TL;DR—decentralization, dispersion of decision-making power, on its own might not be that useful. Instead, consider an equilibrium between an appropriate dispersion of authority among agents in the system, and cost-efficient economics.
Blockchain technology spent fifteen years making the case for why decentralization was important. But is it important, and more than that, what is it, anyway?
In the end I’ve come to see it as quite simple. There’s two answers, the minimal level, which is a simple binary check (the complexity comes in the context, as we will see with a stratified model), and the complex, which in certain contexts is about stable cartel formation, incentives, and governance.
The latter can only really be analysed at runtime, ex-post. This adds a time dimension, further complicating matters. This is true of the former, really, but you can make some basic ex-ante protocol attempts to enforce minimal decentralization. BFT (Byzantine Fault Tolerance) is one of these, but the existence of Sybils (if run permissionlessly) can erode the ex-ante guarantees and require runtime audit and analysis.
A Side Note on AI and Decentralization
As I will explore in other articles, due to discussions over AI governance, decentralization theory has become very timely again.
While I was editing the draft of this post, somebody forwarded me this preprint (Qu et al., 2026), where four key constraints of sovereign AI are discussed: operational independence, resource autonomy, persistence and adaptive capability. It’s an interesting thesis (though I suspect it is incomplete), but what I like in particular is that they provide mechanisms for this to be achieved, including the requirement for “economic self-sustainment.”
Obviously this is confirmation bias, but the first list sounds a lot like a very pure, or idealised view of decentralization,1 while the tension between decentralization and economic viability is I would say, at this point, well-documented and discussed. After all, ‘Scalability’ in Buterin’s original blockchain trilemma assumed an economic component.
Indeed, that tension is the executive summary of this very article. So, let’s talk about decentralization.
Minimal Decentralization
So, the minimal level is simply that of the EDI MDT (Edinburgh Decentralization Index’s Minimum Decentralization Test) methodology—can a single entity, directly or indirectly (ie via delegations, contract agreements, coercion or other means) take control of finality?
If the answer is yes, the MDT is failed.
In terms of a blockchain, one entity can turn it off. In terms of a swarm of AI agents, one entity can turn them off. Simple.
The paper linked above goes on to analyse a network in a stratified set of layers. It’s an interesting approach, but I can’t really adopt their full analysis. It’s both too complex and special-cased to be a general framework for thinking about socio-technical systems. It’s a similar critique to that I’d lodge against other analyses like Gochhayat et al.
Minimal decentralization must be analysed in every category and at every level, otherwise the MDT is failed, and we can judge that a network is at immediate risk of takeover or halt. It’s a very minimal state, a low bar.
Self-promotion time. In my work, we conceive of a network as being composed of technical affordances (network topology) such as data centre, provider, client software implementation, etc, and governance (governance topology).2
With fewer categories or layers, the MDT is less useful perhaps in our model. Plus, the need to analyse systems at runtime (especially in the governance domain) results in the need for something that describes more than just a system that is in a binary state of either “at risk of immediate takeover or halt,” or, “not at risk of immediate takeover or halt.”3
Note that in the AI case, the sword is double-edged; there is one takeover, where an entity turns the AI agents off, yes, but there is another—where the AI agents themselves take over.4 In both cases, the system got subverted and taken over.
Sustainable Liveness—Decentralization in Equilibrium, Over Time
In order to reason about decentralization, we also need to reason about incentives, which means bringing in economics. In order to analyse governance, we need to analyse operational systems, which means adding a time dimension.
These requirements move the goalposts a bit.
Thus, for describing a (de)centralized socio-technical system in equilibrium,5 we propose sustainable liveness, which is a more outcome-focussed state, based on the economics of the system (themselves an output of the network and governance topologies) and its agent graph and incentives.
If decentralization is about dispersion of authority, then there is a tension between that and cost-efficient transactions. Thus, sustainable liveness captures the tension between the need for economic sustainability and dispersion of authority, which itself results in increased governance costs. The most governance-efficient system, after all, is a centralized one.6
Sustainable liveness is more subtle than the MDT—you can hypothesise about it ex-ante but not judge it until runtime.
Put simply it is about two things, of which only the second is directly relevant here:
The ability of the network’s economics to support its infrastructure, or, to put it another way, its network topology.7
The ability for stable cartels to form within subsets of the agents that make up the network.8
The concentration of a cartel-sized block is simple, and just extends the logic of the MDT. A problematic cartel is one that can seize control of finality. The definition differs from the MDT only insofar as the cartel may act collectively but cannot be described as a single entity in the sense of the MDT.
Final thought before we wrap up: the goal of governance in most socio-technical systems is an appropriate dispersion of decision making power (authority, that is, or agency) to agents in the system.9 Meanwhile, in order for it to be operational, it must be live and economically sustainable in the time dimension (a tautology, I know).
Just like any small business, a system must balance the value it adds, and the cost it creates, otherwise eventually we would expect it to be wound down. Thus this is kind of a pedestrian thing, in the end.
In a sense, sustainable liveness is obvious, but it’s also something that isn’t just decentralization, and isn’t just economics. It’s the equilibrium between the two.
Summary
Thus, if decentralization is dispersion of authority, then a minimal decentralized system that is able to function for any extended period must also satisfy the conditions of sustainable liveness.
Whatever its economics and agent graph look like, it must balance cost-efficient governance and economic sustainability, which is part incentives and part finding a balance between systemic value and cost.10
For some reason, I’m reminded of Response D4 in the Appendix of our decentralization fieldwork, “In my mind, I have my own fantasy view of decentralization, which is where every validator on a particular blockchain is somehow evenly distributed across all the different countries in the world, so you have one validator in every single country, each validator is run by a different person or company, and that validator is on slightly different hardware. Perhaps it’s also a different operating system, and on different fibre-optic carriers. In that sense, you have a highly uncorrelated portfolio of validator nodes which, in theory, statistically reduces the probability of any of those validators being sanctioned or turned off. Or if they do go down, from a natural disaster, you’re not really going to be impacted. Now that’s my fantasy view of decentralization. There’s multiple different dimensions, geographical, operator, hardware, and then internet carrier.” The paper with appendices is available as an archive here (the Proceedings/published link does not have full appendices, hence the archive link).
Our paper, presented at FCiR last year is here. You will have noticed this can be applied not just to blockchains, but to any social-technical system. AI rollout in an org? You betcha.
With the caveat that even establishing this is pretty darn hard.
I’m going to say Byzantine Fault Tolerance again, wait for it… Byzantine Fault Tolerance.
I add the caveat to (de)centralized to point out that centralized systems are merely a point on a continuum, and that this analysis can be applied to them as well. Think about government regulation—on some level it’s about moral questions, but often it’s about things like offsetting externalities, which in the end, is about attempting to secure a productive economic environment for things like efficient transactions (if you’re being very clinical about things).
Though I don’t doubt that exceptions exist, as with all things.
This is part, in a sense, of its governance topology, if it can be altered by governance, which in the general case we would assume. This is one of the reasons the governance topology takes precedence over the network topology.
I’ve done some work digging into the OECD cartels dataset as part of a paper we’re working on, as well as other academic literature, and a reasonable heuristic seems to be that both average and median cartels often de-cohere between seven and ten members due to defections and coordination complexity. The idea that agent sets make up the network has been discussed in other posts.
Referencing the work of JP Vergne on the subject, as I’ve done on this blog many times.
If you’re thinking “can you formalize that?” then yes, you can formalize it into an abstract model, which is exactly what we are working on at the moment. Once there’s a draft paper, I’ll present the model in a blog post, but it’s relatively easy to intuit the rough shape of it.