The frantic rush to power and stand up AI data centers has spurred a range of speed hacks—from building AI tent cities, to firing up jet engines, to taking over abandoned industrial sites. What’s most surprising is all of the different ways the biggest AI data centers have improvised and scrounged to meet the staggering electricity demands of AI.Here are some of the more resourceful hacks companies are using to scale up their computing power as fast as they can.Hack 1: Sweat your own real estate.In early 2023 when the AI data center boom was just getting started, both Google and Meta Platforms were sitting on valuable power rights in New Albany, Ohio, near Columbus. I wrote about New Albany and its high-voltage transmission network in my last column. Its electric arteries can deliver so much power that the area has become a new AI capital. Google and Meta’s main campuses are across the boulevard from each other.To take better advantage of the power available there, Meta expanded its AI data centers. To do that, it sent a specialist team of engineers into five of its traditional data centers to replace racks of chips such as central processing units with graphics processing units, which are used for AI. It created a huge cluster of AI computing power, “all in a matter of months!” the company said in a blog post last September. Meta also expanded by building an AI tent city on some of its existing New Albany land, sheltering precious computing and cooling inside temporary structures often used for sporting events and conventions. “A year before that, if somebody said, ‘We’re going to put a bunch of very expensive GPUs in a tent,’ that would have been kind of hard to imagine,” former Meta executive Chris Malone said in a talk last fall. He had moved by then to Open AI to head its data center build-out. Google has squeezed more computing power out of its chips by connecting multiple data centers in Ohio and a string of locations across Nebraska and Iowa with massive fiber optic cables. Effectively, this pooled hundreds of megawatts—likely adding up to gigawatts—of precious power into one thinking brain, which could keep training running even with individual chip failures.Hack 2: Acquire someone else’s power.Microsoft knew about AI’s insatiable appetite for power in 2022 because it provided training capacity for Open AI’s ChatGPT. So when it discovered that one of the largest electrical substations in the country had just been completed and was sitting unused, it jumped. A utility had built the substation in Mount Pleasant, Wis., for Taiwanese manufacturer Foxconn, which then reneged on plans for a giant LCD factory there.“Right now the answer to speed to power is to find somebody that had a power connection and hadn’t fully gotten their project off the ground,” said Landon Lill, energy regulatory law partner for the law firm Baker Botts, who arranges power for AI sites.Microsoft’s new data center, Fairwater, is now running over 350 MW in AI workloads with a path to reach or exceed 2 GW (1 to 2 GW is roughly the capacity required to power a medium-size city), according to SemiAnalysis. Microsoft built an identical data center near Atlanta and now directly connects the two with a new type of fast data connection.Hack 3: Tap green power projects orphaned by President Donald Trump’s energy policies.Data centers have been tapping power supplies developed for crypto miners for the last few years, but another set of power assets has gone untapped. Green hydrogen sparked excitement as a source of clean fuel for transportation and manufacturing. But it remains prohibitively expensive because the process used to make it requires large amounts of renewable energy. Once it’s made, it’s very difficult to transport. Then Trump curtailed tax credits for building it.Data center developers are now tapping sites where solar and wind farms have been built or received permits to serve green hydrogen projects. We already flagged such a hack when we explained why Google acquired renewable energy company Intersect. AI isn’t running yet on these sites, but it’s coming.A pivot from hydrogen also enabled Google’s new agreement with power producer AES, which investors including Blackrock are acquiring. AES will now build renewables to drive a Google supercomputer in Wilbarger County, Texas. The site was originally intended to produce green hydrogen with industrial gas provider AirProducts. Air Products walked away, but AES had teed up land for power construction and grid connections. That allowed it to take advantage of a rule change in Texas giving priority in the queue to data centers bringing in some of their own power. Hack 4: Dust off reliable machinery from another era.Elon Musk’s xAI, now part of SpaceX, is tapping refurbished General Electric aircraft engines that had powered Boeing 747s and 767s to generate electricity for its second Colossus data center project outside Memphis in Mississippi. It won permits this month to run them permanently, despite concerns about air pollution. Proenergy, which refurbishes decades-old aeroderivative engines for use on land by utilities or industrial sites, says 1,000 such engines will be retired in the next decade. Proenergy makes its own replacement parts because some GE patents have expired.Refurbishing old engines is a faster option than a flashy move by AI data center developer Crusoe to buy new jet engines. It signed on as the launch customer for supersonic jet maker Boom Supersonic, which pivoted to AI just months ago when it needed capital to build its engines. It will adapt its engine for land use in a new Superpower Superfactory in Denver. Crusoe has ordered 29 turbines for 1.21 GW of capacity. The idea is that when (and if) they’re ready in one to two years, Boom’s engines stand to produce more power in hot weather than older turbines designed for chilly altitudes. Hack 5: Do it yourself. Applied Digital, which is building huge data centers in North Dakota for CoreWeave and others, recently started its own new independent power producer, Base Electron. A few days ago, Base placed an order with longtime coal-power engineering firm Babcock & Wilcox to build gas power plants using boilers and steam turbine technology adapted from the coal era. Applied Digital and Babcock & Wilcox claim this option will be faster to market than the more typical simple-cycle and combined-cycle gas turbines, which are all on back order.Newsworthy:Emerald AI is announcing results of demonstrations with Nvidia today: Its workload-shifting software, which helps data centers flex power consumption in real time, was able to keep heavy AI workloads running while still reducing power use 20% or more at times of grid stress. The tests simulated plunging overnight temperatures and a multiday heat dome.KKR is eyeing a multibillion-dollar sale of data center cooling company Cool IT Systems.The Nuclear Regulatory Commission gave its first construction permit for a next-generation nuclear reactor to Bill Gates–backed startup TerraPower in Wyoming.BlackRock has launched a $100 million philanthropic fund for workforce development in skilled trades needed for AI and other infrastructure.Upcoming EventsMonday, April 27 — Financing the AI RevolutionJoin The Information at the New York Stock Exchange on Monday, April 27, to hear from top executives and investors on how the rapid buildout of AI is reshaping tech, finance, and capital marketsMore details
5 Ingenious Hacks That Are Boosting AI Data Centers
By Creator Economy
·
·
6 min read
·
598 views
Read in:
aa
ace
af
ak
alz
am
ar
as
awa
ay
az
ba
ban
be
bew
+191 more
bg
bho
bik
bm
bn
brx
bs
bug
ca
ceb
cgg
ckb
co
crh
cs
cv
cy
da
de
din
doi
dv
dyu
dz
ee
el
en
eo
es
et
eu
fa
ff
fi
fj
fo
fr
fur
fy
ga
gd
gl
gom
gn
gu
ha
haw
he
hi
hil
hne
hmn
hr
hrx
ht
hu
hy
id
ig
ilo
is
it
ja
jam
jv
ka
kab
kbp
kg
kha
kk
kl
km
kn
ko
kri
ku
ktu
ky
la
lb
lg
li
lij
ln
lo
lmo
lt
ltg
lua
luo
lus
lv
mai
mak
mg
mi
min
mk
ml
mn
mni-mtei
mos
mr
ms
mt
my
nd
ne
nl
nn
no
nr
nso
nus
ny
oc
om
or
pa
pag
pam
pap
pl
ps
pt
pt-br
qu
rn
ro
ru
rw
sa
sah
sat
sc
scn
sg
si
sk
sl
sm
sn
so
sq
sr
ss
st
su
sus
sv
sw
szl
ta
tcy
te
tg
th
ti
tiv
tk
tl
tn
to
tpi
tr
trp
ts
tt
tum
ty
udm
ug
uk
ur
uz
ve
vec
vi
war
wo
xh
yi
yo
yua
yue
zap
zh
zh-hk
zh-tw
zu
