Some interesting twitter analysis here:

https://x.com/bleysg/status/2073937651150029084

Note that Chinese companies are free to rent from GB300 clouds internationally. There are large datacenter hubs in Singapore and Malaysia serving chinese and other customers.

Though there is also reported [1] significant smuggling of Nvidia chips into China as well.

[1] https://epoch.ai/publications/chip-smuggling

But on the other hand, these companies are essentially paying for the service of taking the debt off books (by paying the leasing premium to the SPV partners). I guess I'm wondering what they really gain from doing so, if again sophisticated investors can see through the games?

https://archive.ph/20260720174223/https://asia.nikkei.com/bu...

Also to the authors last point (extremely long time scales causing degradation), it seems like we'd want high thrust capabilities regardless. i.e. maybe a small gravity well doesn't gain us anything, since we'd need big engines to get up to speed anyway.

In the latter case spinning doesn't get you far.

They don't expect to keep the prices flat over time, and everyone involved will have planned for this. Prices are highest when they're the newest and greatest (part of why it's valuable for neoclouds to be first in line for new models), and drop year by year as newer GPU models can do equivalent work at lower cost.

You can see a pretty cool dataset of this at [1]; H100 prices where $3/hr in 2023, and dropped linear-ish to $1.75/hr by 2025. And also the notable exception that prices are up this year due to shortage.

[1] https://semianalysis.com/gpu-pricing-index/

Also a lot prompt spent feeding it strategies, which feel like they should/will eventually be deduced by the model itself, not explicitly stated. That's not to take away from the outcome in any way; rather, it feels sort of like when you would prompt GPT 4, "think through your answer step by step," as a sort of proto-chain of thought.

Amusing that they use A100e as the reference point to sound impressive. Different ways you could make that conversion, but based on FP4 FLOPs (yes it's disadvantageous to A100, that's the point), that's something like 200hr on a GB300 NVL72 rack.

Not nothing either, but far less astounding sounding than 700k hrs.

Though yes in my experience, at those frequencies people stop using ITU designations, and switch to IEEE (S,C,X-band etc).

You can put some memory on the logic wafer (SRAM) but it's area inefficient, which is wasteful on your expensive N2 wafer. So a dedicated DRAM process is vastly cheaper per bit, even at current elevated prices.

The new game is finding a single sentence with the most instances of "safe" or "safety". My current high score is 4..