I just used Instinct to book my flights to Austin for @TomorrowXSummit (including exit aisle seats). And just like that I’ll never book a flight on my own again. If you’re planning on attending and haven’t registered yet, please use my link to sign up. It helps me out:

Credit to @danielnewmanUV for bringing this up but I think it’s a great point Thanks to Instinct and Muse there’s probably alpha in upward CPU revisions Muse&Instinct are using tons of tool calls and relying heavily on CPUs

I really feel Andrew nails it here with describing what makes Nvidia so special, and why the company has the admiration of even their competitors It’s the organizational grit and resilience and relentless self-belief across an incredible length of time that’s what makes Nvidia so special. Most founders would have left a long time ago

My conversation with Penguin Solutions ( $PENG) and Astera Labs ( $ALAB) from Ai Infra Summit: Both working on CXL related solutions Astera is a $50B market cap company, and Penguin is a $2B market cap company, so there’s a huge size difference between the two I’m lumping them together because they do similar things at least in one of the lines of their business (Astera’s “Leo” CXL product, and Penguin’s compute solutions division) Penguin uses Astera Labs “Leo” controllers for their CXL offering (SMART) (although I guess the next generation will use Qualcomm if I remember correctly), but they layer on top more of a finished product experience primarily for enterprise users (not CSPs) who need some extra support beyond what Astera will provide. So their product is more suitable for Fintech/financial services, health care, and other industries where enterprise on-prem is in-demand These boards have DIM slots for LPDDR5 memory, which can be used to augment memory from GPUs (meaning a server tray can have less compute vis expensive GPUs by pooling in this board’s memory via LPDDR5) instead of having to either A. Have more GPUs ($$$) or B. Going all the way up to memory on the CPU which causes a big slowdown in memory access speeds. The next generation Astera “Leo 2” and Penguin’s next generation equivalent will both also be compatible with LPDDR4 Cool product. Can see the need, makes sense. I don’t know how defensible Penguin’s business here is though, I guess it’s niche enough that Astera is focused elsewhere, but I imagine that if on-prem got lucrative enough maybe Astera would come after it (many such cases across the showroom floor)

Alright… Back in Santa Clara for AI infra summit day 3. My thoughts so far after Day 1-2: The energy at AI infra summit is nothing short of frenetic Every vendor I talk to is running now as hard as they can on their product development roadmap. If you place each companies roadmap on a Venn diagram, many overlap with each other. All these advancements seem to be accelerating (probably with the help of AI), so leads that exist today feel wobbly to me when considering where we are going. I think the conversation I had with Astera labs yesterday was a good example of this Basically, the guys at alab are crushing it and they’re moving rapidly, but so is everybody else and other people are working on CXL. It’s just so hard right now to say with confidence that they are going to remain ahead. I think people are really misleading a lot of investors with claims about technical differentiation between companies while not mentioning other companies roadmap, the complete breakdown of trade-offs between them and factoring in how much things are changing and how quickly they’re changing I’ve been covering “the industry” now coming on 3 years, and I feel a distinct sense of acceleration relative to the past across all these product categories from SSD / storage, energy, networking, software, on and on. It’s like everyone hit the FFWD button all at the same time

Think about it Dylan can simply throw $300k (probably less) at 10 accounts with strong influence In return he gets: - top of funnel - influence (aura) - control For $3M annual cost Very cheap imo 🤷🏼‍♂️

This sound bite from @sk7037 was crazy - NVIDIA Rubin Racks gave a 3x performance improvement over Blackwell just right out of the box - then because of the building blocks NVIDIA created to optimize its hardware, OpenAI was able to point Astra at it and over 72 hours it gained another 2x in performance on top of the first 3x fully autonomously 🤯

@sk7037 - “The way we will make sure frontier models are safe and aligned is by spending compute. So the counterintuitive point is that we’ll need even more compute to make sure future models are more safe and aligned” Yeah… slow down my a** 😂 OAI wants even more compute