BlackRock’s digital assets research team published a report in October 2026 called The Machine-Native Economy, and while it reads like an investor thesis, it contains a practical warning for anyone shipping agents: the payment rails most of us build on were never designed for what agents do.
The core argument, stripped down
The report’s framing is that AI is machine-native intelligence and digital assets are machine-native money, and their shared logic is tokenization. LLMs break language into tokens a machine can process; blockchains break real-world assets into tokens a ledger can verify. Because both sides use the same tokenized structure, an agent can read a financial record as directly as it reads your prompt — no conversion layer, no human reconciling a spreadsheet.
That’s the thesis. Whether you buy the investment framing or not, the operational point stands: agents that act on your behalf eventually need to buy things, and the current rails fight them at every step.
Why credit cards can’t carry a sub-cent economy
Agent traffic is largely machine-to-machine: $0.001 per API call, compute billed by the second. Cards carry fixed fees that dwarf those amounts, and ACH settles in at least a day. Neither supports a system that transacts continuously at fractions of a cent.
The report points to a stack emerging on top of agent connectivity standards — Anthropic’s MCP (late 2024) and Google’s Agent2Agent protocol (2025). Coinbase’s x402 protocol builds on these, reviving the HTTP 402 “Payment Required” status code and pairing it with stablecoin wallets so a payment, on-chain verification, and service unlock complete in seconds without a human. Stripe, OpenAI, Google, and Visa have each shipped their own security and authorization protocols (MPP, ACP, AP2, TAP) to let agents use existing settlement systems safely.
The scale numbers from the report are worth knowing: stablecoin circulation passed $300 billion as of September 2026, and adjusted 2025 stablecoin volume hit $11.2 trillion — between Visa ($16.7T) and Mastercard ($10.6T), with an 80% CAGR since 2020 versus ACH’s 8.5%.
Compute is becoming a tradeable asset
The second half of the report is about inference economics. Citing McKinsey’s projections, it expects inference to overtake training as the largest draw on data-center power and compute by 2030 — and inference demand is fragmented and spiky in a way training never was. Markets are responding with GPU-backed financing, compute futures, basis markets, and tokenized compute contracts, so supply and demand can price and hedge capacity.
A signal the report highlights: Stripe’s August 2026 acquisition of OpenRouter, which routes across 80+ providers and 400+ models. When a payments company buys a model router, it’s betting that compute procurement, routing, and programmatic settlement converge. If your product already routes across providers for cost and latency, that decision is quietly becoming a payments decision too — something I touched on when looking at how routing changes your model tier economics.
If you haven’t read it, this connects directly to what one sales agent’s 600 hours a month says about where agents pay off: the agents with real ROI are the ones doing high-volume, small-unit work — exactly the transaction profile these new rails are built for.
What’s still unresolved
The report is candid that this is early. Compute commoditization lacks standards: chip generations differ in effective performance, regional energy prices vary, and nobody has settled how physically-settled compute contracts should deliver. Compliance is being handled with a two-layer design — AML, KYC, and “Know-Your-Agent” checks run off-chain, with verified results passed to smart contracts for on-chain execution. Whether regulators accept that split at scale is an open question.
The practical takeaway: if your roadmap includes autonomous purchasing — agents buying API access, data, or compute on their own — start watching the agentic payment protocols now, and design your pricing to be machine-readable. You don’t need to adopt stablecoins today, but unit prices in the sub-cent range are coming, and card-based billing won’t survive them.