Hook
Xi Jinping's address at the BRICS summit did not contain the words blockchain, stablecoin, or token. Two of those three do not appear anywhere in the bloc's published AI language. That absence is precisely why the trading desks I sit near started asking different questions the following week.
I lead market operations at a mid-tier exchange, and the first signal I registered was not a price. It was counterparty behavior. Treasury managers in Johannesburg, São Paulo, and Jakarta began asking, unprompted, whether cross-border compute invoices could be settled outside the correspondent banking chain. Eighteen months ago that question was ideological. Now it is operational, because AI capacity is bought by the hour, across borders, by buyers who do not want a five-day settlement window stapled to a two-hour workload.
The second signal came from a category of token I rarely trade: decentralized compute. Lease durations shortened. Utilization climbed. The order book stopped behaving like a narrative and started behaving like a utility. The AI race is being financed in dollars and routed through systems no single AI power fully controls — and that gap is where crypto already lives, whether or not any policymaker intends it to.
We are in a chop market. Nothing here is a directional trade yet. It is positioning, and the positioning is happening underneath the headline.
Context
The policy scaffolding is more concrete than the rhetoric suggests. At the Kazan summit in October 2024, Xi proposed a BRICS AI development and cooperation center. At the Rio de Janeiro summit in July 2025, leaders folded AI governance into the bloc's formal declaration. Weeks later, in Shanghai, Beijing floated a Global AI Cooperation Organization headquartered on Chinese soil — an institutional home for the Global AI Governance Initiative it first published in October 2023.
Strip the diplomacy and three material facts remain. BRICS now spans roughly forty-five percent of the world's population and a third of global output at purchasing power parity. Its New Development Bank has committed to lending thirty percent of its portfolio in local currencies by 2026. And every one of those ambitions has a payments leg, a compute leg, and a data leg.
Sit with the payments leg first, because it is the least glamorous and most decisive. Settling an AI service contract in rupees, reais, or dirhams requires a settlement rail that is not a dollar correspondent chain. That is the same problem decentralized finance has been solving, badly and loudly, for five years.
Now add the compute leg. Washington restricted Nvidia's H20 to China in April 2025, then structured an unusual revenue-share arrangement that August. DeepSeek's R1 release in January 2025 reset assumptions about how cheaply frontier-adjacent models can be trained. Huawei's Ascend line and Cambricon's accelerators are the strategic answer, and they are not available to permissionless networks on any terms.
Bring in the data leg. Harmonized AI governance across ten-plus jurisdictions means model registration, data localization, and audit trails. Those are, structurally, identity and provenance problems. Crypto has spent a decade building exactly those primitives, mostly for people who did not need them yet.
The overlap is not a coincidence. It is an architectural consequence.
Core
The settlement layer is where the contradiction lives
The BRICS de-dollarization narrative runs directly into a wall that my desk sees every day: the only stablecoin liquidity deep enough to settle real invoices is pegged to the dollar. The GENIUS Act in the United States and MiCA in Europe did not weaken that; they entrenched it, by giving dollar tokens a legal runway while leaving local-currency alternatives in regulatory limbo.
mBridge is instructive. The BIS Innovation Hub exited the multi-CBDC bridge in late 2024, handing it to participating central banks. That step was framed as maturity. Read from a market-structure angle, it was a transfer of an unfinished rail to institutions that cannot print liquidity into it. Cross-border CBDC bridges without deep two-sided FX markets are plumbing without water.
So the practical outcome for AI procurement is likely to be uncomfortable for everyone: BRICS compute contracts settled in tokenized instruments that are, functionally, dollar liabilities wrapped in sovereign branding. That is not de-dollarization. It is dollarization with better user experience. A rail that settles in someone else's unit of account is not a neutral rail — it is a rented one.
Decentralized compute is not sovereign compute
This is the part of the AI-and-crypto trade that gets marketed hardest and examined least. Networks offering permissionless GPU capacity — Akash, io.net, Render, Bittensor subnets, and the newer inference marketplaces — are genuinely useful. I have watched them absorb workloads that hyperscalers price out of reach for small researchers.
But follow the silicon. The overwhelming majority of that capacity is Nvidia hardware, purchased through jurisdictions that sit inside the American export-control perimeter. Permissionless compute networks are not an alternative to American silicon. They are its secondary market. That is a real business. It is not sovereignty.
And the economics are subsidy-shaped. Subnet emissions pay suppliers more than inference revenue does in most cases. When emissions compress, supply thins within weeks. I have watched this pattern before — it is the same treadmill that Layer 2 proving markets have been running on since 2023, where the cost of generating a validity proof still exceeds the value of the transactions it secures unless gas returns to bull-market levels. Operators are not profitable. They are venture-funded.
Verifiable inference inherits the proving bill
The most under-priced technical question in this whole discussion is what happens when regulators demand proof that a model ran as specified. The answer people reach for is zero-knowledge machine learning. The answer people forget is the invoice.
Based on my own benchmarking of zk proving pipelines, and my doctoral work in cryptographic protocol efficiency, the numbers are brutal. Proving a single inference pass over a mid-sized model in a general-purpose zkVM costs orders of magnitude more than the inference itself. For batch verification of model weights and training provenance, the math works. For per-request inference attestation at production volume, it does not, and no amount of recursive proof composition changes the asymptotics fast enough to matter this cycle.
Designers should stop pretending otherwise. The honest architecture is a hybrid: cryptographic proofs where the value at stake justifies the cost, and economic staking plus slashing where it does not. That is less elegant and far more deployable.
The oracle latency problem gets worse with agents
AI agents that transact on-chain need data feeds with sub-second finality. What they get, in most deployments, is a pull-based oracle with finality measured in tens of seconds and a node set whose decentralization rests on reputation rather than cryptographic guarantee. This has been my position for years, and agentic workloads make it structurally worse: an agent that prices risk on stale data does not make a bad decision, it makes thousands of them, in parallel, automatically.
Chainlink's dominance here is real and earned. It is also a governance question, not a cryptographic one. And when AI governance frameworks in any jurisdiction start mandating data provenance, the entities most likely to be certified as feeds of record are regulated ones — not anonymous node operators.
Community Pulse
The mood I am reading from builders is split, and the split is geographic. Developers in Nairobi, Buenos Aires, and Bangalore describe the BRICS AI push as the first credible path to compute they can afford without a US entity in the loop. Developers in Singapore and Zurich describe the same push as a compliance maze that will require model registration before deployment.
Both are correct. Anxiety is running higher than it was in the second quarter, and it is not price anxiety. It is jurisdictional anxiety — the specific fear that the rules will arrive before the tooling does.
Ethical Impact
Apply the metric honestly. Who captures the value of a state-aligned AI stack? Compute suppliers in sanctioned-adjacent jurisdictions, state champions, and a narrow layer of licensed integrators. Who absorbs the risk? Data laborers, whose contributions are rarely provenance-tracked, and end users, whose interaction data becomes a compliance asset rather than a personal one.
The ethical pulse of the decentralized economy beats loudest exactly where transparency is cheapest to fake. A BRICS AI framework that publishes audit standards but not data-ownership rights is a governance framework, not an ethical one. The distinction will matter within eighteen months, when the first enforcement actions land.
Contrarian
The consensus take is that a BRICS AI bloc is bullish for crypto, because it fractures American technological hegemony and creates demand for neutral rails. I think that reading is backwards, and it is the single most important thing I would tell an institutional client right now.
Sovereign AI and permissionless AI are competitors, not allies. Both want the same thing: models that run outside US jurisdiction. But they want it under opposite governance. A BRICS AI development center is, by construction, a state registration regime — harmonized data localization, model licensing, content filters, and audit trails with legal persons attached. Permissionless networks have no legal person to register.
Every framework published so far, in Washington and in Shanghai, assumes an accountable operator. Open weights and anonymous nodes are not addressed, because they cannot be addressed without admitting that the registry is porous.
The real contest is not Washington versus Beijing. It is jurisdiction-bound AI versus borderless AI — and both great powers are on the same side of that line. Building bridges in a fragmented digital frontier is only worth doing if the bridges connect somewhere. Building bridges in a fragmented digital frontier between two sovereign registries is a different job entirely.
The second unspoken truth: the neutral rails are not neutral. Settlement runs through dollar-pegged liquidity. Compute runs through American silicon. Model training runs through hyperscaler regions that sovereign clouds resell under local branding. Anyone pricing a sovereign AI stack without those three dependencies in the model is pricing a story, not a system.
Takeaway
Watch three things over the next two quarters. Local-currency stablecoin liquidity, which is the only honest measure of whether de-dollarization is real. The first BRICS-linked compute procurement contract settled on-chain, which will reveal whether the rails work under audit. And whether any AI governance framework grants legal standing to a permissionless model.
That third one is the whole game. If the answer is no, then the borderless AI future gets regulated out of existence before it gets built, and crypto's most useful contribution becomes a compliance tool for the very registries it was designed to route around.
So here is the question I am sitting with, in a chop market, with the desks quiet: what does a borderless model actually run on, when every border has a registry, every registry has an owner, and every owner settles in someone else's currency?