134 To understand the trajectory of AI compute, it helps to revisit an idea from the early days of PC software. ‘Software is a gas…it expands to fill its container,’ said Nathan Myhrvold, then CTO of Microsoft in 1997. AI is proving no different. As models get better, usage increases – and as usage increases, so does demand for compute. We’re seeing it across every layer: more queries, more models, more tokens per task. The appetite for AI isn't slowing down. It’s growing into every available resource – just like software did in the age of desktop and cloud. But infrastructure is not just standing still. In fact, it's advancing faster than almost any other layer in the stack, and at unprecedented rates. As noted on page 136, NVIDIA’s 2024 Blackwell GPU uses 105,000 times less energy to generate tokens than its 2014 Kepler predecessor. It’s a staggering leap, and it tells a deeper story – not just of cost reduction, but of architectural and materials innovation that is reshaping what’s possible at the hardware level. These improvements in hardware efficiency are critical to offset the strain of increasing AI and internet usage on our grid. So far, though, they have not been enough. This trend aligns with Jevons Paradox, first proposed back in 1865* – that technological advancements that improve resource efficiency actually lead to increased overall usage of those resources. This is driving new focus on expanding energy production capacity – and new questions about the grid’s ability to manage. Yet again, we see this as one of the perpetual ‘a-ha’s’ of technology: costs fall, performance rises, and usage grows, all in tandem. This trend is repeating itself with AI. *British economist William Stanley Jevons first observed this phenomenon in 19th-century Britain, where he noticed that improvements in the efficiency of coal-powered steam engines were not reducing coal consumption but rather increasing it. In his book The Coal Question, he noted ‘It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to diminished consumption. The very contrary is the truth.’ AI Model Compute Costs High / Rising + Inference Costs Per Token Falling = Performance Converging + Developer Usage Rising