OpenRouter at 10T+ tokens/day: from dismissed wrapper to critical AI infrastructure

Latent Space · rss · 2026-09-26

Latent Space interviews OpenRouter co-founder & CEO Alex Atallah and AMP's Anjney Midha on OpenRouter's journey from the open-model wave to a neutral routing layer for 10M+ developers, now joining Stripe. Key takeaways:

The multi-model bet: Stanford's $600 Alpaca fine-tune of Llama in early 2023 convinced them no single model would win; Discord's early AI deployments exposed closed-model limits.
Distribution beats training: labs spend billions on checkpoints but struggle to reach developers; OpenRouter became the neutral distribution layer, refused adjacent expansions (fine-tuning, memory), and turned focus into its biggest advantage despite VCs dismissing it as "just a wrapper".
Inference marketplace validated: the Mistral price war proved the model; its leaderboard became a live map of AI usage; the platform now serves 10+ trillion tokens per day.
Missed turns: the early MOM (Mixture of Models) fusion experiment failed and was deleted in 2024, only becoming viable again recently.
Joining Stripe and token fraud: Atallah argues token fraud may define the AI economy's security problems, with the next wave coming from autonomous agents attacking valuable token flows — making Stripe's fraud infrastructure strategically complementary.

Original post →

OpenRouter at 10T+ tokens/day: from dismissed wrapper to critical AI infrastructure

More from Companies & People