Não acho que as pessoas apreciem totalmente a quantidade de demanda que a inferência criará para a infraestrutura de IA. O treinamento de modelos de trilhões de parâmetros empalidece em comparação com os volumes de trilhões de tokens simultâneos funcionando 24 horas por dia, 7 dias por semana, 365 dias por ano em todos os negócios, organizações e atividades do consumidor. Estou tão otimista com essa construção. E é claro que $NVDA se sai bem, mas nomes em chips, servidores, energia, rede e plataformas agenciais aumentarão com ele. 💪🏻 Marque 😎🚀

Highlights from AMZN call: -AI Infrastructure Capacity Shortfall: AWS has "more demand than we have capacity" with power being the single biggest constraint; expects supply issues to persist for "several quarters" despite $31.4 billion quarterly CapEx in "chips, data centers and power" -Multi-Year Infrastructure Buildout: Acknowledges building sufficient AI capacity will take "several quarters" but expects improvement each quarter; views AI as "biggest technology transformation in our lifetime" requiring sustained heavy investment -Inference Economics Driving Strategy: Expects 80-90% of AI costs will shift from training to inference at scale; positions custom silicon as critical advantage since customers will anchor for better price performance like they did with CPUs -Massive Revenue Opportunity Constrained: AWS at $123 billion annualized run rate but "could be doing more revenue and helping customers more" if capacity existed; generative AI already a "triple-digit year-over-year percentage multibillion-dollar business"