Structural Analysis 7

NVIDIA's Collapse

Selling Shovels to Gold Miners. What If There's No Gold?

The Shovel Seller's Business Model

In every gold rush, the biggest winners were not the miners. They were the people selling shovels.

NVIDIA is the modern shovel seller. It sells GPUs (shovels) to Google, Microsoft, Meta, and OpenAI as they rush toward the AGI gold mine.

NVIDIA's revenue structure: Data center GPU sales → Over 80% of total revenue Main customers → Google, Microsoft, Meta, Amazon, OpenAI Why they buy → To build AGI NVIDIA's market cap → Assumes AGI demand continues indefinitely

But as Chapter 6 showed, AGI has three fundamental problems: data centers depend on petrochemical materials, the geniuses inside them have limited applicable work, and past data cannot reveal the future.

And one more — AGI is unlikely to break even.

If AGI Doesn't Pay, NVIDIA Collapses

If current AI (Claude, etc.) is sufficient for most tasks, AGI's additional value does not justify its enormous additional cost.

NVIDIA's collapse scenario:
Companies invest trillions in AGI development Massive GPU purchases from NVIDIA
Even if AGI is achieved, what companies pay is barely more than for current AI
Investment cannot be recouped AI investment begins to shrink
Additional GPU purchases stop
NVIDIA's revenue plummets

NVIDIA does not make end products. It does not sell AI services like Google does. If customers have no reason to keep buying GPUs, NVIDIA has nothing left.

Gold Rush AI Rush
The gold mine (goal) Gold AGI
The shovel seller Levi's, Wells Fargo NVIDIA
The miners Gold miners Google, Microsoft, OpenAI
Was gold found? Most miners found nothing AGI is unlikely to break even
Shovel seller's fate Demand vanishes when the mine is exhausted Demand vanishes when AI investment shrinks

The CUDA Lock-In Collapses — AI Kills NVIDIA

NVIDIA's real competitive advantage is not GPU hardware performance. It is CUDA.

What is CUDA: The software platform for running programs on NVIDIA GPUs. AI researchers have spent the past 15 years accumulating an enormous body of code, libraries, and tools written for CUDA. PyTorch, TensorFlow, and all major AI frameworks are designed with CUDA as their foundation. "Even if you want to switch from NVIDIA, your CUDA code won't run elsewhere" — this is the lock-in.

AMD's GPUs approach NVIDIA in raw performance. Google's TPUs have their own strengths. But the massive wall of the CUDA ecosystem has prevented customers from switching.

Then Claude enters the picture.

CUDA lock-in collapse scenario:
Hand CUDA code to Claude
Claude rewrites it for AMD ROCm / Intel oneAPI / generic GPU targets
Claude generates CUDA-compatible libraries for other hardware
Existing CUDA code assets now run on non-NVIDIA hardware
CUDA lock-in disappears
There is no longer a reason to choose NVIDIA GPUs

This is not hypothetical. Claude itself is developed using Claude. The ability to comprehend and rewrite codebases of hundreds of thousands of lines already exists. Porting from CUDA to ROCm is a structural code transformation — exactly the kind of work AI excels at.

The ironic structure: NVIDIA built the infrastructure for the AI industry. The AI (Claude) produced by that industry destroys NVIDIA's competitive advantage (CUDA lock-in). NVIDIA is undermined by the very industry it created.

Once CUDA lock-in collapses, the GPU market becomes a price war. AMD, Intel, Google TPU, and Chinese custom chips all enter. NVIDIA's profit margins plummet, and the premise behind its $3 trillion market cap crumbles.

GPU Raw Materials Are Petrochemical

NVIDIA's problem is not just its business model. Its products depend on petrochemical materials.

Petrochemical materials essential for GPU manufacturing: Photoresist → Photosensitive resin (petrochemical). Essential for semiconductor patterning. No alternative. Packaging → Epoxy mold compound (petroleum). Protects and connects chips. No alternative. Substrate → Epoxy resin + glass fiber (petroleum). Carries signals. No alternative. Coolant hoses → Synthetic rubber (petroleum). Essential for high-performance cooling. Thermal paste → Silicon-based compounds. Contains petroleum-derived ingredients.

The most advanced GPUs (H100, B200, etc.) cannot be manufactured without cutting-edge photoresist. Photoresist can only be made from petrochemicals.

NVIDIA's material dependency:
NVIDIA designs GPUs TSMC manufactures them
TSMC needs photoresist
Photoresist needs petrochemical feedstock
Petrochemical feedstock needs naphtha
Naphtha needs crude oil refining
When oil runs out, GPUs cannot be manufactured

NVIDIA calls itself "the center of the AI revolution," but the physical foundation of that revolution is petrochemical materials. The tools for building "a future independent of oil" cannot themselves be built without oil — the same structural contradiction as fusion reactors.

The Premise Behind a $3 Trillion Market Cap

NVIDIA's market cap reached approximately $3 trillion (as of 2024). What does this number assume?

Assumptions behind $3 trillion:

  1. AGI development investment continues expanding → GPU demand persists
  2. NVIDIA's GPUs have no substitute → Competitors cannot catch up
  3. Data center construction accelerates → Physical infrastructure expands
  4. AI service revenue justifies the investment → AGI breaks even If any of these four fails, $3 trillion cannot be sustained.

But what does reality show?

Assumptions breaking down:
Assumption 1 If AGI doesn't break even, investment shrinks
Assumption 2 AMD's MI300X, Google's TPU, custom chip development advancing
Assumption 3 Data centers depend on petrochemical materials. Fossil resources are finite
Assumption 4 If current AI is sufficient for most domains, AGI's additional value is limited
All four assumptions carry risk

The Lesson of Every Gold Rush

Gold rushes lasted 3 to 5 years. Not because the gold ran out — but because extraction costs exceeded the value of the gold.

The AI rush has the same structure. When the cost of mining the AGI gold (data center construction, operations, electricity, semiconductors) exceeds the value AGI produces, the rush ends.

NVIDIA's market cap assumes the rush continues forever. In all of history, no rush has ever lasted forever.

NVIDIA sells shovels to gold miners.
No gold mine, no shovel business.
If AGI doesn't break even, GPU demand disappears.
And GPUs themselves cannot be made without petrochemical materials.
The shovel seller's future depends on the gold mine existing and the shovel's raw materials being available.
Neither will last.

← Prev: Microsoft's Collapse Next: Eliminating Enterprise IT Taxes →

With nature, we can live.

The AI gold rush will end. What remains is soil.

AISeed — 生物多様性・食料・AIと暮らし(Facebook)