Rare signals carve out distinct subspaces in the latent space without dilution
United States
Joined July 2015
- Tweets1.5K
- Following1.4K
- Followers112
- Likes14.5K
AI 2027 investing roadmap
1) Elon wants to launch 10Gw computer by end of 2027 - at 50Bn (max) that’s 500Bn @elonmusk
2) Nvdia helps financing with 500Bn underway and assuming many such future tranches with backstop provided by them @JensenHuang
3) at glossy returns that’s 250bn (2 year breakeven) per year that “AI compute clients have to pay”
4) say Anthropic, OpenAI, SpaceXAI etc “pay 250bn” to the compute providers as cost of tokens, they obviously are gonna add a thick margin on top of that to sell it to their customers
5) assuming conservative 50% margins that becomes 375bn
6) how do “real world companies and compute users” conjure up 375Bn per year
7) with CoGS of S&P500 companies being 11Tn (real world is much bigger), the 375Bn represents only 3.5% additional input cost, which is easily surpassable through value addition
8) Overall HBM memory available already is estimated to support ~20Gw and ~35Gw in 2027 and 2028 in total across all players
9) rounding back to the investors of 500Bn - thus there’s ample room for returns and ample rounds of such deals possible in coming years within ‘existing’ memory output available, even without considering the ramp ups
10) Over time period of 3-10 years, as the exponential curves increase in terms of memory supply, power supply, the corresponding RoI curves take similar shapes too with the following ranking highest to lowest - Power (SpaceX), Memory, physical Chip Makers, Inference/NeoClouds, Model providers - with exponential returns attributable to equity investors followed by the static returns to the project/tranche based debt investors
@DavidSacks @chamath
Given the scale of things Chamath will hardly be sitting on a couch sipping wine, so here’s to ending it on a lighter note with meme of the year :-)
🤖 Made with AI
