Source item

Quoted context

These are the quoted author’s words.

One of the coolest things about Altimeter is how open @altcap encourages us to be with our research process. In this interview w/ Jensen we asked him the questions we (as analysts) wanted to know - and sharing it with the community Some of my takeaways: 1/ Nvidia is a full stack company. This works across the entire company from how the product is physically built (GPU, CPU, Networking, Rack, DC), to how it is run (CUDA, cuDNN, backwards compatibility, future support, etc), to business model – general purpose accelerated computing that is supported across all ecosystems that runs training AND inference workloads AND improves generation to generation Easy for pundits to say hyperscalers make 50% of NVDA DC revenues -- but who do you think is truly steering the ship? 2/ GenAI for leading labs is a multi-step process – all of which requires accelerated computing. People focus on pre-training, post-training, inference as the major buckets - but between all of this there are MANY workloads that all need accelerating. It will take some time for AI Labs to migrate - if at all - because it is a utilization question. When you migrate to new bleeding edge GPUs (GB200) to train a model, you've now left behind massive clusters of H100s that can be used for inference. And when you are thinking about the datacenter, the right metric to focus on is lifetime utilization 3/ On ASICs - "I am very situationally aware of all the things our competitors are doing... we are just doing something different" Remember everything Nvidia is innovating on is at the bleeding edge. Not sure whether a hardware company has ever had as wide an advantage - perhaps IBM in early days - but that was also when the competitive landscape was much softer. In many ways, Nvidia is leading on hardware (NVLink) - but also writing software that pushes the bleeding edge. Unless there is a 10x better product or a 5x cheaper product (for the same performance), there are great natural disincentives to switch No doubt a lot of fierce competition ahead - the prize is so great - but let's not get ahead of ourselves on ASICs. It took Google a decade to build TPU and even now - it is not flexible enough to be general purpose accelerated compute. Over time, can fit ASICs to internal workloads Other cool takeaways: Market maker, not share taker New models on the horizon @elonmusk is a force of nature Nvidia is not single threaded - the whole point is to build a flexible architecture that will meet how the world develops (and they have visibility into that future) Inference scaling - huge Everyone will be their own CEOs of agents On a human level, this was one of the coolest experiences of my life - being able to spend time with two of the smartest and down to earth people who I deeply admire. I cringe watching myself vs Jensen and Brad just flow naturally. A lot of work from the team and a lot of support from a lot of unsung heroes went into putting this all together so hope you all enjoyed!

_clarktang ↗