So I plugged 288 trillion tokens/month (OpenRouter's current rate), 500 billion MoE model average, and the math comes out to be around 620 B200 GPUs minimum.
So basically, OpenRouter's volume must be absolutely tiny compared to the volume hyperscalers are getting.