An AI agent calls an API, fetches a dataset, pays for it — and moves on. No account, no API key, no human in the loop. This is already happening at scale in 2026: a Keyrock report (with Coinbase, Tempo and Virtuals) counts roughly $73M across 176M transactions settled by AI agents. The median payment is between $0.01 and $0.10; 76% of all payments fall below the $0.30 floor where card fees even begin.

That single number decides almost everything. When the typical payment is a few cents, card rails are mathematically dead — the fee is bigger than the purchase. Which makes the blockchain architecture underneath the real question. Which Layer 1 is built best for this era?

The honest answer has no single winner. It splits by the layers of the stack.

What the job actually demands

Before comparing chains, the spec. A rail for agent payments needs: sub-cent fees (or overhead eats the micropayment), sub-second finality (agents work synchronously, not in batch cycles), programmable spend limits (an agent must not be able to drain the whole wallet), and identity plus reputation (who is accountable when an agent is wrong or malicious?). On top of that, stablecoin-native settlement, because value moves in USDC, not in a volatile chain token.

Over all of it sits the standard that ties it together: x402 — launched by Coinbase in May 2025, reviving the long-dormant HTTP 402 ("Payment Required") status code. It now lives under the Linux Foundation umbrella, with Coinbase, Cloudflare, Google, Visa, Mastercard, Stripe, Circle and the Solana Foundation on board. Google folded x402 into its Agent Payments Protocol (AP2); more than 100M payments have already run across chains. So x402 isn't chain-bound — the interesting question is which chain carries settlement best.

Solana: the settlement layer that's winning now

When it comes to where the money actually moves, Solana is ahead today. It handles around 65% of all x402 payments and processed roughly $650B in stablecoin volume in February 2026 alone. Finality near 400 milliseconds, fees around $0.00025 — the order of magnitude at which cent-sized payments make sense. Firedancer added a second client for resilience; the upcoming Alpenglow upgrade targets finality in the ~150-millisecond range.

The strongest evidence isn't a number, it's a decision: Google Cloud picked Solana as the rail for Pay.sh — agents paying in USDC for services like Gemini, BigQuery and Vertex AI, with no account or API key. When a hyperscaler chooses the rail, that's an architecture verdict. Solana's weakness remains the historical one: a long-single-client setup (now addressed by Firedancer) and an account-based execution model that's more contended under load spikes than parallelized designs.

Ethereum & Base: the trust and authority layer

Solana wins settlement — but settlement is only half the job. The other half is the question the hype skips: who is this agent, and what may it do?

Here Ethereum is uniquely positioned. ERC-8004 (live on mainnet since January 29, 2026, built by the Ethereum Foundation, MetaMask, Google and Coinbase) gives agents an on-chain identity with a reputation and validation registry — deliberately leaving out payment mechanics so it composes with x402. Tens of thousands of agent identities are already registered, across chains (Base, Arbitrum, Polygon, BNB). Add ERC-4337 plus EIP-7702: account abstraction with session keys, letting an owner scope authority tightly — "this agent may spend $50/day on data, anything above $100 needs a human signature." That's the programmable spend limit from the spec, in its purest form.

And x402 itself was born on Base, Coinbase's Ethereum L2. L2 fees (around $0.01–0.05) sit above Solana's level, but Ethereum's role is different: not the cheapest settler, but the most neutral, most decentralized substrate for identity, reputation and accountability. Settlement can happen elsewhere — accountability lives here.

Sui: the cleanest architecture — with the smallest crowd

If you ask the question purely technically — which general-purpose L1 is structurally best built for agents — the answer is Sui, and it's not close.

Sui's object-centric data model treats assets as first-class objects with ownership written into the code (via the Move language). Independent transactions run in parallel, and payments using "owned objects" skip the full consensus round entirely — producing sub-second finality at sub-cent fees without thousands of agents sharing one sequential queue. That parallelism is exactly what swarms of agents structurally need. Add an AI-adjacent toolkit: Walrus for scalable storage, SEAL for encrypted access, Nautilus for verifiable compute — primitives a multi-agent system actually uses.

The catch is just as real: Sui's ecosystem is smaller than Ethereum's and younger than Solana's, DeFi TVL hovers around $1B, and in concrete x402/stablecoin adoption Sui trails the two giants. The best architecture means little if the liquidity, the tools and the payment flows still sit elsewhere.

The leapfrog: chains built only for payments

The ETH-vs-SOL-vs-SUI debate misses an entire category: Layer 1s that exist for nothing but stablecoin payments.

Tempo (by Stripe and Paradigm, led by Paradigm co-founder Matt Huang) is the sharpest example. EVM-compatible, but with payment-specific moves: no native gas token — the fee is paid in stablecoins, a built-in AMM swaps into the validator's preferred currency, making the chain issuer-agnostic. Dedicated payment lanes, predictable blockspace, account abstraction, and ISO-20022-compatible metadata fields (for invoice IDs and cost centers) bridge crypto and the banking world. Tempo also co-created the Machine Payments Protocol (MPP), explicitly for agent payments, and was built with design input from OpenAI, Anthropic, Visa, Shopify and Deutsche Bank. Target: up to 100,000 TPS. Public testnet since December 2025, mainnet expected in 2026.

Arc (by Circle, the USDC issuer) is the counterpart: a USDC-native L1 for institutional settlement, treasury and tokenized assets — the stablecoin issuer building its own rail. Alongside them are Tether-aligned chains like Plasma and Stable for USDT, and Stellar, an established payments network that has also integrated x402.

Purely against the job, these chains may be the best-fit architecture of all. The catch: their mainnets are barely live, adoption near zero. Best-built is not the same as most-adopted.

The comparison at a glance

The honest answer

"Which L1 is best positioned?" can't be answered with one name without being dishonest. For settling high-frequency micropayments today, it's Solana. For the trust and authority layer — identity, reputation, programmable spend rules — it's Ethereum with Base. For the cleanest general-purpose architecture, it's Sui, which still has adoption to catch up on. And for purpose-built payment architecture, it's Tempo and Arc, possibly the furthest-thought-through technically and the least proven.

The real risk isn't picking the wrong chain. It's concentration: a single asset, USDC, settles roughly 98.6% of agent payments. A single standard, x402, carries most of the traffic. A handful of rails holds almost everything. A monoculture is efficient — until it fails. That's the variable worth watching, long before "SOL or SUI?" is decided.

From the trenches

This take isn't from the outside: we run an x402 agent API ourselves, on Base — pay-per-call in USDC, no account, no bearer token. The tech works; it's not a promise, it's in production. Which is exactly why the sober view is warranted: the rails are real, and so is part of the hype — and the difference is whether an architecture solves the job or just sounds good.

Study the Past — Improve your Future. 🥋