Can you explain how that would work in this case? VAST isn’t an investor in CoreWeave, nor is CoreWeave an investor in VAST.

Boy I might’ve set expectations too high. I hope VAST brings the socks to #SC25 now because I don’t have enough in my personal stock to satisfy demand.

I will hold back a pair for you too. But you might have to visit the VAST booth (or our secret restaurant) to actually pick them up!

Hopefully this isn’t the year St. Louis is empty because of US dysfunction instead of COVID.

Tasty, but not as nice as these beauties.

Agreed - optics are the only way to overcome physics here. But their flakiness in practice is undeniably greater than copper. Just like with HBM, people will endure increasing pain over reliability in the pursuit of better performance

I will set you up. Still owe you for the EPCC Highland cow.

It’s easier to claim “higher performance pod” based on a spec sheet, but that’s a very HPC (not AI) way of thinking about performance. Says nothing about realized training/inferencing throughout, power and opportunity cost, or cost/reliability at scale. Optics create non-perf challenges at scale.

Yeah that’s a good point. It’s a new world out there. Nobody is monogamous in the AI infrastructure business.

Once software catches up to the trajectory of QLC/PLC, no AI infrastructure provider is going to waste their time trading GPUs for HDDs. A HAMR HDD’s ROI/TB is lower than a 100+ TB SSD’s because its TCO (esp. opportunity cost) is so much higher.

I don’t say this just because I work at a company that is all-flash; the reality is that AI is power-constrained, and every large AI installation has to decide how may MW to spend on GPUs vs. storage. 128-256TB SSDs let you deploy more GPUs than the best HAMR drives, period.

Well, it’s true. Most of what I do day-to-day isn’t storage. There are plenty of other people at the company who have that covered.

Come to my talk at SC (sc25.conference-program.com/presentation...) if you want to hear about the non-storagey parts of my work.