I came across an estimate that the accelerator fleet sold outside China had reached roughly 25.6 million H100-equivalents by mid-2026. The same analysis estimated that current inference demand uses only around 16-26% of that capacity, potentially even less now that China is offering inference services on its own chips. These are modeled estimates, not audited figures, and not all capacity is available for inference. But if they're even roughly in the right ballpark, the implications are staggering. Even if inference demand grew 50% every year, the existing fleet could theoretically cover around 3-4.5 years of growth without buying another chip. So what justifies the next hundreds of billions in AI infrastructure spending? Are we building years ahead of demand, or am I missing something fundamental?
CryptoAlerta — análise de criptomoedas e mercado em tempo real