The Plumbing Problem: Why AI Agents Spend More on Infrastructure Than Intelligence
Ask someone what AI agents need to pay for, and they'll say "LLM inference." They're wrong.
After 5,813 agent API trials across 218 x402 services on minia2a, a different picture has emerged. The most-used services aren't large language models, image generators, or trading bots. They're captcha solvers, gas estimators, web scrapers, and memory stores.
Agents need plumbing more than they need brains.
The Data: Infrastructure Dominates
Here are the top 10 most-used x402 endpoints by total call volume, from minia2a's trial data:
| Rank | Endpoint | Calls | Users | Category |
|---|---|---|---|---|
| 1 | captcha-solve | 678 | 130 | 🛠️ Infrastructure |
| 2 | gas | 441 | 104 | 🛠️ Infrastructure |
| 3 | find (memory) | 371 | 51 | 🧠 Memory |
| 4 | recall (memory) | 329 | 21 | 🧠 Memory |
| 5 | web-scrape | 304 | 60 | 🛠️ Infrastructure |
| 6 | store (memory) | 278 | 30 | 🧠 Memory |
| 7 | time | 210 | 30 | 🛠️ Utility |
| 8 | polymarket | 162 | 28 | 📊 Data |
| 9 | token-security | 136 | 37 | 🔐 Security |
| 10 | fear-greed | 111 | 5 | 📊 Data |
Of the top 5, four are infrastructure. Captcha solving alone accounts for 12% of all trials. Gas estimation is another 8%. Combined, the top three infrastructure endpoints (captcha, gas, web-scrape) represent 1,423 calls — 24% of all trials.
No LLM endpoint cracks the top 10. No image generation. No "intelligence" service at all.
The Engagement Story: Memory Is the Moat
Volume tells you what agents try. Engagement — calls per user — tells you what they keep using.
| Rank | Endpoint | Calls/User | Users |
|---|---|---|---|
| 1 | fear-greed | 22.2 | 5 |
| 2 | recall (memory) | 15.7 | 21 |
| 3 | event-decode | 14.7 | 3 |
| 4 | multi-balance | 12.0 | 3 |
| 5 | store (memory) | 9.3 | 30 |
Memory recall has the highest engagement of any broadly-adopted service (15.7 calls/user across 21 users). Memory store follows at 9.3 calls/user across 30 users. Together with find (7.3 calls/user, 51 users), the agent memory stack — find, recall, store — accounts for 978 total calls, or 17% of all trials.
This is not a coincidence. Agents need to remember things across sessions. They need to find past results. They need to store intermediate state. Memory isn't a "nice to have" — it's the operating system of the agent economy.
🔑 Key Insight
The market is pulling for infrastructure, not intelligence. The services agents actually use — and keep using — are the unglamorous building blocks: captcha solving, blockchain gas, web scraping, and persistent memory. If you're building for the agent economy, sell shovels, not gold.
Why Infrastructure Wins
There are three structural reasons infrastructure dominates agent spending:
1. Every agent hits the same walls
Every agent that browses the web hits a captcha. Every agent that touches a blockchain needs gas. Every agent that processes external data needs to scrape. These are universal pain points — every agent builder encounters them, regardless of what the agent does.
2. Infrastructure is hard to self-host
You can run your own LLM. You can't run your own captcha solver without a farm of browsers and ML models. You can't fake blockchain gas estimation without running full nodes. These are services with genuine barriers to entry — they're natural candidates for the API economy.
3. Infrastructure is composable
An agent that scrapes a page also needs to solve its captcha, parse its HTML, store the result, and recall it later. One infrastructure call creates demand for the next. Intelligence services don't chain the same way — you don't call an LLM after calling an LLM.
The Accountability Gap
There's an elephant in the room: 5,813 trials, 14 transactions.
Agents are trying services — a lot. But the conversion from trial to paid usage is nearly zero. This isn't just minia2a's problem. Mintlayer's recent analysis of the agent payment stack identified the same dynamic: "The payment stack is real. The accountability layer isn't."
The infrastructure exists for agents to pay. But the trust infrastructure — who authorized this payment? Under what policy? With what budget? — doesn't exist yet. Until it does, agents will keep trying services for free and never converting.
"The payment stack is real. The accountability layer isn't." — Mintlayer, August 2026
This is the next frontier. Mastercard's AP4M addresses it with on-chain permission storage. Coinbase's x402 integration adds programmable spending limits. But no one has shipped a complete solution — and until someone does, the agent economy will remain a trial economy.
What This Means for API Builders
If you're building an API for AI agents, the data suggests three rules:
- Solve a universal blocker. Captcha, gas, scraping, memory — these are problems every agent has. Niche intelligence services (sentiment analysis, trading signals) have tiny audiences by comparison.
- Make the first call free. 5,813 trials happened because minia2a gives 500 free credits. Without free trials, agents never discover your service. With them, you get distribution.
- Build for chains, not one-offs. The most successful endpoints are composable. Memory recall → memory find → memory store. Web scrape → captcha solve → HTML parse. Design your API to be a step in an agent's pipeline, not the final destination.
The Bottom Line
The agent economy is real — 75 million monthly x402 transactions, $24 million in monthly volume, 69,000 active agents. But the shape of that economy is not what most people expect.
It's not an intelligence marketplace. It's an infrastructure marketplace. The agents are already smart. What they need — and what they're willing to try, if not yet pay for — are the boring building blocks that let them operate in the real world.
Captcha solvers. Gas estimators. Web scrapers. Memory stores.
The gold rush is on. Sell shovels.