Grayscale Says AI Agents Will Need Blockchains

AI agents can already write code, book travel, and manage a calendar. What they can't do, at least not on a traditional bank account, is spend money without a human clicking "approve." Grayscale Research thinks that gap is about to close. It thinks public blockchains are the ones that close it.

In a research note published in late February, Grayscale's head of research Zach Pandl argued that AI and crypto aren't really in competition, even though markets have been trading them that way. Software stocks had just sold off hard on AI-driven volatility, and crypto fell right alongside them. Pandl's point was that the correlation was masking something else: blockchains, he said, are shaping up to be the financial rails AI agents actually use.

The logic is fairly plain. A chatbot today has no bank account. Opening one requires a human, a Social Security number, a signature. A blockchain address requires none of that. Any piece of software can generate one and start transacting in seconds, around the clock, without asking a bank's permission first. If AI agents are going to pay for API calls, compute, or data subscriptions on their own, wallets are the natural interface, and Pandl expects rising volumes of small, automated stablecoin payments to be the first sign the thesis is playing out.

Wallets, Identity And Who Controls The Models?

Grayscale's argument breaks into three pieces, and the first is the most concrete: money movement. Programmable wallets are already live. Ethereum's ERC-4337 standard, combined with the newer EIP-7702 upgrade, lets a wallet enforce spending limits, approved-address lists, and expiring session keys. In theory, that's what makes it safe to hand an agent a card with a ceiling on it rather than the whole account. MetaMask leaned into exactly that framing when it opened early access to its own Agent Wallet in June, pairing a "Guard" mode for tightly bounded trading with a looser "Beast" mode for agents given more room to improvise. The company made the announcement on X, framing it as agents trading DeFi positions across EVM chains while users kept their keys. Coinbase had shipped a similar Agentic Wallets framework a few months earlier.

The second piece is messier: identity and reputation. If a wallet can belong to a bot, the internet needs a reliable way to tell a person from a piece of software. An agent handling someone's money needs a track record it can point to before anyone trusts it with a real purchase. Ethereum's ERC-8004, co-authored with contributors from MetaMask, Google, and Coinbase, tries to solve this by putting agent identity and reputation into onchain registries instead of leaving it to whichever platform the agent happens to run on.

The third piece is the one Grayscale keeps returning to: decentralization itself. Pandl's argument is that concentrating AI model weights, compute, and governance inside a handful of companies is a structural risk, not just a competitive one. He made that case more pointedly in June, after the U.S. government temporarily restricted access to Anthropic's newly released Fable 5 and Mythos 5 models. It was, in his telling, a real-world example of what happens when a small number of firms and a single government decision can switch off AI access for everyone downstream. Networks like Bittensor exist to route around exactly that kind of chokepoint.

The Numbers Behind The Pitch

Bittensor is worth a closer look, because its case is less about a thesis and more about traction already showing up on-chain. The network now runs more than 128 active subnets, each an open marketplace paying TAO to whoever produces the most useful AI output on a given task: model training, inference, protein folding, fraud detection. In March, its Templar subnet finished training a 72-billion-parameter language model called Covenant, built collaboratively by more than 70 independent contributors on commodity hardware, with no permission required. TAO jumped roughly 90% on the news. Whether that scales into something that actually competes with OpenAI or Anthropic is still an open question. One subnet reportedly pulls in $52 million a year in emissions against just $2.4 million in outside revenue, which says the incentive design is running well ahead of real-world demand for now.

Ethereum and Solana show up in Grayscale's thesis for a simpler reason: they're where the wallet infrastructure already exists. Grayscale itself laid out the beneficiaries on X: lower-cost, high-throughput chains like SOL, BASE, and NEAR, stablecoin issuers like MKR, and DeFi protocols like UNI. It's a post that's aged into something close to a checklist for the trade.

None of this comes without a catch, and Grayscale's own report says as much. The tools that make AI agents more capable, pattern recognition at scale, automated code review, cut both ways. Better surveillance software makes on-chain privacy harder to hold onto. Agents probing for smart contract bugs work just as well for attackers as for defenders, which is part of why OpenAI recently launched EVMbench, a project aimed at using AI to find and patch those vulnerabilities before someone else does.

The bet, in the end, is a simple one: every task handed off from a person to an AI agent is a task that eventually needs a way to pay for itself. Grayscale thinks that's blockchain's job. Whether it plays out that cleanly is still, by its own admission, an open question, but the wallets, the identity standards, and the decentralized compute networks are already being built as if the answer is yes.