The token throughput improvements are impressive. This has direct implications for usage-based billing in AI products — faster inference means lower cost per request, which changes the economics of credits-based pricing models significantly.