Qwen 3.8 Max, DeepSeek e Kimi: cosa sta succedendo davvero nell'AI

The transcript presents a comprehensive analysis of the current state of AI competition, particularly between Chinese and American companies. Chinese laboratories including DeepSeek, Alibaba (Qwen), and others are releasing frontier-level open-source models approximately weekly, achieving performance parity with American models like GPT-4o at significantly lower costs—DeepSeek V4 Flash costs 8 cents per million output tokens versus GPT-4o Luna's $1.20. This price competition has forced OpenAI to depreciate token costs by 80% and is driving a fundamental restructuring of the AI industry's business model. The speaker argues that Large Language Models have become commodities, meaning performance alone no longer differentiates products. Consequently, the competitive battlefield has shifted from "who has the smartest model" to infrastructure, hardware efficiency, and software optimization (referred to as "harness"). Chinese companies maintain cost advantages because they develop their own chips, control manufacturing, and operate their own data centers without relying on U.S. semiconductor vendors. In response, American technology companies including Nvidia, Microsoft, Meta, and even startups like Mistral and Anthropic are increasingly releasing open-source models, recognizing that the real profit lies in hardware sales rather than token subscriptions. Nvidia's DJX Station and rumors of Apple's M7 Ultra with 1.5TB of unified memory represent attempts to capture the local inference market. The speaker presents financial data showing that major AI companies are currently unprofitable (losses ranging from billions to hundreds of billions), while hardware manufacturers like Micron and Nvidia are highly profitable, validating the thesis that hardware, not models, is where value accumulates. The analysis also addresses consumer hardware pricing increases, Apple's shift toward device leasing models, and speculation about future consolidation where companies will offer complete packages combining custom hardware, proprietary models, and software ecosystems. The speaker expresses concern about the "rental economy" approach to hardware ownership and discusses AI capability expansion, citing Alibaba's Qwen 3.8 Max achieving 16-day autonomous operation (versus previous single-day capabilities) and implications for prompt engineering transitioning to "loop engineering." The transcript concludes with predictions that model sizes will continue growing exponentially for data centers while simultaneously developing smaller, optimized models for edge devices and consumer hardware, fundamentally reshaping how AI infrastructure and business models evolve.