1) Hypothesis: hyperscale cloud co’s alternate annually between spending on data centers (real estate, buildings) and compute/network/storage (CPUs, GPUs, Accelerators, Memory, HDD, etc.) to fill those data centers based on the timing of new semiconductor manufacturing nodes.
2) i.e. 2018 was a year of CPU/GPU/Accelerator/Memory spending based on the 10/12 nanometer node, 2019 was a year of data center (real estate) spending and 2020 will be another year of CPU/GPU/Accelerator/Memory spending based on the 7 nanometer node.
3) The CPU/GPU/Memory spend drags network (switches and fabrics) and HDD with it despite those not necessarily being on the same technological upgrade cycle as even though Moore’s Law has slowed down, nothing else in hardware technology is as powerful.
4) This would be the efficiency/NPV maximizing strategy for the cloud co’s and seems(ish) to fit the data going back several years. Caveat would be that crypto related shortages and associated inventory builds may have inflated 2018. Thoughts from Twitter on this hypothesis?
5) Datacenter revenue for HDD/CPU/GPU players who disclose this was down similar(ish) amounts 2018 peak to 2019 trough with HDD down the most (share loss to NAND) and GPU down the least (taking share of workloads).
6) Cloud computing companies have also begun to speak about capex being more focused on building data centers in 2019 after their capex was more focused on “servers” in 2018.
7) Also - completely understand that the cloud co’s are getting more efficient all the time and continuously driving utilization up which should *slowly* mute this pattern as they move towards planetary scale computing.
8) Offset to this is that traditional enterprise is not on this spending pattern to the same degree and as cloud takes share of workloads, the effect of this pattern should become steadily more pronounced on suppliers.
9) Smart friend who wishes to remain anonymous pointed out that Spectre/Meltdown accentuated this pattern in 2018 - had not thought of this but clearly correct
