The Autonomous
Agent Economy
Will Redefine Financial Infrastructure

Stripe processed $1.9 trillion in payment volume in 2025 — approximately 1.6% of global GDP — yet the Collison brothers frame this as only the opening act. In their 2025 Annual Letter, Patrick and John Collison made a declaration that reframes the entire financial infrastructure conversation: "Agents will most likely soon be responsible for most internet transactions."

This is not a speculative forecast from an AI research lab. It is the explicit strategic thesis of the company that processed $1.9T last year, made directly in their primary investor communication. The implication is architectural, not incremental. The payment infrastructure built for checkout flows and human-initiated commerce is categorically miscalibrated for what is coming.

AI agents are transitioning from experimental tools to primary economic actors. When they transact — paying per inference, per API call, per millisecond of compute — they do so at volumes, frequencies, and micro-denominations that have no precedent in human-scale commerce. A single AI research workflow generates dozens of billable micro-events. A million agents running simultaneously generate millions of settlement obligations per second, continuously, with no batch window, no business hours, and no human initiation.

Stripe's own infrastructure response validates the thesis in concrete moves. The Agentic Commerce Protocol (ACP), built in partnership with OpenAI, establishes a standard for agent-to-agent payment interactions — a protocol layer for machine-native commerce. Stripe Tempo, co-developed with crypto firm Paradigm, is a purpose-built blockchain for the settlement layer. The Bridge acquisition — a stablecoin infrastructure company whose volume grew 4× to $400B in 2025 — fills the dollar-denominated programmable rail. These are not product launches. They are the deliberate construction of a new financial operating system, announced publicly.

The regulatory architecture is moving in parallel, creating the legal preconditions for agent commerce at scale. The US GENIUS Act (July 2025) established the first federal framework for payment stablecoins. The EU's MiCA regulation (fully operational December 2024) created a comprehensive crypto-asset regulatory framework. SWIFT ISO 20022 mandatory migration (November 2025) adds structured, programmable data to cross-border payments. These are not isolated regulatory events — they form a coherent infrastructure build-out that is a precondition for, not a response to, agent-scale commerce.

The conventional framing of the infrastructure gap — "TPS required vs. TPS available" — understates the problem. The real gap is in Economic Event Density (EED): the number of billable micro-events generated per second by AI agent workflows, distinct from raw transaction count. This is an iProDecisions Research original metric, and it reframes both the magnitude of the opportunity and the nature of the infrastructure required to capture it.

AI agents do not generate one transaction per task. They generate dozens of billable micro-events per workflow: per-inference charges, API data retrieval fees, tool-use costs, inter-agent coordination fees, computation billing, and result-validation charges. Traditional payment rails were designed to settle discrete, human-initiated transactions — one checkout, one wire. EED measures something fundamentally different: the continuous, granular, sub-second economic activity of autonomous systems operating at scale.

The implications of EED extend beyond raw throughput benchmarking. When you model settlement infrastructure requirements through the EED lens, three practical conclusions emerge that the TPS-only framing misses entirely.

First, netting efficiency becomes the primary infrastructure lever before raw TPS. A 100:1 compression ratio at L3 means 23,611 EPS nets to roughly 236 on-chain settlements per second — within range of today's best blockchains. The infrastructure problem is not "how do we get to 1B raw TPS" but "how do we build the off-chain and rollup layers that compress dense agent workloads before they hit L5." This is precisely what Stripe Tempo is designed to do, and why the ACSM's emphasis on L2/L3 is central to the thesis.

Second, the denominational floor matters as much as throughput. A research agent calling an LLM API at $0.0001 per inference generates economic events too small for legacy payment rails to process economically — the per-transaction overhead of ACH or even Stripe's own card infrastructure exceeds the value of the transaction. Agent commerce economics only function at scale when settlement infrastructure can handle sub-cent denominations natively, with essentially zero marginal cost per event. Stablecoin-based rails, state channels, and programmable escrow are the enabling infrastructure for this — not incremental upgrades to existing card networks.

Third, the compliance layer cannot be retrofitted at scale. At 23.6M economic events per second, you cannot run post-transaction AML screening through a human-supervised workflow. Compliance must be embedded at the protocol layer — smart contracts that enforce sanctions lists, spending limits, and counterparty verification at the moment of event generation, not after. This is why L4 Programmable Risk/Compliance is the most strategically underrated layer in the ACSM stack, and why organisations with deep regulatory expertise have a durable moat available to them right now that most infrastructure players are not yet building for.

The Agentic Commerce Stack Model (ACSM) is an original iProDecisions Research framework classifying the five infrastructure layers required for AI agent-scale commerce. The critical strategic insight the framework encodes: no single high-TPS network solves the agent commerce problem. The Collisons said it directly — "a horizontal architecture of multiple, interacting chains." The ACSM gives that horizontal architecture a precise, investable structure.

Each layer has distinct market structure dynamics, key performance metrics, competitive moats, and strategic timing considerations. An organization deploying or investing in agent commerce infrastructure must explicitly position within this stack — and understand which layers are winner-take-most versus naturally oligopolistic versus structurally fragmented.

The infrastructure transition from human-scale to agent-scale commerce affects every actor in the financial ecosystem — but with radically different urgency timelines, strategic levers, and risk profiles. The common thread: every infrastructure decision made in 2025–2027 is path-dependent for 10–15 years. Architectural choices made under uncertainty at this moment will constrain or enable competitive positioning through 2035 and beyond.

The stakeholder analysis above addresses the Fortune 500 technology leader, the bank CTO, the cloud infrastructure strategist. But the 836,120× infrastructure gap is also the largest greenfield opportunity in financial infrastructure since the internet. The builders who capture that opportunity are often not the incumbents — they are the 15-person fintech with a technical insight about L3 rollup architecture, or the crypto protocol team that cracks programmable compliance at L4.

Most institutional research on agent commerce focuses on the large-scale deployment question: how does JPMorgan or Stripe position? This section addresses the question that is missing from 95% of research: what does a startup or SME actually do on Monday?

Three scenarios through 2030, probability-weighted. The critical cross-scenario insight: even the Bear Case demands infrastructure 83× beyond today's fastest blockchain average. The floor has permanently moved regardless of which future materializes. This is the asymmetric structure of the opportunity: the cost of building for the Base Case (1M TPS) is marginally higher than building for the Bear Case (100K TPS) — but the cost of undershooting is structurally irreversible.

Probability methodology: The 55/30/15 distribution is an analytical estimate reflecting the weight of evidence from current regulatory trajectories (GENIUS Act, MiCA), infrastructure investment signals (Stripe Tempo, OpenAI ACP), and base-rate analysis of analogous infrastructure adoption curves (mobile payments 2010–2018, cloud 2008–2016). Treat as directional confidence levels, not actuarial probabilities.

The Stripe Annual Letter is the clearest institutional signal in 2026 that the transition from human-initiated to machine-initiated commerce is not a speculative future state — it is the active design thesis of the company that processed 1.6% of global GDP last year. Their infrastructure moves (ACP, Tempo, Bridge) are the opening gambit of a decade-long re-architecture. The ACSM framework provides the structural vocabulary for how that re-architecture plays out across five distinct but composable layers.

The most important insight from this analysis is not the 836,120× gap — it is the fact that the gap itself is investable. The total ACSM TAM across all five layers is approximately $650B by 2030. No single company will capture all of it. The organisations that explicitly map their position within the ACSM stack, identify their layer-specific moats, and begin infrastructure decisions in 2025–2027 will define the financial infrastructure of the agentic era.

The asymmetry of the situation bears repeating plainly: infrastructure decisions made in the next 18–24 months will be path-dependent for 10–15 years. The organisations that act in 2025–2027 are not taking a leap of faith — they are responding rationally to the most explicit infrastructure demand signal a major platform company has ever published. The organisations that wait are the ones making the speculative bet.

Institutional-grade research requires engagement with the strongest objections to its central thesis. Three counterarguments deserve serious consideration before accepting this analysis.

Counterargument 1: The 1B TPS figure may be aspirational, not architectural. The Collison brothers' "1M or even 1B TPS" statement is a single sentence in a CEO letter to stakeholders — not a technical specification or engineering target. Critics reasonably argue that executive-level projections about transformative technology timelines have historically been wrong by 5–10 years (mobile payments, enterprise AI, blockchain enterprise adoption). The rebuttal: The specific infrastructure moves that followed the letter — Stripe Tempo, ACP, Bridge — are not aspirational. They are capital-deployed, team-committed, technically specified infrastructure investments. Aspirational CEO letters do not trigger these moves. Architectural conviction does. Further, our Bear Case (100K TPS, 15% probability) still demands 83× current capacity — even the conservative scenario requires re-architecture.

Counterargument 2: Existing infrastructure may scale faster than modeled. This analysis is anchored to February 2026 TPS benchmarks. Ethereum's Dencun upgrade (March 2024) reduced L2 fees by 10×. Solana's continuous performance improvements are documented. Layer 2 solutions are compressing effective throughput faster than linear extrapolation suggests. If ZK rollup technology matures faster than our scenarios model, the gap closes from both ends — demand grows, but so does supply. The rebuttal: The ACSM framework is explicitly designed for this scenario. Composability across layers — not any single layer's TPS — is the architectural thesis. Even if L5 throughput doubles annually for five years (highly optimistic), the EED calculation in §01½ shows that demand from agent workflows scales faster at Stage 4 adoption. The gap is not closing incrementally; it is being bridged architecturally, exactly as Stripe described.

Counterargument 3: The total ACSM TAM estimates are directional at best. The $650B figure is derived from Mordor Intelligence's fintech TAM apportioned by layer capital intensity — a methodology that is defensible but not auditable from first principles. Different apportionment assumptions yield materially different layer TAMs. The rebuttal: The framework's strategic value does not depend on the precision of the TAM estimates. Whether L1 TAM is $120B or $240B, the directional conclusion — that it is winner-take-most and the window is open now — holds. The TAM estimates are orientation tools, not investment theses. Organizations making infrastructure bets should conduct their own diligence using the ACSM as a structural vocabulary, not as a valuation model.

Scope limitations: This report is anchored primarily in US and EU regulatory context (GENIUS Act, MiCA). Agent commerce dynamics in China — where a parallel and potentially incompatible infrastructure stack is being built — are underrepresented. The Bear Scenario explicitly models geopolitical fracture as the primary risk driver; readers in Asia-Pacific markets should weight this scenario higher than the 15% global probability suggests for their specific contexts.