/biz/ - Business & Finance

>>19566664
>>19566821
>>19566862
>... While we have sympathy for this concern (especially for NAND which we view as more elastic), we would make several counter-points:
>1) DRAM is more inelastic than elastic, especially in the cloud. The primary function of DRAM is to keep a processor fed – when processors are mostly contained in PCs, tablets, handsets where applications are “good enough” and utilization are low – i.e., processors are not eating often. DRAM content has a low ceiling; however, when you move processing power to the cloud, hyperscalers’ primary economic driver is utilization – the more highly they can utilize the datacenter, the more profitable they will be. High utilization means processors want to eat all the time – increasing DRAM becomes a point of leverage at any price.
>2) We have yet to see the impact of AI/ML workloads on DRAM demand – these workloads are extremely memory/DRAM intensive. There is no cognition without memory. The DGX2, NVDA’s ML server – has a maximum configuration of 1.5 TB of DRAM versus the average server of just 50 MB – while the DGX2 has ~1/2 the processing power of the human brain (btw the DGX2 needs 3200 watts, the brain 32 watts – ie biology is still multiple orders of magnitude more efficient), the human brain still has been 100-300x the memory capacity of the DGX2. Most all of the private start-ups in semis are focused on solving the memory wall.
>3) NAND is clearly driven by elasticity and a longer-term flattening of the cost curve likely put downward pressure on LT NAND demand assumptions of 30-40% -- however, we see a significant upgrade cycle coming in handsets in 2H20/2021 – as 5G phones become more readily available. Data would suggest that consumers are delaying their upgrades and that the installed base of handsets is 12-18 months older than it was just a few years ago. Even without a killer app we see 5G as a catalyst for consumer to upgrade.