This piece has a generous 3000+ word introduction, because I want as many people to understand NVIDIA as possible. The (thousands of) words after the premium break get into arduous detail, but I’ve written this so that, ideally, most people can pick up the details early on and understand this clusterfuck.

Please do subscribe to the premium! I really appreciate it.

I've reached a point with this whole era where there are many, many things that don't make sense, and I know I'm not alone. I've been sick since Friday last week, and thus I have had plenty of time to sit and think about stuff.

And by "stuff" I mean the largest company on the stock market: NVIDIA. 

Look, I'm not an accountant, nor am I a "finance expert." I learned all of this stuff myself. I learn a great deal by coming to things from the perspective of being a dumbass, a valuable intellectual framework of "I need to make sure I understand each bit and explain it as simply as possible." In this piece, I'm going to try and explain both what this company is, how we got here, and ask questions that I, from the perspective of a dumbass, have about the company, and at least try and answer them.

Let's start with a very simple point: for a company of such remarkable size, very few people — myself included, at times! —  seem to actually understand NVIDIA.

NVIDIA is a company that sells all sorts of stuff, but the only reason you're hearing about it as a normal person is that NVIDIA's stock has become a load-bearing entity in the US stock market.

This has happened because NVIDIA sells "GPUs" — graphics processing units — that power the large language model services that are behind the whole AI boom, either through "inference" (the process of creating an output from an AI model) or "training" (feeding data into the model to make its outputs better). NVIDIA also sells other things, which I’ll get to later, but it doesn’t really matter to the bigger picture.

As an aside, NVIDIA makes other things unrelated to the chips that power large language models, like the consumer graphics cards you'd find in a gaming PC or gaming console, but the reason I'm not going to discuss these things is that 90% of NVIDIA's revenue now comes from selling either GPUs for LLMs, or the associated software and hardware to make all of that stuff run.

Back in 2006, NVIDIA launched CUDA, a software layer that lets you run (some) software on (specifically) NVIDIA graphics cards, and over time this has grown into a massive advantage for the company.

The thing is, GPUs are great for parallel processing - essentially spreading a task across multiple, by which I mean thousands, of processor cores at the same time - which means that certain tasks run faster than they would on, say, a CPU. While not every task benefits from parallel processing, or from having several thousand cores available at the same time, the kind of math that underpins LLMs is one such example. 

CUDA is proprietary to NVIDIA, and while there are alternatives (both closed- and open-source), none of them have the same maturity and breadth. Pair that with the fact that Nvidia’s been focused on the data center market for longer than, say, AMD, and it’s easy to understand why it makes so much money. There really isn’t anyone who can do the same thing as NVIDIA, both in terms of software and hardware, and certainly not at the scale necessary to feed the hungry tech firms that demand these GPUs.

Anyway, back in 2019 NVIDIA acquired a company called Mellanox for $6.9 billion, beating off other would-be suitors, including Microsoft and Intel. Mellanox was a manufacturer of high-performance networking gear, and this acquisition would give NVIDIA a stronger value proposition for data center customers. It wanted to sell GPUs — lots of them — to data center customers, and now it could also sell the high-speed networking technology required to make them work in tandem. 

This is relevant because it created the terms under which NVIDIA could start selling billions (and eventually tens of billions) of specialized GPUs for AI workloads. As pseudonymous finance account JustDario connected (both Dario and Kakashii have been immensely generous with their time explaining some of the underlying structures of NVIDIA, and are worth reading, though at times we diverge on a few points), mere months after the Mellanox acquisition, Microsoft announced its $1 billion investment in OpenAI to build "Azure AI supercomputing technologies."

Though it took until November 2022 for ChatGPT to really start the fires, in March 2020, NVIDIA began the AI bubble with the launch of its "Ampere" architecture, and the A100, which provided "the greatest generational performance leap of NVIDIA's eight generations of GPUs," built for "data analytics, scientific computing and cloud graphics." The most important part, however, was the launch of NVIDIA's "Superpod": Per the press release: 

A data center powered by five DGX A100 systems for AI training and inference running on just 28 kilowatts of power costing $1 million can do the work of a typical data center with 50 DGX-1 systems for AI training and 600 CPU systems consuming 630 kilowatts and costing over $11 million, Huang explained.

One might be fooled into thinking this was Huang suggesting we could now build smaller, more efficient data centers, when he was actually saying we should build way bigger ones that had way more compute power and took up way more space. The "Superpod" concept — groups of GPU servers networked together to work on specific operations — is the "thing" that is driving NVIDIA's sales. To "make AI happen," a company must buy thousands of these things and put them in data centers and you'd be a god damn idiot to not do this and yes, it requires so much more money than you used to spend.

At the time, a DGX A100 — a server that housed eight A100 GPUs (starting at around $10,000 at launch per-GPU, increasing with the amount of on-board RAM, as is the case across the board) — started at $199,000. The next generation SuperPod, launched in 2022, was made up of eight H100 GPUs (Starting at $25,000-per-GPU, the next generation "Hopper" chips were apparently 30x times more powerful than the A100), and retailed from $300,000.

You'll be shocked to hear the next generation Blackwell SuperPods started at $500,000 when launched in 2024. A single B200 GPU costs at least $30,000.

Because nobody else has really caught up with CUDA, NVIDIA has a functional monopoly (edit: I wrote monopsony in a previous version, sorry), and yes, you can have a situation where a market has a monopoly, even if there is, at least in theory, competition. Once a particular brand — and particular way of writing software for a particular kind of hardware — takes hold, there's an implicit cost of changing to another, on top of the fact that AMD and others have yet to come up with something particularly competitive.

Anyway, the reason that I'm writing all of this out is because I want you to understand why everybody is paying NVIDIA such extremely large amounts of money. Every year, NVIDIA comes up with a new GPU, and that GPU is much, much more expensive, and NVIDIA makes so much more money, because everybody has to build out AI infrastructure full of whatever the latest NVIDIA GPUs are, and those GPUs are so much more expensive every single year.

NVIDIA’s Latest Generation Blackwell GPUs Require Entirely New Servers - And If You Want To Run Lots Of Them, An Entirely New Data Center, Because They Require So Much More Power And Cooling

With Blackwell — the third generation of AI-specialized GPUs — came a problem, in that these things were so much more power-hungry, and required entirely new ways of building data centers, along with different cooling and servers to put them in, much of which was sold by NVIDIA. While you could kind of build around your current data centers to put A100s and H100s into production, Blackwell was...less cooperative, and ran much hotter.

To quote NVIDIA Employee Number 4 David Rosenthal:

The systems are estimated to be more than half the capex for a new data center. Much of its opex is power. Just as with mining rigs, the key feature of each successive generation of AI chips is that it is more efficient at using power. But that doesn't mean they use less power, they use more but less per operation. The need for enhanced power distribution and the concomitant cooling is what has prevented new AI systems being installed in legacy data centers. Presumably the next few generations will be compatible with current state of the art data center infrastructure, so they can directly replace their predecessors and thereby reduce costs.

In simple terms, Blackwell runs hot, so much hotter than Ampere (A100) or Hopper (H100) GPUs that it requires entirely different ways to cool it, meaning your current data center needs to be ripped apart to fit them.

Huang has confirmed that Vera Rubin, the next generation of GPUs, will have the same architecture as Blackwell. I would bet money that it's also much more expensive.

Anyway, all of this has been so good for NVIDIA. As the single vendor for the most important component in the entire AI boom, it has set the terms for how much you pay and how you build any and all AI infrastructure. While there are companies like Supermicro and Dell who buy NVIDIA GPUs and ship them in servers to customers, that's just fine for NVIDIA CEO Jensen Huang, as that's somebody else selling his GPUs for him.

NVIDIA has been printing money, quarter after quarter, going from a meager $7.192 billion in total revenue in the third (calendar year) quarter of 2023 to an astonishing $50 billion in just data center revenue (that's where the GPUs are) in its most recent quarter, for a total of $57 billion in revenue, and the company projects to make $63 billion to $67 billion in the next quarter.

Now, I'm going to stop you here, because this bit is really important, really simple, yet nobody thinks about it much: NVIDIA makes so much money, and it makes it from a much smaller customer base than most companies, because there are only so many entities that can buy thousands of chips that cost $50,000 or more each. 

$35 billion, $39 billion, $44 billion, $46 billion and $57 billion are very large amounts of money, and the entities pumping those numbers into the stratosphere are collectively having to spend hundreds of billions of dollars to make it happen.

Let Me Explain To You How Expensive And Arduous It Is To Actually Build A Data Center, And Remind You That Every Single Time Anybody Buys NVIDIA GPUs They're Probably Dealing With This

So, let me give you a theoretical example. I swear I'm going somewhere with this.You, a genius, have decided you are about to join the vaunted ranks of "AI data center ownership." You decide to build a "small" AI data center — 25MW (megawatts, which in this example, refers to the combined power draw of the tech inside the data center). That can't be that much, right? OpenAI is building a 1.2GW one out in Abilene Texas. How much could this tiny little thing cost?

Sidenote: It’s a minor thing, but I want to clarify something. I said “in this example” in the previous paragraph because when we talk about the power capacity of a data center, we could be referring to one of two things. The first is the power draw of the servers in the facility — which is called the IT Load — or the total amount of power that can be provided to that facility.

Here’s where it gets tricky. A facility that can draw, say, 25MW of power from the grid can’t just use all of that in one go. You need a reserve for what’s known as the “design day,” which is the hottest day of the year, when the facility’s cooling systems are under the most strain, and when power transmission losses are at their highest. That reserve is, from what I’ve been told, around 30% of the total available electricity. 

Cooling systems are power-hungry! Who knew? Me. I did. I’ve been telling you for over a year.

Okay, well, let's start with those racks. You're gonna need to give Jensen Huang $600 million right away, as you need 200 GB200 racks. You're also gonna need a way to make them network together, because otherwise they aren't going to be able to handle all those big IT loads, so that's gonna be another $80 million or more, and you're going to need storage and servers to sync all of this up, which is, let's say, another $35 million.

So we're at $715 million. Should be fine, right? Everybody's cool and everybody's normal. This is just a small data center after all. Oops, forgot cooling and power delivery stuff — that's another $5 million. $720 million. Okay.

Anyway, sadly data centers require something called a "building." Construction costs for a data center are somewhere from $8 million to $12 million per megawatt, so, crap, okay. That's $250 million, but probably more like $300 million. We're now up to $1.02 billion, and we haven't even got the power yet.

Okay, sick. Do you have one billion dollars? You don't? No worries! Private credit — money loaned by non-banking entities — has been feeding more than $50 billion dollars a quarter into the hungry mouths of anybody who desires to build a data center. You need $1.02 billion. You get $1.5 billion, because, you know, "stuff happens." Don't worry about those pesky high interest rates — you're about to be printing big money, AI style!

Now you're done raising all that cash, it'll now only take anywhere from 6 to 18 months for site selection, permitting, design, development, construction, and energy procurement. You're also going to need about 20 acres of land for that 100,000 square foot data center. You may wonder why 100,000 square feet needs that much space, and that's because all of the power and cooling equipment takes up an astonishing amount of room.

So, yeah, after two years and over a billion dollars, you too can own a data center with NVIDIA GPUs that turn on, and at that point, you will offer a service that is functionally identical to everybody else buying GPUs from NVIDIA.

Your competitors are Amazon, Google and Microsoft, followed by neoclouds — AI chip companies selling the same thing as you, except they're directly backed by NVIDIA, and frequently, the big hyperscaler companies with brands that most people have heard of, like AWS and Azure.

Oh, also, this stuff costs an indeterminately-large amount of money to run. You may wonder why I can't tell you how much, and that's because nobody wants to actually discuss the cost of running GPUs, the thing that underpins our entire stock market.

There're good reasons, too. One does not just run "a GPU" — it's a GPU in a server of other GPUs with associated hardware, all drawing power in varying amounts, all running in sync with networking gear that also draws power, with varying amounts of user demand and shifts in the costs of power from the power company.

But what we can say is that the up front cost of buying these GPUs and their associated crap is such that it's unclear if they ever will generate a profit, because these GPUs run hot, all the time, and that causes some amount of them to die.

NVIDIA's Future Success Is Predicated On The Ability For Everybody To Keep Feeding It Tens Of Billions Of Dollars Every Quarter, Requiring Endless Debt, Space, and Actual AI Demand To Exist

Here are some thoughts I have had:

• A 25MW data center costs about $1 billion, with $600 million of that being GPUs — 200 GB200 racks, to be specific.
• It needs about 20 acres — 100,000 square feet for the data center, roughly.
• NVIDIA sells about $50 billion of GPUs and associated hardware in a quarter, so let's say that $40 billion of that is just the GPUs and $10 billion is everything else (primarily networking gear), so around 13,333 GB200 racks. I realize that NVIDIA sells far more than that (GB300 racks, singular GPUs, and so on).
• Deep-pocketed hyperscalers like Microsoft, Google, Meta and Amazon representing 41.32% of NVIDIA's revenue in the middle of 2025, funneling free cash flow directly into Jensen Huang's pockets...
  • ...for now. Amazon ($15 billion), Google ($25 billion), Meta ($30 billion) and Oracle ($18 billion) have all had to raise massive amounts of debt to continue to fund AI-focused capital expenditures, with more than half of that (per Rubenstein) spent on GPUs.
• Otherwise, basically anybody buying GPUs at any scale has to fund doing so with either venture capital (money raised in exchange for part of the company) or debt.
• NVIDIA, at this point, is around 8% of the value of the S&P 500 (the 500 leading (meaning they meet certain criteria of size, liquidity (cash availability) and profitability) companies on the US stock market). Its continued health — and representative value as a stock, which is not necessarily based on its actual numbers or health, but in this case kind of is? — has led the stock market to remarkable gains.
• It is not enough for NVIDIA to simply be a profitable company. It must continue beating the last quarter's revenue, again and again and again and again, forever. If that sounds dramatic, I assure you it is the truth.
• NVIDIA's continued success — and its ability to continue delivering outsized beats of Wall Street's revenue estimates — depends on:
  • The willingness of a few very large, cash-rich companies (Microsoft, Meta, Amazon and Google) to continue buying successive generations of NVIDIA GPUs forever.
  • The ability of said companies to continue buying successive generations of GPUs forever.
  • The ability of other, less-cash-rich companies like Oracle to continue being able to raise debt to buy massive amounts of GPUs — such as the $40 billion of GPUs that Oracle is buying for Stargate Abilene forever.
    • This is becoming a problem.
  • The ability of unprofitable, debt-ridden companies like CoreWeave, AI "neoclouds" that use the GPUs they purchase from NVIDIA as collateral for loans to buy more GPUs, to to continue raising that debt to buy more GPUs.
  • The ability of anybody who buys these GPUs to actually install them and use them, which requires massive amounts of construction...and more power than is currently available, even to the most well-funded and conspicuous projects.
  • In simple terms, its success depends on the debt markets to continue propping up its revenues, because there is not really enough free cash in the world to continue pumping it into NVIDIA at this rate.
  • And after all of this, large language models, the only way to make any real money on any of these GPUs, must prove they can actually produce a profit.
    • Per my article from September, I can find no compelling evidence (outside of boosters speciously claiming otherwise) that it's profitable to sell access to GPUs.
    • Based on my calculations, there's likely little more than $61 billion of actual AI revenue in 2025 across every single AI company and hyperscaler.
    • Note that I said "revenue." Absolutely nobody is making a profit.

The NVIDIA situation is one of the most insane things I've seen in my life.

The single-largest, single-most-valuable, single-most-profitable company on the stock market has got there through selling ultra-expensive hardware that takes hundreds of millions or billions of dollars (and years of construction in some cases) to start using, at which point it...doesn't make much revenue and doesn't seem to make a profit. 

Said hardware is funded by a mixture of cashflow from healthy businesses (see: Microsoft) or massive amounts of debt (see: everybody who is not a hyperscaler, and, at this point, some hyperscalers). The response to the continued proof that generative AI is not making money is to buy more GPUs, and it doesn't appear anybody has ever worked out why.

This problem has been obvious for a long time, too. 

Today I'm going to explain to you — simply, but at length — why I am deeply concerned, and how deeply insane this situation has become.