AI × Blockchain - GeneseAI

11.1Why pair AI with blockchain

Diagram11.1. The encounter, in theory. For each of AI's needs, blockchain claims to offer an answer. The question, taken up throughout the chapter, is to determine in which cases that answer is genuinely superior to centralized solutions.

11.2Autonomous economic agents

The emblematic player is Fetch.ai, which merged in 2024 with SingularityNET and Ocean Protocol within the Artificial Superintelligence Alliance (ASI), whose token is FET. In late 2025, Fetch.ai demonstrated one of the first fully autonomous payments from one AI to another, with no human in the loop. In 2026, the ecosystem is developing a chain dedicated to agents (ASI:Chain) and verifiability tools, such as AEVS (June 2026), which attaches to each agent action a signed, tamper-proof "receipt," compatible with the frameworks seen in Chapter 6 (LangChain, MCP).

Around this realization, payment protocols for agents emerged in 2025-2026. The most striking, x402 (launched by Coinbase), revives a forgotten web code, the HTTP status "402 Payment Required": when an agent requests a paid resource, the server responds with "402," the price and the destination address; the agent signs a stablecoin payment, attaches it to its request and obtains the resource, with no account or API key. At Google, AP2 (Agent Payments Protocol) plays a complementary role: it is not a payment rail but a framework that defines how an agent obtains authorization to pay (signed mandates, audit trail), compatible with cards, transfers and stablecoins. The topic now attracts the heavyweights: the consortium backing x402 (the x402 Foundation, hosted by the Linux Foundation) brings together names as diverse as Cloudflare, Circle, Stripe, Visa and Amazon Web Services. Above all, Stripe, the online payment giant, launched in 2026 its own machine payment protocol (MPP, built on its Tempo blockchain) and an agent wallet, allowing software to settle micropayments in stablecoin or by card; the long-established networks Visa and Mastercard are rolling out their own systems (Trusted Agent, Agent Pay). The shared conclusion: financial infrastructure was designed for humans, and it needs to be re-tooled for agents.

A new category of tools accompanies this shift: agent wallets, equipped with programmatic access, spending caps, rules and an auditable log, precisely to govern a piece of software that handles money. But it is worth keeping a sense of proportion: despite tens of thousands of active agents and tens of millions of dollars exchanged in early 2026, payments made by agents represent only a tiny fraction of stablecoin activity, which is itself overwhelmingly non-commercial. The phenomenon is therefore real but embryonic, and it raises open questions: security (a hacked agent that drains a wallet), identity (knowing which agent is paying, hence standards such as ERC-8004), refunds, and regulation (the U.S. framework known as the 2025 GENIUS Act on stablecoins). It is the most concrete illustration of the "agentic economy"—and the one to handle most cautiously.

In context

The oracle problem (connecting the chain to the real world)

A fundamental limitation of smart contracts (Section 11.1) is that they only "see" what happens on the blockchain: by construction, they have no direct access to the outside world (a stock price, the result of a match, the delivery of a parcel, or the output of an AI model). To bridge this gap, oracles are used: services that route external data onto the chain in a verifiable way. The challenge, known as the oracle problem, is to avoid reintroducing a single trusted third party that would ruin decentralization; decentralized oracle networks are therefore used (the best known being Chainlink), where several independent sources agree on the value to record. For convergence with AI, the stakes are twofold: an oracle can feed a contract with the output of a model (for example a risk score), and it can allow an autonomous agent (Chapter 6) to trigger on-chain actions in response to real-world events. The oracle is thus the indispensable bridge—and the most delicate link—between AI, the physical world and blockchain.

11.3Decentralized networks for compute and intelligence

In plain terms

Faced with the concentration of compute among a handful of giants (Chapter 8), some networks are trying to create decentralized markets.

Decentralized compute. Networks like Render or Akash pool the idle GPUs of individuals and companies, rewarded in tokens, to offer an alternative to the giants' cloud. The argument is economic: democratizing access to computing power. The open question is reliability (quality of service, latency) compared with centralized providers.
Decentralized intelligence. The most ambitious project is Bittensor (token TAO, one of the largest market caps in the sector, on the order of 3 to 4 billion dollars). Its principle: a network of specialized "subnets" where AI models "compete and collaborate," rewarded in TAO according to the value of their contribution, assessed by validators (a mechanism dubbed "proof of intelligence"). In March 2026, one of its subnets (Templar) ran what is presented as the largest model training ever carried out on a decentralized network: Covenant-72B, a 72-billion-parameter model trained by more than 70 contributors on consumer hardware connected over home internet, with no data center at all. The feat relies on a technique for compressing the exchanges (SparseLoCo) that cuts the required bandwidth by roughly 146 times, and the weights are released open source (Apache license). It is worth keeping a sense of proportion, however: with a score of around 67 on the MMLU benchmark, the model sits at the level of a Llama 2 70B (2023) and remains very far from the frontier, which illustrates both the viability and the current limits of decentralized training.
Data markets. Ocean Protocol aims to enable the secure exchange of data (the fuel of AI), including through federated learning (models learn without centralizing the raw data).

In context

Decentralized physical infrastructure networks (DePIN)

These shared-compute networks belong to a broader family, designated by the acronym DePIN (Decentralized Physical Infrastructure Networks): projects that reward, in tokens, the individuals or companies pooling a real physical resource, whether computing power (idle GPUs), storage, bandwidth or wireless coverage. The appealing idea is to build infrastructure "from the bottom up," with no central operator, by aggregating dormant capacity. Applied to AI, it aims to pool GPUs for training or inference at lower cost. The limits are the same as those mentioned above (verifying the quality of contributions, holding up under scale, rivaling the efficiency of the large data centers), and a share of these projects is more speculative token than real use. But the model illustrates a recurring avenue: pitting decentralized pooling against the concentration of compute resources (Chapter 8).

In context

Blockchain for training medical AI (the Galeon case)

A concrete example links this chapter to healthcare (Chapter 14). The French company Galeon publishes an electronic health record deployed in nineteen hospitals, and underpins it with a technique called Blockchain Swarm Learning: a variant of federated learning (seen above) in which the AI model travels from one hospital to another to learn, while the patients' data, for its part, never leaves the institution. The blockchain does not store the medical data there: it records each training cycle in a tamper-proof way, in order to guarantee hospitals control over their data. To this is added, as is often the case in this field, a crypto dimension (a token, GALEON, and a decentralized autonomous organization meant to fund research). The interest is real—reconciling medical AI and confidentiality—but it is worth keeping a sense of proportion: this is a niche player whose promise remains to be confirmed at scale, and the speculative side of the token calls for the same caution as the other projects in this chapter.

11.4Cryptography: zkML and homomorphic encryption

In plain terms

Here is a quieter but technically profound convergence, one that could solve two major problems of AI: confidentiality and verifiability.

Fully homomorphic encryption (FHE). This is a near-magical form of cryptography: it allows computation on encrypted data without ever decrypting it. Applied to AI, it would let a service analyze your data (a medical record, for example) without ever having access to it in the clear. Companies such as the French firm Zama are working to make this technology practical.
Zero-knowledge machine learning (zkML). Built on "zero-knowledge proofs," zkML makes it possible to prove that a model did indeed produce a given result, without revealing either the model or the input data. Useful for guaranteeing that an AI ran honestly, or for certifying a result without disclosing trade secrets. Players such as Polyhedra Network are exploring this path.

11.5Prediction markets and the data economy

Key takeaways (Chapter 11)

AI needs compute, data, payments, coordination and trust; blockchain claims to offer decentralization, incentives, ownership and traceability. The pairing is real but heavily oversold.
Autonomous economic agents pay one another in programmable money: stablecoins and dedicated protocols (Coinbase's x402, Google's AP2) are the emerging rail (Fetch.ai/ASI, but also Visa, Mastercard, Stripe). A promising convergence, but still embryonic and risky (security).
Decentralized networks for compute (Render) and intelligence (Bittensor and its subnets) offer an alternative to the giants, but still struggle to match the centralized approach and to prove their real usefulness.
Cryptography (Zama's FHE, Polyhedra's zkML) promises a confidential and verifiable AI, at the cost of a heavy computational overhead.
Prediction markets (Polymarket) and data tournaments (Numerai) connect crowd, data and AI.
Honest verdict: plausible use cases do exist (agent payments, provenance, decentralized compute, confidentiality), but the sector remains dominated by narrative, and the value of tokens is often disconnected from utility.

From the distributed ledger, let us move on to a rupture of an entirely different physical nature: quantum computing.

11.1Why pair AI with blockchain#

11.2Autonomous economic agents#

11.3Decentralized networks for compute and intelligence#

11.4Cryptography: zkML and homomorphic encryption#

11.5Prediction markets and the data economy#

Key takeaways (Chapter 11)

11.1Why pair AI with blockchain

11.2Autonomous economic agents

11.3Decentralized networks for compute and intelligence

11.4Cryptography: zkML and homomorphic encryption

11.5Prediction markets and the data economy