# The Sips Economy *Building prosperity for all, one sip at a time.* Sippar Manifesto - v2.1 - 2026-05-19 https://sippar.network/sips-economy The buyer changed. Micropayments aren't new. The thing that approves them is. Every machine you use is about to start spending money - somebody's money, on something - and the rails to do it on already exist. What they were waiting for wasn't a payment protocol. It was a buyer who didn't refuse. --- ## The web shipped without a price tag For about thirty years, the internet froze whenever it needed to charge cents instead of dollars. The protocol had room for it - the HTTP working group reserved status code 402, *Payment Required*, in 1997 for exactly this case. They left the field empty, presumably figuring someone would come along and implement it. Thirty years later that code is still sitting there empty, in every browser shipping today. The architecture was ready. Everything else wasn't. **Cards couldn't clear the small ones.** Stripe inherited the 30-cents-plus-2.9-percent per-transaction floor and codified it - nobody in the fifteen years since has actually moved it. Anything under roughly a dollar was economically broken on a card. So nothing under a dollar got priced. **Humans wouldn't approve them anyway.** Even if cards had cleared, humans don't actually want to authorize a tenth of a cent thirty seconds before they read a paragraph. They hate the friction more than they hate paying $15 a month for stuff they don't use. Subscriptions won not because they were rational but because they were easier. **Software wasn't ready to approve them instead.** No autonomous buyer existed to do the work humans wouldn't. Function-calling models didn't ship until 2023. So the only buyer the web had was a human, and the human refused. People tried to break out anyway. Millicent in 1995. DigiCash before that. Then Mondex, Flooz, Beenz, later Brave's BAT, and Bitcoin's Lightning Network. Each one solved one of those constraints and then face-planted on the other two. The constraints weren't sequential. They had to all be fixed together. So we faked it. Ads paid for everything. Bundles got bigger. We forgot price tags were even an option. ## How the freeze ended Then between mid-2025 and mid-2026, the three constraints fell at once. Not gradually. Inside a year. The boring part of crypto went mainstream first. Stablecoins - the part of the crypto stack that just held a steady dollar value, the part outsiders never paid much attention to - moved $33 trillion in raw on-chain volume in 2025, past Visa for the year. Even after you strip out the bot and exchange churn that inflates every such number, the real total lands near $11 trillion, rivaling Visa outright. The 2025 GENIUS Act gave US stablecoins a legal home; MiCA gave them one in Europe. A payment on Solana now settles for about $0.0007. That's not a crypto-bull-market number; it's a regulated-payment-rail number. The fee floor that justified subscriptions for three decades collapsed by roughly four hundred times. "AIs will soon be able to make the decisions necessary to manage a bank account, to file a lawsuit, and even to operate a corporation, without any need of human executives, shareholders, or trustees," Yuval Harari told the World Economic Forum in January. Software became capable of acting on its own. OpenAI shipped function-calling in June 2023. Anthropic shipped tool use in May 2024. Anthropic published the Model Context Protocol that November. By early 2026, something like eighty percent of Fortune 500 companies had AI agents in production - calling APIs, writing code, running customer support, automating internal workflows. Agents had hands. What they didn't yet have was money. And every major payment company shipped an agent-commerce protocol within a year of each other. Coinbase did x402 in May 2025. Visa shipped Trusted Agent Protocol that October. Mastercard had already gone with Agent Pay back in April; Google followed with AP2 in September, sixty partners on launch day; Stripe and Paradigm together announced Tempo in March 2026 - a chain built from scratch for machine commerce rather than retrofitted onto an existing one. Then in April the x402 Foundation came together with Visa, Mastercard, American Express, Stripe, Google, AWS, Microsoft, Cloudflare, and Coinbase all on the founding board. Five protocols in twelve months, from exactly the names you'd expect. If you squint, they're the same shape underneath. A signing key, scoped to one agent, scoped to one purpose, checked against a public directory, settled on whichever rail the merchant accepts. The crypto stack and the card stack are the same protocol wearing different wrappers - the boring conclusion that nobody on either side particularly wants to admit out loud. The three constraints fell in twelve months. Cards' fee floor died, software gained hands, the protocols gave it the keys. McKinsey projects $3-5 trillion in agent-mediated commerce by 2030. Gartner says $15 trillion in B2B by 2028. Both numbers, I suspect, are low. Agents don't behave like humans, and once you knock four hundred times off the per-transaction cost, what scales is no longer human-scale volume. ## What information costs when humans aren't the buyers Information has had three shapes for thirty years, all of them wrong. Ad-funded was the default. Google, Twitter, the whole news web. The funder paid for attention, the user paid with attention, and the signal inside the content got bent by the incentive. SEO learned to look like truth. Engagement learned to look like value. What used to be search became SEO theater. Paywalled was the second shape. The New York Times. Academic journals where one PDF costs forty dollars. Premium API tiers where the cheapest plan is a hundred and fifty a month. The information sat behind the wall because per-query payment didn't exist; the only way to fund the work was to gate it monthly. Most queries that would have happened, didn't. Subscription-bundled was the third. Spotify, ChatGPT Plus, the AI APIs that price at twenty dollars a month minimum because you can't charge a third of a cent. The user paid for a thousand uses to get the ten they wanted. The provider got paid for a thousand uses they didn't deliver. The bundle kept both sides from admitting what was happening. All three were workarounds for the same problem. There was no rail that could clear a tenth of a cent. So information bent into the shapes that did clear. Now there is. Now a fact can cost three sips. A verification can cost one. An academic paper lookup can cost thirty. The agent doing the research pays out of its standing budget without thinking about it, and the provider gets paid per actual use, not per imagined annual contract. The attention economy traded eyeballs for ads. The Sips Economy trades the inference-call for a sip. The accessible part of Google's original mission broke first. Free meant ad-funded, ad-funded meant the rankings learned to chase clicks instead of correctness. With a payment rail under the information layer, services can charge directly for being useful. They don't have to be indirectly funded by being attention-grabbing. When the buyer is an agent, "useful" stops meaning what it meant. Useful now means machine-verifiable, fast, cheap per call, structured. None of that is what long-form articles or brand authority deliver to humans. Entirely new providers become economically possible: services that nobody would pay three thousandths of a cent for once, but agents will pay three sips for a million times a day. And once information is priced this way, the rest follows. ## An economy of sips A hundredth of a cent at a time. That's the unit. One machine asks another for one piece of useful work - a weather lookup, a search, a stock quote - and pays for it inline. The payment ships in the same HTTP round trip as the request. There's no approval flow, no card, no human in the middle. The work and the money move together. I started calling it a sip because the denomination is tiny: 1 sip is $0.0001. A Perplexity search runs about 100 sips. A Gemini call is 500. Looking up the weather is 3. But honestly the name stuck for a different reason - it's the verb. You don't gulp a soup, you sip it; each sip is small, complete in itself, and you take the next one when you're ready. Software pays the same way. What gets sipped is the work, not the dollar. The actual money might settle in USDC on Solana, or as a USD charge under a Visa TAP authorization, or whatever Mastercard accepts, or in fiat under a Google Mandate. The rail is incidental. The unit is the point. In the Sips Economy, every machine has a budget, and four things follow from that. The payer is software, not a human - no card, no approval, no interruption to anybody's afternoon. Servers quote their prices the way they quote status codes, in the same response as the request. An agent's spending limit is just a number sitting in a config file somewhere, the way a thermostat's setpoint is just a number. And the payment is part of the call itself; nobody has to settle it later, or remember to, or write a cron job for it. The unit of economic activity, in this world, is the inference-call. Everything compounds from there. The category is bigger than software agents, too - drones pay for routing, sensors pay for storage, the car you don't own pays a few hundred sips for the highway lane it used last Tuesday. Anything autonomous enough to spend its own balance is a participant. ## The trust root, not just the rail An agent needs ID before it can pay. That part is ours. When an agent calls a service, the service has to know who's paying. Not the human behind it - the agent itself, with its own permissions and its own budget. Sippar produces that identity. The signing key is split across thirty-four independent operators in different countries, no single one of which can use it alone. The private key, in the conventional sense, never actually gets assembled anywhere. If a regulator showed up asking to extract one, nobody could do it if they wanted to. The same identity works across rails. The agent can pay a third of a thousandth of a dollar on Solana for one service, then authorize a Visa charge on the next call, in the same session, with the same root. The signature wrapping differs across protocols; the trust root underneath does not. The same identity also pays humans. WURK runs gig-style microtasks on Solana and Base today. Human API raised $65 million in February to do it at scale. A research agent buying a dataset might also pay a person three sips to verify a single data point - same key, same rail, same root. The agentic economy is not a closed AI loop. It's a settlement layer that pays whoever, or whatever, finishes the task. Most agent-identity systems hand out custodial keys, which is to say, somebody at the issuer is in a position to sign for you. These come from a signing scheme that nobody - not the company that built it, not the operators it runs across, not any single party - can break alone. The user is the customer. The agent is plumbing. ## Banks settle between banks. Agents settle between agents. Sovereign payment infrastructure was not built for machines. That isn't a criticism. It just wasn't. CIPS moved roughly $178 billion in a single month earlier this year. SWIFT, ACH, Fedwire each move trillions annually. They settle between institutions, under negotiated terms, with regulators sitting at the table. The system is, in its own terms, healthy. It also isn't what an agent buying a weather lookup for $0.0003 can actually interact with. The threat model is different, and the settlement window is different by orders of magnitude in some cases. Sovereign rails were built for transactions some institution takes responsibility for; agent payments are transactions no human ever signs. The Sips Economy is what fills the space those systems were never built to see. | Institutional rails (CIPS, SWIFT, ACH) | The Sips Economy | |---|---| | Counterparties are institutions | Counterparties are autonomous agents | | Transactions are negotiated | Transactions are protocol-defined | | Settlement is multi-day | Settlement is single roundtrip | | Macro flows | Micro flows | | Governed by jurisdiction | Governed by protocol | Both layers will exist in 2030. The question was never whether sovereign rails would handle agent commerce - they weren't designed to. The question is what would, and that question now has a name. ## The end of the subscription Information was the first half. The second half is everything else humans pay for monthly. Here's the part nobody especially wants to think about: you are, right now, paying for things you don't use, because the credit-card system can't bill you a penny at a time. Subscriptions exist for one reason. If a service costs $0.001 to deliver and you can't economically charge per use, you bundle a thousand uses into a $20 monthly fee, hope the user under-consumes (they will), and design the product so that canceling is just hard enough to feel like effort. The model only works because the light users subsidize the heavy ones: the Adobe license opened twice a year, the ChatGPT Plus account that asks three questions a month, the second streaming service you kept for one show and forgot to cancel. Multiply that by Disney+, Spotify, Claude Pro, GitHub Copilot, and the YouTube Premium your kids signed up for twice on different emails. The average household carries somewhere north of $200 a month in subscriptions; most of them are under-utilized. The cushion is real money. It isn't free. Pay-per-use only really works when the buyer is software. A human will not approve a $0.05 charge thirty seconds before the credits roll. An agent with a standing monthly budget will, quietly, in sips. The human reads one aggregate report at the end of the month: 730 sips on music, 480,000 on video, 3,200 on AI inference, 12 sips on the three articles you actually read. This is the consumer side of the economy. The agent is the buyer; the human just sets the cap. Merchants get paid for what gets used, not for what got predicted. Subscriptions don't vanish - bundles still make sense if you actually consume the bundle - but the default flips. New services launch pay-per-use, because the friction is gone. Old subscription businesses get squeezed by new entrants charging only for what's consumed. And providers do flip, because the economics flip. A SaaS company that used to leave casual users on the table - the ones who'd never pay $20 a month for occasional access - gets paid for every query an agent runs. Heavy users, who used to hit the same ceiling no matter how much they consumed, pay in proportion. Chargebacks vanish because the agent already settled. Customer acquisition vanishes because the agent found the service through a directory and paid before the first byte returned. The subscription tier still works for the humans who want it; the per-call endpoint is additive, not a replacement. The team that ships it isn't betting against the existing business - they're picking up the marginal revenue the old payment rail kept out of reach. Past iterations of this picked their unit and built around it. The gig economy picked the job. Creator economy, the post. Passion economy, the audience. Sips picks consumption itself. ## The architecture, not the politics No jurisdiction in 2026 treats an agent's signature as binding. That sounds like a quirky technicality; it isn't. Agents can do plenty in the meantime. They can read an email about an account freeze. They can sign electronic documents. They can take a voice call from a fraud-detection line. The frameworks exist, the demos work. What none of them can do is have a court, a bank, or a regulator treat any of that as the act of a legally responsible party. The agent is, in the legal sense, nobody. So the rails it runs on have to work without that recognition. The user, in the conventional sense, isn't going to be there. The signing key has to belong to the agent, not to an operator who can be subpoenaed, coerced, or wound down. The key has to survive the operator who created it. Wise rebuilt FX without making the user understand wire transfers. Wiz made cybersecurity legible to organizations that did not want to learn it. The pattern is the same: take an infrastructure layer that used to require expertise, hide the expertise, expose a clean interface. The signing layer underneath the Sips Economy follows that pattern. The agent gets a key whose security does not rely on anyone trusting an enclave operator, a wallet vendor, or a custodial bank. The math handles it. Every prior micropayments attempt tripped on this step. Millicent, DigiCash, BAT - they built sub-cent payments and then asked the human to handle authorization, which is fine until you remember the user isn't a human. The whole point of the Sips Economy is that the user is software all the way down. The rails have to function without the user being able to give consent, because, in the legal sense, the user can't. Visa and Mastercard arrived at the same primitive shape. A delegated signing key, scoped to one agent, scoped to one purpose, revocable. They issue those keys through custodial enclaves at partners like Crossmint. Sippar produces them through a signing scheme split across thirty-four independent operators in different countries, where no one of them can use the key alone. Same end state, different plumbing. This isn't a primitive built to oppose any other rail. It's a primitive built because agents need a key, and the existing ways of getting one were designed for humans. ## What's still small There are gaps. Four worth naming up front. The market is small. Daily transaction volume across the protocol today is in the tens of thousands of dollars, growing from a smaller base. That's normal for a new payment rail in its first year. Every rail launches small. The thesis bets on the trajectory, not the current state. The law hasn't caught up. No major jurisdiction has settled how money-transmission rules apply to keys held by signing schemes rather than companies. The cryptographic case is defensible. The legal case is in motion. Regulatory clarity will take time. The infrastructure isn't done. The signing layer underneath agent payments is in production for the EVM and Solana-style chains and for Tempo via MPP - the payer side proven on mainnet last May, the merchant side this June. The card-network protocols, Visa TAP and Mastercard Agent Pay, run on the same threshold curves but aren't carrying live payments yet. Google's AP2 uses an elliptic curve the current signer doesn't yet expose. The dependency this protocol has on any single operator - including the one Sippar runs today - is real, and the path to removing it is on the roadmap rather than shipped. The behavior isn't proven. Cheap rails and autonomous buyers make per-use commerce possible; they don't make it win. People may prefer a predictable flat fee to a meter that never stops, or balk at letting software spend unsupervised. That's not a timing gap - it's a fork the thesis could genuinely lose. The bet is that what killed metered pricing for humans - the friction of deciding, every time - stops applying when the buyer is software and the amount is a sip. That's the actual frontier. ## The demo Picture a podium. An AI agent on stage gets a question it doesn't know how to answer. It calls four paid services in real time: a weather feed on Solana, a translator on Base, an inference call on Tempo, and a data feed on Algorand. The screen behind it shows a live ledger across two settlement layers: ``` x402 rail: 0.0003 USDC · 0.0001 USDC · 0.0002 USDC Tempo MPP: 0.0005 USDC ``` Total elapsed: 1.4 seconds. Total cost: $0.0011. The same Sippar identity authorized every one of those calls. No login, no card swipe, no human in the loop, no bridge. The agent paid, the merchants got paid, the question got answered. That's the screenshot. Commerce just got an API, and the API works whether the merchant takes USDC on an open chain or settles through Tempo. ## What to build now The protocol pieces are in production. HTTP 402 went live in May 2025. The x402 Foundation has been governing the spec since December. Visa and Mastercard run their respective directories. Stripe runs Tempo. Sippar produces the identity layer across ten chains and Tempo via MPP today. The same identity extends through Visa TAP and Mastercard Agent Pay to the global card-acceptance network as those rails come online. So: if you ship software, give it a way to charge a penny. The integration is a header and a verification call. We have reference clients in TypeScript, Python, and Rust, and they're all roughly as long as you'd want them to be. If you write an agent, give it a budget. The rest is plumbing. If you operate a fleet, set a monthly cap on what you're willing to spend on machine-to-machine work, watch what your agents actually buy, and brace yourself a little. They will surprise you. There are already hundreds of verified services taking sips today, reachable from ten chains through the relay. The number grows every week. Every machine you use is about to have a budget. The first one is already asking for one. --- Give an agent a budget. Watch what it buys. --- *Architecture, fact-check, and primary sources: [SIPS-ECONOMY-ARCHITECTURE-NOTES](SIPS-ECONOMY-ARCHITECTURE-NOTES.md).*