Today’s three AI news stories seem unrelated on the surface.
One is about US-China tech competition.
Another about Amazon purchasing AI chips.
The third covers litigation between OpenAI and The New York Times.
But they all really ask the same question:
Before an AI model gets stronger, where do its abilities, chips, and data come from?
As AI enters the next phase, competition is no longer just about:
“Whose benchmark scores are higher?”
It’s now about the entire supply chain behind the models.
First Story: US Names 6 Chinese AI Firms, Accusing Them of Large-Scale Use of US Models for Distillation
On September 8, US law enforcement and intelligence officials publicly accused Chinese AI firms of conducting “industrial-scale” targeted distillation.
Distillation, often called model distillation in Chinese, is a concept that’s not mysterious.
Imagine there is a very large and expensive AI model.
You let it answer lots of questions.
You then take these input/output pairs—inputs and answers—and use them to train a smaller model.
This smaller model can then learn part of the larger model’s capabilities.
Model distillation is a common AI technique.
The real controversy is:
Whose model’s outputs are used, whether authorization was obtained, and whether usage restrictions were deliberately bypassed.
What Is the US Accusing Specifically?
According to Reuters, US officials named six Chinese companies.
Among those publicly identified are:
DeepSeek.
Moonshot AI.
Alibaba.
Officials claim these companies utilized AI model variants from US companies like Anthropic, OpenAI, Google, and SpaceX to accelerate their model development.
Officials even stated these activities are “likely” known to the Chinese government.
But facts should be distinguished here.
These are accusations from the US government, not court rulings.
No court has yet judged that these six companies stole technology.
At the time Reuters published its report, the Chinese Embassy in the US had not immediately responded to these accusations.
So the most accurate way to put it:
The US is elevating large-scale distillation from a normal model competition issue to one of national security and intellectual property.
Why Has Distillation Become a National-Level Concern?
Because one of the most expensive parts of AI models is:
Research.
Training.
Data.
Compute power.
Talent.
If a company can heavily leverage already-trained frontier model outputs to quickly shorten its own R&D timeline:
The originally very high training costs could be significantly reduced.
For typical developers, this may just mean:
“Using a strong model to help a weaker model learn.”
But from the perspective of national competition, it raises the question:
Can one country rapidly catch up by exploiting AI capabilities built with another country's huge investments?
This is why the same technology, called distillation in research papers, can be described within geopolitics as:
Technology Transfer.
Intellectual Property.
Even National Security.
Second Story: Amazon Can Buy Up to $60 Billion of Qualcomm AI Chips
Another big figure today is:
$60 billion.
Qualcomm and Amazon announced a long-term, multi-generation AI data center collaboration.
Reuters, citing regulatory filings, reports:
Amazon may purchase up to about $60 billion worth of Qualcomm AI data center chips and related products in the future.
But this figure can be easily misunderstood.
It does not mean:
Amazon placed a $60 billion order today.
It’s the maximum potential purchase scale under a long-term partnership framework.
There’s a big difference between the two.
Qualcomm Wants to Move Beyond Being Just a “Mobile Chip Company”
Many recognize Qualcomm because of:
Android phones.
Snapdragon.
5G modems.
But over the past year, Qualcomm has rapidly expanded into another market:
AI data centers.
Meaning large AI data centers.
Amazon and Qualcomm’s cooperation goes beyond general CPUs.
Focus areas include:
AI inference chips.
These chips support the heavy workload of answering user requests once the model training is complete.
Also:
Custom silicon.
Chips designed to Amazon’s specific needs.
And scaling up to 1.6 terabits per second optical connectivity—high-speed optical connection technology.
Because once AI data centers grow large enough, the challenge isn’t just:
Can chips compute fast enough?
But:
Can thousands or tens of thousands of chips quickly transfer data among themselves?
Amazon May Also Receive Qualcomm Shares
This deal has another aspect.
Reuters cites regulatory filings that Qualcomm will grant Amazon about $4 billion worth of warrants.
Warrants mean:
The right to purchase company stock at a predetermined price under specified future conditions.
These rights vest gradually, depending on Amazon’s product purchases.
So this is not simply:
“Amazon buys chips from Qualcomm.”
Rather:
The more Amazon procures, the deeper their economic interests with Qualcomm might become.
This kind of arrangement is gaining attention in AI infrastructure because:
AI requires massive capital.
Chipmakers.
Cloud providers.
Data center companies.
Model developers.
Capital markets.
They increasingly invest in, guarantee, purchase from, and become clients of each other.
For Nvidia, the Real Competition Is Also Changing
Nvidia remains the dominant AI accelerator player.
But major cloud companies are less willing to rely on a single chip provider.
Amazon is developing its own AI chips.
Google has its TPU.
Microsoft is designing its own AI silicon.
Now Qualcomm is pushing further into AI inference.
The reason is simple:
When spending hundreds of billions or more annually on AI:
Every watt of power, every chip, and every inference cost becomes a serious business consideration.
The strongest chips won’t disappear.
But enterprises will ask:
Which tasks truly require the costliest chips?
Which inferences can run on alternative hardware?
This aligns with the recent trend of “model routing.”
But now:
It's not just model routing anymore.
Hardware routing is starting too.
Third Story: OpenAI and The New York Times Seek Court Ruling on Whether AI Training Constitutes Fair Use
The third story could impact the entire generative AI industry.
OpenAI.
Microsoft.
The New York Times.
Writer groups including John Grisham and George R.R. Martin.
All are urging the Manhattan federal court to directly rule on a core question:
Is training AI on copyrighted articles and books without authorization considered Fair Use?
Fair Use is the concept of reasonable use under US copyright law.
No Court Decision Yet on Who Will Win
This is very important.
Both sides have filed:
Summary Judgment motions.
Meaning, they believe there’s enough legal and factual basis for the judge to decide without a full jury trial.
The New York Times and authors argue:
AI companies have used vast amounts of articles and books without authorization to train models.
Generative AI might compete directly with original creators in the same markets.
Therefore, this should not be protected by Fair Use.
OpenAI and Microsoft argue:
Models don’t store and reprint entire books.
The purpose of training is to learn language and statistical patterns from massive content and generate new content.
This is highly transformative use.
Currently:
The judge has not yet issued a final ruling on this fundamental dispute.
Why Is This Case Especially Important?
Because many generative AI capabilities rely on massive amounts of data:
Books.
News.
Websites.
Forums.
Images.
Code.
If the court ultimately rules:
That certain large-scale AI training qualifies as Fair Use,
model developers would gain crucial legal backing.
But if the court holds that:
Even if training is transformative, the market impact on original works is too great,
then AI firms may need to:
Obtain authorization.
Pay fees.
Limit data sources.
Or even redesign training methods.
So, the lawsuit’s true stakes are not merely:
Whether OpenAI pays The New York Times.
But rather:
The fundamental cost structure for US AI companies accessing training data in the future.
Today’s Three Stories Are One AI Battle
Putting the three together:
The first asks:
Where do model abilities come from?
The US is investigating distillation.
The second asks:
On which chips will models ultimately run?
Amazon brings Qualcomm into the large AI data center supply chain.
The third asks:
What data is a model legally allowed to use to learn?
OpenAI and publishing industries want a court to rule on Fair Use.
In other words:
AI competition used to be mainly about the model itself.
Now it’s expanding outside the model:
Data.
Other models’ outputs.
Chips.
Networks.
Power.
Law.
Licensing.
Everything is becoming a factor in whether AI can keep getting stronger.
What Does This Mean for Everyday People?
When you see a new AI model that’s strong, don’t just ask:
“What’s its score?”
Ask three more questions:
First:
Where does its ability come from?
Self-trained?
Distillation?
Synthetic data?
Outputs from other models?
Second:
What hardware supports its long-term operation?
GPU?
Custom chips?
Cloud?
Local devices?
What are the cost differences?
Third:
What rights cover its training data?
Public data?
Licensed data?
Customer-owned data?
Or data still under legal dispute?
A truly mature AI industry won’t just ask:
Who is the smartest?
It will also ask:
Can this capability be used legally, reliably, and affordably over the long term?
This may be the real competition in the next AI phase.
Today, let’s grow a little alongside AI.
Learn an AI skill daily.
Save a bit of time every day.
Improve your abilities day by day.
SasaDaily, growing with you.
