The AI Gold Rush
The four biggest cloud companies plan to spend up to $745B on AI this year. Like 1849, the surest money so far is in selling the shovels, and the biggest risk is for the miners who borrowed to buy them.
In 1848, a San Francisco shopkeeper named Sam Brannan ran up and down the streets shouting “Gold! Gold on the American River!”
Before he started shouting, he had quietly bought up every pick, pan and shovel he could find. A pan that cost him 20 cents sold for $15.
About 300,000 people poured into California. Most of the miners went home with little. Brannan became the state’s first millionaire, selling them the tools.
Many people call AI the next dot-com bubble. We think it looks more like a gold rush.
Picks, shovels and store credit
Every gold rush has the same cast. Here is who plays each part today:
- The miners: big labs and cloud giants, spending heavily to build AI.
- The shovel sellers: Nvidia, Micron and the other chip makers.
- The store credit: loans and long-term leases that pay for the buildout.
- The price of gold: the cost of AI itself, which keeps falling.
The question for investors is the same one it was in 1849: who is actually getting paid, and who is digging on credit?
The miners are digging fast
The four biggest cloud companies plan to spend $720–745B in 2026. That is about 78% more than last year.
The gold is real. Azure grew 43% last quarter. OpenAI and Anthropic are each reportedly near or above $65B a year in sales.
But sales are still much smaller than the spending, and much of the digging is done on credit. Roughly $1.4T of data-center lease promises have been signed, and about $1.1T of that is not yet on company balance sheets.
The gold is real. The question is whether it pays for the digging.
How long do the shovels last?
AI chips are bought on the bet that they stay useful for five or six years. Investor Michael Burry says it is closer to two or three.
Same chip. Half the life, if the skeptics are right.
Why it matters: the shorter a chip’s life, the more of its cost hits each year’s profits.
Charged against profits each year
Halve the life and the yearly cost doubles. Today's profits would look smaller than they are.
Nobody knows yet. The test is simple: do older chips keep earning money?
How deep is the seam?
We used to argue that bigger language models would hit a wall, because reasoning has to come before language. The evidence since then doesn’t back that up. Labs added reasoning on top of language models, and it works well.
What is still unknown is how deep the gold goes. Some research suggests the training only sharpens what models already knew. If so, the miners are digging a seam that thins out.
We were wrong about the wall. Nobody yet knows where the bottom is.
The gold is getting cheaper
Free, open models now trail the best paid ones by only about four months. Cheaper AI is like cheaper gold.
So far it hasn’t hurt. Cheap AI means more people use it, and more use means more chips. Here is how the shovel sellers and the miners look:
- Nvidia: $96.2B in sales last quarter, more than double a year ago.
- Micron: about $37.7B in profit in one quarter.
- Microsoft: we once called it a loser. Not so fast. Azure is booming.
- SpaceX: it bought xAI in February, so it is now a miner too.
- Tesla and Apple: long-term bets. Tesla’s next AI chips arrive in 2027 and 2028, and we haven’t re-checked Apple’s story in this update.
Shovel sellers do well while the miners keep digging. They get hurt fast when the digging slows.
When the miners pump the brakes
A gold rush usually ends when the gold stops paying for the digging. This time, the miners themselves are talking about slowing down.
On September 14, Anthropic’s CEO called for the industry to slow down, and OpenAI’s and SpaceX’s CEOs agreed.
Here is what we are watching:
- Cloud giants cutting their spending plans.
- Data centers canceled.
- Chip orders weakening.
- Older chips still earning money.
- Growth that does not depend on one big customer.
The general store
Not everyone in 1849 was a miner or a shovel seller. Some ran the general store, selling flour, coffee, boots and lamp oil, and got paid whether or not anyone struck gold.
Berkshire Hathaway is that kind of business. It earns from railroads, utilities and factories, not from any one miner’s luck. And it is starting to spend:
It is still a stock, so a broad market drop would hit it too. Short-term Treasuries and money-market funds are the other common places to hold cash.
The Kuberin view
Sam Brannan’s story has a second half. He poured his fortune into land and grand projects, then had to sell off his properties to pay a divorce settlement. When he died in 1889, he didn’t have enough money to pay for his own funeral.
Selling shovels made him rich. Overreaching undid him.
That is the lesson for AI. The tools are real, the gold is real, and so far the shovel sellers are getting paid. But the bill for the buildout is large, much of it borrowed, and the miners are already talking about slowing down.
In a gold rush, follow the cash, not the shouting.
Calculation note. The chip-cost example spreads $100 evenly across the chip’s useful life (straight-line depreciation): $100 ÷ 6 years ≈ $17 a year, and $100 ÷ 3 years ≈ $33 a year. It is an illustration of the debate, not an estimate for any company. Company figures are as reported in the sources below and rounded.
Disclaimer. Educational content only. Not investment, legal, tax, or financial advice. The analysis provided is for illustrative purposes and does not constitute a recommendation to buy or sell any specific security. All investments involve risk, including possible loss of principal.
Sources. Gold Rush history (Britannica) · Sam Brannan (Wikipedia) · Hyperscaler spending (FactSet) · 2026 spending tracker (Value Add VC) · Microsoft results (SEC) · Nvidia results (SEC) · Micron results (CNN) · Lease obligations (J.P. Morgan Asset Management) · Chip depreciation debate (Value the Markets) · Open vs. closed models (WhatLLM, citing Epoch AI) · OpenAI revenue (Axios) · September selloff (CBS News) · Berkshire results (Yahoo Finance)