AI Agents Don’t Run on Hype. They Run on Power, Wallets, and Settlement Rails

Everyone is talking about AI agents. Investors, analysts, enterprise software executives. Everyone is focused on what agents do . What workflows they replace. What software categories they disrupt. Which SaaS stocks they pressure. Almost nobody is asking what agents run on . I lead EMCD, one of the world’s top ten Bitcoin mining pools and a crypto-fintech platform active across 120+ countries. That means I look at this question from a different angle: not as a software demo, but as an infrastructure problem involving uptime, wallets, payouts, settlement, risk checks, and users moving value across borders at any time of day. When I look at what the agentic economy actually needs to function, not at the demo layer, but at the infrastructure layer, I do not see a neat Silicon Valley software story. I see a story about the kind of plumbing crypto has been building for more than a decade. Not hype. Not dashboards. Infrastructure. The number nobody is talking about There is a recent report that deserves more attention than it has received. Keyrock , in collaboration with Coinbase, Tempo, and Virtuals Protocol, tracked AI agent payment activity from May 2025 through April 2026. The result: agents settled more than $73 million across 176 million on-chain transactions. That number is interesting. But the payment size is the real story. The report found that 76% of agent payments were below Visa’s $0.30 minimum fixed fee threshold. Many payments were not just small. They were too small for traditional card rails to process economically. Read that again. Three quarters of agent payments were below the point where card economics start to make sense. Some of this activity is still experimental. Some of it will disappear. But the payment pattern is already visible. Agents do not behave like human buyers. They do not sit down, compare two SaaS plans, click ‘subscribe’, and expense the invoice at the end of the month. They may pay for a data request. A model response. A compute task. An API call. A verification check. Another agent’s output. They may do this thousands of times a day. The machine economy is not just about automation. It is about transaction granularity. And legacy payment rails were not designed for that. Agents need more than cloud credits SaaS was built around human users. A human logs in. A human has a seat. A human has a company email, a corporate card, a manager, and an expense policy. A human can approve a payment, explain a decision, and be held accountable when something goes wrong. Agents are different. They are software actors that can call APIs, trigger workflows, move value, create obligations, and coordinate with other systems without waiting for a person to click through every step. That changes the infrastructure requirement. Here is the question every enterprise tech article should be asking. When an AI agent executes a task at 2 a.m. on a Sunday, uses a third-party data source, pays for the request, verifies the output, logs the transaction, and triggers the next step, what exactly does it need? Not a Salesforce seat. Not a corporate card in a human employee’s name. Not a monthly subscription tied to a user who may never touch the workflow. It needs identity. It needs authorization. It needs a wallet or payment method with spending rules. It needs settlement. It needs audit trails. It needs compute. It needs limits. It needs someone to be accountable when something goes wrong. That sounds less like SaaS and more like infrastructure. The emerging agent payment stack is trying to solve this. Coinbase describes x402 as an open payment protocol that enables automatic stablecoin payments directly over HTTP, allowing human and machine clients to pay for access without accounts, sessions, or complex authentication. Google launched the Agent Payments Protocol, or AP2, as an open protocol for agent-led payments across platforms, built with more than 60 payments and technology partners. The point is not that one protocol will win. The point is that serious players have already understood the same thing: agents need payment rails, identity rails, and authorization rails that were not built into traditional software. The $0.30 problem The $0.30 card-fee threshold is a small number with a large implication. If agents transact in payments of a few cents, or even fractions of a dollar, most traditional rails become economically awkward. A human can pay $20 once a month. An agent may need to pay $0.02 a thousand times. That difference breaks the model. Blockchains and stablecoins are not perfect. They have risks, congestion, UX problems, regulatory questions, and fragmentation. But they have one thing agents clearly need: programmable, always-on settlement. In the same Keyrock dataset, USDC dominated agent payment settlement. That is not surprising. Agents executing large numbers of small transactions need a settlement asset that is relatively stable, liquid, and widely supported across crypto infrastructure. For a human user, stablecoin settlement may sound like a feature. For an agent, it may be oxygen. Agents need passports, not passwords The identity problem is just as important as the payment problem. A human employee has a login. A company email. A role. A manager. A payroll record. A compliance profile. An AI agent has none of that by default. If an agent acts, who owns it? Who gave it permission? What can it spend? What systems can it touch? Who can revoke its authority? What happens if it behaves incorrectly? A password is not enough. Agents need verifiable identity and permissioning. They need spending rules. They need transaction caps. They need allowlists. They need logs. They need auditability. They need to be stopped if their behavior becomes risky. The ERC-8004 proposal is one early attempt to make agents discoverable and verifiable across organizational boundaries. That matters because agent identity cannot rely on a single company directory forever. If agents are going to coordinate across APIs, wallets, protocols, and enterprises, they need a way to prove who they are, who authorized them, and what they are allowed to do. Whether ERC-8004 becomes the dominant standard is not the point. The point is that the market is now treating agent identity as an infrastructure problem. This is where crypto-native architecture becomes relevant, not because every agent will trade tokens, but because wallets, programmable permissions, signed messages, and on-chain activity give us building blocks for machine-to-machine trust. An uncontrolled agent with access to money is not innovation. It is an incident report waiting to happen. Why miners are closer to agents than SaaS vendors think There is another infrastructure story unfolding in parallel, and most enterprise tech coverage still treats it as a side plot. Bitcoin miners are pivoting into AI and high-performance computing. CoinShares described more than $70 billion in cumulative AI and HPC contracts announced across the public mining sector in its 2026 Q1 mining report. Core Scientific, TeraWulf, Cipher, Hut 8, and others are increasingly being valued not only as Bitcoin miners, but as power-backed infrastructure operators. That shift is not random. AI needs power. Cooling. Land. Grid access. Uptime. Operational discipline. Facilities that can run machines continuously. Teams that understand what happens when downtime is not a small inconvenience, but the whole economic model. Bitcoin miners have been dealing with those realities for years. No, mining pools do not automatically become AI-agent infrastructure companies overnight. That would be too neat. Mining hardware is not GPU infrastructure. Bitcoin mining facilities and AI data centers are not the same thing. Converting one into the other is expensive and technically demanding. But the operational mindset maps more closely than many people think. Bitcoin mining is a business of energy, uptime, machines, margins, and global coordination. The agentic economy will also be a business of energy, uptime, machines, margins, and global coordination. Different machines. Similar pressure points. The infrastructure gap is physical, not theoretical The software world likes to talk as if scaling AI is mostly a model problem. It is not. It is a power problem. A data center problem. A cooling problem. A grid problem. A settlement problem. A compliance problem. A reliability problem. Goldman Sachs Research forecast in 2025 that global power demand from data centers could rise 50% by 2027 and as much as 165% by the end of the decade compared with 2023. This matters because the agentic economy multiplies demand from both sides. On one side, agents need compute to reason, plan, and execute. On the other side, agents may create huge volumes of small transactions that need settlement, verification, logging, and risk controls. That is not just ‘more software’. That is more infrastructure at every layer. What crypto gets right, and what it still has to prove Let’s be honest. Crypto infrastructure is not ready to declare victory. Not every agent wallet is meaningful. Not every on-chain payment is useful. Some activity is experimentation. Some is noise. Some protocols will disappear. Some narratives will age badly, as crypto narratives often do. We have all seen that movie. It usually has too many Telegram channels and not enough product-market fit. The infrastructure is also fragmented. x402, AP2, agent wallets, identity standards, stablecoin settlement, smart accounts, and chain-specific tooling are moving quickly, but they are not yet a single mature stack. Enterprise adoption will not happen just because something is technically possible. It will require trust, auditability, support, compliance, risk management, and plain old reliability. Crypto has strong primitives for machine payments. But primitives are not products. Products need uptime, support, documentation, monitoring, customer protection, and a reason for normal businesses to care. The agentic economy will not be built by vibes. It will be built by the teams that can turn promising primitives into dependable infrastructure. Why EMCD sits closer to this than it looks EMCD started with mining. That matters. Mining teaches a company to think in operational realities, not slideware. Hashrate. Uptime. Payouts. Energy conditions. Cross-border users. Wallet flows. Risk checks. Support. Systems that need to work when no human is sitting there gently refreshing a dashboard. Today, EMCD is more than a mining pool. We operate across mining, wallet, Coinhold, P2P, swaps, and infrastructure tools. Our users do not operate on one country’s banking hours. They do not all use the same currency. They do not all have the same local payment environment. They need access to crypto infrastructure that works across time zones, markets, and user behaviors. That experience does not mean EMCD is suddenly an AI company. It means we understand the infrastructure requirements that the agentic economy is beginning to expose: Settlement rails that operate when no human is watching Wallet and payment logic that can support high-frequency activity Multi-market operations where local access and compliance matter Systems designed around uptime, not office hours Support processes for when money, identity, and automation collide Infrastructure that treats small payments as first-class events, not rounding errors This is why the agentic economy looks familiar from where I sit. Not because it is the same as mining. Because it has the same intolerance for weak infrastructure. The part investors should stop ignoring AI agents are usually discussed as software. But once agents can pay, settle, verify, and coordinate work, they become economic actors. Small ones at first. Limited ones. Often experimental ones. But economic actors nonetheless. Economic actors need infrastructure. They need power. They need compute. They need identity. They need payment rails. They need wallets. They need auditability. They need governance. They need systems that work before, during, and after the transaction. That is why the mining-to-AI pivot matters. The $70 billion shift inside public mining companies is not just a mining story. It is a signal that capital is starting to understand where the bottlenecks are. Compute does not float in the cloud like magic. It sits in facilities, plugged into grids, cooled by real systems, operated by teams who know what downtime costs. The future of AI agents will not be decided only by who has the best model. It will also be decided by who controls the infrastructure those agents run on. What comes next The first wave of the AI-agent conversation was about capability. Can agents book meetings? Write code? Handle support? Update CRM systems? Trade tokens? Order services? Chain tasks together? The next wave is about infrastructure. Can agents identify themselves? Can they pay economically? Can they settle reliably? Can they be audited? Can they operate across jurisdictions? Can the power grid support the compute demand? Can payment rails support transactions smaller than a card network wants to process? That is where the conversation gets serious. For developers, the signal to watch is whether agent tooling moves from demos to reliable identity, wallet, and payment standards. For investors, the signal is whether capital keeps flowing into power-backed infrastructure, not only model-layer startups. For enterprise architects, the signal is whether agents can be governed, audited, rate-limited, and shut down without breaking the workflow they were built to automate. The agentic economy does not run on hype. It runs on power, uptime, identity, and settlement rails that never sleep. Not because the future became crypto. Because the future became infrastructure. \