Twenty-nine nations signed an agreement in Shanghai to form the World Artificial Intelligence Cooperation Organization (WAICO). The blockchain industry barely noticed. That’s a mistake.
Most crypto natives are fixated on the next L2 launch or the latest MEV exploit. But this is the kind of infrastructure play that will shape the regulatory and technical environment for decentralized AI—a sector I’ve been building in since 2026, when I prototyped a ZK-circuit to verify off-chain model outputs for a leading AI lab. That project taught me that the line between “open” and “trusted” is razor-thin. WAICO threatens to blur it entirely.
Context: What WAICO Actually Is
The organization claims to “lower the barrier to AI adoption” through open-source models, technical training, and shared computing resources. Its 29 founding members include China, Russia, Cuba, and a majority of African and Asian nations. Notably absent: the U.S., EU, Japan, and South Korea. The headquarters is in Shanghai.
On the surface, it’s a multilateral tech cooperation forum. In practice, it’s a geopolitical vehicle to export China’s AI stack—including its model architectures, training pipelines, and potentially its chip ecosystem—to the Global South. The framing is “AI for all,” but the subtext is “AI by us.”
For the crypto world, this matters because WAICO’s open-source models will become the default choice for developers in those 29 countries. Many of them will build DeFi dApps, NFT marketplaces, and even on-chain AI agents using these models. If those models are opaque, biased, or backdoored, the entire Web3 stack built on top is compromised.
Core Analysis: Open Source ≠ Verifiable
WAICO promises “open-source models.” The blockchain community loves open source. But we need to ask: open source in what sense? The source code of the model architecture may be public, but the training data, the weights, the inference environment—none of that is verifiable on-chain or with zero-knowledge proofs unless explicitly designed.
During my 2026 audit of an AI-oracle system, I discovered that even fully open-source models (like those from Meta’s Llama family) can have latent biases introduced during training that are invisible to anyone without access to the full dataset. The difference is that Llama’s weights are available for independent third-party verification via ZKML—something I had to build from scratch because existing tools couldn’t handle the proof size.
WAICO’s models, by contrast, will likely be distributed under licenses that restrict redistribution or commercial use. Or worse, they’ll be open-source in name but require connecting to a centralized inference API run by a Chinese state-owned cloud provider. That’s not openness—it’s vendor lock-in with a freedom sticker.
Math doesn’t negotiate. If the model weights aren’t publicly auditable, any zero-knowledge proof generated from that model is suspect. The prover could be cheating by using a different model. This is the same trust assumption that plagues centralized oracles. WAICO’s “open source” could actually set back the verifiable AI movement by creating a false sense of transparency.
My hands-on experience with ZKML
In 2022, I spent six months building a Groth16 prover in Rust. That taught me that proving a model’s output is correct requires the entire model to be committed to in a public, immutable way. During my 2025 collaboration with a legal-tech startup, I integrated a ZK-proof circuit that verified creditworthiness without leaking personal data. The key insight: the model itself had to be hashed and stored on-chain. If the model is controlled by a single entity (China through WAICO), then the security of any application relying on that model is only as strong as the trustworthiness of that entity.
Privacy is a feature, not a bug. WAICO’s narrative of “lowering barriers” could be used to justify data localization requirements or mandatory use of Chinese-hosted models. For a DeFi protocol that needs to run AI-based risk assessments on user portfolios, relying on a WAICO model could mean handing sensitive transaction data to a foreign government. That’s not just a buzzkill—it’s a compliance nightmare under GDPR and emerging crypto-specific privacy regulations.
Contrarian Angle: Could WAICO Actually Benefit DeAI?
One could argue that WAICO’s standardization of AI infrastructure might accelerate the adoption of verifiable computation. If models are widely deployed under a unified framework, projects like EZKL or Modulus Labs could build ZK circuits that prove inference against those specific models. The network effects of a single model family could make it economically viable to produce high-quality ZK proofs for inference—something that currently remains expensive for the hundreds of bespoke models floating around.
But this requires WAICO to commit to open, auditable standards. Everything about the organization’s structure—its membership, its location, its timing—suggests the opposite. The absence of Western oversight, the inclusion of Russia and Cuba (both under sanctions), and the focus on “technology training” (a euphemism for dependency creation) all point to a closed ecosystem.
Code is law, but bugs are reality. The most dangerous scenario is not that WAICO’s models are malicious, but that they are buggy in ways we can’t detect. A subtle numerical instability in a transformer layer could cause a lending protocol to misprice risk, leading to a cascade of liquidations. Without cryptographic verification of the inference path, we would blame the smart contract, not the model. I’ve seen this pattern before: during the 2021 LUNA crash, the Anchor Protocol’s withdrawal function had an integer overflow that amplified the death spiral. That bug was in the code, but the root cause was an oracle design flaw. With WAICO, the “oracle” is an entire AI model—and we won’t even see its full code.
Takeaway: The Clock Is Ticking
The blockchain industry has a choice. We can ignore WAICO and let it become the default AI backend for billions of users, or we can proactively build the verification infrastructure that turns any AI model—centralized or not—into a trustless component. The tools exist: ZKML, TEEs for inference, and on-chain model registries. What’s missing is demand. WAICO’s creation should be the wake-up call.
Over the next six months, I will be auditing the first open-source model released under WAICO’s banner. I’ll publish a full forensic analysis of its architecture, training data provenance, and inference API. If you’re building on top of AI in crypto, pay attention. The future of verifiable intelligence depends on it.