We are building a $1 trillion infrastructure on the assumption that everyone needs an Nvidia GPU. What happens when Google proves you don't?
Inspiration: The quiet panic in Jensen Huang’s recent tweets, and the realization that “proprietary” hardware might be the ultimate moat.
Everyone thought the “AI Race” was about who could buy the most Nvidia H100s.
It was a simple equation: More GPUs = Better AI.
Then Google dropped Gemini 3.
It wasn’t just “good.” It was state-of-the-art. And crucially, it was trained entirely on Google’s custom TPUs (Tensor Processing Units). Not a single Nvidia GPU was needed for the core training.
This shattered the narrative. It proved that you don’t need the “arms dealer” (Nvidia) to win the war. You can build your own weapons.
The “Nvidia” Overreaction: “We Are Not Enron”
Nvidia didn’t take this lying down. They issued defensive blog posts, aggressive “benchmark comparisons,” and CEO interviews emphasizing their “software moat” (CUDA).
When a monopoly starts explaining why they are still relevant, the monopoly is cracking.
Here is the risk: If the Hyperscalers (Google, Amazon, Meta, Microsoft) all successfully move to custom silicon (TPUs, Trainium, Maia), Nvidia’s margin collapses. They go from being the “only game in town” to being a “commodity provider” for everyone else.
The “Chinese” Decoupling
We assumed US sanctions would kill Chinese AI. “No H100s, no progress.”
The reality? It forced innovation. Chinese models (like DeepSeek and Qwen) are climbing the leaderboards.
They aren’t using smuggled H100s. They are building massive domestic chip clusters using thousands of lower-power, locally made chips (Huawei Ascend). They are solving the problem with architecture, not just brute force.
The global market is bifurcating. The East doesn’t need Nvidia. The West’s biggest players (Google/Amazon) don’t want Nvidia. Who is left holding the bag?
Meta Crossing the Rubicon
If you needed proof that the monopoly is cracking, look at the latest bombshell.
Meta is reportedly in talks to buy Google’s TPUs.
(Is Google becoming a hardware company from a software one?)
Let that sink in. Meta is Nvidia’s single biggest customer. They are planning to spend $72 billion on AI infrastructure this year alone. They are the whale.
And they just signaled that they are looking for an exit ramp.
According to The Information, Meta is negotiating to rent Google’s TPU clouds next year and potentially deploy them in their own data centers by 2027.
Why does this matter?
Validation: It proves Google’s TPUs aren’t just an internal tool; they are a commercially viable alternative to the H100.
The “Nvidia Tax” is Optional: If Meta—the company with the deepest pockets—is refusing to pay the premium forever, the pricing power of Nvidia evaporates.
Nvidia knows this. Their defensive tweet (“We are delighted by Google’s success… but we are a generation ahead”) wasn’t confidence. It was damage control.
The “moat” just got bridged.
The “Burry” Prophecy
Michael Burry (The “Big Short” guy) has been betting against the semiconductor index (SOXX) and Nvidia specifically.
Like 2008, he might be early. The bubble can inflate longer than you can stay solvent.
But his thesis isn’t against AI. He is betting against the CapEx Bubble.
We are spending hundreds of billions on data centers assuming a perpetual 50% profit margin for hardware. If AI efficiency improves (smaller models, better code) or if hardware becomes commoditized (TPUs, Chinese clusters), the valuation of the entire infrastructure layer collapses.
Conclusion: The Efficiency Paradox
Here is the irony: The better AI gets, the less compute it might need.
We are moving from “Training” (massive compute) to “Inference” (efficient compute) and “Distillation” (making small models smart).
Final Take: AI will change the world. But the companies selling the picks and shovels for the gold rush might be the ones who get buried in the mine.