Jensen Huang's leather jacket has become a kind of diplomatic uniform. So when he walked into the รlysรฉe Palace to sit across from Emmanuel Macron, the photograph traveled faster than any joint communiquรฉ ever could. Two men, one table, and โ depending on which feed you scrolled โ either the dawn of a European technological renaissance or the quiet consolidation of a new kind of empire.
What struck me, watching the coverage ripple across crypto Twitter, was how little anyone in my corner of the industry paused to notice the irony. Here was the most centralized compute supply chain in modern history being toasted as a triumph of sovereignty. Here was a company holding an estimated ninety percent of the AI accelerator market being celebrated by a nation that has spent the better part of a decade insisting it wants to reduce its dependence on American technology.
Nobody asked the question that should have been on every decentralization advocate's lips: if sovereignty means controlling your own infrastructure, how exactly do you achieve it by buying the entire stack from a single foreign vendor?
That question is the thread I want to pull. Because the Paris meeting isn't really about France, and it isn't really about NVIDIA. It's about a fork in the road that the blockchain industry has been quietly walking for years without admitting where it leads.
The Doctrine of Sovereign AI โ and Its Hidden Cost
To understand why this matters, you have to understand what NVIDIA has actually been selling for the past two years. It is not chips. Or rather, it is not only chips. The product is a narrative called sovereign AI โ the idea that every nation deserves its own domestic compute capacity, its own models, its own data, running on its own soil. And the delivery mechanism is the AI factory: an integrated, rack-scale data center that NVIDIA will help you build, power, cool, and populate with its latest systems.
The logic is seductive, especially in Brussels and Paris. Europe has watched the United States dominate cloud infrastructure and China dominate manufacturing, and it has concluded that whoever controls compute will control the next fifty years of economic power. France, in particular, has positioned itself as the continent's AI champion. It is home to one of the few credible European frontier model labs. It has a nuclear power grid that produces some of the cheapest and lowest-carbon electricity in the developed world โ a genuine advantage when the binding constraint on AI data centers has shifted from silicon to wattage. And it has a political class that has learned to speak the language of technological sovereignty fluently.
So when Macron sits down with Huang, both men are performing a ritual that has become standard in the sovereign AI era. The nation says: we want to be a player, not a customer. The vendor says: I can make you a player โ with my help, my systems, and my timeline.
The media framing follows a well-worn template: powerful CEO meets powerful leader, handshake, implication of grand partnership. I have learned to read these moments with a specific kind of skepticism, born from watching the 2017 ICO wave turn press releases into valuations. A meeting is a signal. It is not a contract. The gap between the two is where most of the industry's worst decisions get made.
So I want to be disciplined about what we actually know. We know two people met. We do not know what was signed, promised, or refused. Everything else here is inference โ careful, informed inference, but inference nonetheless. The value of writing it anyway is that the underlying dynamic is real regardless of this single meeting's outcome. The question of who owns compute, and whether sovereignty can be rented, is the defining infrastructure question of this decade. The Paris photograph is just the occasion to ask it.
Here's what the ritual obscures. The blockchain industry has spent the same period building a genuinely different answer to the same question โ an answer rooted in distributed infrastructure, cryptographic verification, and the radical idea that compute need not flow through a single corporate chokepoint. And almost nobody in the mainstream conversation is comparing the two approaches honestly.
The Architecture of Dependency
Let me start with the architecture, because the architecture is where the philosophy hides.
An NVIDIA AI factory is not a collection of interchangeable parts. It is a tightly coupled system: GPUs linked by high-bandwidth interconnect, networked through specialized fabrics, orchestrated by software that assumes CUDA at every layer. The moat is not the transistor. The moat is the integration โ the fact that removing any single component degrades the whole. This is deliberate. It is also brilliant. And it means that when a nation buys an AI factory, it isn't buying hardware. It is adopting an ecosystem, with all the lock-in that word implies.
I spent part of 2024 inside a very different room, designing a crypto literacy program for a European bank's digital assets desk. The executives I trained asked the same question in a dozen different ways: if this technology is supposed to reduce dependence on centralized intermediaries, why does every serious implementation seem to depend on a handful of companies? It was a fair question then, and it is a fairer one now.
Because look at what sovereign AI actually produces. A nation builds a data center. The data center runs one vendor's systems. The systems require that vendor's software, support contracts, and upgrade cycles. The nation has achieved sovereignty over the building and the land and the electricity โ and nothing else. The layer that matters, the layer where the intelligence is actually computed, remains foreign.
This is not a conspiracy. It is an incentive structure, and incentive structures are the most honest things in technology. The business model requires large, concentrated, high-margin deployments. A thousand small buyers are harder to serve, harder to support, and less profitable than one sovereign customer signing a multi-billion-dollar framework agreement. The entire go-to-market motion โ CEO-level diplomacy, government frameworks, national AI strategies โ is optimized for exactly this. Huang doesn't visit Paris because he likes the croissants. He visits because the sovereign customer is the highest-leverage customer in the world right now.
And here is the uncomfortable symmetry. The European Union's own political logic pushes in the same direction. AI sovereignty is a national project. National projects require national champions. National champions require scale. So even the forces that claim to resist one company's dominance end up reinforcing the concentration they fear โ because the only counterweight they can imagine is a European version of the same thing, not the absence of it.
The blockchain industry looked at this exact problem years ago and asked a different question. Instead of 'how do we build a national champion,' it asked 'what if there is no champion at all?'
The Mistral Question
If sovereign AI has a poster child in Europe, it is the Paris-based frontier lab that has become the continent's most credible answer to the American labs. Its rise has been genuinely impressive. A small team, a fraction of the budget, models that hold their own on benchmarks that matter.
But here is the question that the sovereign AI narrative never quite answers: where does the compute come from?
The honest answer, for most of its history, has been a combination of partnerships and rented capacity โ much of it, inevitably, touching the same accelerator hardware everyone else uses. This is not a criticism of the lab. It is the structural reality of building frontier models in 2025: you need enormous, concentrated, homogeneous compute, and there are very few places on Earth to get it. Europe's champion is, in the most literal sense, running on the infrastructure its politicians say they want to escape.
I find this genuinely instructive rather than hypocritical. The lab is doing the rational thing. It is using the best available tools to compete at the frontier. But it illustrates the core tension of the entire sovereign AI project: you cannot have technological sovereignty without owning the compute layer, and owning the compute layer requires either building a competitor to the incumbent โ which no European firm is close to doing โ or accepting that your sovereignty is rented.
There is a third path, and it is the one I keep coming back to. You can reduce your dependence on frontier-scale compute by specializing. Not every valuable model is a frontier model. Domain-specific, privacy-preserving, verifiable, small models can run on distributed infrastructure today. Europe's regulatory culture โ obsessed with data protection, suspicious of centralization โ is actually a better fit for distributed, verifiable compute than for a gigawatt arms race it will probably lose. The sovereignty Europe wants may be available, but only if it stops trying to win a race defined by someone else's architecture.
The CUDA Moat and the DePIN Dilemma
DePIN โ decentralized physical infrastructure networks โ is the closest thing the crypto world has to a sovereign AI counterproposal. The premise is simple and, to a decentralization believer, beautiful: instead of building one giant data center, coordinate thousands of independent operators who contribute GPUs, storage, bandwidth, and energy, and let a token mechanism settle the economics. Rendering networks, compute markets, storage protocols, wireless meshes. The list grows every cycle.
I want to be honest about this category, because honesty is the only thing that makes an evangelist credible. DePIN is real. It is also, today, a rounding error next to the sovereign AI pipeline. And the reason is not marketing or token price. The reason is the moat I described above โ the software stack.
Here is the technical heart of it. Almost every serious AI model in production is trained and served through software that assumes one vendor's programming model. The kernels, the libraries, the compilers, the profilers โ they are all built for it. A decentralized network that aggregates consumer GPUs is, by definition, aggregating hardware that speaks a dozen different dialects, connected by links that are orders of magnitude slower than a purpose-built interconnect, coordinated by scheduling software that cannot assume homogeneity. Training a frontier model across that kind of mesh is not impossible. It is just brutally, economically uncompetitive.
This is where the DePIN conversation usually goes wrong. Enthusiasts point to the falling cost of inference and declare that the moat is crumbling. They are half right. Inference โ running a model, rather than training it โ is far more amenable to distribution. A quantized model served on a single GPU can be replicated across thousands of nodes, and the coordination overhead is manageable. That is why so many DePIN compute networks have pivoted toward inference. It is the honest beachhead.
But here is the part that gets left out. The moment you distribute inference across untrusted nodes, you inherit a verification problem. How does the customer know the node actually ran the model it claimed to run, on the data it claimed to use, and returned an honest result? In a sovereign AI factory, you don't ask that question because you own the whole machine. In a decentralized network, the question is the entire product.
Which is why the most interesting work in this space is not about GPUs at all. It is about proofs.
The Cultural Moat
I need to say something that is usually said only in engineering circles and rarely in strategy ones: the moat is not just code. It is a generation of human beings.
For roughly fifteen years, virtually every serious machine learning engineer has been trained on one stack. Their intuitions, their debugging habits, their mental models of how computation should be organized โ all of it is shaped by it. This is a form of lock-in that no competitor can replicate with a faster chip or a cleverer compiler, because it lives in people, not in silicon.
I saw an echo of this in 2020, when I was running DeFi workshops and the EIP-1559 fee-burning debate was raging. The confusion wasn't really about the mechanism. It was about the mental models people had already built โ and once those models were set, no amount of documentation could easily dislodge them. Technical education is a form of infrastructure, and it compounds.
For decentralized compute, this is both the hardest problem and the most hopeful one. The hardest, because you cannot out-engineer a decade of collective habit. The most hopeful, because habits can shift, and the shift often starts in exactly the places centralized vendors ignore โ the students, the hobbyists, the developers in regions that big vendors don't prioritize. When I distributed five hundred whitepaper summaries to university clubs back in Bonn, I wasn't building a product. I was building a habit of reading critically. That is the same work decentralized compute needs now: not a better chip, but a generation of engineers who reach for a distributed network by default because someone taught them how.
Verifiable Compute: The Missing Primitive
Verifiable compute is the primitive that decentralized AI has been missing, and it is finally arriving.
The idea is straightforward to state and extraordinarily hard to implement. If a computation happens on a machine you don't control, you need a cryptographic guarantee that the computation was performed correctly โ without re-running it yourself, because re-running it defeats the purpose of distributing it. Zero-knowledge proofs offer one path: prove that a computation was executed correctly without revealing the inputs or forcing the verifier to redo the work. Trusted execution environments offer another: hardware enclaves that attest to what code ran. Optimistic systems offer a third: assume honesty, allow challenges, slash the dishonest.
Each approach has a cost curve that most crypto commentary ignores. Zero-knowledge proofs for general computation are still, in most cases, orders of magnitude more expensive than the computation they verify. Trusted execution environments introduce a hardware trust assumption that sits awkwardly next to the whole ethos of trustless systems. Optimistic verification introduces latency โ you can't finalize a result until the challenge window closes.
I have watched this trade-off play out in a very concrete way. In 2025, I led a human-centric AI initiative for a Frankfurt startup, and one of our working groups spent months debating how to embed ethical constraints into smart contracts that would govern autonomous agents. The naive answer was to put a verifier on-chain for every decision. The realistic answer was to accept that verification has a budget, and to spend it only where the stakes justify the cost. That is not a compromise of the ideal. That is engineering.
So when I look at the sovereign AI buildout, I see a world that has solved the verification problem by eliminating the need for it โ by owning the whole stack. And I see a decentralized world that is slowly, painfully, learning to solve it cryptographically. The second world is more elegant. The first world is faster. And in a bull market, faster usually wins the headline.
But headlines and infrastructure are different things. The infrastructure question is: which approach can actually scale to serve the demand that sovereign AI is creating? And here, the honest answer is that decentralized compute is not trying to replace the AI factory. It is trying to do the jobs the AI factory is too expensive and too centralized to do well โ the long tail of inference, the privacy-sensitive workloads, the applications where verifiability matters more than raw throughput.
The Energy Question Nobody Wants to Answer
The binding constraint on AI infrastructure has shifted, and almost no one in the crypto conversation has internalized it. It is no longer the chip. It is the electron.
France's nuclear fleet is the real reason for the Paris visit, and everyone knows it even if no one says it. Modern AI data centers draw power at a density that traditional grid planning was never designed for. A single rack of the latest systems can pull more power than a small neighborhood. Multiply that across a national AI program and you are talking about gigawatts of firm, low-carbon, always-on capacity. France has that. Most of Europe does not.
This is where the sovereign AI story and the decentralized compute story diverge most sharply, and where I think the crypto world has been slow to notice its own advantage. A centralized AI factory must find one enormous source of power and build around it. A decentralized network can aggregate thousands of small sources โ rooftop solar, idle industrial capacity, stranded energy, even the waste heat that data centers currently dump into the atmosphere.
The efficiency gap is real and I won't pretend otherwise. Large, purpose-built facilities are simply more efficient per watt than a mesh of heterogeneous small operators. But efficiency is not the only metric that matters when the alternative is a decade-long permitting queue and a political fight over every transmission line.
I watched this dynamic play out in miniature during the 2022 bear market, when I founded a support network for displaced Web3 workers. The lesson from that period was not that big organizations fail and small ones succeed. It was that resilience comes from redundancy โ from having many paths to the same outcome rather than one perfect path. Centralized AI infrastructure optimizes for the perfect path. Decentralized infrastructure optimizes for having many. In a world of geopolitical shocks, export controls, and grid constraints, the second property is worth more than it looks.
AI Agents and the Ethics of On-Chain Autonomy
Now let me connect this to the thing that actually keeps me up at night: what happens when AI agents start transacting on-chain at scale, and what that means for the sovereign AI debate.
In 2025 I organized a summit in Frankfurt โ a thousand participants โ to debate how to embed ethical constraints into smart contracts. The premise was that autonomous agents would soon be moving real value through on-chain systems, and that the rules governing them could not be an afterthought. The debate was messy, as it should have been, but one conclusion emerged with unusual clarity: the ethics of autonomous systems cannot be bolted on at the application layer. They have to be part of the infrastructure.
Here is why this connects to the Paris meeting. Sovereign AI is being built as a national asset โ closed, controlled, accountable to a state. That is not necessarily bad. States have legitimate interests in the systems that run on their soil. But the same architecture that gives a state control also gives it the power to decide which agents may operate, which models may be served, and which transactions may settle. When the infrastructure is centralized, the ethical constraints are whatever the operator decides they are.
A decentralized compute layer changes that calculus, for better and for worse. For better, because no single operator can unilaterally censor an agent or freeze its ability to think. For worse, because 'no single operator can censor it' is exactly the property that makes illicit agents harder to stop. I am not going to pretend this tension resolves cleanly. It doesn't. It is the same tension that has shadowed every permissionless system since Bitcoin's genesis block.
What I will say is this: the choice is not between centralized ethics and decentralized chaos. It is between two different accountability models. In one, accountability flows from ownership. In the other, it flows from transparency and cryptographic proof. Both are legitimate. Only one of them is compatible with a world in which the most important infrastructure is owned by a handful of firms and a handful of states. And if we drift entirely toward the first model without building the second, we will have solved the coordination problem and lost the reason we cared about decentralization in the first place.
The Data Availability Analogy
Let me make a comparison that I have found clarifying, even if it makes me unpopular in certain rooms. The sovereign AI buildout reminds me of the data availability layer race in rollups.
For a couple of years, the Ethereum scaling conversation was dominated by the assumption that every rollup would eventually need its own dedicated data availability layer โ a specialized, high-throughput chain just for publishing data. Enormous capital flowed into the premise. And then, quietly, the reality set in: the overwhelming majority of rollups never generate enough data to justify dedicated DA. They are throughput-light. The bottleneck was never where the marketing said it was.
I have a strong view on this, and I'll state it through the evidence rather than as a slogan: most infrastructure demand is smaller than the narrative that surrounds it. The AI factory is sold as a national necessity, and for a handful of frontier labs it genuinely is. But the long tail of AI workloads โ the inference, the fine-tuning, the agent orchestration, the privacy-sensitive processing โ does not need a gigawatt facility. It needs cheap, verifiable, distributed capacity. That is precisely the gap decentralized compute is built to fill.
The analogy is not perfect, because the stakes in AI are higher and the demand curve is steeper. But the structural lesson holds. Whenever an industry converges on a single, enormous, capital-intensive answer to every question, it is usually over-building for the median case and under-building for the tail. And the tail, in aggregate, is often bigger than anyone admits.
Export Controls as a Preview
China is the control group in this experiment, and Europe should be studying it carefully.
When the United States restricted advanced chip exports, it did not merely inconvenience Chinese AI labs. It forced an entire ecosystem to confront what dependence actually costs. Chinese firms responded by investing heavily in domestic alternatives, by optimizing aggressively for the hardware they could get, and by restructuring their roadmaps around constraints they did not choose. The efficiency losses were real. So was the strategic lesson: the supplier of your compute has leverage over your future.
Europe, as a US ally, believes it is exempt from this dynamic. I would gently suggest that this belief is a bet, not a fact. Alliances are political, and politics changes. The same export-control machinery that was aimed at one country can, in principle, be aimed at another if interests diverge. A continent that has built its AI ambitions on a single foreign vendor's hardware has not eliminated its vulnerability. It has deferred it, and probably at a higher price.
This is where I think the blockchain framing earns its keep. Export controls are a sovereign act โ a state deciding what may cross its borders. Decentralized infrastructure is, at its best, an attempt to make such chokepoints less decisive by distributing the capacity they control. I am not naive enough to think a token network can route around a superpower's export regime. But I am realistic enough to note that the only infrastructure that survives a chokepoint is infrastructure that was never concentrated behind it in the first place. Europe is building the opposite, and calling it sovereignty.
The Token Incentive Problem
Let me spend real time on the mechanism that makes decentralized compute possible and fragile at the same time: the token.
Every DePIN network faces the same bootstrapping problem. It needs supply โ compute, storage, bandwidth โ before it has demand, because no customer wants to use a network with no capacity. The standard solution is to pay early suppliers in tokens, which creates supply ahead of revenue. This works, spectacularly, as long as the token appreciates. It fails, spectacularly, when it doesn't.
The failure mode is well documented and I have watched it from the inside. When token prices rise, contributors flood in, lured by yield rather than by utility. They are mercenary capital, and there is no shame in that โ they are responding rationally to incentives. But when prices fall, the same capital leaves, and the network's capacity contracts just as it needs to prove itself to paying customers. The result is a network that is most fragile exactly when it most needs to be robust.
I have a strong view about how to fix this, and it comes from the bear market rather than the bull. In 2022, running a support network for displaced Web3 workers, I learned that the durable communities were never the ones with the best incentives. They were the ones with the best reasons to stay when the incentives stopped working. The same principle applies to compute networks. A network that pays contributors purely in speculation will lose them to speculation. A network that pays them in genuine, recurring demand โ customers who actually need the compute and pay for it โ builds something that survives the cycle.
This is why I am more interested in DePIN networks with real enterprise customers than in ones with the highest yields. The yield is a symptom of a network that hasn't yet found its demand. The customer is a symptom of a network that has. In a bull market, that distinction is easy to miss, because everything looks like it's working. The bear market is where the difference between a protocol and a promise becomes visible, and I would rather build for the bear.
The Last Mile Problem
There is one more piece of the puzzle, and it is the piece that determines whether any of this matters to actual users.
I have spent years watching the cross-chain experience, and I want to describe it honestly because it is a useful warning for the AI infrastructure race. Ethereum's Dencun upgrade dramatically lowered the cost of moving data between rollups. On paper, interoperability got an order of magnitude cheaper. In practice, the user experience of moving assets across chains is still so bad that most people simply withdraw to a centralized exchange and re-deposit on the other side โ a workaround that defeats the entire point of decentralized infrastructure.
That gap between theoretical capability and lived experience is the most important lesson the crypto industry has to offer the AI infrastructure race. You can build the most elegant distributed system in the world, and if the last mile is painful, users will route around it through a centralized chokepoint. Sovereignty at the base layer means nothing if the interface layer funnels everyone back through a custodian.
For decentralized compute, the last mile is developer experience. Can a team of three engineers actually deploy a production model on a distributed network without hiring a cryptography PhD? Can they get predictable latency and pricing? Can they verify results without building their own proof infrastructure? Today, in most DePIN networks, the answer to at least one of those questions is still no. And until it is yes, the sovereign AI factory will keep winning by default โ not because it is philosophically superior, but because it is frictionless.
What I Learned Translating Whitepapers
I want to bring in something from my own history, because it shaped how I read moments like the Paris meeting.
In 2017, while finishing an applied mathematics degree in Bonn, I built a tool that took dense cryptographic whitepapers and reduced them to plain-language summaries for non-technical students. I distributed five hundred copies to university clubs. The point was not to make the math less rigorous. The point was to make the claims legible enough that a reader could tell the difference between a real protocol and a fraudulent one.
That skill โ translating complexity into clarity without stripping it of truth โ is exactly what is missing from the current AI infrastructure debate. Sovereign AI is presented as a self-evident good. Decentralized compute is presented as either a utopian fantasy or a solved problem, depending on who is selling tokens. Neither framing survives contact with the actual engineering.
What I learned in 2017 is that clarity is a form of accountability. When you force a claim into plain language, the weak parts become visible. Apply that test to sovereign AI and you get: 'We will achieve independence by buying a dependent stack.' Apply it to decentralized compute and you get: 'We will achieve scale by coordinating untrusted nodes, and we are still learning how to prove they behaved.' Both statements are honest. Both reveal a gap between the pitch and the reality. The difference is that one of those gaps is being papered over by a trillion-dollar market cap, and the other is being argued about in the open.
A Concrete Scenario
Let me make all of this concrete, because abstraction is where bad decisions hide. Here is what a realistic decentralized AI workload looks like end to end, and where the friction actually lives.
Imagine a mid-sized European hospital that wants to run a diagnostic model on patient data. It cannot send that data to a US hyperscaler without navigating a thicket of data protection rules, and it cannot afford to build its own AI factory. So it turns to a distributed compute network. The hospital submits the job, the network schedules it across a set of nodes that meet the hospital's jurisdiction and privacy requirements, and the model runs. The result comes back with a cryptographic attestation that the computation was performed correctly and that the data never left the permitted boundary.
That is the pitch, and it is a genuinely good one. Now the friction. The hospital's IT team needs to trust the attestation mechanism, which means someone has to explain zero-knowledge proofs or hardware attestation to people whose job is patient care, not cryptography. The network needs to guarantee latency and availability, which means it needs a layer of scheduling sophistication that most DePIN projects have not yet built. The pricing needs to be predictable, which means the token's volatility cannot leak into the customer's bill. And the whole thing needs to survive a node going offline at the wrong moment.
None of these problems is unsolvable. All of them are unglamorous. And this is precisely the work that the sovereign AI narrative lets everyone ignore, because the factory approach makes the hospital's problem disappear โ not by solving it, but by moving the hospital's data into a facility the hospital does not control. For some institutions, that trade is acceptable. For a growing number, it is not. That 'not' is the entire market for decentralized compute, and it is a market built on trust, verification, and the last mile โ not on raw throughput.
What Sovereign AI Gets Right
I have spent most of this article arguing that sovereign AI is more dependency than sovereignty. Let me steelman the other side, because an argument that only attacks is not an argument.
The case for sovereign AI is not stupid. It rests on a legitimate fear: that the countries which control compute will control the terms on which everyone else participates in the AI economy. That fear is grounded in history. The nations that controlled oil in the twentieth century shaped the fortunes of everyone else. The nations that control compute in the twenty-first may do the same. A state that wants to avoid being a price-taker in that world is behaving rationally, not paranoidly.
There is also a legitimate public-interest case. If AI systems are going to make decisions about healthcare, benefits, and security, then some level of public accountability over the infrastructure is reasonable. A fully permissionless compute layer offers no obvious mechanism for that accountability. Sovereign AI, for all its faults, at least puts a name and a jurisdiction behind the systems that matter.
And there is a pragmatic case that cuts against my own enthusiasm. Distributed networks are hard to govern, hard to audit, and hard to hold responsible. When something goes wrong โ a model serves a biased result, an agent causes harm โ the centralized operator is easy to identify and the decentralized network is not. That is not a feature. It is a cost, and it is one the decentralization movement has been slow to price in.
So I will say this clearly: sovereign AI is answering a real question. My objection is not to the question. It is to the answer. The question is 'how do we ensure that the infrastructure of intelligence serves the public good.' The sovereign AI answer is 'by owning it.' The decentralization answer is 'by making it verifiable and accountable without requiring ownership.' I think the second answer is better. I do not think it is obviously better. And the honest version of this debate holds both truths at once.
The Pragmatism Test
Now the pragmatism test, because an evangelist who only preaches is useless.
Here is the uncomfortable claim I want to put on the table: decentralized compute, as it exists today, is not a serious competitor to sovereign AI infrastructure, and pretending otherwise is a disservice to the very community it claims to serve.
Run the numbers the way a hard-nosed operator would. A frontier training run requires sustained, homogeneous, high-bandwidth interconnect at a scale no distributed network can currently match. The economics of a purpose-built AI factory โ amortized power, optimized cooling, single-vendor support โ beat a mesh of heterogeneous nodes on every axis that a CFO cares about. The token incentives that make DePIN networks work also make them fragile: when the token price falls, contributors leave, and the network's capacity evaporates precisely when demand might be rising. That is not a theoretical risk. It is the observed history of every compute-sharing network that has tried it at scale.
So no, I do not believe decentralized compute is about to replace the AI factory. But that is the wrong comparison, and letting it stand is the trap. The right question is not 'can DePIN beat the incumbent at the incumbent's game.' The right question is 'which jobs will never be served by a centralized factory, no matter how cheap it gets?'
The answer is a real and growing list. Workloads that cannot leave a jurisdiction. Workloads where the customer needs cryptographic proof of correct execution rather than a vendor's assurance. Workloads too small or too bursty to justify a contract with a hyperscaler. Workloads that must run on hardware the customer already owns. These are not consolation prizes. They are the structural gaps that centralized infrastructure, by its very design, cannot fill โ and they are where the community that believes in this technology should be building.
The trap is believing that the only meaningful victory is total replacement. That belief turns every honest assessment of DePIN's limitations into a betrayal of the cause. It shouldn't. The community is the only chain that cannot be broken โ and a community that cannot hear bad news is already broken, just quietly.
The Only Chain That Cannot Be Broken
So let me end where the photograph began, and look forward rather than backward.
The Paris meeting will be read by most people as a story about one company's unstoppable rise. I think it is more usefully read as a story about a question that has not yet been answered: who owns the intelligence that runs the world? Sovereign AI answers 'the nation.' The vendor answers 'whoever can afford the factory.' Neither answer is complete, and both leave out the possibility that the answer should be 'no one, verifiably.'
That third answer is being built right now, imperfectly, by people who will not be in the photograph and will not sign the framework agreement. Whether it matures into something that matters depends less on ideology than on the unglamorous work of verification, developer experience, and the last mile. The community is the only chain that cannot be broken โ but only if it stays long enough to finish the work.
The nations are buying. The vendors are selling. And somewhere in the noise, a quieter question waits: when the bill comes due, will sovereignty have been worth the dependency?