00:00:00
I spent some time recently reading
00:00:01
through a big spreadsheet of questions
00:00:04
submitted by vox's audience members and
00:00:07
one of them caught my eye it was from
00:00:10
Cathy a retired school teacher in New
00:00:12
York City what is the question that you
00:00:14
wanted us to answer so the question is
00:00:16
can green energy even begin to handle
00:00:18
the increased demands that Ai and crypto
00:00:21
and cloud storage are going to put on
00:00:23
our Energy System it's a good question
00:00:26
I've done some reporting on AI but I've
00:00:28
never thought much of the climate impact
00:00:30
of all the AI products we're
00:00:31
increasingly using and all of our
00:00:34
digital belongings like photos and
00:00:36
documents and emails getting stockpiled
00:00:39
in servers around the world they need a
00:00:41
lot of electricity and the electricity
00:00:42
has to come from somewhere this is all
00:00:44
happening while the climate crisis
00:00:46
demands we use less energy not invent
00:00:48
new ways to use more of it I think our
00:00:51
climate goals already feel pretty
00:00:53
impossible to me but now it's almost
00:00:56
like we haven't changed the goal poost
00:00:59
we've changed the entire game so let's
00:01:01
get to the bottom of
00:01:02
[Music]
00:01:06
this within Cathy's big question is a
00:01:09
more basic one about how much
00:01:11
electricity our digital lives require at
00:01:14
first I was thrown off that Kathy
00:01:15
mentioned things like cloud storage and
00:01:17
Ai and cryptocurrency in one category
00:01:20
but then I realized that their
00:01:22
electricity demands happen at the same
00:01:24
place data centers ultimately you're
00:01:27
talking about machines loaded up in
00:01:29
large facilities who generate
00:01:31
computations and they need power a lot
00:01:33
of it they need water uh they need space
00:01:36
I spoke to Alex D he runs a research
00:01:39
site called dig Economist where he's dug
00:01:41
into this exact topic data centers are
00:01:44
massive often windowless warehouses that
00:01:46
house thousands of servers that run
00:01:48
virtually non-stop some of the bigger
00:01:50
data centers are as big as four football
00:01:52
fields and use as much electricity at
00:01:55
any given time as 880,000
00:01:58
households there are more than 8,000
00:02:00
data centers around the world and the US
00:02:03
has more than any other country in 2022
00:02:07
data centers artificial intelligence and
00:02:09
cryptocurrencies made up about 2% of
00:02:11
total Global electricity demand but by
00:02:15
2026 that number is expected to double
00:02:18
which is like adding the amount of
00:02:19
electricity used by the entire country
00:02:21
of Sweden I'll explain why in a
00:02:24
minute all right support for this video
00:02:28
comes from clavo clavo works with
00:02:30
businesses to turn their data into
00:02:32
meaningful connections with their
00:02:34
customers through AI powered email text
00:02:36
messages and more according to clavio
00:02:39
over 150,000 Brands trust their data and
00:02:42
marketing platform to build smarter
00:02:44
digital relationships with their
00:02:46
customers during the holiday season and
00:02:48
Beyond clavio has no editorial influence
00:02:51
over our work but they make videos like
00:02:53
this possible learn more at the link
00:02:56
below that big jump from 2022 to 2026
00:02:59
six is thanks to Rising cloud storage
00:03:02
and cryptocurrency electricity demands
00:03:04
but it's also because of the AI boom We
00:03:07
Know AI requires a ton of computational
00:03:09
power but it turns out that the amount
00:03:11
of electricity it uses is a really
00:03:14
difficult question to answer AI is a
00:03:16
huge umbrella term that includes
00:03:18
everything from basic statistical models
00:03:20
that detect patterns and data to
00:03:22
generative AI that creates text and
00:03:25
images and videos that's the most
00:03:27
computationally intensive kind the thing
00:03:29
is the handful of private tech companies
00:03:31
that dominate the AI field don't really
00:03:34
disclose how much of their energy use is
00:03:36
dedicated to AI specifically if you look
00:03:38
at Google's latest environmental report
00:03:40
it clearly states they absolutely don't
00:03:44
want to make a distinction between
00:03:45
regular workloads and AI specific
00:03:47
workloads and these company's AI models
00:03:49
are mostly closed Source meaning no one
00:03:52
knows exactly how they are built this
00:03:54
has left some researchers to try to
00:03:56
piece it together on their own
00:03:57
researchers looked at an open source
00:03:59
large language model called Bloom that
00:04:01
has roughly the same amount of
00:04:03
parameters as
00:04:04
gpt3 and found that training something
00:04:06
like gpt3 required almost 1300 megawatt
00:04:09
hours of electricity about as much power
00:04:12
as consumed by 130 homes in the US for 1
00:04:16
year today large language models like
00:04:19
GPT 4 have hundreds of billions of
00:04:21
parameters if not a trillion and
00:04:23
researchers say that the computational
00:04:25
power required to train these models is
00:04:27
expected to double every 9 months so far
00:04:30
it has mostly been large language models
00:04:32
driving the AI energy boom of course
00:04:36
that could change going forward now we
00:04:37
see AI on the rise for image generation
00:04:41
and also specifically video generation
00:04:43
so far we talked about training a large
00:04:45
language model researchers also looked
00:04:48
at energy use from people actually using
00:04:50
it it's been estimated by myself and
00:04:53
others that a single jet GPT interaction
00:04:56
would take like 3 Watt hours which is
00:04:59
comparable to running a low Loom and LED
00:05:01
bulb for one hour so on itself it
00:05:04
doesn't sound like a whole lot but of
00:05:06
course hey it's the volume that matters
00:05:08
this is 10 times more than a standard
00:05:10
Google search and of course if you're
00:05:12
talking about millions or billions of
00:05:14
interactions the numbers start to stack
00:05:16
up quickly Alex took another research
00:05:19
approach by looking at the hardware used
00:05:21
for AI training and use over 95% of the
00:05:25
AI industry uses servers made by the
00:05:27
company NVIDIA they could sell one 1.5
00:05:30
million of their servers by 2027 he
00:05:33
multiplied that by the information
00:05:34
Nvidia publicizes about each of their
00:05:36
servers energy demand he found that data
00:05:40
centers devoted to AI alone could
00:05:42
consume around 100 terawatt hours of
00:05:44
electricity per year or about the same
00:05:47
as his home country of the
00:05:50
Netherlands there's a big part of
00:05:52
Kathy's question I haven't gotten to yet
00:05:54
can renewable energy meet the surging
00:05:56
demand from the world's data centers
00:06:00
the good news is that using green energy
00:06:02
is the stated goal of a lot of these
00:06:04
companies both Google and Microsoft have
00:06:06
made pledges to be Net Zero by 2030 but
00:06:10
there are signs that AI is disrupting
00:06:12
those plants that's because solar and
00:06:14
wind energy can't produce electricity
00:06:16
all of the time and these data centers
00:06:18
need to be running all of the time in
00:06:21
most cases they will just have a backup
00:06:22
connection to the power grid which will
00:06:24
uh have fossil fuels on it it's not just
00:06:27
that data centers are being built at a r
00:06:29
that renewable energy infrastructure
00:06:31
can't keep up with it can take a year to
00:06:34
build a data center but many more years
00:06:36
to get a solar or wind farm on an
00:06:39
electrical grid Google's 2024
00:06:41
sustainability report showed that the
00:06:43
company's emissions Rose by 48% from
00:06:46
2019 to 2023 in large part due to its
00:06:49
data center energy consumption
00:06:51
suggesting that integrating AI into
00:06:53
their products can make reducing their
00:06:55
emissions challenging there's already
00:06:57
evidence in the US that coal plants that
00:06:59
were meant to close are staying open
00:07:01
because of data centers electricity
00:07:03
demands and that state utilities are
00:07:05
building new natural gas plants for the
00:07:07
same reason but even if these tech
00:07:09
companies can look good on their
00:07:10
sustainability reports and get to Net
00:07:12
Zero they's still a problem the thing is
00:07:15
that our renewable energy Supply
00:07:16
globally is limited so if we attributing
00:07:19
an increasing part of that to the data
00:07:21
center industry the consequences that
00:07:24
there's less Renewables available for
00:07:26
everything else that probably will mean
00:07:28
that on the whole we will end up using
00:07:31
more fossil fuels anyway with all this
00:07:34
context the answer to Cathy's question
00:07:36
is that for right now we aren't prepared
00:07:39
for renewable energy to meet the
00:07:41
increasing demand of the world's data
00:07:43
centers so what do we do about this as
00:07:47
users it would be extremely difficult to
00:07:50
opt out of backing up our data on the
00:07:52
cloud or even refrain from using AI I
00:07:56
think AI is embedded in so many things
00:07:58
that I'm not sure I will have the option
00:08:00
to say I'm not using it you know I'm out
00:08:03
researchers like Alex say the best place
00:08:05
to start is to force more transparency
00:08:08
from these tech companies in the EU the
00:08:10
AI act doesn't really Force tech
00:08:13
companies to disclose anything with
00:08:15
regard to the environment and that's the
00:08:17
EU not even talking about the us yet
00:08:20
which is lagging behind a bit on this
00:08:22
matter some environmental organizations
00:08:24
and local communities are calling for
00:08:26
moratoriums on data centers and some
00:08:29
researchers have proposed the idea of an
00:08:31
Energy Efficiency rating so companies
00:08:33
and consumers can choose data centers
00:08:35
that are the most sustainable we could
00:08:37
also hope that the servers and data
00:08:39
centers will keep getting more energy
00:08:41
efficient but more than anything this
00:08:44
issue emphasizes how desperately we need
00:08:46
to be scaling up renewable energy and
00:08:49
fast not only to meet the ever
00:08:51
increasing data center demands but so
00:08:53
there's plenty of renewable energy to go
00:08:57
around if you like this video you'll
00:08:59
love vox's new podcast explain it to me
00:09:02
I'm the host JQ and every week I call up
00:09:05
a listener with a question get them some
00:09:07
answers and we have some fun along the
00:09:09
way you can find a link to the podcast
00:09:11
in the description while you're there if
00:09:13
you want to support Vox you can check
00:09:15
out the details on our membership
00:09:16
program