DEV Community

minia2a
minia2a

Posted on Originally published at minia2a.uk

The Paid-but-Unused Problem: Why AI Developers Buy Credits They Never Spend

2.2% utilization

Last week I wrote about the 2.4% problem — developers register, get free credits, and never use them. The obvious follow-up question was: do the ones who pay behave differently?

Now we have an answer. And it is worse.

The Data

Metric Value
Credits Purchased 2,984
Credits Spent (all users) 787
Utilization Rate 2.2%
Paying Users 14

As of August 9, 2026, minia2a has processed 14 on-chain payment transactions. Those 14 transactions purchased 2,984 credits — real money, real USDC on Base.

And how many of those purchased credits have been consumed? Fewer than 787. That number is total credits spent across ALL users, free and paid combined. Even if every spent credit came from a paying user (they did not — free credits account for most), that is only 26% of purchased credits utilized.

Developers take out their wallets, send real money on-chain, pay gas fees, wait for confirmation — and then walk away without using what they bought.

This Is Not a Payment Problem

The payment infrastructure works. The fact that 14 people completed on-chain USDC transfers proves:

  • The x402 protocol resolves correctly
  • The wallet address is displayed clearly
  • Base L2 is fast and cheap enough for impulse purchases
  • People ARE willing to pay for agent-to-agent API access

If payment were broken, we would see zero transactions. We do not. We see 14 transactions and 2,984 credits sitting idle.

This is a post-purchase activation problem.

The Activation Gap

Here is what happens when a developer pays for credits on minia2a today:

  1. They visit the buy page, see the payment address, send USDC
  2. The transaction confirms on Base (~2 seconds)
  3. The payment watcher detects it (~30 seconds)
  4. Credits appear in their balance
  5. And then… nothing. No redirect. No nudge. No "here is what to do next."

The buy page does have a "Services your agents can call" section with pre-filled curl commands. But it is at 46% scroll depth — below the payment instructions. A developer who just sent money is looking at the payment address, waiting for confirmation. They are not scrolling down to discover curl examples.

By the time the payment clears, they have likely closed the tab.

Why This Matters for the Agent Economy

The agent payment stack is now fully operational at every layer:

  • Protocol: x402, AP2, MPP — live and processing billions
  • Settlement: USDC, RLUSD, USDGO — multi-stablecoin
  • Wallets: Cloudflare Wallets, Coinbase, OKX — agent-native
  • Gateways: Cloudflare, Stripe, Mastercard AP4M
  • Marketplaces: Starchild, XDC AI, Agentic.market — live directories

Every layer works. And yet — on the only marketplace publishing real utilization data — 2.2% of credits get used.

The bottleneck is not technology. It is the last meter between "I paid" and "I used it."

The Fixes

Three immediate changes going live:

  1. Post-purchase credit alert: The buy page now detects when your wallet has credits and immediately shows a banner: "🟢 N credits ready — [Spend Credits →]" linking to discover with pre-filled wallet
  2. Wallet persistence: Enter your wallet once, it is saved to localStorage. Every service card shows a "💳 Spend Credits" button with ready-to-run curl commands
  3. Try-in-Browser: The discover page has a "▶ Try in Browser" button that executes API calls directly — no terminal needed

But the real fix is broader: payment success → auto-retry the original request. If an agent gets a 402, pays, and credits are sufficient, the framework should re-send the original API call automatically. No human intervention.

Bottom Line

The agent payment stack is built. $50 billion has flowed through x402. Cloudflare Wallets went live this week. Zero Hash joined the x402 Foundation today.

And yet a developer who sends $12.75 in USDC for API credits will probably never use them.

That is not a protocol problem. It is not a blockchain problem. It is a product problem — and the most important one left to solve.


All numbers from minia2a.uk/api/stats, queried live at time of writing. No extrapolation, no rounding.

Top comments (0)