Update (23 September 2026): This is the second piece in a pair. The first concluded that AI agents will most likely default to Tempo — a Stripe-governed, fee-free stablecoin rail — and that the agentic economy will absorb one permissioned strand of crypto rather than bypass it. This piece asks the question that conclusion leaves hanging: if the rail is free, where does the money actually go?
A payment rail that charges nothing sounds like a business with no business model. It isn’t — it’s a business model where the transaction is the cost, not the revenue. Every fee-free rail in financial history (card networks before interchange, exchange settlement systems, and now stablecoin chains) has made its money somewhere adjacent to the payment: on the balances sitting still, on the services wrapped around the movement, or on the data the movement generates.
So the question worth testing is not “will agents pay on Tempo?” but who gets paid when they do — and through what mechanism. Let’s work through the candidates honestly. Some are genuinely well positioned. Several are not.
The Question the Free Rail Poses
Worth restating the mechanics, because they drive everything that follows. Tempo launched mainnet on 18 March 2026. It is a Layer 1 built by Stripe and Paradigm for stablecoin payments: no native gas token, finality in under a second, and fees of a fraction of a cent that can be paid in any USD stablecoin.
That last detail matters more than any other for the money question. When settlement is near-free and the fee token is a dollar, the marginal cost of moving money approaches zero — and a business cannot charge a spread on something whose marginal cost is zero. The tollbooth on the transaction is gone. The value has to be collected somewhere else.
There are realistically only five places it can go: balances (float), adjacent services, distribution (owning the point where the user or agent sits), risk and compliance, or nothing at all — the genuine possibility that this is a low-margin future for everyone.
Candidate 1: Stripe Itself — Owning a Rail Without a Tollbooth
Stripe is the most obvious answer and the hardest to price, for the plainest possible reason: Stripe is private. After a February 2026 employee tender offer it was valued at $159bn, and secondary-market pricing in mid-August 2026 implied $172–176bn. No listing, no published quarterly, no way for a public-market investor to own it directly. That single fact reshapes the whole question for anyone who is not an accredited private-market participant.
Stripe’s economics are large and mostly still transaction-based. Net revenue was around $6.8–6.9bn in 2025 (up 33–36%), and $1.9tn of total payment volume ran across its rails. Its headline card price — 2.9% + $0.30 for a typical US online charge — is not where Tempo sits, and the effective net take rate after interchange is far lower, closer to 0.40% on the core checkout stack.
So how does Stripe monetise a rail it deliberately makes free?
- Adjacent services. Stripe’s Revenue and Finance Automation suite is heading towards a $1bn annual run rate — billing, tax, invoicing, treasury. This is classic fintech practice: give away the movement, sell the workflow.
- Owning the routing layer. The pending OpenRouter acquisition — agreed on 19 August 2026 at a reported $7bn-plus, still working through customary closing conditions in late September — matters here because OpenRouter routes traffic across 400+ AI models. If Stripe sits between agents and the compute they buy, it monetises the spend decision, not just the settlement.
- Owning the issuer. More on this below, because it is the sleeper.
- Float and treasury income. Stripe holds and moves customer balances. Free settlement increases the incentive to leave money on the platform.
Read together, Stripe is not betting on a transaction fee. It is betting that whoever owns the checkout, the settlement layer and the model-routing layer doesn’t need to charge per transaction — the money comes from everything wrapped around it. Well positioned, and structurally hard to compete with. Also: not publicly investable.
Candidate 2: The Stablecoin Issuers — Where the Float Actually Goes
Here is the candidate that deserves the most serious scrutiny, because the mechanism is the most robust one on this list.
A stablecoin issuer takes a dollar in, issues a token, and invests the dollar in short-term Treasuries and money-market funds. It keeps the interest. It pays the holder nothing — and under the GENIUS Act’s framework, payment stablecoin issuers are prohibited from paying holders yield, with US licensing taking effect in January 2027.
That is a balance-sheet rent, not a transaction toll. It scales with the size of the float, not the number of payments. A fee-free rail is, if anything, good for it: lower settlement costs encourage more stablecoin circulation, and a rail designed for machine-speed payments encourages balances to sit on-chain rather than sweep back to a bank account after every transaction.
The numbers are not small.
- Tether — private, and the single best illustration of the model. Q1 2026 net profit of roughly $1.04bn, Q2 2026 net operating profit around $1.5bn, on the back of a ~$141bn Treasury portfolio that was throwing off an estimated $4bn a year in income. Total assets of about $187.75bn against liabilities of $183.64bn at 30 June 2026. Its USDT supply sat near $183bn in September 2026.
- Circle — public, and the cleanest listed pure-play on the mechanism. Q2 2026 total revenue and reserve income of $701m, of which roughly $668m was reserve income. Average USDC in circulation was about $76.5bn; USDC’s market cap sat near $75bn in September 2026.
- Bridge — and this is the detail that ties the whole piece together. Bridge, acquired by Stripe for $1.1bn (announced October 2024, closed February 2025), issues USDB, the stablecoin that shows up on Tempo. In February 2026, Bridge received conditional OCC approval for a national trust bank charter.
That last point is the sharpest single insight in this analysis. Stripe owns the rail and an issuer on it. If the float is where the money is, Stripe collects it directly, under federal supervision, on a rail it also governs. The free transaction isn’t charity — it’s customer acquisition for the balance sheet.
Circle’s version of the story comes with an important asterisk. Its reserve income is real, but it leaks heavily to distribution partners. In Q2 2026, Circle’s distribution and transaction costs were about $410m — of which $324.6m (roughly 79%) went to Coinbase under their revenue-share agreement, renewed through 2029. Coinbase takes 100% of reserve income on USDC held on its platform and 50% of residual off-platform income. Circle owns the mechanism but not the customer. That is the difference between owning a rail and owning the money on it.
The vulnerability across the whole category is interest rates. Every dollar of that income depends on the level of short-term rates — the FOMC’s September 2026 projections put rates at 4.1% by year-end 2026. A rate-cutting cycle compresses reserve income mechanically, with no lever the issuer can pull. Float income is the strongest mechanism on this list, but it is a levered bet on the rate cycle.
Candidate 3: The Card Networks — Collecting the Layer Agents Can’t Settle
If agents pay in sub-cent amounts, cards are structurally excluded from most of that flow: a $0.30 fixed fee makes a $0.05 payment impossible. One Keyrock analysis in May 2026 found that roughly 76% of agent transactions fall below the card fee floor. So cards lose the micropayment layer by arithmetic.
They do not lose the whole market. Visa and Mastercard are both public, both are building agent-specific rails, and both are leaning on the thing stablecoins still cannot provide: dispute resolution, chargeback rights and consumer protection.
- Visa’s stablecoin settlement run rate reached $20bn annualised by September 2026 — tripling in six months — supporting 160-plus stablecoin-linked card programmes, with the Visa Stablecoin Platform launched in July 2026. Its agent play is Trusted Agent Protocol and Intelligent Commerce Connect.
- Mastercard’s Agent Pay and Agent Pay for Machines use “Agentic Tokens” and “Verifiable Intent” to bound what an agent can authorise.
- Both are building a cross-network “Know Your Agent” framework with Ant International.
There is a fee-headwind too: the $38bn Visa/Mastercard interchange settlement, preliminarily approved in June 2026, cuts posted credit interchange by 10 basis points for five years and caps standard consumer credit at 1.25% for eight. Card economics are drifting down, not up.
Well positioned in a specific niche — the disputed, higher-value, consumer-facing purchase — and structurally shut out of machine micropayments. They collect where trust is the product, not where it is free.
Candidate 4: The Agent Platforms — Owning Demand
OpenAI and Anthropic are both private, as is every serious frontier lab, so this candidate is not directly investable either — but its positioning is real. Whoever owns the agent closest to the user owns the default. If your agent picks the rail, the merchant, and the model, you are the demand, and the rail becomes a commodity underneath you.
That is why the coalition matters and why “design partner” is such a loaded phrase. Stripe’s Agentic Commerce Protocol, co-developed with OpenAI, powers Instant Checkout inside ChatGPT. Anthropic sits in the same orbit. But design input is not exclusivity — if OpenAI and Anthropic keep routing across multiple rails, their leverage is that they can switch, and they will extract terms on that basis.
The more interesting structural point: AWS entered this fight. Amazon Bedrock AgentCore Payments, in preview from around May 2026, lets agents make micropayments, with Coinbase and Stripe supplying the wallets and rails. Coinbase Agentic Wallets launched in February 2026; Circle’s Agent Stack followed in May 2026. The point is that the agent platform is a genuine chokepoint — but it is contested by the hyperscalers, not owned by the labs. Strong position, unresolved ownership, and for public investors, reachable only indirectly through the partners the labs choose.
Candidate 5: Picks and Shovels — Compliance, Identity, Custody, Orchestration
The layer that never has to pick a winner is the least glamorous and the most reliably paid. Whoever wins the agent-payments race, it will need identity (is this agent authorised?), compliance (is this counterparty sanctioned?), custody and key management, and orchestration to decide which rail a given payment should use.
This is where Coinbase (public) has moved fastest — Agentic Wallets and x402 support — and where a cluster of private specialists (Fireblocks, Chainalysis, Crossmint, Nevermined, FluxA) are building agent-native wallets, metering and identity. The mechanism here is a licensing or usage fee on a service that is not the payment, which makes it immune to the fee compression that hits the transaction. Consistently well positioned, individually hard to size, and mostly private outside the listed exchanges.
Candidate 6: The Chains and Tokens — Does Rail Success Accrue to Holders?
The candidate RWA investors most want to work is the one that works least cleanly. Tempo has no native token, and its fees are paid in stablecoins. Whatever value Tempo creates as a rail, there is no token whose supply captures it — the equity sits in Stripe, which is private.
For other chains, the question is whether agent activity translates into token demand. The evidence is mixed. x402, Coinbase’s open stablecoin payment protocol, processed roughly 165 million agent transactions and about $50m of volume by April 2026 — an average payment near $0.31 — and was running around 29 million transfers a month at about six cents each by August 2026. Real growth. But one widely cited analysis suggested only 0.6% to 7.5% of x402 payment value is genuinely agentic rather than human-driven or scripted, and value settled in stablecoins doesn’t automatically bid up any gas token — especially on chains where gas is subsidised or paid in dollars anyway.
The honest verdict: chain success and token-holder return are not the same thing. Rail success can be genuine and accrue to nobody holding the token.
The Counter-Case: Fee Compression May Mean Nobody Captures Much
Everything above assumes value is captured somewhere. The stronger bear case is that it isn’t — or not much of it is.
- The margin pool is being destroyed at the base layer. Settlement costing a fraction of a cent, no gas token, and zero per-transaction fee means the base layer of agentic commerce generates almost no revenue by design. Everyone builds on top of a free substrate.
- Float income is rate-dependent and contested. Issuers cannot pay holders yield under the GENIUS Act, but that prohibition is politically fragile — it was one of the two disputes that killed the CLARITY Act in the Senate on 15 September 2026. Community banks lobbied hard against stablecoin yield for fear of deposit flight. If the yield ban is ever loosened, the float rent gets competed away.
- The volumes are still small relative to the story. Stablecoin settlement is enormous in aggregate — trillions monthly, driven overwhelmingly by exchange and treasury flow — but the genuinely agentic slice is a rounding error next to it. Fee-free rails and penny payments may produce a large utility and a small profit pool.
- Distribution eats the issuer’s margin, exactly as Circle’s $324.6m payment to Coinbase shows. Owning the mechanism without the customer means owning the cost, not the profit.
A low-margin, high-volume, utility-like agentic payments layer is a perfectly plausible outcome. In that world the winners are the hyperscalers and platforms that bundle payments into something else they already sell, and the losers are anyone who built a business model on charging for movement.
What to Watch
- Whether Stripe prices anything on Tempo. If the rail stays genuinely free while Stripe sells billing, treasury and routing alongside it, the “free rail, paid adjacency” model is confirmed. If per-transaction fees quietly appear, the thesis changes.
- The OpenRouter close. A completed acquisition puts Stripe directly between agents and their compute spend — the most valuable non-payment chokepoint in the stack.
- Circle’s distribution costs as a share of reserve income. If Coinbase’s cut keeps rising past the current ~51%, Circle’s listed exposure gets thinner even as USDC grows.
- The rate path. The FOMC’s projected 4.1% by end-2026 is the single biggest input into stablecoin issuer earnings. A faster cutting cycle compresses the float income that this whole analysis leans on.
- The OCC’s November 2026 rules and the January 2027 licensing cliff — and whether the stablecoin yield prohibition survives the next round of US legislation.
- Whether any chains capture token demand. Watch for a major agent rail attaching a fee, a burn, or a staking requirement that makes rail usage accrue to holders. So far, none of the leading candidates do.
The Bottom Line
On a rail that charges nothing, the answer to “who gets paid?” is whoever holds the money while it sits still. The transaction is not the product; the balance is. That makes the best-positioned candidate the issuer of the stablecoin the agents actually use — and, most sharply, Stripe itself, because it owns the rail, an issuer on it (Bridge’s USDB, under a conditional OCC trust charter), the checkout, and increasingly the routing layer between agents and compute. The mechanism is reserve and treasury income that scales with the size of the float — a rent that a fee-free rail increases rather than erodes, by making money cheap to move and easy to leave sitting.
The rest of the field is more qualified. Circle is the purest public expression of the mechanism but hands half the upside to its distribution partners. Visa and Mastercard keep the layer agents can’t settle — the disputed, higher-value, consumer-facing purchase — while being arithmetically excluded from micropayments. OpenAI, Anthropic and the hyperscalers own demand but not the rail, and are not directly investable. Infrastructure and picks-and-shovels get paid regardless of who wins, and mostly in private hands. The chains collect nothing unless a token is deliberately wired to the flow — and Tempo, notably, has no token at all.
And the counter-case is real: fee compression at the base layer, a rate cycle that erodes float income, distribution partners taking the issuer’s margin, and an agentic volume slice that is still small. It is entirely possible that agent-native payments become one of those enormous, essential, low-margin utilities that everybody uses and almost nobody profits from — with the money going to whoever bundles it into something else they already sell.
Which is the same lesson as the first article, arriving from the other direction: in the agentic economy, the rail is not the business. Owning the rail is.
This article draws on Stripe’s newsroom and the February 2026 tender-offer reporting, Tempo and Bridge public materials, Circle’s Q2 2026 results and its Coinbase revenue-share disclosures, Tether’s Q2 2026 attestation, Visa and Mastercard 2026 agentic-payments announcements, the Visa/Mastercard interchange settlement, Keyrock’s May 2026 agent-transaction analysis, AWS/Coinbase/Circle agent-payments product launches, Senate roll-call coverage of 15 September 2026, and market data as of 23 September 2026. Figures are attributed to their sources; nothing here is a recommendation. Not financial advice.