What’s More Expensive, Coffee or AI?

Engineers often celebrate breakthroughs that save a few picojoules per bit. But what does that really mean?

A picojoule is one trillionth of a joule. Boiling a cup of coffee takes roughly 84 quadrillion picojoules. So a picojoule is an extraordinarily small unit of energy. The important words, however, are per bit. There are a lots of bits in AI. On a single AI rack there is enough bandwidth to host 250 million people on a zoom call. 

Across a conservative forecast of 124,000 AI racks in the United States, eliminating just 1 pJ/bit could free approximately 124 MW of continuous power – about 1.1 TWh annually, worth roughly $76 million in electricity.

But in a power-constrained AI factory, electricity is not the real prize. The freed capacity can run more compute and generate more tokens. At an illustrative energy requirement of 1 joule per token, that power could produce roughly 3.7 quadrillion additional tokens per year. At a blended price of $1 per million tokens, that represents approximately $3.7 billion in annual revenue capacity.

That is why the industry invests billions in optical technologies such as CPO, LPO, NPO and XPO to recover individual picojoules. $1B to save 1pj/bit with a $3.7B annual return – is a good move any day. 

At Resolight, we believe the opportunity goes further: redesigning the scale-up network could eliminate up to 40 pJ/bit. Under the same assumptions, that represents as much as $149 billion per year in additional theoretical AI revenue capacity across the projected U.S. fleet.

If you would like to learn more about this visit our https://resolight.ai/ and ask for a link to the webinar that explains the business and technical details behind this. 

I still owe you an answer:  what is more expensive: coffee or AI?

The energy needed to heat one cup of coffee could generate approximately 84,000 AI tokens. At $1 per million tokens, those tokens would cost about 8 cents – far less than the average cup of coffee. So yes, coffee is still more expensive than AI (from a value add on energy perspective). But who knows what the future holds?