The real hardware story isn't the models—it's the cost spiral no one can see. China's 2.8T-parameter Kimi K3 and Alphabet's $120B profit surge both point to the same problem: we're scaling inference without understanding the actual power and cooling footprint per token. That startup estimating corporate AI emissions isn't sexy, but it's the only item here that matters for hardware planning—you can't architect efficient chips if you don't know what you're actually powering. Everything else (neuro-symbolic reasoning, Persian OCR datasets, HIPAA LIMS workflows) is real work, but marginal; the infrastructure cost crisis will determine whether those models even run economically in 18 months.