Amazon's $100M Moonraker Leak Reveals How Far Alexa Lags Behind GPT-L…

Amazon's leaked Moonraker project reveals plans to spend $100M+ on GPU costs to add multi-request task chaining to Alexa, the same day OpenAI launched GPT-Live with full-duplex voice and agentic capabilities. The gap exposes a fundamental architecture and compute economics gap: agentic AI requires 12-60x more tokens per interaction than single-task voice commands, and Amazon's own Nova models are insufficient—internal documents show Amazon is testing with Anthropic's Sonnet instead. If you're selling inference compute, reasoning models, or voice stack components to consumer AI platforms, Amazon just proved the economics of the next generation. Moonraker's $100M+ annual burn at 10k-50k tokens per task is the new table stakes for voice agents—and Amazon chose Anthropic's Sonnet over its own Nova, signalling which models win when cost per task explodes.