Only two AI stories are truly worth noting today.
The first:
Amazon is reportedly preparing to spend about $7 billion
to directly acquire an AI startup.
The second:
An important AI company’s data center
has been physically attacked and forced offline for several consecutive days.
On the surface,
one is an M&A deal,
and the other is warfare.
But viewed together,
they both answer the same question:
After AI becomes fundamental infrastructure, who controls its capabilities, and can the system continue to operate?
1. Amazon Reportedly in Talks to Acquire Decart for Around $7 Billion
On October 11, Reuters cited The Wall Street Journal reporting:
Amazon is negotiating with the AI startup:
Decart
on a potential acquisition.
The deal price is said to be approximately:
$7 billion.
However, it remains at the:
talks stage.
It is not yet a finalized acquisition.
Reuters also clearly stated:
they could not independently verify the news at the time.
So the accurate description today is:
Amazon is reportedly negotiating to acquire Decart.
It cannot be stated:
"Amazon has already acquired Decart."
What Does Decart Actually Do?
Looking only at its AI demos,
one might easily mistake it for just
another AI video company.
But Decart actually operates on three levels.
First Layer: Lucy
Lucy is a:
Real-time Video Editing Model.
Decart’s flagship feature allows videos to be:
edited in real time to change characters,
alter backgrounds,
add objects,
switch styles,
and apply visual effects (VFX).
Lucy is already moving towards true real-time processing,
instead of the previous:
input prompt →
wait tens of seconds →
receive a new video.
Second Layer: Oasis
Oasis is closer to a:
World Model.
It doesn't merely create a single video.
Instead, it enables:
robots,
autonomous vehicles,
manufacturing systems
to perceive scenarios,
take actions,
and receive new environmental feedback in AI-generated environments.
Decart positions Oasis 3 as an:
Interactive World Model for Physical AI.
For example, autonomous vehicles cannot wait for rare accidents,
unusual weather,
or sudden obstacles to occur on real roads
before collecting training data.
The World Model concept is
to massively simulate those
rare,
dangerous,
and costly
scenarios in a synthetic environment.
The Third Layer: Possibly What Amazon Values Most—DOS
Decart also offers a suite called:
Decart Optimization Stack (DOS).
This is not about generating video models.
Instead, it is an:
AI training/inference optimization stack.
Decart claims DOS can make models run:
faster,
cheaper,
and with better utilization
across different hardware.
Moreover, it doesn't only support:
Nvidia.
They explicitly list support for:
Nvidia,
AWS Trainium,
and Google TPU.
This is particularly notable for Amazon.
Why?
Because Amazon doesn't want to rely solely on Nvidia GPUs forever.
AWS has invested heavily in developing its own:
Trainium chip.
The challenge is, enterprises already run many AI workloads within the:
Nvidia CUDA ecosystem.
Convincing customers to switch chips is not as simple as:
"The alternative chip is cheaper."
The real difficulty lies in whether:
models,
kernels,
compilers,
and inference stacks
can be migrated together.
If a company's core capability is:
making it easier for the same AI workload to run across multiple chip platforms,
that value for AWS would be more than just:
"another model."
It could help Amazon solve the problem of:
how to move AI workloads from Nvidia to Trainium.
This remains an industry strategic inference at this time.
Amazon hasn't publicly said:
"We're buying Decart because of Trainium."
But Decart’s own product positioning
indeed crosses:
Nvidia,
AWS Trainium,
and Google TPU.
The Wall Street Journal Also Reveals: Decart Has Early Access to Trainium3
The Wall Street Journal reports that
AWS has already given Decart early access to:
Trainium3.
This indicates Amazon and Decart have been in contact well before today’s news.
Founded in:
2023,
Decart has to date raised:
over $400 million.
Its valuation as of May this year was close to:
$4 billion.
If the acquisition goes through for around:
$7 billion,
it would represent a significant jump in valuation in a short time.
Amazon May Be Buying Not Just an AI Video Tool but an Entire Real-Time AI Stack
Putting the three products together makes the vision clearer:
Lucy:
real-time video.
Oasis:
physical AI/world model.
DOS:
underlying training/inference optimization.
Together, these layers point to:
Real-time AI.
AI no longer:
thinks for a long time before answering,
but must respond immediately during:
robot movement,
camera views,
game worlds,
and autonomous driving systems.
This is fundamentally different from a typical chatbot.
If Amazon ultimately acquires Decart,
it could gain:
models,
world models,
inference optimizations,
and a team already experienced in maximizing efficiency across Nvidia, Trainium, and TPU.
This Aligns with Amazon’s Recent Aggressive AI Infrastructure Investments
Amazon is preparing a massive capital expenditure on AI infrastructure this year.
They are not only selling:
AWS servers,
but also developing:
Trainium,
Bedrock,
Agent,
foundation models,
robots,
and Alexa.
Acquiring Decart would indicate that cloud providers’ competition is increasingly:
vertical.
Not simply delivering:
"a machine for you,"
but integrating from:
chip →
compiler/optimization →
model →
application
all the way up.
But the Biggest Caveat Today: The Deal Isn’t Final Yet
Don’t overlook this.
All current information is from:
The Wall Street Journal citing insiders.
Reuters could not independently verify.
The acquisition talks could:
complete,
change price,
or fall apart completely.
So it is premature to conclude that:
Decart has become part of AWS,
Lucy/Oasis technology will be integrated into Amazon,
support for Google TPU will be discontinued,
or Decart’s product roadmap has shifted.
The only confirmed fact is:
Amazon is reportedly in talks.
What Should General Users Do Now?
Basically, nothing.
But this potential deal signals one important takeaway:
Future AI competition
may increasingly be less about:
which company has the strongest model,
and more about:
which company can control models, hardware, inference, cloud, and actual applications together.
If Amazon does acquire Decart,
this point will become even clearer.
2. Yandex’s Third Data Center Attacked Within Days: AI’s “Cloud” Is Far from Virtual
The second story
brings AI back from the cloud
to physical realities like:
concrete,
electricity,
and servers.
The Russian technology and AI company:
Yandex
announced on October 11 that
its data center in the:
Vladimir Region
was completely shut down after a drone attack.
No personnel were injured,
but some Yandex services became:
unavailable.
What’s more significant is:
This marks the third major data center
disabled within just a few days.
What Happened to the First Two?
The first one was:
Sasovo.
Yandex said after an attack from the evening of October 7 to the early morning of October 8,
the data center was severely damaged and has been:
completely shut down.
Yandex could not confirm at the time whether the equipment:
could be repaired.
The second was:
Kaluga.
After a drone attack on October 9,
part of the infrastructure was:
out of service.
Then on October 11,
Vladimir was attacked.
Why Is This an AI Story?
Because
Sasovo is not just an ordinary web hosting facility.
According to Reuters,
Yandex’s:
three AI supercomputers
used to develop AI models
have:
two of their units located in Sasovo.
This means the attack targeted a site holding not only:
email,
websites,
and general cloud services,
but also core compute power crucial for training AI.
This is a very real aspect
once AI enters strategic infrastructure territory.
Cloud Often Makes People Forget that There’s a Physical Building Behind It
When you open:
AI chatbots,
cloud storage,
search engines,
maps,
or business software,
it feels like they exist in the:
"cloud."
But the truth is that cloud means there are:
buildings,
power,
cooling,
fiber optics,
server racks,
GPUs,
storage,
and networks somewhere physically.
If that site experiences:
power outages,
network failures,
equipment damage,
or disasters,
cloud services could be disrupted.
Yandex described its data centers in an October 9 statement as its:
company’s iron heart.
It noted that millions of people rely daily on its transportation, music, video, shopping, and information services,
as well as hundreds of thousands of businesses.
What’s Most Important Isn’t Whether AI Supercomputers Were Destroyed
That cannot be confirmed at this point.
Yandex has stated it is still assessing damage to Sasovo.
Reuters has not confirmed how much damage those two AI supercomputers suffered.
So it’s inaccurate to say:
"Yandex’s two AI supercomputers have been destroyed."
The confirmed fact is:
The data center housing them:
is closed,
and the ability to recover equipment was unknown at the time.
The Larger Issue: Concentration Risk
Suppose a company has:
three critical AI clusters,
and two are located at the same site.
If that site experiences a problem,
the impact could be severe.
This applies not only to war situations
but also regular risks like:
earthquakes,
floods,
fires,
long power outages,
fiber cuts,
cooling failures,
or cloud region outages.
So the most practical question today isn’t:
"Will my data center get hit by a drone?"
But rather:
Does my AI workflow have a single physical point of failure?
Four Questions Companies Should Review Now
If AI is only used occasionally for copywriting,
this has almost no impact.
But if a company already relies on AI for:
customer service,
order processing,
data analysis,
enterprise search,
coding,
or agent workflows,
they can ask:
1. Model
If the primary model fails,
is there a backup to take over?
2. Provider
If an entire cloud provider has issues,
can operations continue?
3. Region
Are data and compute resources consolidated purely in one region?
4. Data
Are critical assets like:
prompt templates,
knowledge bases,
customer data,
and workflow configurations
backed up independently?
These four questions are far more realistic than:
"Cloud should be reliable, right?"
This Doesn’t Mean Every Company Needs Multi-Cloud
This is a key boundary.
Multi-cloud,
multi-region,
and multiple model redundancy
all add:
cost,
complexity,
synchronization challenges,
and more maintenance work.
If an AI function can tolerate a two-hour outage with no loss,
building expensive disaster recovery may be unnecessary.
The real question is:
How long can this AI workflow be offline before it truly becomes a problem?
If the answer is:
"Not an issue for a whole day,"
there’s no need to design bank-grade infrastructure.
If the answer is:
"Ten minutes impacts sales, customer service, and operations,"
then investment in redundancy becomes worthwhile.
The Yandex Incident Doesn’t Mean "All AI Data Centers Are Unsafe"
These events occurred in a:
war environment.
It’s not reasonable to conclude that
all data centers globally face equal drone attack risks.
The real takeaway for regular businesses
is the fundamental level:
AI service reliability ultimately depends on the physical infrastructure.
The cloud does not eliminate physical risk;
it just hides it the rest of the time.
What Connects These Two Stories?
Amazon/Decart is about:
running AI software faster across different computing platforms.
Yandex is about:
those computing platforms ultimately existing in physical locations.
As AI gains importance, key capabilities now include, besides:
model intelligence,
also:
portability—
can AI workloads run on different hardware?
and
resilience—
if a site goes down, can work continue?
The AI industry once mainly asked:
Which GPU is fastest?
Which model is strongest?
But as AI becomes infrastructure,
these additional questions matter:
If this hardware isn’t available, can I switch?
And:
If this site disappears, can I still run?
This may be the crucial issue enterprises will face as AI matures.
Recommended Reading
Why Is the Market Starting to Relax as Amazon Plans to Invest $220 Billion in AI Over One Year?
