00:00:00
And I actually think Michael Bur's got
00:00:01
this wrong. And so what Michael Bur is
00:00:03
saying is that all of these hyperscalers
00:00:05
have extended their depreciation
00:00:07
schedule or the useful life of their
00:00:10
data centers by roughly 2x which cuts
00:00:13
the operating costs in half when they
00:00:15
report it in earnings and so it's making
00:00:17
their earnings inflate. So he's claiming
00:00:18
they're cooking the books. Google first
00:00:20
made this change in Q1 of 21 where they
00:00:23
said the servers are now going from 3 to
00:00:25
four years. The chart shows the
00:00:26
combination of servers and networking
00:00:28
equipment separately. In 2021, Google
00:00:31
took networking equipment from 3 to 5
00:00:32
years and then in 2023 they took it from
00:00:34
5 to six years. And so this is a result
00:00:36
of this effort where they went in and
00:00:38
did an analysis. So what happened? What
00:00:40
happened in the data centers is that the
00:00:42
data centers transitioned from being
00:00:44
primarily data storage and data transfer
00:00:47
systems where you would use hard drives
00:00:49
and RAM and memory to store data and
00:00:52
then transmit it back out to being data
00:00:55
processing centers because of the AI
00:00:57
boom. So as AI became more important in
00:00:59
the data center, more of the dollars
00:01:01
that are going into data centers were
00:01:03
allocated towards chips from data
00:01:05
storage which initially was hard drives.
00:01:07
And then suddenly when you put these
00:01:08
processors in to process the data to do
00:01:10
AI, that's the majority of the spend and
00:01:13
the majority of the energy is going
00:01:14
towards the processors. I made some
00:01:16
calls and I checked around with some
00:01:18
other friends and everyone says the same
00:01:19
thing that these seven and 8year-old
00:01:21
TPUs and GPUs that are sitting in the
00:01:23
data centers are still being used and
00:01:25
they're being used at 100% utilization.
00:01:27
So that actually justifies and validates
00:01:29
the depreciation schedule being much
00:01:31
longer versus shorter.
