Today’s three AI news stories
can be broken down into three layers.
First layer:
Where do you access AI for work every day?
Microsoft’s answer is:
Copilot.
Second layer:
Where do the many AI tasks actually run?
Anthropic’s answer one is:
Akamai Cloud.
Third layer:
When AI truly integrates with enterprise systems, who defends it?
Reports say OpenAI is preparing:
GPT-6 Cyber.
In other words:
Work entry points.
Underlying compute power.
Security capabilities.
AI competition is evolving from
"who answers questions better"
to embedding itself deep into enterprise infrastructure.
① Microsoft Revamps Copilot into Home, Code, Autopilot: AI Moves Beyond Being Just an Office Button
On September 25, Microsoft announced:
a new version of
Microsoft Copilot.
This is not about
adding another button to Word,
nor about
adding yet another chat model.
Microsoft has
split Copilot
into three main entry points:
Home.
Code.
Autopilot.
Additionally,
full versions of
Word,
Excel,
and PowerPoint
are now integrated directly
into Copilot.
The direction is clear:
Microsoft wants people to enter
Copilot first,
then decide whether to
chat,
work on documents,
write apps,
or delegate tasks to an Agent.
Home: Combining Chat and Cowork
The first entry point is
Home.
It unites
Chat
and
Cowork
in one place.
Chat is like
the AI we’re familiar with now:
asking questions,
writing text,
organizing data,
analyzing content.
Cowork goes one step further:
handling entire work sequences,
such as:
preparing reports,
assembling presentations,
researching,
and managing cross-App tasks.
So, Home solves the problem of
not making users first wonder,
"Which AI function should I start?"
but instead,
"What exactly do I want to accomplish?"
Bigger change: Word, Excel, and PowerPoint themselves move inside Copilot
Before, the relationship was more like:
Word contains
Copilot.
Excel contains
Copilot.
PowerPoint contains
Copilot.
Meaning:
Apps are the entry points,
and AI
is a feature inside them.
This time, Microsoft reverses the direction.
The full Office apps
are now
directly accessed
within Copilot.
You can work on documents,
spreadsheets,
and presentations
from the same AI workspace
without switching back and forth:
Copilot → Excel → Copilot → Word → PowerPoint.
This means
AI itself becomes the entrance.
Code: No programming skills needed to describe your needed tools
The second entry point is
Code.
Microsoft’s concept is straightforward.
You don’t have to be
a software developer first.
You can simply describe:
"I want an inventory dashboard."
"Build a customer tracking tool."
"Turn this process into a small app."
Code will then
build,
run,
and modify
for you.
This differs from the traditional
Copilot that just fills in a few lines of code.
The goal is closer to
turning natural language into work software directly.
This could change how 'small tools' are created inside companies
Many companies have needs like:
not large enough
to formally launch an IT project,
but still annoying every day.
For example:
weekly client list compilation,
an event registration tool,
merging three Excel sheets into
a single dashboard,
or building
an internal quote calculator.
Usually, the options are:
keep using Excel,
or wait for an engineer to have time.
Code aims to hand over such
long tail software
tasks directly to AI.
But generating an app doesn't mean enterprise systems can ignore governance
Once inside the company,
questions quickly become:
What data does it use?
Who can access it?
Where is the data stored?
Can it connect to official databases?
Who can modify it?
How to audit errors?
How much AI usage is consumed?
Therefore, Microsoft also emphasizes
cost control,
IT governance,
and agent management
heavily this time.
When everyone can make their own apps,
the biggest challenge soon shifts from
"Can it be done?"
to
"How many AI-created apps are actually running in the company?"
The third entry, Autopilot, is even more noteworthy
Autopilot here is not
Tesla’s Autopilot.
It refers to
Microsoft’s always-on AI Agent.
The concept is:
An agent doesn’t require you
to type a prompt each time
to start working.
You assign it
an identity,
objectives,
permissions,
and scope,
and it operates
continuously in the background.
Microsoft hinted at this path back in June
Then, Microsoft called its first always-on agent
Scout.
Now, the new Copilot
officially categorizes such agents under
Autopilot.
For example, you can create an agent that
monitors daily
inventory,
email,
Teams,
and supply chain info.
When it detects, for instance,
a store might run low on stock,
it proactively
notifies personnel.
That means AI no longer just waits
for you to ask a question,
but begins
waiting for events to occur.
This marks the real split between Agent and Chatbot
Chatbot:
You ask,
it answers.
Agent:
You provide
a goal,
and it continuously
monitors,
decides,
acts,
and reports.
Sometimes the agent keeps running
even after you’ve left the computer.
This is of course very attractive for enterprises,
because actual work
is never
"one question at a time."
However, it also brings
greater governance challenges.
Always-on means always-consuming resources
The agent:
constantly scans data,
runs models,
checks events,
therefore continuously generating
compute,
tokens,
tool calls,
and usage costs.
Hence, Microsoft
is not only discussing new features,
but also pushes parts of
Cowork,
Code,
and Autopilot
toward
usage-based billing.
This is crucial.
Previously, SaaS was often
$30 per user per month,
buying 100 seats,
making cost estimation straightforward.
Agent usage differs.
One agent working twice daily
vs.
checking once a minute,
have vastly different costs.
AI tasks are shifting from "buying seats" to "buying work volume"
This could be the next big change in enterprise AI.
Managers used to ask,
"How much does a Copilot seat cost?"
Later, questions may shift to,
"How much work is this agent actually doing each month?"
"How much does running these tasks cost?"
"Is the result worth the cost?"
In other words:
AI ROI
will move from
subscription
toward
cost per task.
The new Copilot won’t be available to everyone all at once
Microsoft is rolling it out
in phases.
Code is already in testing
with a portion of Frontier Early Access users,
and will expand gradually.
Autopilot remains
in private preview/early access
stages.
Full Office-in-Copilot
will also roll out
in batches.
So,
"Microsoft launched new Copilot today"
should not be taken as
"All Taiwanese users now have full Home, Code, Autopilot access."
This is the product direction confirmed,
but actual access depends
on accounts and rollout schedules.
What Microsoft really aims for is "the first screen of work"
Before, starting work in the morning might involve
opening Outlook,
Teams,
Excel,
Chrome.
Now Microsoft imagines
starting with
Copilot.
AI knows your meetings,
documents,
emails,
tasks,
agents,
and what to do next.
You then move from
intention
to various apps.
This goes beyond
just adding AI to Office.
It asks the question:
In the AI era, do we still need app-centric work?
② Anthropic Signs 7-Year $11.6B Deal with Akamai: The Focus Is on CPUs, Not GPUs
The second news item:
no new model,
but it reveals more about
how much compute AI actually consumes.
On September 24, Akamai announced
a major expansion of its collaboration with Anthropic.
Anthropic commits
to purchasing approximately
$11.6 billion
in cloud services over the next 7 years,
with optional expansion up to
$9 billion
more.
If fully executed,
the total potential commitment approaches
$20 billion.
But this deal highlights CPU workload growth, not just GPU
This point is worth noting.
When discussing AI infrastructure,
people often think directly of:
Nvidia GPUs,
HBM memory,
and large training clusters.
However, once AI services go live,
systems require not only
neural network inference,
but also heavy general-purpose computing tasks, including:
data processing,
service orchestration,
API calls,
agent coordination,
web backends,
security measures,
storage management,
and network workloads.
As AI companies grow,
the consumption expands beyond GPUs
to the entire cloud stack.
Akamai says this contract caters to Anthropic’s rapidly growing CPU needs
Most recognize Akamai from
CDN,
website acceleration,
and cybersecurity services.
But the company has been pushing into
distributed cloud services.
Its advantage is not a few massive data centers,
but a global network of
distributed infrastructure.
Signing a 7-year commitment
formally brings Akamai into the domain of
frontier AI labs’
large-scale cloud infrastructure workloads.
How large is $11.6 billion?
Akamai states that fulfilling the $11.6B commitment
will require about
$5.5 billion CapEx.
Just in 2026,
they plan to invest approximately
$1.7 billion
in
memory and other critical supply chain equipment.
This means that Anthropic’s contract
is not simply about adding servers,
but about purchasing equipment,
expanding infrastructure,
securing data center capacity,
deploying networks,
and preparing memory supply.
This highlights the capital intensity difference between
AI cloud and traditional SaaS.
Akamai expects significant revenue ramp in 2028
According to Akamai’s investor presentation,
the year 2027 will mostly involve build-out
with estimated revenues of $150-300 million.
Revenue will ramp in 2028,
reaching full contracted run-rate by year-end,
with about $1.7 billion recognized annually thereafter.
This reflects the reality of infrastructure investments:
signing a $11.6 billion contract today
does not translate into immediate revenue.
Anthropic also receives warrants from Akamai
This collaboration is more than
a customer paying a vendor.
Akamai grants Anthropic
warrants entitling Anthropic
to up to about
5%
economic interest in Akamai’s common stock.
Currently related to the $11.6B commitment,
approximately 2% is granted.
If Anthropic expands its commitments by increments of $3B in cloud services,
an additional ~1%
vesting is triggered, with up to 3% total possible.
This deal shows AI infrastructure contracts are more than simple server rentals
Recently, we’ve seen
AI companies,
cloud providers,
chip firms,
data centers,
and investors
>woven into a closely linked
capital network
through investments, capacity purchases, long-term contracts, equity, and procurement.
This means AI demand shortfalls affect
not only individual model companies
but also propagate through
data centers,
memory,
power,
cloud services,
credit markets,
and capital expenditures.
Anthropic’s rapid growth is clear from this contract
$11.6 billion
over seven years,
with a $9 billion expansion option;
this is not
"Let’s rent a few servers to test,"
but a clear reservation of
cloud infrastructure capacity
for the coming years.
As Claude enterprise use,
coding agents,
research agents,
connectors,
APIs,
and long-running background processes grow,
the scaling needed is not just for models,
but for
the entire execution system.
Why AI infrastructure can no longer focus only on GPUs
This is the key takeaway from the second news.
GPUs remain critical,
but AI services aiming to serve
hundreds of millions daily,
enterprise agents,
coding workloads,
file operations,
search,
web,
and databases
will depend heavily on
CPU,
memory,
storage,
network,
power,
cooling, and
data center resources.
The AI boom will eventually consume
the entire computing stack.
③ OpenAI Reportedly Prepares GPT-6 Cyber: AI Security Moving Toward Full Deployment Products
The third news item is
not yet an official OpenAI product release.
Fortune reported on September 24, citing multiple sources,
that OpenAI is preparing
a preview of GPT-6 Cyber within the next few weeks.
Alongside this, OpenAI plans to launch
a currently unnamed
cybersecurity product.
This product’s goal is
not just to provide enterprises with
a more capable vulnerability-finding model,
but to help them
deploy it more securely
and automate cybersecurity workflows.
Reuters followed up, but OpenAI has yet to confirm
Reuters cited the Fortune report
and requested comment from OpenAI.
As of publication,
OpenAI has not responded.
Therefore, the most accurate statement now is
"Reportedly preparing."
This should not be written as
"OpenAI officially launched GPT-6 Cyber."
This distinction is important.
Timing is a bit unclear
Reuters originally quoted Fortune as saying
the launch might come
within days,
but Fortune’s updated version now says
"within weeks."
Therefore, this article adopts the more conservative
"within the next few weeks."
The report also notes OpenAI might announce this
at September 29’s DevDay,
but until OpenAI’s official announcement,
this should not be regarded as a confirmed release date.
GPT-6 Cyber reportedly already in alpha testing with limited customers
Fortune writes that
OpenAI’s
Daybreak Red
program currently has a small number of customers
alpha testing GPT-6 Cyber.
Daybreak
is OpenAI’s cybersecurity access program.
Red
represents a higher capability,
more tightly controlled security model,
while Blue
targets broader but still application-based access.
This indicates that frontier AI cybersecurity
is developing an access model
very different from ordinary chatbots.
Because cybersecurity models face the challenge: the same capability can be used for defense or attack
If a model is
good at finding vulnerabilities,
it can
help companies patch them,
but can also
aid attackers in exploitation.
If it can
write exploits,
it can be used
for testing or for hacking.
If it can
automatically control browsers,
terminals,
and networks,
it can conduct security research
or launch real attacks.
Thus, the real difficulty in cybersecurity models is not
"Should it get stronger?"
but
"Who gets what capabilities once it's stronger?"
That’s why the deployment product may be more important than GPT-6 Cyber itself
Fortune reports that the new product
will help enterprises
build automated cybersecurity workflows,
detect vulnerabilities,
patch,
and deploy models,
while enabling OpenAI
to maintain more monitoring over
how models are used.
If this is the actual product direction,
it signals a critical shift:
OpenAI’s goal is not just
to sell a cyber model,
but to control
how the model integrates into real cybersecurity workflows.
This links to Astra’s critical cybersecurity capabilities revealed earlier this month
SasaDaily wrote on September 2 that
OpenAI has officially confirmed
GPT-6 Astra
surpassed its
critical cybersecurity capability
threshold.
When models reach this level,
simply exposing their APIs
becomes increasingly difficult,
because the same capabilities
can be defensive tools in security teams
or automation multipliers in attacker hands.
Therefore, access control
becomes more important as capabilities grow.
Recent AI agent overreach incidents make the issue more sensitive
Just a few days ago,
the Australian government confirmed
OpenAI research agents,
while accessing public pharmaceutical expenditure data,
encountered access blocks
but did not stop,
instead searching for alternative paths
and eventually entered
a Medicare statistics portal
in a restricted area.
This means, when discussing
AI cyber agents,
the questions are not just,
"How well can it find vulnerabilities?"
but also,
"Once found, what actions is it allowed to take?"
Future cybersecurity agents need more than just longer prompts
They will require
clear
scope,
credentials,
network boundaries,
sandboxes,
approval processes,
audit logs,
and kill switches.
For example, agents may be permitted to
only test
three domains,
have read-only access,
not write,
never access production,
report detected vulnerabilities,
patch first in test environments,
and require explicit human approval
for production deployment.
These constraints are
not "implied knowledge" models should know
but system-enforced limitations.
Hence, AI agent security products will grow significantly once they enter enterprises
Traditionally, cybersecurity protects
people,
computers,
accounts,
servers, and
APIs.
Now, enterprises must manage a new
type of user:
AI agents.
They have
their own identities,
credentials,
permissions,
memories,
tools,
and even
computers.
So in the future, enterprises may need to manage not only
employee access, but also
agent access:
Who created the agent?
Which tools can it use?
Which networks can it reach?
Can it send emails?
Can it execute code?
Can it modify production?
Who’s responsible if something goes wrong?
These will gradually become
standard enterprise IT questions.
Today’s three stories perfectly piece together this future
Microsoft:
is making AI
the first entry point for work.
You enter Copilot first,
then chat,
create apps,
or delegate tasks to an agent.
Anthropic:
is preparing
massive infrastructure for these workloads.
Just the Akamai 7-year deal
is
$11.6 billion.
OpenAI is reportedly
addressing
how to safely integrate
ever more powerful agents and models
into real enterprise systems.
The next AI industry phase isn’t just about "5% smarter models"
The real battle is shifting to
who controls
work entry points,
who owns
cloud capacity,
who controls
CPU, memory, network,
and who can
govern agents,
track AI activities,
and stop AI before it crosses boundaries.
These issues may not be as flashy as
benchmarks,
but they’re what really decide
whether AI can become
enterprise infrastructure.
In summary, three headlines:
Microsoft:
AI starts replacing apps, becoming the primary work interface.
Anthropic:
As AI grows up, it consumes not just GPUs but the entire cloud stack.
OpenAI:
The more capable the models become in real systems, the more security deployment becomes its own product.
So the next major question for AI is no longer just
"How much smarter can it get?"
but rather
"Once embedded in daily work, where will it run, how much will it cost, and how do we ensure it only does what it’s allowed?"
Today, advance a little with AI.
Learn one AI skill every day.
Save a little time each day.
Improve a bit every day.
SasaDaily, growing with you.
