The Coming AI Bailout?

During the dot-com boom, I was a programmer...

Which, by default, meant I was family tech support.

My father-in-law had just retired and started investing as a hobby.

He turned roughly $200,000 into $10 million watching CNBC and buying dot-com stocks.

One day I was helping him with his computer, and he asked me what I thought about a stock.

I heard: Cisco.

That made sense. Cisco made Internet routers.

He had made a lot of money but didn't want to sell because of the tax bill.

I grew up poor, so my answer was instant: "You won't regret taking profits. If it were me, I'd sell half."

He looked at me and said: "But I see their semis everywhere."

Sadly, his words didn't register with me at the time.

It wasn't until years later, after the dot-com crash, that I realized we had been talking about two completely different companies.

He meant Sysco. S-Y-S-C-O. The food distributor. And he was right.

Their semis (trucks) were everywhere.

I often think about this when I eat bland chicken fingers...

And I thought about it again today because investors are once again trying to figure out the top AI stocks.

I'm Still an AI Bull

Last week, I called compute "the new oil."

I argued that cheaper intelligence will create more AI tasks...

And it will happen fast enough that compute demand can keep rising even as costs fall.

I also gave you my two favorite stocks to invest directly in that compute boom: Nvidia and SpaceX/xAI (plus CoreWeave, the most direct bet on scarce compute).

I haven't changed my mind. But no one knows the future.

Investing isn't just being right about what happens next. You have to be more right than everyone else.

So, let's see the world as AI doomers see it.

First, remove the obvious gray swans: China doesn't invade Taiwan. TSMC stays online. No critical-mineral shock derails the supply chain.

So, the logic of technology continues. Models get better. Agents get better. AI is allowed to progress unfettered.

But what happens if AI stops growing exponentially for a few years?

The Internet Wasn't the Bubble

This is similar to the dot-com fiber rollout.

The Internet bulls were right. Internet traffic exploded. We eventually needed vastly more bandwidth than almost anyone imagined.

But telecom companies poured more than $500 billion into fiber, switches and wireless networks in the years after the 1996 Telecommunications Act.

Cisco stock ran up 3,800% in about five years.

Capacity grew faster than demand, prices collapsed and companies failed.

Cisco crashed 88%.

Brookings estimated telecom shareholders lost roughly $2 trillion.

The technology won, but few investors took profits in time.

So, AI can transform civilization and still produce terrible returns for some of the companies financing the transformation.

The number Wall Street cares about is ROIC: return on invested capital.

How much economic value do you get back for every new dollar you put into the machine?

Somebody Has to Buy the Chicken Fingers

Let's start with the AI doomers' strongest concern.

Nvidia sells chips to computing providers. OpenAI buys those computing services. Nvidia invests in OpenAI.

Imagine an equipment company investing in a restaurant that buys its fryers.

The equipment company records a sale. The restaurant gets a fryer.

But customers have to buy the chicken fingers.

In February, OpenAI announced funding including $30 billion from Nvidia and $50 billion from Amazon, alongside expanded infrastructure relationships.

The AI titans' businesses are all connected by investments, loans and purchase agreements.

Trouble at one node can travel through that network. You don't need AI to fail for investors to lose money.

You need the bills to arrive before enough customers paying for AI do.

This is where the bear argument is strong.

AI is technology connected to credit... connected to data centers... connected to utilities... connected to chips... connected to government policy.

And many of those networks depend on the same assumption: AI demand keeps compounding fast enough (exponentially) to pay everyone.

Anthropic, for example, reportedly has hundreds of billions of dollars of infrastructure commitments over the coming decade.

One of those deals reportedly gives it up to $84.5 billion of Nvidia-based compute through xAI.

OpenAI has enormous commitments too. Oracle is spending aggressively to serve them.

The more systems connected to the same assumption, the larger the shock if that assumption is merely a little wrong.

On that point, few would disagree.

The $30 Trillion Knowledge-Worker Market

So, to predict the future we need to see who those paying customers are.

ARK's July 2025 study uses about $1.25 trillion in 2024 software and cloud spending as its starting point...

ARK models three annual software-market outcomes for 2030:

And -- importantly -- it estimates knowledge-worker wages at roughly $30 trillion.

My bet: AI is coming for a lot more.

We could even see a white collar wipeout.

AI Demand

Today, AI demand looks extraordinary.

CoreWeave's revenue backlog recently reached roughly $104 billion.

Nvidia's data-center business produced about $89 billion in one quarter.

SpaceX's AI segment is becoming a meaningful part of its business.

But let's imagine AI revenue growth falls from triple digits to 35%.

Nothing breaks. Claude and ChatGPT get better, and agents become more useful.

But, at the same time, models also become dramatically more efficient.

Open-source systems handle a large percentage of the work.

Competition drives inference prices lower.

Companies take longer than expected to deploy autonomous agents.

People could use ten times more AI... without spending ten times more money.

The bubble can burst even if AI keeps working.

It can burst because monetization grows slower than the infrastructure built to serve it.

COREWEAVE: The First Crack?

CoreWeave is probably where I'd look first because it's the most direct bet of my three on scarce compute.

And right now, scarcity is the rule, so revenue is growing rapidly, backlog is enormous, and some new capacity has been contracting at higher prices.

But CoreWeave also reported a $626 million net loss last quarter.

Debt costs money. Hardware wears out (how fast is a hotly debated topic), and new chips displace old.

Now imagine utilization slips from "we need every GPU you have" to merely "very strong."

Rental prices soften, customers delay capacity and financing gets more expensive.

The GPUs are still useful.

The problem is what CoreWeave paid, and how it financed them.

In a severe version, the stock gets crushed, expansion stops and the capital structure gets reworked.

Maybe CoreWeave restructures.

Creditors can take losses, shareholders can be diluted or wiped out, and another operator can own the data centers.

Every GPU can keep humming.

NVIDIA: A Great Company Can Still Be a Bad Stock

Nvidia is almost the opposite business.

It designs the scarce machinery everyone else wants, wraps it in CUDA and networking, and sells the system.

That's why Nvidia remains my favorite long-term holding.

But Nvidia doesn't need AI to fail for NVDA to get crushed.

Its customers only need to discover that they collectively bought 2029's compute in 2027.

Orders get pushed out and lead times collapse. Scarcity premiums disappear for a while, margins fall and earnings estimates reset.

Nvidia survives and AI keeps advancing, but the stock can still get demolished.

That's where Cisco becomes useful. Cisco wasn't a fraud, and the Internet wasn't fake. Cisco remained important after the dot-com crash.

Investors had simply paid too much for a future that arrived more slowly than the stock price required.

SpaceX/xAI: What If the Crash Is an Opportunity?

My other favorite AI stock is interesting.

SpaceX/xAI's business is spending aggressively and currently loses money.

So if compute economics weaken, xAI gets hurt too. But SpaceX/xAI has something CoreWeave doesn't: another extraordinary business underneath it.

Starlink and launch operations give it different cash flows and strategic patience.

Now imagine everyone else pulls back.

Data centers trade below replacement cost and GPUs become available. Power contracts get renegotiated, and weak operators disappear.

SpaceX/xAI could become a buyer of distressed compute.

So my three compute picks would behave very differently in a bust:

CoreWeave needs scarcity.

Nvidia sells the machinery into scarcity.

SpaceX/xAI may have enough staying power to buy when scarcity temporarily disappears.

The long-term demand thesis can remain intact while the ownership economics change violently along the way.

And what happens if it all falls apart?

The Coming AI "Bailout"?

Don't be surprised if Washington steps in.

Compute is already a national-security asset...

But that doesn't mean the government writes a giant check to make Nvidia or CoreWeave shareholders whole.

It has different tools.

The Pentagon can sign long-term compute contracts.

The Department of Energy can expand capacity.

Washington can guarantee strategic loans and restrict sales of distressed infrastructure to overseas adversaries.

A hypothetical "$300 billion AI bailout" could mostly consist of guarantees and procurement rather than $300 billion of taxpayer losses.

And here's the irony: While not great for investors... that could be great for AI.

If investors overbuild and lose money, the infrastructure doesn't disappear.

Someone buys it cheaper.

GPU rental prices fall and inference gets cheaper. Startups gain access to compute they couldn't previously afford, and agent economics improve.

Yesterday's investors subsidize tomorrow's revolution.

That's what happened with fiber.

The Real Bet

I still think AI is one of the great investment opportunities of my lifetime.

I see more adoption (only 3% of users even pay for access). I see more monetization options (ad revenue). I still see the proliferation of AI agents. I see efficiency creating demand. I see robotics initiating a future compute cycle. I see AI discoveries creating new markets.

That's the bet I'm making.

But being bullish on AI isn't the same thing as being bullish on every dollar spent building AI infrastructure.

Will the world need more intelligence?

(You already know the answer to that.)

The question is what you're paying for the right to provide it... and what return that capital earns. That's ROIC.

I'll keep an eye on this for you as it plays out.

The Bighorn River

My father-in-law passed away six years ago. I thought of him on December 10, 2025.

That day Cisco finally closed above its dot-com-era record high.

I also think of him sometimes when I pass a Sysco semi on the interstate... Or get served a basket of bland chicken tenders.

But mostly I remember Montana. The Bighorn River.

He loved fly fishing a Woolly Bugger with a dropper behind it. We spent quiet days together floating down that river.

I miss fishing with him.

And years later, his accidental lesson still seems useful...

Before you decide to invest in tech...

Make sure you understand what you're investing in.

Then you can stand up tall and enjoy the drift.

Always be prospering,

Harry Seldon
Moonshot Minute

🔓 Premium Content Begins Here 🔒