Capital Markets Infrastructure for Compute Traders

SF | NYC | London
Filter
Exclude
Time range
-
Minimum likes
Replying to @mayazi
Quick calc: B300 server costs $550K, it's 80% financed over 3 years in the neocloud case at 12%, in the hyperscaler case at 7%. Respectively: $158K of interest for a neocloud, against $92.4K of interest for a hyperscaler. Delta is = $66K. Now, what does a B300 server fetch for a ~3yr reservation assuming ~= $4.5/gpu/hr equivalent to $946K for the 3 years. The delta in financing cost explains 66/946 ~= 7% of the difference. This is a small but important contribution to the bulk indexes like the ones @ComputeDesk publishes. Does it show up in the indexes ? Indirectly, yes, but practically it it is absorbed by the margins of the cloud companies. Now what about everything else. Practically the GPUs are used for longer than 3 years. So the 7% figure shrinks down to sub 5% at best. At this level, spot pricing is dominated by supply and demand, and not by cost of credit. Consider that the cost of hardware increased by 20% over the last month, which means the at least one of the confounding factors is greater larger than the credit factor. What about open source demand ? So, the answer is, eh, credit does affect the spot price of compute, a bit in the first few years, and then not much. But twitter is not a place for nuance.
1
15
1,236
Replying to @andrawesbahou
Bet @michaeljburry and @edzitron would find this funny
2
69
Thank you for the #DESKB300 mention !
1
2
190
Thank you for the $DESKB300 mention !
1
2
130
Oh ! @ComputeDesk in the wild. ticker DESKB300 through the roof, and it's not stopping. Deepseek just made things worse. Wait till monday's B300 print.
While neocloud investors got absolutely wrecked last week, GPU rental prices skyrocketed. B300 rates are now up >20% over the past month. Know what you own... H/T: @matthew_sigel $IREN $NBIS $CRWV
1
1
17
1,551
Replying to @creatine_cycle
And Compute dealer
46
Replying to @ren_stocks
18 days today ! ;)
3
721
Replying to @b3labs @a16z
And a B300 more than 40% more than 3 months ago !
Nvidia hardware is an appreciating asset, with B300 value up 40% in 3 months. "The B300 Summer". In July, @Kimi_Moonshot released KimiK3, expanding B300 demand and value. By late September @Supermicro "Gold SKU" supply on 4-week lead time was depleted, triggering a continued value increase, with no sign of slowing down. Residual value for Nvidia hardware is increasing. @NVIDIAAIInfra @JensenHuang
3
79
Needless to say, the reason for the $DESKB300 gpu price ticker is exploding (We're calling it the "B300 Summer") is this. @Kimi_Moonshot, did this with KimiK3
Blackwell Ultra (B300) GPU rental prices up 17 consecutive days to a new all-time high, according to Compute Desk's index of verified transactions from U.S. neoclouds.
1
1
8
803
Replying to @matthew_sigel
Thanks for there reference. Open source models, and Supermicro capacity shortage to blame. @ComputeDesk
Nvidia hardware is an appreciating asset, with B300 value up 40% in 3 months. "The B300 Summer". In July, @Kimi_Moonshot released KimiK3, expanding B300 demand and value. By late September @Supermicro "Gold SKU" supply on 4-week lead time was depleted, triggering a continued value increase, with no sign of slowing down. Residual value for Nvidia hardware is increasing. @NVIDIAAIInfra @JensenHuang
1
10
1,111
Replying to @RMladek
Bonjour
1
3
29
Replying to @Melt_Dem
They don't know
2
205
Replying to @zerohedge
And it continues to explode
Nvidia hardware is an appreciating asset, with B300 value up 40% in 3 months. "The B300 Summer". In July, @Kimi_Moonshot released KimiK3, expanding B300 demand and value. By late September @Supermicro "Gold SKU" supply on 4-week lead time was depleted, triggering a continued value increase, with no sign of slowing down. Residual value for Nvidia hardware is increasing. @NVIDIAAIInfra @JensenHuang
1
20
Nvidia hardware is an appreciating asset, with B300 value up 40% in 3 months. "The B300 Summer". In July, @Kimi_Moonshot released KimiK3, expanding B300 demand and value. By late September @Supermicro "Gold SKU" supply on 4-week lead time was depleted, triggering a continued value increase, with no sign of slowing down. Residual value for Nvidia hardware is increasing. @NVIDIAAIInfra @JensenHuang
11
12
77
715,608
Replying to @andrawesbahou
*Blushing*
5
114
How do we know compute is a commodity ? We track >500 GPU compute providers. Hyper-fragmentation is a hallmark sign of a commodity offering. @ComputeDesk is making compute tradable.
The number of GPU compute providers exploded from 3 to 500 and now buyers and sellers negotiate through private texts. One of the world’s most important commodities and buyers still have to call “their guy” for pricing.
3
1
28
2,864
Replying to @Decentralisedco
Great article, and we actively collaborate with our friends at @prlnet. Minor correction, however: Unlike the other indexes, Compute Desk indexes are based on executed bulk compute transactions. To take an oil analogy, our indexes are the only ones which measure the price of "crude oil" (big reserved compute contracts), instead of going to the gas station, and collecting the receipts from people who just filled up their petrol tank (on-demand).
4
321
Great article, and we actively collaborate with our friends at @prlnet. Minor correction, however: Unlike the other indexes, Compute Desk indexes are based on executed bulk compute transactions. To take an oil analogy, our indexes are the only ones which measure the price of "crude oil" (big reserved compute contracts), instead of going to the gas station, and collecting the receipts from people who just filled up their petrol tank (on-demand).
2
1
21
3,132
Inference companies are caught between: - a rock (pay a lot more for compute) - a hard place (commit to compute for a much longer duration). Compute derivatives offer a lot more flexibility. Get in touch to learn more about compute options, futures, forwards and swaps @ComputeDesk
Pay 2x for GPUs or commit for 3 years. That’s the compute squeeze facing fast-growing AI companies: pay up for speed or lock into capacity before you even know your long-term demand. @andrawesbahou
1
42
4,072