the smart contract, the cryptographic wallet - none of these could have been built or sustained without the infrastructure and users that Web 2.0 created.
AI fits the same pattern. Its current centralisation may be a necessary detour rather than a terminal destination. The compute concentration that troubles us is the same concentration that produced models capable of reasoning and coordinating at scale. The centralised phase may be producing the capability that will eventually make decentralised AI viable - through the open weights it releases, the inference efficiencies it discovers, and the applications it proves are worth building.
This does not make present concentration acceptable. It makes it explicable. The goal is not to prevent centralisation from having occurred - it is to ensure it does not become the terminal state.
Universal centralisation - a web in which all infrastructure, all intelligence, and all governance converges into the hands of a small number of actors - is the genuinely unwanted outcome. Not because decentralisation is an end in itself, but because concentrated control over the coordinating layer of human communication, commerce, and identity is a structural precondition for the abuse of power at civilisational scale.
Decentralisation, understood this way, is best seen not as a technical specification but as a regulative ideal. It may never be fully achieved. Pure decentralisation is almost certainly utopian: governance requires coordination, coordination requires authority, and authority tends to concentrate. But the ideal is productive precisely because it is not yet fully true. It sets a direction, creates accountability, and generates the friction that prevents consolidation from becoming permanent.
The most defensible goal for what comes next is to decentralise access to, and governance of, AI - not necessarily AI itself. Access decentralisation means the capability is available to all, not merely those who can afford API pricing or comply with terms of service. Governance decentralisation means the values embedded in AI systems are determined by accountable, diverse, and contestable processes - not by the preferences of a small number of corporations.
The web has been here before. It centralised, produced enormous value, generated intolerable concentrations of power, and provoked the movement that is still attempting to correct it. AI will likely follow the same arc.
The paradox is real. So is the precedent. Computing has broken this kind of stranglehold before - not by distributing what already existed, but by imagining it differently.
The Von Neumann moment for AI is still ahead of us. The question is whether we will recognise it when it arrives.
We love to hear your comments on this article.