Agreed, it was really cool listening to Andy break down the economics. Have you done the IRR math per Andy's comments, how does it check out? I tried to model it out year-by-year but the IRR comes out too low
Simplistically,
If breakeven in 3 years, 33% return per year
$0.33*5 = $1.65 total cash flow
$1.65 - $1 (initial investment) = $0.65 (FCF)
IRR = $0.65/5 = 0.13 (13%) straight line return p.a. (compounded/IRR should be c. 10% p.a.)
This excludes DC building (GOOG & MSFT said servers / DC split roughly 40%/60%), which would drag IRR despite longer life
And there is this: The AWS model is higher abstraction with agent harnesses while leveraging existing customers and data assets that need that locality, so is both sticker and more expensive than more commodity-like and substitutable token factories. Also this leverages their years of enterprise software hosting experience. Genius. They are really capturing the agent market.
Without Moore's Law making silicon noticeably better, ROI could get even better once those original servers get depreciated as they may not need replacing for 10 years, or until they literally fail. The depreciation life of a server used to be three years, then it moved to four and now prices for even 6 year old chips are holding up as improvements slow to a crawl. The flip side is this could hurt operating costs because old silicon runs hotter to produce less output than newer stuff.
Great post. 🤘
Agreed, it was really cool listening to Andy break down the economics. Have you done the IRR math per Andy's comments, how does it check out? I tried to model it out year-by-year but the IRR comes out too low
Simplistically,
If breakeven in 3 years, 33% return per year
$0.33*5 = $1.65 total cash flow
$1.65 - $1 (initial investment) = $0.65 (FCF)
IRR = $0.65/5 = 0.13 (13%) straight line return p.a. (compounded/IRR should be c. 10% p.a.)
This excludes DC building (GOOG & MSFT said servers / DC split roughly 40%/60%), which would drag IRR despite longer life
And there is this: The AWS model is higher abstraction with agent harnesses while leveraging existing customers and data assets that need that locality, so is both sticker and more expensive than more commodity-like and substitutable token factories. Also this leverages their years of enterprise software hosting experience. Genius. They are really capturing the agent market.
Without Moore's Law making silicon noticeably better, ROI could get even better once those original servers get depreciated as they may not need replacing for 10 years, or until they literally fail. The depreciation life of a server used to be three years, then it moved to four and now prices for even 6 year old chips are holding up as improvements slow to a crawl. The flip side is this could hurt operating costs because old silicon runs hotter to produce less output than newer stuff.