Hardware substrate (Moore + algorithm density) net tempo?

Chip density and AI algorithm efficiency compound faster than power consumption grows between now and end-of-2028. Nvidia's transition from Hopper through Blackwell to Rubin, paired with TSMC shrinks from N3 to N2, delivers enough transistor density gain that each new GPU generation cuts cost-per-token by 40–50% despite rising absolute power draw. Epoch's public efficiency metrics show algorithms (sparse attention, better quantization) improve as fast as hardware does. By 2028, the net effect is that training and inference become measurably cheaper per unit of compute, pulling down the real cost of running large language models and powering new model sizes that weren't economically viable in 2024.

Power walls, manufacturing yield limits, and algorithm gains flatten out faster than chip designers can shrink transistors. TSMC's N2 transition slips or costs more than roadmap; AMD MI400 and Nvidia Rubin ship but with lower-than-expected clock speeds or higher defect rates. Epoch efficiency curves decelerate—sparse attention and quantization hit diminishing returns. By 2028, cost-per-token falls only 15–25% from 2024 levels, and rising electricity costs eat gains. The hardware-software loop slows: new model architectures offer only marginal wins, and hyperscalers pause scaling plans because the economic math no longer closes.