Token Arbitrage Not Proxy

This distinction says the money is in controlling order flow between model buyers and model supply, not in the routing software alone. A proxy is just the plumbing that lets requests move across providers. The business becomes arbitrage when a company aggregates many models, sets the retail price or takes a spread, and wins by steering demand toward whichever provider is cheapest or underused at that moment.

• OpenRouter fits this marketplace pattern. It offers one API for 400 plus models, has been described as processing 3T plus tokens per day, and its economics are framed around a take rate on gross model spend, not around selling gateway seats or enterprise control software.
  2 sacra 6 sacra 7 sacra
• Kong is describing a different business. Inside large companies, the gateway is bought as internal infrastructure to enforce security, routing, caching, rate limits, and chargebacks across many LLMs. The customer is paying for governance and cost control behind the firewall, not for discounted token resale.
  1 sacra 5 sacra
• Public gateways can blur the line. Vercel explicitly says bring your own key carries 0% markup, which makes it closer to software and observability. Cloudflare ties AI Gateway usage to Workers AI token billing, which shows how public gateways can also become distribution channels for underlying model consumption.
  3 vercel 4 cloudflare 8 vercel

Going forward, AI gateway companies will separate into software vendors and token exchanges. The software vendors will win on security, reliability, and internal control. The exchanges will win on liquidity, price discovery, and demand aggregation. The biggest platforms will try to combine both, then use the software layer to capture the transaction layer above it.