November 13, 2024 • 9 mins
Can data centers sustain their rapid growth without jeopardizing our energy future? In this episode, we promise to uncover the intriguing dynamics of power supply challenges faced by data centers and the innovative solutions needed to address exponentially increasing energy demands. As data centers become ever more essential, they transform into energy-hungry giants, especially during intensive AI training and inference processes. Our discussion highlights collaborations with nuclear power plants, like Amazon's link with Talon Susquehanna and Microsoft's efforts at Three Mile Island, along with groundbreaking alternatives such as modular nuclear plants and geothermal projects. These strategies reveal the intricate landscape of energy procurement, essential to maintain the relentless pace of tech-driven development.
Our exploration also ventures into the emerging trends shaping data center power consumption. Facing mounting pressure to boost computing power, data centers are increasingly reliant on diverse energy sources, including the controversial use of the gas grid, potentially leading to higher carbon footprints. We delve into the vital role of third-party independent power producers in enabling swift expansions, while also addressing the impact of AI workloads on energy prices and market competitiveness. Don't miss the insights on how the bulk power grid might be reshaped by these tech demands and what the future holds for this ever-evolving industry.
Our exploration also ventures into the emerging trends shaping data center power consumption. Facing mounting pressure to boost computing power, data centers are increasingly reliant on diverse energy sources, including the controversial use of the gas grid, potentially leading to higher carbon footprints. We delve into the vital role of third-party independent power producers in enabling swift expansions, while also addressing the impact of AI workloads on energy prices and market competitiveness. Don't miss the insights on how the bulk power grid might be reshaped by these tech demands and what the future holds for this ever-evolving industry.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:01):
In our first two sessions we looked at the
enormous projections for datacenter growth and the amount of
energy that may be required toserve that potential demand.
We then discussed issuesrelated to chips, power draw,
availability of data and otherissues affecting growth in
future power demand.
In today's piece we'll examinethe potential supply options
that data center owners andoperators are chasing.
(00:23):
Later we'll discussimplications for competitive
power markets as well asdistribution utilities, but
those topics are for another day.
So let's talk supply.
Of all the limiting factorsdetermining data center growth,
power is the big one.
A steady supply of electrons isnecessary to feed these
extraordinary machines and thepower-hungry racks of chips they
(00:44):
host.
Let's look at where that energygets used.
Today.
Most of it is used for trainingthe computers, crunching
through data, to makeconnections and create
intelligence.
Training and inference areestimated to consume about 70 to
80 percent of the power used Intraining.
Most of the energy is used inthe initial run model, which can
take weeks or longer.
20 to 30 percent of processingpower use is dedicated to
(01:07):
cooling, though highly efficientdata centers can get those
cooling numbers down into theteens.
The amount of energy used forinference, where those decisions
are made from the trainedmodels.
That's growing rapidly.
Inference doesn't have to beperformed by energy-thirsty GPUs
, but the usage adds up when youask a question to chat or
Perplexity or Gemini or a hostof other applications.
(01:30):
Those queries burn throughabout 10 times the energy it
would take to run a standardGoogle search.
As we increase our number ofqueries and I've performed at
least five just since I startedresearching this section then
the aggregated usage soarsstarted researching this section
.
Then the aggregated usage soars, meta, reported in a 2022 paper
that 30% of their energy usewas devoted to training,
(01:50):
including experimentationaffecting algorithms and
modeling approaches, whileinference for its own internal
Facebook use represented 70% oftotal electricity consumption.
So if data centers cannot escapethe insatiable hunger for power
, where will they get it?
Well, first, as we've alreadydiscussed, they'll grab as much
from the power grid as they can,wherever they can.
(02:11):
Globally, there's a preferencefor the US with its large grid,
stable economy, rules of law andaccess to communications cables
.
Europe would also fit the billin many ways, but the grid is
old and space is at a premium,so some data centers migrate to
Malaysia, singapore, brazilanywhere else they can get power
.
China's its own unique casewith geopolitical implications.
(02:35):
Indeed, as dominance in AIincreasingly becomes a national
security issue, geopoliticsincreasingly enters the fray.
Fortune reported in October,for example, that since the
release of chat GPT two yearsago, the US Department of
Defense has awarded $670 millionto over 300 companies working
on AI-related projects, and AIis already being used to shape
(03:00):
the Ukrainian battlefield.
Sticking with the US, then, thesupply strategies generally look
like this First, get as muchpower as quickly as one can
directly from the utilities andthe grid.
We'll talk more about that in acoming session.
Failing that, tie into existingassets and co-locate where
possible with them.
(03:20):
This is what Amazon WebServices did with Talon
Susquehanna Nuclear Plant, wherethey were able to access 300
megawatts of supply at aco-located data center before
the FERC gave them the Heismanand rejected plans for an
additional 660 megawatts.
It's also what Constellationand Microsoft are playing on
with their 20-year 835 megawattagreement to resuscitate Three
(03:43):
Mile Island Unit 1, shutteredsince 2019 for economic reasons.
Constellation may spend up to$1.6 billion on revitalizing the
plant and is seeking a federalloan guarantee to make it happen
.
Meanwhile, nextera recentlyannounced it's in talks with the
feds and conducting engineeringassessments as they eye
restarting the 600 megawattDwayne Arnold nuclear plant in
(04:06):
Iowa that was taken out ofservice in late 2020.
There are, however, only so manyexisting and recently closed
nukes to go around.
So what then?
Well, we may also see a fewsmaller facilities supplied by
fuel cells.
In May, bloom Energy anddisclosure I own some shares
announced it was significantlyexpanding on its original 6.5
(04:28):
megawatt supply agreement to anIntel data center in Silicon
Valley, though it gave nodetails on size.
In June, bloom announced a15-year 20-megawatt deal with
Amazon Web Services for aplanned Silicon Valley data
center and in its recentearnings call, bloom's CEO
teased more.
But really, who cares?
In the face of the tens ofthousands of megawatts being
(04:50):
sought by data centers, thesequantities are akin to a pimple
on a fruit fly in terms ofsignificance to anybody but
Bloom to advanced and enhancedgeothermal projects and nuclear.
Google kicked off the geothermalgame with a 3.5 megawatt deal
with developer Fervo Energy inNevada that delivered its first
(05:12):
juice in late 2023.
On the heels of that, in Juneof 2024, google committed to
buying 115 megawatts ofadditional 24-7 energy from
Fervo under a long-term contractwith Nevada utility Envy Energy
.
Energy from Fervo under along-term contract with Nevada
utility Envy Energy.
Then, in August, sageGeosystems stepped into the ring
(05:34):
, announcing a 150 megawattsupply agreement with Meta for
geothermal from a site quoteeast of the Rockies, unquote.
Other projects will undoubtedlyfollow as advanced geocompanies
get better at drilling, cutcosts and continue to improve
their technologies.
But these first large projectswon't come online for a few more
years and the industry isn'tgoing to get tens of gigawatts
there anytime soon.
So what about nukes?
(05:55):
Modular nuclear plants, thosesmall ones that can be built in
the factory, shipped andassembled on site.
Here too, we've seen numerousrecent announcements.
Small reactor company Okloinked an April deal with data
center co-location companyEquinix for 500 megawatts, with
start dates to be determined.
Google signed an October 2024deal with startup Kairos Power
(06:18):
for 500 megawatts of modularnukes to be online between 2030
and 2035.
Between 2030 and 2035.
Also in October, amazon WebServices announced it had made
an investment in modular reactorcompany X-Energy as well as
signed an agreement with a groupof Washington State Utilities,
energy Northwest, for 320megawatts to be furnished by
four 80-megawatt reactors, withan option to expand to 960
(06:42):
megawatts.
Eventually, delivery date isexpected to be in the early to
mid-2030s.
And that's the rub here themodular nuclear industry won't
scale and deliver largequantities to support AI data
centers anytime soon.
And, with a singular exceptionof NuScale, none of the numerous
players even have designapproval from the Nuclear
Regulatory Commission.
Nimbyism will undoubtedly alsobe an issue, limiting siting
(07:05):
opportunities, while wastesecurity and disposal will also
need to be addressed.
So where do we then stand?
You've got a data centerindustry desperate for power,
willing to pay nearly any pricefor it, and a remarkably
sclerotic grid.
If new technologies are to bethe cavalry riding to the rescue
, don't hold your breath.
The cavalry's horses haven'teven foaled yet.
(07:27):
The logical outcome, then, is tobypass the power grid entirely
and go right to the source thegas pipeline.
Gas is at historically lowprices and fracking continues
unabated.
Heck.
Much of the US gas productionis a byproduct of oil drilling.
In the prolific Permian Basin,a huge portion of gas production
is associated gas from wellsaimed at lucrative oil, not even
(07:51):
targeting gas.
Not surprisingly, data centeroperators are now looking at
building their own generation,co-located with data centers,
giving the term co-location awhole new meaning.
Numerous gas pipeline companiesreport discussions with data
center operators for direct gashookups to support
behind-the-meter generationwhere a grid would represent the
(08:12):
backup power source, forexample.
Energy transfer, is said, is indiscussions with data centers
for new demand in excess of 3billion cubic feet per day For
perspective 2023.
Us average daily consumptionwas 89 BCF per day.
The picture is still murky sinceits early days, but the trend
(08:33):
seems quite obvious.
It's likely that the datacenter crowd will get as much as
they can from the existing grid, while tapping into the gas
grid as quickly as possible,carbon emissions be damned.
It's also quite possible thatthird-party independent power
producers will take on thelion's share of this development
task.
But if the race is really allabout speed to power, the
(08:55):
urgency to quickly develop newcompute capabilities at huge
scale this is likely a hashtag.
Can't not happen.
Type of outcome.
All things being equal, new AIload, whether directly supplied
or dependent on the grid, willlikely have the effect of
raising prices.
In our next session, we'll talkabout the potential impacts on
competitive markets and the bulkpower grid.
(09:16):
Thanks for watching.
In our first two sessions we looked at the
enormous projections for datacenter growth and the amount of
energy that may be required toserve that potential demand.
We then discussed issuesrelated to chips, power draw,
availability of data and otherissues affecting growth in
future power demand.
In today's piece we'll examinethe potential supply options
that data center owners andoperators are chasing.
(00:23):
Later we'll discussimplications for competitive
power markets as well asdistribution utilities, but
those topics are for another day.
So let's talk supply.
Of all the limiting factorsdetermining data center growth,
power is the big one.
A steady supply of electrons isnecessary to feed these
extraordinary machines and thepower-hungry racks of chips they
(00:44):
host.
Let's look at where that energygets used.
Today.
Most of it is used for trainingthe computers, crunching
through data, to makeconnections and create
intelligence.
Training and inference areestimated to consume about 70 to
80 percent of the power used Intraining.
Most of the energy is used inthe initial run model, which can
take weeks or longer.
20 to 30 percent of processingpower use is dedicated to
(01:07):
cooling, though highly efficientdata centers can get those
cooling numbers down into theteens.
The amount of energy used forinference, where those decisions
are made from the trainedmodels.
That's growing rapidly.
Inference doesn't have to beperformed by energy-thirsty GPUs
, but the usage adds up when youask a question to chat or
Perplexity or Gemini or a hostof other applications.
(01:30):
Those queries burn throughabout 10 times the energy it
would take to run a standardGoogle search.
As we increase our number ofqueries and I've performed at
least five just since I startedresearching this section then
the aggregated usage soarsstarted researching this section
.
Then the aggregated usage soars, meta, reported in a 2022 paper
that 30% of their energy usewas devoted to training,
(01:50):
including experimentationaffecting algorithms and
modeling approaches, whileinference for its own internal
Facebook use represented 70% oftotal electricity consumption.
So if data centers cannot escapethe insatiable hunger for power
, where will they get it?
Well, first, as we've alreadydiscussed, they'll grab as much
from the power grid as they can,wherever they can.
(02:11):
Globally, there's a preferencefor the US with its large grid,
stable economy, rules of law andaccess to communications cables
.
Europe would also fit the billin many ways, but the grid is
old and space is at a premium,so some data centers migrate to
Malaysia, singapore, brazilanywhere else they can get power
.
China's its own unique casewith geopolitical implications.
(02:35):
Indeed, as dominance in AIincreasingly becomes a national
security issue, geopoliticsincreasingly enters the fray.
Fortune reported in October,for example, that since the
release of chat GPT two yearsago, the US Department of
Defense has awarded $670 millionto over 300 companies working
on AI-related projects, and AIis already being used to shape
(03:00):
the Ukrainian battlefield.
Sticking with the US, then, thesupply strategies generally look
like this First, get as muchpower as quickly as one can
directly from the utilities andthe grid.
We'll talk more about that in acoming session.
Failing that, tie into existingassets and co-locate where
possible with them.
(03:20):
This is what Amazon WebServices did with Talon
Susquehanna Nuclear Plant, wherethey were able to access 300
megawatts of supply at aco-located data center before
the FERC gave them the Heismanand rejected plans for an
additional 660 megawatts.
It's also what Constellationand Microsoft are playing on
with their 20-year 835 megawattagreement to resuscitate Three
(03:43):
Mile Island Unit 1, shutteredsince 2019 for economic reasons.
Constellation may spend up to$1.6 billion on revitalizing the
plant and is seeking a federalloan guarantee to make it happen
.
Meanwhile, nextera recentlyannounced it's in talks with the
feds and conducting engineeringassessments as they eye
restarting the 600 megawattDwayne Arnold nuclear plant in
(04:06):
Iowa that was taken out ofservice in late 2020.
There are, however, only so manyexisting and recently closed
nukes to go around.
So what then?
Well, we may also see a fewsmaller facilities supplied by
fuel cells.
In May, bloom Energy anddisclosure I own some shares
announced it was significantlyexpanding on its original 6.5
(04:28):
megawatt supply agreement to anIntel data center in Silicon
Valley, though it gave nodetails on size.
In June, bloom announced a15-year 20-megawatt deal with
Amazon Web Services for aplanned Silicon Valley data
center and in its recentearnings call, bloom's CEO
teased more.
But really, who cares?
In the face of the tens ofthousands of megawatts being
(04:50):
sought by data centers, thesequantities are akin to a pimple
on a fruit fly in terms ofsignificance to anybody but
Bloom to advanced and enhancedgeothermal projects and nuclear.
Google kicked off the geothermalgame with a 3.5 megawatt deal
with developer Fervo Energy inNevada that delivered its first
(05:12):
juice in late 2023.
On the heels of that, in Juneof 2024, google committed to
buying 115 megawatts ofadditional 24-7 energy from
Fervo under a long-term contractwith Nevada utility Envy Energy
.
Energy from Fervo under along-term contract with Nevada
utility Envy Energy.
Then, in August, sageGeosystems stepped into the ring
(05:34):
, announcing a 150 megawattsupply agreement with Meta for
geothermal from a site quoteeast of the Rockies, unquote.
Other projects will undoubtedlyfollow as advanced geocompanies
get better at drilling, cutcosts and continue to improve
their technologies.
But these first large projectswon't come online for a few more
years and the industry isn'tgoing to get tens of gigawatts
there anytime soon.
So what about nukes?
(05:55):
Modular nuclear plants, thosesmall ones that can be built in
the factory, shipped andassembled on site.
Here too, we've seen numerousrecent announcements.
Small reactor company Okloinked an April deal with data
center co-location companyEquinix for 500 megawatts, with
start dates to be determined.
Google signed an October 2024deal with startup Kairos Power
(06:18):
for 500 megawatts of modularnukes to be online between 2030
and 2035.
Between 2030 and 2035.
Also in October, amazon WebServices announced it had made
an investment in modular reactorcompany X-Energy as well as
signed an agreement with a groupof Washington State Utilities,
energy Northwest, for 320megawatts to be furnished by
four 80-megawatt reactors, withan option to expand to 960
(06:42):
megawatts.
Eventually, delivery date isexpected to be in the early to
mid-2030s.
And that's the rub here themodular nuclear industry won't
scale and deliver largequantities to support AI data
centers anytime soon.
And, with a singular exceptionof NuScale, none of the numerous
players even have designapproval from the Nuclear
Regulatory Commission.
Nimbyism will undoubtedly alsobe an issue, limiting siting
(07:05):
opportunities, while wastesecurity and disposal will also
need to be addressed.
So where do we then stand?
You've got a data centerindustry desperate for power,
willing to pay nearly any pricefor it, and a remarkably
sclerotic grid.
If new technologies are to bethe cavalry riding to the rescue
, don't hold your breath.
The cavalry's horses haven'teven foaled yet.
(07:27):
The logical outcome, then, is tobypass the power grid entirely
and go right to the source thegas pipeline.
Gas is at historically lowprices and fracking continues
unabated.
Heck.
Much of the US gas productionis a byproduct of oil drilling.
In the prolific Permian Basin,a huge portion of gas production
is associated gas from wellsaimed at lucrative oil, not even
(07:51):
targeting gas.
Not surprisingly, data centeroperators are now looking at
building their own generation,co-located with data centers,
giving the term co-location awhole new meaning.
Numerous gas pipeline companiesreport discussions with data
center operators for direct gashookups to support
behind-the-meter generationwhere a grid would represent the
(08:12):
backup power source, forexample.
Energy transfer, is said, is indiscussions with data centers
for new demand in excess of 3billion cubic feet per day For
perspective 2023.
Us average daily consumptionwas 89 BCF per day.
The picture is still murky sinceits early days, but the trend
(08:33):
seems quite obvious.
It's likely that the datacenter crowd will get as much as
they can from the existing grid, while tapping into the gas
grid as quickly as possible,carbon emissions be damned.
It's also quite possible thatthird-party independent power
producers will take on thelion's share of this development
task.
But if the race is really allabout speed to power, the
(08:55):
urgency to quickly develop newcompute capabilities at huge
scale this is likely a hashtag.
Can't not happen.
Type of outcome.
All things being equal, new AIload, whether directly supplied
or dependent on the grid, willlikely have the effect of
raising prices.
In our next session, we'll talkabout the potential impacts on
competitive markets and the bulkpower grid.
(09:16):
Thanks for watching.
Popular Podcasts
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Therapy Gecko
An unlicensed lizard psychologist travels the universe talking to strangers about absolutely nothing. TO CALL THE GECKO: follow me on https://www.twitch.tv/lyleforever to get a notification for when I am taking calls. I am usually live Mondays, Wednesdays, and Fridays but lately a lot of other times too. I am a gecko.