How Kong Was Born: APIs, Hustle, and the Future of AI Infrastructure
Summary
Marietti turned a $600, 90-day tourist-visa gamble into $51,000 of angel capital; Kong’s story is ultimately one of refusing to die. The three founders then lived on $1,000 a month in San Francisco, sharing mattresses and eating rice, beans, and tuna pasta. Casado’s corrective to Marietti’s “lucky” framing lands: “That sounds pretty unlucky.”
The API marketplace failed for structural reasons, not merely weak execution. Supply lacked exclusivity, only about 30 APIs mattered while 3,000 formed a weak long tail, quality failures damaged trust, and AWS costs erased margin. The valuable asset was the gateway underneath it, built three times to power 20,000 APIs.
Open-sourcing Kong in April 2015 was a near-death extraction of the company’s best technology. A $2 million insider extension kept it alive after it ran “out of gas”; Casado says only two weeks remained after the bridge. Marietti’s retrospective is blunt: “We were dead and we just didn’t know about it.”
Once Kong took off, seven years of starvation converted into unusually fast enterprise growth. ARR later moved from $2 million to $10 million in one year against a $6 million-$7 million plan, and Kong subsequently announced it had crossed $100 million. The company memorializes its lean years with a Founders Award of 2,555 stock options—“2,555 days of struggle.”
AI expands Kong’s market because agents consume the internet through programming interfaces rather than human-facing websites. Agents will “programmatically exchange labor” through APIs and MCP, which Marietti calls “Duolingo for APIs.” More machine-to-machine work means more authentication, authorization, routing, governance, and metering.
The immediate AI opportunity is mundane infrastructure with hard enterprise value. Model monitoring, billing, API-key provisioning, rotation, and permissions must work before agents can “roam free”; otherwise authentication repeatedly forces a human into the loop. In Marietti’s formulation, without APIs an AI system lacks “the mouth and the ears for a model.”
Kong’s larger bet is that enterprises will centralize AI connectivity just as microservices centralized common logic in gateways. As companies move from one large model to five, 10, or 100 models, duplicating token rate limits and authentication across every framework stops making sense. Marietti expects an evolutionary two-to-three-year convergence: “API traffic and AI traffic, they’re converging.”
Deep dive
Not yet available upstream; scheduled sync will retry.