If you think a position in Nvidia covers your AI exposure, Franklin Templeton’s head of digital assets has news for you: the trade you’re missing isn’t in chips or model labs at all. It’s in the payment rails that autonomous software will use to buy things — and in the firm’s view, those rails are blockchains.
Sandy Kaul, Head of Digital Assets and Innovation at Franklin Templeton, published a paper this week arguing that agentic AI — software that shops, books, and pays on your behalf without asking permission at every step — will settle its transactions on crypto networks rather than traditional finance, as reported by Decrypt. Coming from a manager overseeing roughly $1.8 trillion in assets, it is one of the most direct institutional endorsements yet of the idea that machine-to-machine commerce is crypto’s long-awaited breakout application.
“Blockchain will be pivotal in allowing agentic AI to realize its potential for consumer transactions, and the growth of agentic AI is likely to become the ‘killer’ use case that drives blockchain adoption,” the firm wrote.
The use case, concretely: software that pays software
The distinction Kaul draws is between chatbots that answer and agents that act. A Capgemini report cited in the paper describes the shift as moving AI “from a reactive, conversational chatbot to an autonomous system that can perceive its environment, devise a plan, and execute multi-step tasks to achieve high-level goals without constant human supervision.”
Play that forward and the transaction pattern changes shape entirely. An agent renewing your grocery order, comparing flight fares across a dozen APIs, buying compute for a rendering job, and paying a data vendor for a single query doesn’t produce one card swipe a day — it produces thousands of tiny purchases per hour, many of them worth cents. A Bain & Company forecast referenced in the paper expects AI agents to account for 15% to 25% of all U.S. e-commerce sales by 2030. McKinsey & Company sizes agentic commerce overall at $3 trillion to $5 trillion by the same year.
Card networks were built for none of this. They were designed around human-scale purchasing: a person, a checkout page, a signature, and a settlement process that finishes days later. Micropayments between machines break every one of those assumptions — the fee structure alone makes a $0.04 API purchase absurd on a card rail.
The speed-gap argument
Franklin Templeton’s core technical claim is that public blockchains already clear and settle faster than the legacy system — and that the difference compounds when the buyer is a machine.
The paper’s numbers: Bitcoin processes roughly 7 transactions per second and Ethereum about 75, but newer high-throughput chains have recorded maximums of 12,933 TPS on Aptos, 6,284 TPS on Solana, and 3,252 TPS on BNB Chain. Visa, by comparison, handles 1,700 to 10,000 transactions per second in normal operations.
Then comes the part Kaul says makes the raw comparison misleading in crypto’s favor: “Blockchains both record and settle their transactions in that TPS window whereas the Visa network only records a transaction. Settlement on the Visa network takes 1-3 business days.”
For a human buying coffee, the difference between recorded-now and settled-Thursday is invisible. For an autonomous agent making thousands of micropayments an hour — where each counterparty is also software that wants finality before releasing data or compute — a multi-day settlement lag is a structural defect, not an inconvenience. That is the “speed gap” the firm believes gives distributed ledgers a durable advantage as agentic traffic grows.
Anyone who has watched priority fees spike on a busy chain will note the caveat: peak-lab-condition TPS numbers are marketing figures, and real-world throughput under adversarial load is lower everywhere. But the settlement-finality point holds regardless of whose benchmark you believe — a card network authorization is simply not money moved.
The rails are already being laid
What separates this paper from the many “blockchain will disrupt payments” decks of past cycles is that the infrastructure it describes is shipping now, with legacy payment giants participating rather than resisting.
The clearest example is the x402 protocol, an open standard that lets software pay software directly over HTTP. The x402 Foundation — around 40 organizations including Visa, Mastercard, and AWS — formally launched on July 14 to steward it. The protocol’s namesake is the HTTP 402 “Payment Required” status code, reserved in the web’s earliest specifications for a payments feature that never materialized. Three decades later it is being put to work for machine-to-machine settlement, with stablecoin transfers filling the slot the original web left empty.
Google, meanwhile, unveiled its Agent Payments Protocol (AP2) in 2025 — a framework developed with dozens of payments and crypto partners for letting AI agents transact with verifiable user mandates, including an extension for stablecoin settlement built with the Ethereum ecosystem. Coinbase has shipped developer tooling that lets agents hold wallets, trade, and pay autonomously. None of these are whitepaper vapor; they are live standards with Fortune 500 members and working code.
The pattern rhymes with how stablecoins themselves grew: first dismissed as a crypto-internal convenience, then quietly adopted as the settlement layer for cross-border flows once the volume was undeniable. Agent payments look set to follow the same path, except the end user this time is not a person at all.
What it means for a portfolio, per Kaul
Here is where Franklin Templeton’s argument turns from infrastructure analysis into an investment thesis — and where it will be most contested.
If agents buy data, compute, and API access on-chain, each transaction consumes the native token of whatever network carries it: SOL on Solana, ETH on Ethereum, and so on. Token demand, in this model, scales with agentic volume the way fuel demand scales with traffic. Kaul’s conclusion is blunt: “I believe what will become increasingly clear in coming years is that in order to capture the value of decentralized networks and businesses, investors will need to buy the cryptocurrencies and alt coins being issued by those entities.”
Translated: equity in AI companies captures the application layer, but the settlement layer’s economics accrue to token holders — and most traditional portfolios have zero exposure to it. For an asset manager of Franklin Templeton’s size (its digital-assets research arm has been building this thesis across multiple papers) to say so in writing is notable, because it reframes token ownership as infrastructure exposure rather than speculation.
The counterarguments deserve equal airtime. Fee markets are ruthlessly competitive, and history shows transaction volume does not automatically translate into token value — throughput can migrate to whichever chain is cheapest, compressing the very fees the thesis depends on. Stablecoin-denominated agent payments could also concentrate value in issuers rather than in the volatile native assets Kaul highlights. And every forecast in the paper — Bain’s e-commerce share, McKinsey’s trillions — is a projection about technology that is still learning not to hallucinate a shipping address.
If you build or hold, here’s the practical read
For builders, the near-term opportunity is unglamorous plumbing. Agent commerce needs spend-limit tooling, receipts an auditor can read, per-agent wallet isolation, and dispute mechanics for when an autonomous buyer purchases the wrong thing — none of which exist in mature form today. Teams already shipping x402-compatible endpoints are effectively squatting on the toll booths of a highway that Visa, Mastercard, and AWS have publicly committed to paving. The historical analogy is Stripe circa 2011: the payments themselves were commoditized, but the developer experience around them built a $50 billion company.
For holders, the thesis suggests a different lens on the assets you may already own. If Kaul is right, the metric that matters for L1 tokens gradually shifts from retail speculation cycles toward machine-originated transaction volume — a number that, unlike sentiment, can be measured on-chain week by week. It also implies the fee-market design of each chain becomes an investment variable: networks that can profitably clear sub-cent payments capture agent traffic, and networks that can’t, don’t. That is a testable claim, which is more than most crypto narratives offer.
What to watch next
Three signals will show whether this thesis is compounding or stalling. First, x402 adoption metrics: if agent-initiated stablecoin payments start appearing at meaningful volume, the “killer use case” claim gets its first hard data. Second, whether AP2-style mandates get integrated into consumer products — an agent that can provably spend only what you authorized is the unlock for mainstream trust. Third, watch whether high-throughput chains prioritize the sub-cent fee tiers machine commerce requires, because a $0.02 fee on a $0.04 purchase kills the economics as surely as a card network would.
The honest summary: an institution with $1.8 trillion under management just argued that the most consequential AI trade may be denominated in tokens, not shares. Whether or not you buy the portfolio advice, the infrastructure it describes — Visa and AWS building open rails for software to pay software over a resurrected HTTP status code — is no longer hypothetical. The machines are getting wallets. The only question left is which ledger they’ll spend from.
