Third-party expert interview w/ Chief Architect at $Microsof...
Third-party expert interview w/ Chief Architect at $Microsoft (MSFT.US)$ was very bullish on $NVIDIA (NVDA.US)$ demand.
NVDA trades at 30x street CY2025 GAAP EPS. Assuming street EPS is 10% too low, that means stock at 27x. Given only green lights on demand hard to see stock trading below 25x so not unreasonable here - although it felt safer buying this closer to $110 where it was ~25x street EPS.
No one knows what the contour of growth will be but if it ends up that NVDA's growth rate stabilizes ~10-12% qq, then both numbers and the multiple are going higher. Seems like some call option upside here since no real reason to question demand here.
I'm not smart enough to figure out all the derivative bets (data centers, racks, cooling, power) which will act like the tail end of a whip. Personally think that expanding investments down the stack to these components is a lot more dangerous than just moving up to the hyperscalers. To each their own.
Highlights
-Contact has been at MSFT for ~6 years focused on Azure infrastructure
-Contact has been at MSFT for ~6 years focused on Azure infrastructure
-Huge supply shortage in AI compute. Have had to negotiate partnerships w/ CoreWeave just to leverage their infra as well.
-Demand anecdote: Turned on newly-installed cluster at 10:30am and wondered when would reach 100% utilization. Went to lunch and by 1pm it was fully utilized.
-“Simply can’t build these clusters fast enough because of the demand that’s out there.” “Have not seen any slowdown in demand. If anything, the demand continues to increase.” [Not sure if that implies absolute demand increasing or rate of growth increasing]
-Have technology behind the scenes called SmartFabric that looks at customer workload and tries to help decompose data / vectors to see how it should be run most efficiently. Have customers that say they want H100 but will show them they can get that done in other ways to free up those GPUs for customer workloads that require the latest / greatest.
-Typical AI cluster is made of 1,024 nodes with each having 8 GPUs. So 8,192 GPUs that take up floors. But also never create 1 cluster, create 3 clusters in different areas of data center to maintain SLAs w/ customers – so if natural disaster happens would only take down 1 cluster.
-3 tiers of data centers: lowest tier provides ability for customers to hook into MSFT dark fiber and light it to get data into GPUs more quickly. Middle tier are edge data centers w/ Azure stack devices and bespoke workloads that can be run like HPC or AI at the edge. Top tier are largest data centers like one in Quincy that is 2.5 miles long.
-Plan data centers several years in advance – takes 18 months to get transformers and substations. Networking equipment RFPs go out 18 months ahead of time.
-GPUs consume ~2.5x more power than traditional CPU and generate 4-5x more heat. Power is 30% of the cost of a data center so focused on efficiency like cooling.
-Optics: Want GPUs to operate at 400G minimally and now getting into 800G and 1.6T speeds
Liquid Cooling
-Liquid cooling: Sounds like NVDA developing own liquid-cooling (?), use NVLink, other vendors and use a lot of our own proprietary designs as well.
-Liquid cooling: Sounds like NVDA developing own liquid-cooling (?), use NVLink, other vendors and use a lot of our own proprietary designs as well.
-NVDA said don’t need to liquid-cool H100 but then showed them 80% less failure rates. NVDA didn’t want us to do it bc more failure means buying more NVDA chips but this was better bc wouldn’t eat up NVDA customer support. They agreed.
-Lot of our designs on liquid cooling and power management went into Blackwell – which is why you can now buy it w/ liquid cooling in a form factor that is pre-racked.
-If I use NVDA liquid cooling then I go w/ their support. But if I go w/ liquid cooling by Delta or Boyd or CoolerMaster, they come up w/ their own thermal compounds.
-Have partnership w/ SpaceX and all data centers are connected so can use as backup
Disclaimer: Community is offered by Moomoo Technologies Inc. and is for educational purposes only.
Read more
Comment
Sign in to post a comment
Simon 5183 : Under the operation of AI big data stock trading software, these technical aspects are no longer effective!
Awuhsawj :
Awuhsawj :