On the Cisco/NVDA comparison, which so many are sick of hearing, but so many still do not understand.

Cisco required customers to buy extensive support and this was contractual (SmartNet) and this also acted like CUDA a bit - it made each CISCO router a huge investment, and a customer would lose all that support,, certifications, etc. These were real switching costs - my best friend worked for Cisco in this area at the time, and I heard all about this in real time. It was not also not so quickly commoditized. Cisco was dominant for about a decade, and increased revenues 4x over the following 15 years while the stock cratered.

NVDA does not have the contractual lock-in layer. CUDA’s ecosystem has some features of Cisco’s SmartNet, but is not contractual.

As well, currently many chips are in development and will be backwards compatible with CUDA, so one will be able to use other chips and still use CUDA if so desired or if it was easier. This basically makes CUDA transferable to newer non-NVDA chips, and that makes CUDA not much of a moat.

The bigger moat for now is NVDA is the biggest market cap in the world and the mafia Don strong arming smaller firms to lock out competitors. Nevertheless, Anthropic for instance has been built(training) on Trainium and TPUs. Apple too on TPUs. The biggest spenders are developing their own chips and will vertically integrate NVDA out of their future cost structure, and if they want to use CUDA, they will, without NVDA chips.