Xi's BRICS AI Vision Is a Settlement Layer Play — And Crypto Isn't Invited

Consider the moment a national technology strategy stops being about models and starts being about rules. Over the past two years, the BRICS agenda on artificial intelligence has made exactly that transition: fewer announcements about parameter counts, far more language about governance, capacity building, and the architecture of trust. The change is procedural, quiet, and almost entirely unread by the people it will eventually govern. Now the same period, seen from the other end. A developer in Lagos or Jakarta pulls a thirty-billion-parameter model out of a public repository, rents two GPUs by the hour, and ships a working product in three weeks. No license negotiation, no cloud contract, no ministry signature. That is the version of AI most people in this industry assume we are building toward. Both of these things are real. One gets covered as a product story, the other gets covered as geopolitics. They are the same story, and the half that determines your future is being written in the room almost nobody in crypto reads closely. This was never about them. It is about us. The bloc has grown faster than its institutions. Egypt, Ethiopia, Iran, the United Arab Emirates and Indonesia now sit at the same table as Brazil, Russia, India, China and South Africa, with more applicants waiting at the door. An alliance that size does not require a shared ideology. It requires shared plumbing, and the AI agenda is quietly delivering exactly that. Three instruments matter more than the summit photography. The first is the Global AI Governance Initiative, which Beijing advanced through the United Nations and paired with a General Assembly resolution on AI capacity building aimed squarely at the Global South. The second is a proposed international AI cooperation body headquartered in Shanghai, unveiled at the World AI Conference and framed as a counterweight to governance designed elsewhere. The third is financing, with a development bank that has started talking about compute corridors the way it once talked about ports and rail. Notice what is missing from that list. It is not about which model is smartest, or which benchmark is highest. It is about who certifies, who finances, and who settles. In every technology stack in history, those three functions are where the rents live. The model itself is the commodity. For crypto, the reflex is to read all of this as regulatory risk — another bloc, another rulebook, another set of compliance costs bolted onto a business that already has too many. That framing is a decade out of date. The consequential question is not whether permissionless systems will be regulated. It is whether they get designed in as a component of the plumbing, or designed around as an inconvenience. To see why, you have to understand what trusted AI actually means in operational terms. It does not mean a model that behaves well. It means a model that carries a verifiable credential. Provenance for content is already being standardized by industry coalitions that lean on X.509 certificate authorities, the same hierarchical trust model that secures the web. Those certificates are issued by entities you did not choose, validated against roots controlled by institutions you cannot audit, and revoked at their discretion. The AI governance debate will be settled by the identity registry, not by the model weights. Whoever operates the registry of which models passed which evaluation, and which media was generated by which system, holds a chokepoint more powerful than any export control. Washington understands this, which is why its own AI diplomacy pushes American stacks, American cloud regions and American evaluation regimes into allied markets. Beijing understands it too. Both are racing to be the certification authority of record, and neither is racing to make that authority replaceable. Here is the part that genuinely surprised me, and the part that almost nobody in the governance conversation wants to say out loud. Export controls on advanced accelerators were supposed to freeze Chinese frontier capability in place. What they actually produced, under real constraint, was an extraordinary push toward efficiency and open distribution. Mixture-of-Experts routing, lower-precision training regimes, aggressive kernel optimization — all of it is the mathematics of constrained optimization, a subject I spent a master's degree on. When you cannot buy your way to scale, you learn to spend less compute per unit of capability. And then, because you cannot win on closed distribution against the incumbent, you give the weights away. The result is that the most-downloaded open model families in the world today are Chinese. DeepSeek's reasoning release under a permissive license, Alibaba's Qwen lineage, and a wave of successors have become the default substrate for developers who will never have a contract with a hyperscaler. Open weights are not decentralization in any rigorous sense. But they globalized access to capability faster than any governance framework has managed to constrain it, and that is a strategic fact with a very short shelf life if nobody defends it. Which brings me to the honest limit of that victory. Open weights are a snapshot, not a process. You get the parameters; you do not get the data lineage, the evaluation harness, the filtering decisions, or the fine-tuning recipe. When I audited incentive models for a Layer 2 project in Shanghai in 2024, the lesson that stuck with me was structural: publishing the parameters of a system is not the same as publishing the system. I have watched the same confusion infect decentralized AI. Publishing weights is not the same as publishing the process, and a downloadable file is not a governance model. If the training pipeline is opaque and the evaluation is self-reported, you have inherited a black box with a permissive license attached. That is an improvement over a black box behind an API. It is not the same thing as verifiability. Verifiable inference is the actual frontier, and it is where the decentralization story gets uncomfortable. Zero-knowledge proofs of model execution exist and are brutally expensive; proving a large transformer's forward pass in a practical latency budget remains an open problem. Trusted execution environments are the pragmatic path, and they work today. But they relocate trust rather than removing it — into silicon vendors who control the attestation chain, and into whoever issues the certificates that chain depends on. Follow that logic to its end and you arrive somewhere strange: in a genuinely multipolar AI world, the most defensible trust root is a hardware manufacturer's signing key. That is a single point of failure wearing the costume of distributed infrastructure. I have watched this industry spend a decade arguing about whether code is law, only to arrive at a moment where the binding commitments are enforced by a certificate revocation list. Then there is settlement, which is where this stops being an AI article and becomes a monetary one. The multi-CBDC bridge connecting several Asian and Gulf central banks reached pilot maturity, survived the withdrawal of its original institutional sponsor, and continues under participant control. Bilateral local-currency arrangements across the bloc are expanding. Nobody serious believes a single BRICS currency is coming. What is coming is a permissioned, centrally issued, policy-configurable corridor for cross-border value, and it is being built without asking anyone in this industry for input. Permissionless settlement is not being banned in that architecture. It is being excluded by default, which is far more effective. You cannot lobby your way back into a design decision that was made before you noticed there was a meeting. A fair objection is that state-led infrastructure always underdelivers. I would point out that most of what gets marketed as sovereign blockchain infrastructure is a permissioned enterprise ledger with a token bolted on afterward — the same rebranding trick the Bitcoin ecosystem has spent years performing with Ethereum-derived rollups that share almost none of Bitcoin's trust assumptions. Institutional chain announcements are mostly this. So the threat is not a competent rival stack. It is a mediocre one that becomes the default through procurement rather than merit. There is a second-order risk that my own community should recognize, because we invented it. We now have dozens of Layer 2 networks competing for what is, generously, the same few hundred thousand genuinely active users. That is not scaling. That is slicing scarce liquidity into ribbons and calling the fragmentation progress. A multipolar AI world has every incentive to repeat that error at civilizational scale: ten sovereign stacks, ten attestation registries, ten evaluation regimes, none of them interoperable, all of them expensive. Interoperability is the demand worth making, and it is a demand almost no government has a reason to resist. So here is where I get uncomfortable with my own side of the argument. The decentralized AI sector's answer to a standards race has been, in too many cases, a token wrapped around GPU rental. That is not a governance alternative. It is a hosting business with a governance token stapled to it, and the market has rewarded it precisely because narrative is cheaper than infrastructure. Where we do have evidence is funding mechanism design. Retroactive public goods funding, which pays for outcomes already delivered rather than promises made in advance, remains the only model I have seen operate at scale without degenerating into committee patronage. Every grant program I have watched allocate by application has drifted toward whoever writes the best deck and knows the right reviewers. If open model development and open evaluation tooling need subsidy — and they do — that is the mechanism with a track record, and it is being ignored by exactly the institutions now writing the standards. And I will not pretend the alternative bloc is our friend. American export controls, diffusion frameworks and cloud-residency requirements are equally centralizing, equally indifferent to permissionless verification, and equally comfortable appointing themselves the arbiters of acceptable computation. The choice on offer is not freedom versus control. It is between two varieties of control, plus a third path that is underbuilt and underfunded. When I spent two years translating MakerDAO governance proposals into Chinese for a small reading group in Shanghai, the thing that struck me most was how much teeth a governance document loses in translation. Nuance about who can veto what, and under which threshold, evaporates. The same evaporation is happening now with AI governance language, at a scale that will not be reversible. This is about us, in the most literal sense. The registry that decides which content counts as human-made will decide whether your identity is portable or rented. The attestation chain that decides which model output is trustworthy will decide whether you can verify a claim without asking an institution for permission. Those are the same question, and it is being answered in documents drafted by ministries, not by protocol designers. So watch the root key, not the summit photos. In the next eighteen months, the standard for machine-generated content provenance will harden into procurement requirements and platform policy. The identity question underneath it — whether a person can prove they are a person without surrendering a dossier to a registry — will either get a cryptographic answer or a bureaucratic one. My community has perhaps one more funding cycle to show that the cryptographic answer works at population scale, not in a demo. After that, the question will not be whether we were right about decentralization. It will be whether anyone building the next layer of the world's infrastructure remembers that we were asked at all.