a very hardware-heavy week in AI, here's the highlights

google encroaches on nvidia's territory with 2 multi-billion dollar TPU deals

news broke that google's selling $10Bs of their latest in-house tpus to Meta and NATO posing direct competition to nvidia gpus

this makes most sense for meta who could save between 30-50% on ai capex spend. also googles tpus specialize in recommender systems which is basically meta's ad system. marriage made in heaven?

google's head of ai infra went on record saying they need to double compute every 6 months just to keep up with demand...

if we factor in the 1M anthropic tpu deal then google's odds of taking on nvidia went from 0-15/25%

china claims new ai chip can beat nvidia's

china's piling on the nvidia hate by releasing a new chip concept that achieves same performance quality as an nvidia a100.

personally don't really care, nvidia's latest gb200/300 smokes these but the recent acceleration of china's chip technology is something that needs to be monitored.

huawei and alibaba have also independently been making their own chips which are advancing at a similar rate. if china ends up catching up to usa chip dominance then we might be in a bit of trouble when it comes to scaling wars (china usually dominates in raw factory output).

robotics ai vision is underpriced

wrote about this a lot this week, basically i think people underestimate how quickly robots will enter the workforce and actually be productive

most obvious sectors are manual labour in factories and fulfillment centers. Amazon's probably the most experienced company in the latter plus they're trimming fat and investing $100B+ in their ai strat - if i had to bet i'm guessing they are undervalued here.

tesla is the other one, their automotive FSD ai model + optimus rollout is going to be a big deal if executed to plan. not an immediate play but something to think about given we havent seen the "chatgpt moment for robotics" yet.