#blowing bubbles

Based on a post by @earth-unicorn

"By 2030, AI companies will need $2 trillion in combined annual revenue to fund the computing power needed to meet projected demand, Bain & Co. said in a report released in September. Yet their revenue is likely to fall $800 billion short of that mark, Bain predicted."

$1.2 trillion dollars is their lowball, failure state estimate? I'd ask if I could have some of what they're smoking but they're finance people so it's probably just cocaine and lack of sleep.

For some context, the US Gross Domestic product is about $30 trillion. They said that for LLMs to break even, they need to be worth about $2 trillion. Let's look at the latest available BEA report, specifically the table on page 10 breaking it down by sector. For LLMs to break even, they need to be worth as much as all motor vehicle and part sales, all furnishing and durable household equipment sales, and all recreational goods and vehicle sales put together. They are claiming that instead LLMs will "only" be worth as much as all motor vehicle and furnishing sales put together. To put it another way and look at the costs from the other direction, lets look at comparisons. The best case scenario for LLMs without AGI fantasies is that they're a game-changing core-to-everything productivity booster for desk work on part with, say, Microsoft Office, which revolutionized office jobs thirty years ago or so by improving efficiency dramatically and becoming a core necessity that made entire subcategories of employment and regular job tasks redundant, so lets use that for a point of comparison. A bit of googling says that Microsoft Office accounts for 23% of Microsoft annual revenue, which was in turn $198 billion, or about $45.54 billion. They are saying that their lowball, failure state estimate for LLM importance is "26x more valuable than not having to write anything by hand anymore and have spreadsheets easy for anyone instead of just accountants and have presentation-making take minutes instead of days" Want to compare to "doing high level research digitally via the internet instead of digging through libraries"? LexusNexis makes $26 billion annually. Want to compare to doing statistical analyses on computer with easily available support and training instead of by hand and complicated mainframes? Python and R are open source so no comp there, but SAS makes about $3 billion in annual sales.