xAI powers a data center with eight jet engines

Welcome to The Signal from Superpower Daily with Maya and Theo

Today we're looking at how the AI industry's desperate thirst for power is leading one company to literally strap jet engines to its data center

Right

We are really witnessing a massive collision right now

Yeah

The digital ambitions of artificial intelligence are just slamming into the physical limits of our electrical grid

Because you know every new capability requires more compute And

every

new GPU cluster requires megawatts of power The

legacy infrastructure simply cannot keep up with this demand Which

brings us to our lead story today XAI

installs eight converted jet engines to power its Mississippi data center amid an emissions challenge Yeah

this

perfectly illustrates that infrastructure limit we were just talking about It

really does I

mean the company essentially bypassed the traditional utility grid

entirely They sourced eight unused General Electric aircraft

engines Just regular jet

engines

Basically

yeah These are the exact turbines typically attached to Boeing passenger

planes And they just install them in a field in Mississippi to run their

servers Oh Well these are actually known in the energy industry as aeroderivative gas turbines

OK

There is a Missouri based company called Pro Energy that handles

the conversion So they modify them for

the

ground Right They remove the massive bypass fan you normally see on the front of a

commercial jet Then they take the core jet engine and connect the central drive shaft directly to

an electrical generator Oh OK

I see So the thrust is no longer pushing a

plane forward It spins a magnetic rotor to generate

electricity instead It is honestly a brilliant piece of engineering brute force I mean you bypass the local utility grid because the local grid is just

too

slow Exactly But it is also a bit like you

know a homeowner bypassing a slow electrician by buying a literal locomotive engine to power

their toaster That is a

fair comparison But for founders

and investors the speed of the traditional grid is a

massive bottleneck The queue to connect a new data center to regional transmission systems can take four to

five

years Wow

Five

years Yeah And AI development cycles operate on an

18 month cadence You simply cannot wait five years for

a substation upgrade So that timeline mismatch is the

core

driver here Exactly That is

the whole strategy To run a frontier

model training cluster you need a staggering

amount of power Each of these converted GE turbines has a maximum generating capacity

of 50 megawatts Let us put that 50 megawatt number into

perspective for you The average home uses a

tiny fraction of a

single

megawatt According to hardware

analysts one of these turbines running at maximum capacity equals the electricity draw of roughly 15

000

homes Yeah And this installation in Mississippi utilizes eight of

these turbines Eight

of

them Yes So that brings the maximum theoretical capacity to

400 megawatts That

is insane It is equivalent to powering

120 000 homes It is basically a midsize American city dedicated entirely to

processing matrix multiplications But we do need to look closely at how these

machines actually operate because that 400 megawatt figure is the

maximum

ceiling

right Correct They do not run at a flat 100 capacity all

day

long No not

at all Data center computational loads are

highly

spiky When a massive training run

begins the GPUs demand a massive surge of

current And when the run

finishes the power drop plummets almost

instantly Which makes

sense And that is precisely why aero derivative turbines are used

here instead of say traditional combined cycle power plants

Because they can react faster

Exactly A traditional natural gas power plant takes hours to heat up and reach full generation capacity

These aircraft engines can ramp from zero to full output in under 10 minutes

Wow 10

minutes Yeah

They

are literally designed to react instantly to a throttle So

they handle the volatile power demands of an AI data center perfectly But

I

mean combusting fuel at that scale creates immense friction with local environments It

absolutely does The

deployment currently holds operational permits from state environmental authorities The

state regulators approve the math basically

They

give XAI the green light to run these engines Right

But

holding a permit does not guarantee immunity from legal

challenges Obviously not The

Southern Environmental

Law Center is actively contesting those state approvals right now What are they

claiming exactly They allege the

initial math provided by XAI was fundamentally flawed They claim the

company failed to guarantee the turbines will meet established emission standards for specific pollutants And the pollutants

in question are pretty severe They're pointing specifically

to formaldehyde carbon monoxide and volatile organic

compounds Yes And volatile

organic

compounds are the chemical precursors to smog Right And these

compounds

are just unavoidable byproducts of high temperature combustion Okay When you

burn

fuel in a jet turbine you are Obviously

aiming for

complete combustion perfect combustion yields

only water vapor and carbon dioxide Which sounds fine

But real world

physics prevents perfect combustion Variations in

fuel chemistry and temperature create unburned hydrocarbons and those

form the

volatile organic compounds They also form carbon

monoxide And formaldehyde is actually

a direct chemical product of this incomplete burning process And that is a

known human carcinogen Exactly So the legal

challenge

really hinges on predictive modeling here Because when you apply

for an environmental permit you do not build

the plant first and then measure the exhaust No you definitely do

not You submit complex air

dispersion models You calculate the projected

emissions based on the manufacturer's specifications and the anticipated runtime Right So the Southern

Environmental

Law Center is alleging XAI just undercounted these projections And we must be

very precise about the nature of this dispute These are allegations contained

within a formal legal challenge They are not administrative

findings of a violation That is an important

distinction Yes No regulatory agency

has

placed a physical sensor on the exhaust stacks and recorded illegal levels of formaldehyde The dispute centers entirely on the mathematical projections used to secure that

initial state permit And this is completely distinct

from previous legal friction at this specific facility right It is Because the

Southern

Environmental Law

Center previously sued XAI That earlier lawsuit alleged the

company was operating temporary gas generators without any permits whatsoever And this new challenge actually

acknowledges

the permits exist It just attacks the foundational data

used to grant them It is wild This really represents

the frontier of

AI infrastructure policy I mean you can iterate a software model overnight but you cannot iterate the atmospheric

impact of a 400 megawatt power plant No you cannot The physical laws

of thermodynamics and

the legal frameworks of environmental protection they move at a fixed speed So what we need to

watch is whether the specific emissions assurances can withstand the legal challenge If the state permits hold

every

major AI company is going

to look at private aeroderivative turbines as a viable shortcut around grid delays But if the challenge succeeds the ultimate

bottleneck for AI scaling remains

local environmental law The physical infrastructure constraint is really forcing

companies to

innovate rapidly Next up OpenAI says two employees used AI to

design a

hardware prototype in under a Yeah this is a fascinating shift We just looked at

the immense physical footprint required to

run these models Now we look at the models attempting to build physical

objects themselves Right Two OpenAI employees ran a side experiment They used

the

GPT 6 Astra model to design a physical device And what was the device supposed to

do It was

engineered to let a user speak directly to

the Codex coding agent Okay The timeline of less than one week is definitely

notable

but the true breakthrough here is the scope of the

engineering attempt Yeah because building hardware is not simple Not at all

It requires navigating a deeply fragmented stack

of disciplines And

the employees actually tasked Astra with managing all four

primary layers of the stack simultaneously Wow So

the first layer is industrial design right That

is

the external appearance and the aesthetic interaction

And the second is mechanical design This involves engineering the

physical

enclosure and the structural integrity of

the chassis itself And then the third layer is electronics engineering Astra had to

lay out the actual printed circuit board Yes

It had to map the electrical pathways between the processor

the memory and the input sensors That is incredibly complex It

is

And the

fourth layer is software

The model had to

write the firmware to allow all those physical

components to communicate with the cloud based Codex agent It kind of reminds me of a baking show like when

a contestant draws a beautiful concept of a cake

on paper but then they forget the flour Right It looks great on

screen but does it actually exist

in

the physical world Because generating

a beautiful 3D rendering of a smart speaker is basically trivial for modern AI now It is But

building a manufacturable object requires obeying the laws of physics

Exactly Astra used specific design references from gaming aesthetics and also from OpenAI's

previous

Work Louder hardware collaboration It attempted to actually translate those visual references into a functional mechanical

assembly And OpenAI is really pushing aggressively into this space They reported that

AI usage across their internal hardware engineering teams

has increased 23 fold since February That is a massive jump But we have to analyze the metric

they used for that claim

Right They measured that 23 fold increase in tokens Okay wait A

token is just a fragment of a word

or a

line of code right Exactly It is a measurement of compute consumption It is

not

a

measurement of engineering output Ah That distinction is vital for operators assessing these tools A

surge

in token usage It just means engineers are heavily querying the model They are

generating countless iterations But it does not

definitively prove they are finalizing products 23 times faster

No

it just indicates massive experimentation

not necessarily massive efficiency OpenAI did release some internal benchmark data to quantify

this

capability though They tested a

newer internal model against

the BenchCAD 2 0 standard and BenchCAD is basically an evaluation

suite

for computer

aided design tasks The new model scored 0 875 The

GPT

6 ASTRA model they used for the one week prototype

sprint scored

0 594 And we need

to clarify what those scores mean A high BenchCAD score measures geometric accuracy

It evaluates whether the AI placed a digital component

in the correct spatial location based on

a

reference design It confirms the digital wires connect the correct digital pins But it does not evaluate physical manufacturability

It completely ignores that And that is the digital trap

right there A computer aided design environment

is

basically a frictionless vacuum

Real world manufacturing is governed by heat friction

and material properties And this is precisely where that

one week sprint failed Wait it failed

Well ASTRA generated

the

blueprint but it completely missed the manufacturing tolerances required

to actually build

it

Oh wow Yeah Tolerance

is the permissible limit of variation in a physical dimension Right When

you inject

molten plastic into a steel mold the plastic shrinks by a specific

percentage as it cools Of course If you design

a mechanical snap fit joint digitally without accounting for the exact thermal shrinkage rate of that specific polymer the pieces will not snap together They will just shatter or

they will be

way too loose Exactly And ASTRA

also failed to comprehend production limits and assembly details What do

you

mean

It did not factor in how a human hand or a

robotic arm at a

contract manufacturing facility would physically access the screws to assemble the chassis Oh I see The geometry was perfectly correct on the screen but the physical realization

of

it was

impossible So the experiment did not actually yield

a factory ready design No it did not

Human mechanical engineers had to intervene Right They conducted several

rounds of engineering review

they manually corrected all the tolerances

and

they rewrote the assembly logistics Okay so

once the human engineers fixed the

AI blueprints a contract manufacturer did

successfully produce a small back to the devices And that

physical manufacturing phase

took less than a month Right So the AI definitely accelerated the initial conceptual phase but humans were strictly

required

to bridge that gap into physical reality Exactly Operators

must watch the evolution of these models carefully The true inflection point

is

not generating CAD geometry It is whether future AI

models can learn to inherently calculate real world manufacturing

constraints They have to natively understand things like thermal dissipation and material shrinkage And

automated assembly logistics Without human intervention That structural understanding is really what

separates a digital toy

from industrial

software But you

know that failure to understand real world constraints extends way beyond physical manufacturing

It also applies to digital ecosystems Meanwhile an AI fan fiction site has shut

down after author backlash over scraped

stories

Yeah this is a big one The Archive

of Our Own is really a

foundational pillar of internet culture For sure The platform is universally known as AO3 It is a completely non

commercial repository

hosting over 18 million fan fiction works That is massive It is And it operates entirely on fan donations

and volunteer labor

But on October 29th the Archive experienced a significant security incident Right The administrators actually had to take the servers offline

to manage the

breach

They communicated very clearly that no user data was lost Yeah The site

was simply inaccessible during the remediation process But that downtime

created

an information vacuum And an independent developer named Ganesh

exploited that vacuum

instantly He really did

He launched a clone site called AO3 life And this

was a highly

sophisticated automated extraction operation He deployed AI scraping agents across 40 separate virtual machines Wait 40

virtual machines We should really unpack the mechanics of a

distributed scrape like that

Because if you just hit a server with 10 000 requests from

one single

IP address the firewall is going to block you immediately Exactly But by distributing the

AI agents across 40 virtual machines you rotate

the

IP addresses constantly Right You basically bypass standard rate limiting

So you can pull text from

web

caches search engine indices and scattered

online drives all at once But the operation

did not stop

at just text extraction The

site

then utilized AI voice synthesis to convert those scraped stories into audiobooks And the developer then listed these unauthorized audiobooks for sale He basically commercialized

a

massive corpus of stolen non commercial writing Well the site existed

for less than a week It shut down completely on

November 2nd Wow that was fast It was forced

by an overwhelming coordinated response from the

writing community The creator

received over 2 000 formal takedown requests in a single 24 hour

period That is incredible But the creator actually attempted to legitimize the platform

with two distinct

arguments right He did First he pitched the site as a preservation tool

The

homepage explicitly

claimed

it was to quote save AO3 before it

slips away Which is demonstrably false Entirely

false

The core archive was never in

any danger of deletion

It was

just undergoing a routine security pause And the second argument he used was financial He offered authors a

91 profit share on the sales of their

AI narrated audiobooks As if that makes it okay I mean offering a retroactive profit share does not magically

equal legal or ethical consent No

it doesn't

It is like taking someone's car renting it out to strangers and then

saying hey don't

worry I left most of the

rental money in the glove

box That is

exactly

what it

is You cannot appropriate someone's intellectual property package it into a

commercial product and then legitimize the

theft by offering them a majority

stake in the illicit revenue But

despite all the ethical

violations the site generated massive instantaneous traffic It really did In its

first 24 hours the clone site

attracted over 126 000 unique users

It consumed terabytes

of bandwidth Well that massive traffic

highlights the sheer demand for the content But it also highlights the ultimate

failure

of his business model How

so Because out of those 126 000 unique visitors the platform

sold exactly 11 audiobooks Wait 11 11 The conversion rate was

practically zero

The operation generated immense infrastructure costs and

total community outrage all for completely negligible

revenue

Wow

But

you know the deepest breach of

trust here actually involved the platform's security mechanisms Right The

archive locks Yes AO3 allows

authors

to

utilize an archive lock This setting restricts access to a story Only registered logged

in users of the

archive

can read a locked work So it is

specifically designed to prevent search engine indexing

and casual

scraping Exactly But the distributed nature of this specific scraping

operation completely bypassed those digital locks Because the AI agents pulled from disparate

caches and secondary sources Right So deeply restricted stories were exposed

and monetized on the public

clone site anyway The perimeter security failed entirely And the verification process

to

remove these stolen works was fundamentally broken from the start It really was The

clone site demanded authors prove

their identity before a story would be taken down Which sounds reasonable until you look

at how they

wanted you to do it The process required the author to log into their real

AO3 account They had to leave a specific verification code on their public profile

page Then the clone site would scan the profile verify the code

and process the takedown But that created an impossible technical paradox Exactly Because

the

real AO3 was offline Right Authors could not

log in They

could not edit their

profiles So they were completely blocked from

utilizing

the only authorized mechanism to reclaim

their own stolen intellectual property It is a nightmare scenario for

independent creators You really must watch how community platforms architect their verification systems moving forward Creators need decentralized

cryptographic proofs of ownership They cannot rely solely on the uptime of a centralized host to prove they wrote a document

Because

when a primary

archive goes dark automated agents will exploit the vulnerability immediately Absolutely And this aggressive exploitation

of infrastructure vulnerabilities scales all the

way up to the enterprise level In

other news Google has detailed stolen AI

access

schemes that ran up 1 million in corporate charges So we're moving from the theft of

digital text directly to the theft of raw computing power Exactly Google's Threat Intelligence Group collaborated with Mandiant

to publish a really deep architectural analysis of a new threat vector They define this vector as LLM

jacking

LLM jacking is the targeted hijacking of large language model infrastructure Attackers are not stealing data from enterprise cloud

environments No they are stealing the cloud

credentials and the API keys Exactly They leverage these stolen credentials

to

access frontier AI models on the victim's dime This

is not some smash and grab operation This is a highly

mature

underground economy The attackers do not consume the stolen compute themselves They package the

compromised API keys and they integrate

them into illicit routing networks The Google

report identifies these networks as LLM mixers And these

mixers

operate exactly like legitimate software as

a service platforms He's wild The report detailed

one specific underground service called Jetknight Okay so Jetknight

offered a subscription model on the dark web They charged users approximately

60 a

month Which is pretty cheap Right And for that

flat

fee subscribers received unlimited access to premium frontier models Unlimited Yeah

And Jetknight fulfilled that unlimited access by

secretly routing the subscriber

prompts through the stolen enterprise API

keys The computational volume of these attacks is

just

staggering

Google telemetry tracked one specific threat actor over an eight hour window Just eight hours Yes That single actor

executed over 500 000 distinct model calls using stolen

credentials Half a million API calls in a standard workday That

translates to catastrophic

financial damage It really does Because cloud providers bill AI usage by the

token Generating massive volumes of premium tokens incurs massive

costs Google actually documented victims facing

over 100 000

in unauthorized API charges in a

single day And

one specific corporate victim saw fraudulent billing approach a

million dollars We really need to examine the precise mechanism of this compromise The attackers

exploit

specific vulnerabilities in the middleware that companies use to manage their AI

deployments Right The report highlights a critical vulnerability designated as CVE

2026 40217 Okay And this vulnerability targets a popular routing gateway

called LightLLM Yes Developers use LightLLM

as

a proxy and allows an application to switch seamlessly

between

different AI providers without having to rewrite the core application code

Exactly

But all the sensitive API keys sit right behind this gateway So the vulnerability permits an attack known as user ID rebinding A hacker obtains a

low

level highly restricted virtual key Okay Then they manipulate the authentication

header sent to the LightLLM gateway The gateway misinterprets the manipulated

header And it binds the low level user ID

to

a root administrative profile Precisely The attacker escalates their privileges instantly They

transition from a restricted user to

a master administrator Wow And now they can access every premium API

key

stored in the company's vault They

export those keys directly to their LLM mixer But wait

security

teams usually rely on volume based alerts to detect

this kind of exfiltration right

They do Like if an API key suddenly generates

10 000

requests a minute web application firewalls trigger an alarm Normally yes But the attackers

bypass this

entirely using a single token stealth tactic This is honestly the

most dangerous mechanical detail of the

whole operation

When an attacker steals a batch of API keys they have to

validate them first Right But instead of running a standard query that

might trigger a billing alert they just request a

single output token A single token

represents

just a fraction of a word Exactly So the computational cost is microscopic

It registers as basic background noise to

standard monitoring tools A web application firewall would completely ignore

a

single token request Yeah But that single ping confirms the key is active Then the attacker

just quietly stages the validated key in their routing network And

the

breach often occurs in environments where AI is not

even actively utilized Yeah In several documented cases the victim companies were traditional

cloud computing customers They'd never even enabled machine learning services Oh wow So the

attackers

used

the stolen administrative credentials

to navigate into the cloud console Right And

they manually activated the AI service tier themselves

The victim's

IT department had zero monitoring configured for AI APIs because they did not even

know

the APIs were turned on That is terrifying Yeah

Security

operators must watch for unauthorized service activations on dormant cloud accounts right now Yeah Yeah You cannot rely on reactive billing alerts

to catch a

breach anymore No Yeah You must configure hard unalterable spending limits on model consumption

at

the

organizational level And you must monitor specifically for those single token validation