The smartest people in the room are currently making the dumbest mistake in financial history. They are treating a hardware manufacturer like a software monopoly.
Everyone is staring at the same chart—a vertical line representing the market cap of the world’s most valuable company—and calling it a "defiance of expectations." It isn't. It is a fundamental misunderstanding of the physics of production. When the financial press screams that the "AI bubble is intact," they are accidentally admitting they don’t understand how bubbles actually pop. Bubbles don't end when people stop believing; they end when the marginal utility of the next dollar spent drops below zero.
We are fast approaching that cliff.
The consensus view is lazy. It suggests that as long as big tech companies keep ordering H100s and Blackwell chips, the party continues. This assumes that demand for compute is infinite. It isn't. Compute is a commodity, and right now, we are overpaying for the luxury version of a product that is rapidly being democratized.
The Margin Trap
Every software analyst on Wall Street wants to believe NVIDIA is the next Microsoft. They see the 70% gross margins and the "moat" of CUDA, and they salivate.
But here is the truth they won't tell you: NVIDIA is a hardware company.
Hardware is subject to the brutal, unforgiving laws of the physical world. Software can be replicated for zero marginal cost. A chip, however, requires silicon, power, logistics, and a supply chain that is currently being squeezed from both ends.
When you look at companies like Apple or Microsoft, their dominance is built on consumer lock-in. Once you are in the ecosystem, you stay. NVIDIA's "lock-in" is purely technical. Engineers use CUDA because it is the fastest way to get work done right now. The moment a competitor provides a more cost-effective alternative—whether it is AMD's MI300 series or the custom silicon being built in-house by Amazon, Google, and Meta—the moat evaporates.
The big tech giants are not NVIDIA's partners. They are its biggest threat. They are currently spending billions on NVIDIA chips because they have to, but they are simultaneously spending billions more to ensure they never have to again.
Why the "Moat" is a Mirage
Let's dismantle the CUDA argument once and for all. People say developers will never switch. I’ve spent twenty years watching developers switch. They switched from Perl to Python. They switched from monolithic architectures to microservices. They will switch from CUDA to Triton or Mojo the second the price-to-performance ratio tilts.
Developers are pragmatic, not loyal.
The moment the cost of porting code becomes lower than the "NVIDIA tax," the migration begins. We are already seeing the first cracks in the foundation. PyTorch 2.0 and the OpenAI Triton compiler are specifically designed to make the underlying hardware irrelevant.
If the hardware becomes irrelevant, NVIDIA’s valuation becomes a joke.
The Efficiency Paradox
We are currently in the "brute force" phase of AI development. The industry believes that if you throw more compute at a problem, the model gets smarter. This is the scaling law fallacy.
History shows that intelligence follows a curve of diminishing returns. To get from GPT-4 to whatever comes next, we are told we need 10 times the compute. To get to the step after that? 100 times.
This is not a sustainable business model. It is a suicide pact.
The true breakthrough in AI won't come from a bigger GPU cluster. It will come from an algorithmic shift that makes the current clusters obsolete. We are already seeing "Small Language Models" (SLMs) that outperform their massive ancestors while running on a fraction of the power.
Imagine a scenario where a company spends $10 billion on a massive data center today, only for a new architecture to emerge in 18 months that allows the same results to be achieved on a $100 million rig.
That $9.9 billion difference is the bubble.
The Ghost of the 2000 Fiber Optic Crisis
To understand what happens next, stop looking at 2021 and start looking at 2000.
Back then, the narrative was exactly the same. "The internet is the future. We need more fiber. Demand is infinite."
Companies like Global Crossing and WorldCom laid thousands of miles of fiber optic cable. They spent billions. The "bubble" didn't pop because the internet failed. The internet became the most important invention in human history.
The bubble popped because we built too much of it, too fast.
We ended up with a massive oversupply of "dark fiber." The price of bandwidth collapsed. The technology won, but the companies that built the infrastructure lost everything.
We are currently laying the "digital fiber" of the AI era. We are building massive data centers at a rate that assumes we will need this level of compute forever. We won't. Once the models are trained, the inference phase begins. Inference is orders of magnitude cheaper and less resource-intensive than training.
We are building a fleet of aircraft carriers to deliver a pizza.
The Real Cost of Energy
While Wall Street ignores the bill, the power grid is screaming.
The most valuable company in the world is effectively a bet on cheap, unlimited electricity. We don't have it.
The moment a carbon tax or a power-use cap hits these massive clusters, the economics of AI training fall apart. You can't "innovate" your way out of the Second Law of Thermodynamics. You can make a chip more efficient, but you cannot make a massive data center stop needing cooling.
The Fallacy of "Infinite Demand"
I hear it every day: "But every company in the world needs AI!"
No, they don't.
Most companies need a slightly better database and a way to automate their customer support emails. They don't need a trillion-parameter model to do that.
We are seeing a massive "fear of missing out" (FOMO) spend from Fortune 500 CEOs who are terrified of being called "behind the curve" in their next earnings call. They are buying tokens they don't use and training models they don't deploy.
This is "shelfware" on a global scale.
When the CFOs finally look at the ROI on these AI projects and see that they’ve spent $50 million to save $2 million in headcount, the budget for NVIDIA chips will dry up overnight.
Stop Asking if the Bubble is "Intact"
The question isn't whether the bubble is intact. The question is who is going to be left holding the bag when the narrative shifts from "growth at all costs" to "utility at a reasonable price."
If you think NVIDIA can maintain its current trajectory while its biggest customers are actively building its replacements, you are not an investor; you are a gambler.
The hardware wars are always won by the person who can make the commodity the cheapest. Right now, NVIDIA is selling the most expensive commodity in history.
That is not a sustainable monopoly. It is a temporary supply chain bottleneck masquerading as a revolution.
Sell the hype. Buy the efficiency.
The world doesn't need more compute; it needs better ideas.
