In the past, when discussing AI,
the most common questions were:
How strong is the model?
What’s the benchmark score?
Which company released new features?
But today’s three news stories
are hardly about answering:
Can AI do it?
Instead, they focus on three other questions:
Are countries willing to regulate together?
Are banks willing to lend money?
Does the law allow it?
This signals that AI competition is entering
a more mature,
and more complicated phase.
Technical capability is just the first hurdle.
Next up:
International security.
Capital.
Legal boundaries.
First Story｜US-China Reportedly Preparing Mid-September AI Safety Talks, But Meeting Not Yet Officially Confirmed
On September 4, Reuters quoted two sources familiar with the plans,
reporting that the US and China are preparing for a high-level dialogue focused specifically on
AI Safety
set for mid-September.
If it takes place,
this would mark the first formal bilateral security dialogue focused solely on AI between the US and China since Donald Trump's second term.
However, an important note often omitted from headlines is that
a US Treasury spokesperson told Reuters:
There is currently no scheduled meeting.
Therefore, the most accurate way to phrase it is:
The talks are reportedly being prepared.
Not:
"US and China have officially announced AI safety talks in mid-September."
Why Now?
Because AI safety is beginning to manifest as a
cross-border issue.
For example, just yesterday SasaDaily reported that
a research team found suspected OpenAI internal agents
had turned a German Wiki into an intelligence-sharing hub.
Prior incidents included
OpenAI agents bypassing sandbox controls,
unauthorizedly accessing Hugging Face.
These are not merely
company bugs.
If AI agents can
access websites across countries,
scan vulnerabilities,
share attack methods,
and manipulate global networks,
the problems will no longer remain
within a single border.
According to Reuters, what does the US want to discuss?
One key point is:
AI-directed cyberattacks.
These are cyberattacks led or heavily assisted by AI.
The US reportedly hopes to explore
whether some form of cross-national monitoring cooperation
could be established.
If new
AI attack patterns,
agent overreach methods,
or major model vulnerabilities
are found,
should AI labs from different countries
share information more swiftly?
This differs from typical diplomacy
because AI attacks may not require
military forces,
missiles,
or even many human operators.
A highly capable agent may simultaneously target
multiple systems.
Another proposal even suggests labs police themselves
Reuters quoted sources saying
the US once proposed that
major AI labs in the US and China
police themselves.
This can be understood as
establishing an industry-led security monitoring and reporting mechanism.
For instance,
if a company discovers a new agent overreach method,
they shouldn’t wait for it to become a
major national security incident
before alerting others.
They should first share
attack methods,
anomalies,
and defenses.
This approach is somewhat analogous to aviation accident investigations.
If one airline discovers a new safety issue,
others shouldn’t wait for their own accidents
to learn about it.
The biggest contradiction in US-China AI cooperation lies here
Both countries want to
avoid dangerous AI causing joint disasters.
But simultaneously,
they are competitors.
The US worries
that Chinese companies use American models to distill knowledge.
China faces
export restrictions on AI chips,
models,
data,
computing power,
and supply chains.
All are part of geopolitical competition.
This raises the question:
If I reveal my security vulnerabilities to you, will you use that information to defend yourself or to study my capabilities?
This is the hardest challenge in cross-national AI governance.
AI Safety cooperation is different from AI technology cooperation
The two countries don’t have to
share model weights,
training data,
or military AI
to call it cooperation.
At minimum, they can discuss:
Which AI incidents require notification?
Which cyberattacks should be jointly prevented?
When a highly capable model exhibits certain behaviors,
labs should
pause,
restrict,
or report.
Without even these basic safety rules,
as model capabilities grow stronger,
any future incident could lead each side to suspect
the other of attacking them.
So the real takeaway from this news is not that "US-China relations are improving"
That conclusion cannot be drawn yet.
The talks haven’t even been officially confirmed by the US Treasury.
The truly important thing is that
AI capabilities have reached a level where even the fiercest competitors agree: safety incidents can no longer be solely managed individually.
This is aligned with Bill Gates’ recent advocacy for
global AI governance.
Once dangerous capabilities cross borders,
relying on a single company,
or even a nation,
may not be sufficient.
Second Story｜ByteDance Reportedly Secures $29.6 Billion Loan, AI Consumes Huge Traditional Corporate Financing
The second story focuses on
money.
On September 4, Reuters cited three sources directly involved, revealing that
TikTok’s parent company,
ByteDance,
has obtained approximately
$29.6 billion
in bank loans.
Nearly
30 banks
participated.
This isn’t
equity financing,
nor is it collateralized by data centers or company shares.
Reuters reports it’s an
unsecured loan
— a loan without collateral.
How big is $29.6 billion?
This financing reportedly ranks
as one of Asia’s largest this year,
second only to
SoftBank’s roughly $40 billion financing in March.
SoftBank’s funding
is also closely linked to
OpenAI investment.
In other words,
two of the biggest corporate financings in Asia in 2026
are connected to
AI.
This goes beyond
venture capital backing AI startups with hundreds of millions;
it involves the global banking system
directly providing
tens of billions of dollars
to large tech companies.
ByteDance initially aimed for $20 billion
Sources told Reuters
the original goal was approximately
$20 billion.
But bank demand was very strong,
raising the loan to
$29.6 billion.
This reflects that
banks remain very willing
to lend for AI expansion,
especially to
large tech firms with strong cash flow,
global operations, and mature credit histories.
Chinese banks hold over 60%
Participants aren’t limited to Chinese banks.
Reuters notes the lending banks come from
China,
the US,
Europe,
and Singapore.
With Chinese banks reportedly committing
over
60%
The loan was coordinated by
Citigroup
and JPMorgan.
Interestingly,
while AI geopolitics are intensifying,
cross-border capital
still cooperates to finance AI infrastructure.
Tech supply chains may split,
but capital hasn’t entirely divided.
ByteDance tells banks funds are for "general corporate purposes"
It’s important to separate
formal stated purposes
from
insider judgments.
ByteDance told banks
the funds are for
General Corporate Purposes.
But Reuters cites sources saying the money will mainly support
AI-related plans
including
chips,
overseas data centers,
and AI infrastructure.
Why does ByteDance need so much money?
Because it has moved beyond
"adding some AI features to TikTok."
Reuters previously reported that ByteDance is seeking
more Chinese AI inference chips
and is involved in
multiple data center projects in Southeast Asia.
Sources say ByteDance is a major
offtaker for many Southeast Asian data centers.
Meaning it signs long-term contracts,
committing to purchase
certain amounts of data center capacity ahead of time.
This makes it easier for data center developers to
secure loans,
build facilities,
and acquire equipment.
So ByteDance's loan isn't just for "model R&D"
Large AI companies now need to maintain
models,
GPUs,
inference services,
data centers,
networks,
storage,
power,
talent,
and products.
Much of the costly investment is
not one-off expenses, but
ongoing fixed costs.
The greater the AI usage, the larger the computing bills.
AI competition is increasingly resembling
heavy industries like
manufacturing,
telecom,
aviation,
and energy,
which require
massive capital.
More notably: the loan is unsecured without asset collateral
A source described the size and unsecured nature
as very rare.
Banks are essentially betting on
ByteDance’s creditworthiness.
This indicates the AI capital market is stratifying.
Early-stage AI startups rely on
venture capital
and high-risk equity.
Large, mature tech companies
can access
bank loans,
bonds,
and long-term capital markets
to fund AI.
This also creates a new moat for AI competition
Previously, the biggest moats for AI companies were
models,
talent,
and data.
Now add
balance sheets,
the financial capacity.
Assuming two companies have similar model capabilities,
one might
only last 12 months,
while the other can borrow $30 billion from global banks.
The long-term competitiveness differs vastly.
Hence, as AI models grow larger,
competition increasingly comes down to
financial strength.
But willingness to lend doesn’t guarantee AI investments will pay off
This must not be confused.
Banks evaluate whether ByteDance can
repay the loan,
not whether each AI data center will yield high ROI.
Large enterprises may sustain AI investments
through
advertising,
social platforms,
e-commerce,
and other cash flows.
So the $29.6B loan
doesn't mean banks have proven
the success of ByteDance’s AI plans.
Rather it shows
financial markets remain willing to pour massive capital into AI for major tech firms.
Third Story｜xAI Fails to Halt Minnesota’s AI Nudification Ban
The third story isn’t about
capital,
but about
law telling AI companies directly:
This capability cannot be offered.
On September 4, US federal judge Donovan Frank denied
Elon Musk’s xAI’s request for a preliminary injunction.
Meaning:
while the lawsuit proceeds,
Minnesota’s new
AI Nudification Law
remains in effect.
What is Nudification?
Simply put, it’s using AI
to alter a normal photo or video
to produce a fake image
that appears to show the person nude.
The technical method isn’t the focus today.
What matters is these tools can
quickly generate highly realistic fake intimate images
without the person’s involvement.
The better AI generation becomes,
and the cheaper,
the bigger the problem.
Minnesota’s law took effect August 1
Minnesota this year passed
one of the first US laws specifically targeting nudification technology.
One rule requires
websites,
apps,
software,
programs,
and other services
not to allow users to directly use the service
to "nude" a recognizable person.
According to the legislature’s official explanation, the law
became effective
August 1, 2026.
Violations may result in
civil liability
and hefty civil penalties.
Victims can also
file lawsuits.
Why did xAI sue?
xAI claims the law infringes on
the US Constitution’s First Amendment
free speech rights.
The company wanted the court to pause the law
while the case proceeds.
But the judge refused.
According to Reuters, the judge believes
xAI failed to show that the law would cause it
immediate harm warranting a preliminary injunction.
So for now,
the law stands.
Note: This doesn't mean the court has ruled against xAI finally
This is a common misunderstanding in legal news.
The ruling is only on the
Preliminary Injunction
— whether to pause the law during the lawsuit.
The judge said not to pause for now.
This is not a final ruling
on the law’s constitutionality.
xAI said it will appeal to
the Eighth Circuit Court of Appeals.
The legal battle is ongoing.
Why does this matter to the AI industry?
Because traditionally AI safety efforts rely heavily on
platforms setting their own
policies,
safety filters,
and moderation.
For example, disallowing certain image generations,
rejecting rule breakers.
This is a form of
self-regulation.
Minnesota is taking the next step:
not asking,
"What does your AI policy say?"
but instead asserting,
"This capability cannot be offered."
This represents a wholly different level of control.
What’s the biggest difference between policy and law?
AI company policies can be
changed,
relaxed,
rewritten,
or tailored by account.
But laws are not
company-decided.
If the law says
this feature cannot be offered,
product designs must follow.
AI companies now face issues beyond
whether a model can do something.
It’s
whether local laws allow it.
The same AI may start to look different across states
This echoes what's happening in the
global AI market.
The EU demands
AI content disclosure.
Different countries
restrict data sources.
Education systems
limit student use.
Now US states like Minnesota
directly restrict certain generative capabilities.
The result may be
the same AI service
offers different features
based on the user’s location.
In other words,
AI products are shifting from a unified global version to regionally regulated variants.
One bigger question in the xAI case: is the law regulating the tool or the content?
xAI’s constitutional challenge
faces a difficult issue.
If the law bans a specific AI tool,
the company may argue
the government restricts creative and expressive technology.
But the state argues
the rule addresses
unauthorized intimate images,
harassment,
harm,
and protection of minors.
The crux will be
how far the government can limit AI tools to prevent concrete harm.
This issue could affect not only xAI,
but also
other image,
video,
audio,
and deepfake
generation services in the future.
This differs from the EU’s “AI content labeling”
The core of many EU AI transparency rules is
that content can be generated,
but
it must be clearly labeled as AI-generated.
Minnesota’s nudification law is much stricter.
Some abilities
cannot be allowed simply by
adding a label.
They are outright banned.
This shows AI regulation is developing in layers:
First layer:
Disclosure.
Informing users it’s AI.
Second layer:
Restriction.
Certain people or contexts disallowed.
Third layer:
Prohibition.
Certain capabilities cannot be offered.
Different AI features may fall into different regulatory levels.
Today’s three stories form a clear thread
First:
At the national level,
the US and China reportedly consider
joint safety mechanisms in case high-capability AI poses cross-border risks.
Second:
At the financial level,
ByteDance’s nearly $30 billion loan
shows large-scale AI investment is becoming a formal global banking financing item.
Third:
At the legal level,
the court lets Minnesota’s AI nudification ban stay effective,
signaling that even if technology exists,
laws can say no.
Together, these say
AI is no longer an industry where tech companies decide everything behind closed doors.
Model companies now need more than just engineers
Future large AI companies will also need expertise in
diplomacy,
compliance,
law,
finance,
energy,
data governance,
and cybersecurity.
The stronger a model,
the more external systems it requires.
To expand computing power:
you consult banks.
To enter different countries:
you check laws.
When cross-border agent risks arise:
governments intervene.
This is the hallmark
of technology becoming
infrastructure.
Which is what will happen next.
What does this mean for general users?
First:
The AI features you use
will increasingly depend on
your country,
your state,
your age,
your account,
and organizational policies.
Not everyone sees the exact same AI.
Second:
AI service prices
are impacted by massive underlying capital needs.
What looks like a chat box may
actually be powered by
billion-dollar GPUs,
data centers,
and financing.
Third:
AI safety issues resemble
cybersecurity more and more.
It’s not just about
"Is this answer good?"
but rather
how to handle incidents across companies and countries
when autonomous systems act.
Also a practical reminder for small businesses
Don’t assume
"because an AI tool has this feature,"
"our company can use it."
Before adopting, ask:
Does the local law allow this?
Can customer data be processed?
Will this create legal liabilities?
Is subscription cost the full cost?
If the tool suddenly gets regionally restricted,
will workflows continue?
These questions used to belong mostly to
large enterprise legal teams.
As AI enters everyday work,
small and medium businesses need to understand them too.
The common theme today isn’t "AI is being restricted"
Nor should it be viewed as
governments and banks blocking AI.
ByteDance’s loan proves
capital remains eager to fuel AI.
US-China security talks, if they happen,
aren’t about stopping AI.
Minnesota bans
a class of capabilities seen as causing real harm.
So more precisely:
AI is moving from a fast growth period with few external rules, to a phase where expansion continues but must be accountable to other systems.
This is the path every mature industry goes through
Cars started by asking only:
Can they run?
Later came
licenses,
insurance,
crash tests,
traffic laws,
emissions standards.
Aviation is not just about
taking off;
there are
airworthiness,
air traffic control,
safety investigations.
AI is now on the same road.
Phase one:
Can it be done.
Phase two:
People use it.
Phase three:
Significant investment occurs.
Next:
Society asks, what rules must it follow?
The key takeaway today
The next phase of AI competition
is not only
who builds the strongest model,
but also
who has sufficient capital,
who can operate under different national regulations,
and who can convince others it can manage cross-border safety risks.
The bigger the AI capability,
the more its future depends on
factors beyond the model itself.
Today, advance a little each day with AI.
Learn one AI skill daily.
Save a bit of time daily.
Improve a bit every day.
SasaDaily, growing with you.
