Thanks for the write up. The point on Huawei’s potential lock-in strategy through CANN and MindSpore deserves more attention, especially as domestic developer ecosystems mature. I write on the Chinese military’s integration of AI on my Substack if folks are interested. https://ordersandobservations.substack.com
You mentioned 3 legs of the moat. Although the software leg is most easily analyzed, the other 2 legs are just as powerful. The GPU advantages are well understood, history with Intel shows how long that can last if properly executed. But people also forget how important (datacenter) networking is. In my experience (40 years) processor people don't understand networking. I would also claim storage, memory and IO are also important ingredients to a successful AI datacenter that are not to be underestimated. Although I generalize, server design and datacenter design are also highly specialized and therefore differentiated. These are long term CS problems that have to be rebalanced every time there's a new workload.
Very interesting. US policy, in one stroke, has essentially resolved previously dominant qualms about the delays and inconveniences of bringing an alternative to maturity.
The battle of the AI stacks will be one to watch very closely. Looking forward to hear how this develops over the next few years.
Thanks for the write up. The point on Huawei’s potential lock-in strategy through CANN and MindSpore deserves more attention, especially as domestic developer ecosystems mature. I write on the Chinese military’s integration of AI on my Substack if folks are interested. https://ordersandobservations.substack.com
come for the PLA AI, stay for the civil war history!
You mentioned 3 legs of the moat. Although the software leg is most easily analyzed, the other 2 legs are just as powerful. The GPU advantages are well understood, history with Intel shows how long that can last if properly executed. But people also forget how important (datacenter) networking is. In my experience (40 years) processor people don't understand networking. I would also claim storage, memory and IO are also important ingredients to a successful AI datacenter that are not to be underestimated. Although I generalize, server design and datacenter design are also highly specialized and therefore differentiated. These are long term CS problems that have to be rebalanced every time there's a new workload.
Very interesting. US policy, in one stroke, has essentially resolved previously dominant qualms about the delays and inconveniences of bringing an alternative to maturity.