There are only two AI updates worth noting today.
But looking at them together makes more sense than forcing in a third.
They precisely represent AI computing power moving in two completely opposite directions:
On one side:
Large-scale AI is so expensive that even GPUs are becoming financial assets.
On the other side:
Agents that once had to run in the cloud are now starting to run locally on desktops.
1. Report: Amazon plans to sell $8 billion worth of Nvidia chips then lease them back
Let’s first look at the more counterintuitive update.
Amazon has obviously invested heavily in AI chips, so why would it now prepare to:
Sell those chips to others?
According to people familiar with the matter cited by Financial Times, Amazon is currently discussing with investors the idea of moving approximately:
$8 billion
worth of Nvidia advanced AI chips into a Special-Purpose Vehicle, or SPV.
These are not future orders yet to be delivered.
The report says this includes thousands of Grace Blackwell chips already installed across more than ten data centers in five U.S. states.
After selling, Amazon continues to use the chips
This is the crucial part.
Amazon isn’t:
Selling chips → Stopping AI operations.
Rather, it might use a:
Sell → Lease Back
model.
Simply put:
External investors own the chips held in the SPV,
and then Amazon leases the same equipment back from the SPV.
Operationally,
the chips remain in the data centers,
continuing to run AI for Amazon.
What changes is:
Who officially owns these expensive assets on the balance sheet.
Why would such a wealthy company do this?
Because the scale of AI infrastructure costs has grown so large that the question is no longer:
“Does the company have money to buy GPUs?”
But rather:
“How to properly account for hundreds of billions in capital expenditure on the balance sheet?”
AI chips themselves are very expensive.
At the same time, hardware inevitably:
depreciates,
gets replaced,
and becomes outdated by next-generation products.
If a cloud company permanently owns all its GPUs, it must not only upfront substantial cash but also bear:
asset depreciation,
technological obsolescence,
and declining asset value.
So placing some hardware in an outside SPV and continuing to use it through leasing
helps achieve a lighter asset structure.
This approach is common in industries like aviation, real estate, and infrastructure.
Now,
AI GPUs are starting to follow the same path.
Investors are not simply buying ‘a graphics card’
If this deal goes through,
external investors are effectively betting that
this Nvidia AI hardware will still be in demand for years to come.
Amazon would pay lease fees to continue using it.
Financial Times also reports that the SPV could raise funds by issuing bonds,
and Amazon might offer up to around 10% equity stake.
In other words,
AI compute is starting to become a kind of infrastructure asset that can be:
held,
financed,
bonded,
and leased.
This is quite far from the ordinary notion of:
“Tech companies buying some GPUs.”
But the deal is not finalized yet
This must be clarified.
Currently, Financial Times cites insiders saying Amazon is discussing such an arrangement with investors.
Reuters also reported that neither Amazon nor Nvidia immediately responded to requests for comments.
The proper way to say this now is:
Amazon is reportedly exploring or negotiating this structure.
It’s not correct to say:
“Amazon has officially sold $8 billion worth of chips.”
This isn’t the first sign of AI financing complexity recently
Just the day before, we covered:
Broadcom agreeing to provide up to $42 billion in financing to Anthropic,
helping fund its long-term compute capacity.
Earlier, the market has seen collaborations involving:
chipmakers,
cloud providers,
AI labs,
and investors
providing funding,
guarantees,
chips,
data center capacity.
Now Amazon introduces another direction:
financializing already deployed GPU assets.
This means competition for AI infrastructure no longer revolves just around:
who can get chips.
It’s increasingly about:
who can hold compute power long-term with the lowest capital costs.
2. Meanwhile, Nvidia is doing the complete opposite: bringing AI compute back to the desktop
On the same day,
Nvidia unveiled a new product:
DGX Spark 64GB.
This is not a large data center rack.
It’s a desktop-level AI computer.
The new model features:
64GB Unified Memory.
Starting October 23, it will be offered by partners including:
Acer,
ASUS,
Dell,
Gigabyte,
HP,
and MSI.
Nvidia’s official starting price is:
$4,999.
Why is 64GB considered modest compared to many high-end computers? What’s the real point?
The key isn’t just the RAM size.
The DGX Spark uses Nvidia’s GB10 Grace Blackwell superchip,
and the whole stack including:
DGX OS,
CUDA-X AI Libraries,
Agent Toolkit,
and Local Model Runtime
is integrated.
Nvidia isn’t selling simply:
“A computer with lots of memory.”
But rather:
A desktop system ready out of the box to develop local AI agents.
Officially supported environments include:
Ollama,
vLLM,
PyTorch,
and llama.cpp,
among common local AI tools.
Nvidia says a single machine can run models up to roughly 100B parameters
According to Nvidia, one 64GB DGX Spark can support local workloads on models with approximately:
100 billion parameters
The actual smoothness depends on factors like:
model quantization,
context length,
KV cache,
number of agents,
inference speed,
and other variables.
So “supporting 100B” does not mean every 100B model can run fast in all setups.
But the direction is clear:
tasks at this scale traditionally meant:
renting cloud GPUs.
Nvidia wants to shift more of these tasks
to the desktop.
Even more interesting: two units can be directly connected
If 64GB isn’t enough,
Nvidia’s Sync Cluster Assistant lets two DGX Spark units connect directly.
The company states that two 64GB machines can pool to:
128GB memory
and support models with up to roughly:
200 billion parameters.
The Cluster Assistant can automatically:
detect the second device,
verify settings,
and configure ConnectX-7 networking,
so developers don’t have to set up the cluster manually.
What really matters here is not simply “two computers connecting.”
It’s that Nvidia is attempting to bring what used to be:
small-scale data center
functionality
to
standard desktop equipment features.
Why is local AI suddenly important now?
Not because cloud AI is bad.
But Agents now create new demands.
For example, a coding agent may:
read repositories all day,
run tests,
and revise code repeatedly.
A research agent may:
analyze large document sets over long periods.
Businesses may have data they:
simply don’t want to send continuously to external clouds.
Local AI offers several practical advantages:
First:
Data can stay local.
Second:
No recurring cloud inference fees.
Third:
Agents can run continuously over long periods.
Fourth:
Operations won't halt if network or API services drop.
But there are trade-offs:
You must buy hardware upfront,
maintain systems,
update models,
handle capacity,
and manage security yourself.
So local and cloud AI will coexist rather than completely replacing each other.
What’s truly interesting today is this contrast
In the large-scale AI world, Amazon is facing:
How to financially manage $8 billion in chips?
At the personal developer level, Nvidia faces:
How to pack enough AI compute into a desktop computer?
One end turns GPUs into:
financable infrastructure assets.
The other turns GPUs into:
personal AI infrastructure at developers’ desks.
This suggests AI’s future may not be just:
cloud,
or entirely:
local,
but rather a three-tiered model.
First layer: Super large cloud AI
Handles:
cutting-edge frontier models,
massive training,
and huge inference demands.
Requires data centers,
power,
and large-scale financing.
Second layer: Enterprise or team private compute
Contains sensitive:
company documents,
codebases,
research data,
and long-term agents
likely running on company-controlled hardware.
Third layer: Typical cloud API
Used for:
latest models,
occasional big jobs,
and elastic capacity.
Then calls cloud.
In other words,
not all workloads must go to the same place.
What does this mean for everyday users?
When you use ChatGPT, Claude, or Gemini,
you normally don’t need to think about:
where the GPUs powering them actually sit.
But as agents become more common,
this will start to affect:
cost,
privacy,
speed,
service reliability,
and even what equipment your company procures.
In the future, companies won’t just ask:
“Should we use AI?”
But rather:
“Which tasks require frontier cloud models?”
“Which data should remain in-house?”
“Which agents should run locally year-round?”
Today’s stories from Amazon and Nvidia present both ends of that future.
Recommended Reading
When chip sellers start guaranteeing buyers: Is the AI boom entering a risky cycle?
