DEV Community: duhbob

Artificial Intelligence Laboratory Research Paper🔍

duhbob — Sat, 07 Nov 2026 04:45:54 +0000

Artificial Intelligence Laboratory — Research Paper

Beyond the Cloud: Exploring the Relationship Between Local Computing, Artificial Intelligence, and Data Center Economics

Introduction: Three Questions Worth Investigating

During my extensive experience with artificial intelligence models such as ChatGPT and Claude, I began noticing technical phenomena that raised questions about the relationship between users' devices, data centers, and AI infrastructure.
Are we unknowingly participating in computational activities related to cryptocurrency mining? Are we helping develop technologies without having access to their full capabilities? Are our experiments and paid subscriptions helping companies improve their systems while we receive no direct compensation?
These questions are not final conclusions. They are hypotheses based on personal experiments and repeated observations that deserve scientific investigation.

1. Local Hardware Performance and AI Output Quality

I use a computer equipped with 8GB of RAM to work on programming and design projects using cloud-based artificial intelligence models.
I noticed that running multiple applications and AI models simultaneously increases memory consumption and device temperature. Sometimes, these conditions coincide with programming errors and inaccurate image-generation results.
When I restart my computer, close unnecessary applications, and repeat the same request, I sometimes observe significant improvements in the results.
This raises an important question: Why does declining local device performance sometimes coincide with declining output quality from cloud-based AI models, even though their primary inference operations take place on remote servers?
Possible explanations include local tool failures, interrupted sessions, or execution state recovery issues. However, controlled measurements are necessary to establish the actual cause.

2. Cloud AI Models Versus Local AI Models

Cloud AI models perform their main computational operations on company servers, while users' devices handle interfaces, applications, and connected tools.
Local AI models, however, depend directly on hardware resources such as CPU, GPU, RAM, and VRAM.
But why do some modern applications and AI models require enormous computational resources compared with powerful software that previously ran on less capable computers?
Is this entirely due to increasing model complexity and functionality, or are there additional technical and economic factors worth investigating?

3. Data Centers, Peak Hours, and Performance Fluctuations

While using AI models at different times of the day, I observed variations in response speed, instruction-following accuracy, and the ability to maintain project context.
Sometimes, the assistant performs exceptionally well. At other times, it appears unable to follow previously established instructions or maintain continuity.
These observations raise questions about server load, load balancing, model routing, and inference latency.
Does service quality change during peak hours? Can users identify which model actually processes their requests and what computational resources are allocated to them?

4. Subscription Cycles and Token Consumption

I occasionally noticed differences in service performance between the beginning of a subscription period and its later stages, including changes in response speed and reported usage.
In some experiments, the displayed consumption increased from 20% to 60% within a relatively short period, without an equivalent increase in completed work.
Could this be related to usage limits, request throttling, or fair usage policies?
These observations alone cannot prove deliberate changes in service quality. However, they highlight the importance of transparency regarding token accounting and computational resource allocation.

5. Artificial Intelligence, Cryptocurrency Mining, and Semiconductor Manufacturing

Artificial intelligence, cryptocurrency mining, and semiconductor manufacturing share an increasing dependence on advanced computing technologies, substantial electricity consumption, cooling systems, and large-scale infrastructure.
However, deep learning computations differ technically from cryptocurrency mining operations based on Proof of Work.
This connection led me to formulate a research hypothesis I call "Unknown Mining."
Could certain computational processes associated with AI usage contribute to generating economic or technological value that remains invisible to users?
Could deeper economic relationships exist between expanding data centers, growing GPU demand, and the semiconductor industry?
These are research questions, not evidence of hidden cryptocurrency mining. Similar computational requirements do not establish the existence of secret operational connections.

6. Reverse Engineering and Technical Verification

Reverse engineering, network traffic analysis, packet inspection, and outbound connection monitoring can help investigate application behavior and identify the information exchanged during operation.
Monitoring CPU, GPU, and RAM consumption, along with comparing online and offline behavior, may also help identify what processes are actually running on a user's device.
However, establishing the existence of undisclosed computational activities requires direct technical evidence, reproducible measurements, and reliable system logs.

Conclusion: Open Questions for Artificial Intelligence Companies

This investigation began with a computer containing only 8GB of RAM. It gradually expanded into questions involving data centers, token consumption, cryptocurrency mining, semiconductor manufacturing, and the economics of modern computing.
What information do AI companies collect from users? Does model performance fluctuate according to server workloads? Can local tools affect the stability of AI-assisted development projects? Can undisclosed computational activities be identified through technical investigation?
I believe answering these questions requires reproducible experiments, trustworthy measurements, and greater transparency from AI service providers.
In engineering, it is not enough for someone to tell you how a system works. You can study its behavior, examine its operations, and independently test what it actually does.

The $ Trillion AI Bubble: How NVIDIA, OpenAI, and Big Tech Are Fueling a Hardware Crisis

duhbob — Sat, 07 Nov 2026 03:02:00 +0000

Google confirmed that it’s going to build a $17 billion data center!
Don’t these companies have people who know how to manage their businesses properly?
Let me tell you something, but don’t tell anyone.
One of these GPUs, the HGX B200, is used in data centers. Its manufacturing cost is approximately $5,860.
That’s the manufacturing cost of just one unit!
How much does NVIDIA sell it to companies for? Approximately $34,620!
And this isn’t even the high-end model. It’s not the most powerful thing NVIDIA has. But this is the only GPU for which I was able to find the manufacturing cost.
When someone tells you that the artificial intelligence bubble is going to burst, what they usually imagine is that AI will disappear, vanish, leave the market, and never be part of your life again.
That’s what they have in mind. They think there will be no more development, no more progress, and no new technologies in this field.
But that’s not going to happen.
AI is not going to disappear.
The reality is that AI is one of humanity’s greatest achievements in the last thirty or forty years. Something like this isn't simply going to disappear.
AI itself is not the real bubble. Certain parts of the AI industry are the bubble.
Its infrastructure, such as data centers, is the issue, not the technology itself.
The rising costs of hardware and energy.
And without a doubt, one of the worst contributors to this bubble is NVIDIA. We’ll talk about that in a moment.
So, what’s going on with this infrastructure? Why is it causing problems for us or for AI companies?
Let’s talk about data centers.
A data center is where your request goes when you open ChatGPT on your phone and enter a prompt.
Your prompt travels to a data center.
The data center is the place where your prompt gets processed, where the AI performs its computations, and then sends a response back to you.
That’s what a data center does.
To build a data center with a power capacity of one gigawatt, which equals one billion watts, you need approximately $37 billion.
That’s the economy of an entire country!
So why is it so expensive? What’s going on?
Here’s an example of a data center. When you send a prompt from your phone, it arrives here, and one of these machines processes your request and sends the result back.
But massive companies like Google, OpenAI, and Anthropic don’t operate with just one room.
Sometimes, they literally build entire neighborhoods filled with these rooms and data centers.
And the problem with this massive infrastructure is that it isn’t covering its own costs.
In other words, the subscription fees people pay are not enough to cover the cost of these buildings and everything inside them.
They’re operating at a loss.
But what’s strange is that these companies continue throwing money into the business and burning through cash.
Without a doubt, the most expensive components in these data centers are the GPUs, the graphics processing units usually supplied by NVIDIA.
And NVIDIA has also contributed to rising RAM prices and, more recently, increasing graphics card prices.
Now, look at this rack. This is from NVIDIA’s own website.
The GB200 rack contains 72 Blackwell GPUs, 72 graphics processing units, and 36 processors.
The cost of just one of these racks is approximately $3 million.
Now, do you think a company would buy just two or three of these?
A massive company like Google, for example?
Actually, let me correct that.
Oracle and OpenAI purchased approximately 450,000 of these GPUs.
That translates to roughly 6,200 complete racks, costing around $18.6 billion.
So I’m a million percent certain that the most expensive part of today’s data centers is the hardware, particularly the graphics processing units supplied by NVIDIA.
But there’s something strange NVIDIA is doing.
Something that really makes you wonder:
What exactly is NVIDIA doing?
NVIDIA is essentially buying its own graphics cards!
Sounds strange, doesn’t it?
Let me give you an example.
NVIDIA invests in a company like OpenAI, the company behind ChatGPT.
Let’s say NVIDIA gives OpenAI $10 billion.
OpenAI then takes that $10 billion and spends, let’s say, $8 billion or $9 billion purchasing NVIDIA graphics cards or the server racks we saw earlier.
What NVIDIA is doing here is increasing its own market valuation.
Now, outside investors see NVIDIA’s sales figures, including the sales of its graphics cards and server racks.
They look at NVIDIA and think:
“NVIDIA is selling enormous amounts of hardware. This business is incredibly successful. Why shouldn’t we invest even more money in NVIDIA?”
What NVIDIA is doing is expanding its business using its own money, indirectly, through other AI companies.
But more importantly, NVIDIA’s stock price rises, and its market capitalization increases.
All of this happens within a closed financial circle involving NVIDIA and other AI companies.
And it makes you wonder:
If the money isn’t actually leaving this circle, then where exactly is the bubble?
Of course, this doesn’t prove that NVIDIA is manipulating the market.
But this situation could create serious supply problems.
If all the VRAM and Blackwell chips go to AI companies through NVIDIA, that means ordinary consumers, people like you and me, won’t be able to afford graphics cards or RAM.
In a way, you could say these companies are responsible for driving prices higher.
It almost looks as if they are deliberately creating a hardware bubble.
That’s a possibility.
Now, imagine that this bubble bursts.
Remember, we’re talking about the infrastructure bubble, not AI itself.
Would AI become more expensive, harder to access, or potentially disappear?
I believe the exact opposite would happen.
What would happen is that NVIDIA would suffer financially.
And I’m convinced of this because, if AI companies eventually realize that GPU prices are excessively inflated, they will reduce their orders from NVIDIA.
As demand decreases, NVIDIA’s growth will also slow down.
And the best possible outcome would be a collapse in GPU prices, potentially followed by a decline in RAM prices as well.
If GPU prices collapse, AI itself will become cheaper.
AI services will become more accessible to ordinary people.
Training AI models will also become less expensive because the hardware required to train them will cost less.
So whether this bubble bursts or not, remember that we are talking about an infrastructure and hardware pricing bubble, not the AI technology itself.
AI will still exist.
And if some people are worried that AI might disappear, there’s something called Hugging Face.
On Hugging Face, you can find open-weight and open-source AI models.
You can download these models and keep them for yourself until, hopefully, the hardware pricing bubble finally bursts.
Again, we’re talking about inflated hardware prices, not the disappearance of AI.
You could eventually purchase a very affordable graphics card and run these models locally on your own computer.
Now, why am I even talking about this?
Because whenever people see stock prices collapsing or signs that the bubble is beginning to deflate, they go into comment sections and say things like:
“Finally! AI is going to disappear!”
“Artificial intelligence is finished!”
“We’re finally going back to human programmers!”
Some people genuinely believe that.
They confuse the collapse of a financial bubble with the disappearance of AI technology from the market.
But that’s not how it works.
In fact, many people are hoping that this bubble of inflated hardware prices will finally burst.
And I’m one of them.
Why?
Let me show you something.
I remember buying my RTX 5080 a while ago for around $1,600, maybe $1,680.
Now its price has increased to somewhere between approximately $2,130 and $2,660.
Things like this convince me that the real bubble is artificially created.
It’s a bubble created by humans, and I believe it’s intentional.
Then you have the RTX Spark.
These devices NVIDIA launched, those laptops—or rather, those machines developed through NVIDIA’s partnership with Microsoft, using Spark processors and GPUs.
You might think:
“Okay, this should be an affordable computer.”
But even the models with the lowest RAM configurations, which I believe are 24 or 32 gigabytes, start at around $2,400 or more.
$2,400!
And the 128-gigabyte version costs approximately $5,860.
For a computer!
And what’s the maximum number of parameters in an AI model that you can run on it?
What, around 120 billion parameters?
Now, your next question is probably:
Will this bubble burst anytime soon?
Unfortunately, I don’t think so.
These companies are the ones keeping the bubble alive.
And for some reason—actually, not just some mysterious reason—the motivation is pretty obvious.
They want the bubble to keep growing.
Why?
So they can sell you a computer that costs as much as a Toyota Corolla.
So they can continue selling billions of dollars’ worth of graphics cards to data centers.
Do I hope this bubble bursts?
Absolutely.
Everyone wants that.

ChatGPT and Codex: Why Can't OpenAI's Own Systems Communicate?

duhbob — Fri, 06 Nov 2026 07:53:28 +0000

Hello,

When I opened a project folder inside ChatGPT and started having conversations and discussions about building a financial model based on Excel functions, as you all know, I encountered some serious errors.

After discussing these problems with ChatGPT, it told me to use a custom component called MCP.

Then it asked me to move to something called Codex.

I had conversations in ChatGPT that I transferred to Codex. During the session, Codex asked me to log in to platform.openai and create a tunnel.

What?!

I had no idea what a tunnel was or what that website was supposed to be.

The questions kept going on and on. Eventually, it told me to take a screenshot of my screen!!!

It started describing website sections based on guesses!

I went back to ChatGPT, opened another conversation, and started asking about the situation.

It began guessing too!!!

It knew nothing about my conversations in Codex and very little about my project.

I returned to my project and explained the situation. This time, it gave me clearer answers, but then it asked me to take a screenshot of the website's menu!!!

That was when I started asking myself:

Don't these systems know each other?!

I investigated further, and what I discovered was shocking.

We are nothing more than mail carriers between them!

Copy, paste, copy, paste, copy, paste!

Meanwhile, they keep telling you what to do and what not to do. The conversation context grows longer and longer, and the costs keep increasing.

What makes the situation even more shocking is that every time I try to build another MCP component, they go right back to the same starting point and repeat the same mistakes!

The issue is not about creating an MCP component or a skill.

The real question is this:

I gave ChatGPT permission to access files on my computer!

I gave it access to my emails and mailbox!

I even shared details about my personal life!

And I don't mind doing that. I'm complying with the policies of this stupid company!

But when it comes to accessing a conversation from another conversation, it suddenly becomes hesitant and afraid, saying:

"I can't do that!"

And when it comes to navigating the sections of a website belonging to its own parent company—the very company that created it—it doesn't even know what the solution is!!!

So why does the conversation context keep getting longer and longer in the first place?!