Why the AI Shovel Seller Looks Cheaper Than the AI Labs

Transcript

From the video captions, split by chapter. Timestamps link to that moment on YouTube. Corrections are marked inline.

6 passage(s) with price targets or valuation scenarios were removed from this transcript (see Methodology).

0:00 Why does the shovel seller look cheap?

In May, investors valued Anthropic, an AI lab that still reported an operating loss for the first half of the year, at nine hundred sixty-five billion dollars. [Sentence removed on this site: price target or valuation scenario. See Methodology.] Now look at the company selling the chips. [Sentence removed on this site: price target or valuation scenario. See Methodology.]

Its own long-run average is above thirty times. So here is the puzzle. The seller of the shovels made about sixty billion dollars of profit in a single quarter. The labs buying them are still in the red or burning cash.

Why does the market price the profitable one like a value stock, and the unprofitable ones like the future? The usual answer is that it is a bubble and the labs are simply overpriced. Maybe. But that skips the interesting part.

The two prices are measured with different rulers, and they are bets on different things. This is Bet The Decade. Today: what the shovel seller earns, how the labs are priced, some simple payback math, and the best case for the labs. [Personal holdings statement omitted on this site. The author may hold shares in companies discussed. See Disclosure.]

This is not financial advice.

1:22 What is the shovel seller actually earning?

In the quarter that ended in late July, Nvidia reported revenue of about ninety-six billion dollars, up one hundred six percent from a year earlier. About eighty-nine billion of that came from data centers. Its gross margin was seventy-five percent, and its net income was about sixty billion dollars in three months. For the quarter that ended in October, it guided to about one hundred eight billion dollars of revenue.

A forward price to earnings ratio divides the share price by the profit analysts expect over the next year. [Sentence removed on this site: price target or valuation scenario. See Methodology.]

2:12 How are the labs being priced?

Anthropic raised sixty-five billion dollars in May, at a valuation of nine hundred sixty-five billion including the new money. In February, it had been valued at three hundred eighty billion. That is about two and a half times higher in roughly three months. OpenAI closed a round of one hundred twenty-two billion dollars at the end of March, at a valuation of eight hundred fifty-two billion.

Among the investors were Amazon, SoftBank, and Nvidia itself, with thirty billion dollars. In late October, reports said OpenAI was trying to raise more at about one point two five trillion dollars, down from an earlier target of one point four trillion, with about eleven billion committed so far. That round has not closed, so treat those as reported numbers. Here is the first ruler problem.

Nvidia's price is set every trading day by millions of buyers and sellers. A private lab's price is set in a funding round, by a small group of investors buying new shares. Anthropic's filing shows those investors bought convertible preferred stock, a different class from what the public would buy.

3:25 How do you compare a profit to a promise?

Because the labs have little or no profit to divide by, the fairer ruler is revenue. Anthropic said its annualized revenue run rate passed forty-seven billion dollars in May. OpenAI's was reported at about fifty billion dollars at the end of September. Now the rough math: value divided by yearly revenue.

Nvidia, using its last quarter times four, sits at about thirteen times. Anthropic, at the May round, about twenty. OpenAI, at its March price, about seventeen, and about twenty-five at the price it is reportedly asking now. These use different dates, so read them as ballpark numbers.

So on revenue, the gap is smaller than the headlines suggest. The real difference is cost. In its second quarter, Anthropic spent about five point six billion dollars on compute while bringing in about eleven point six billion. Nvidia keeps about seventy-five cents of gross profit on every dollar it sells.

4:28 How long would it take to earn the price back?

Let us do some simple payback math. This is illustrative, not a forecast. [Sentence removed on this site: price target or valuation scenario. See Methodology.] Now try Anthropic at nine hundred sixty-five billion.

Assume its forty-seven billion of revenue stays flat and one day earns a twenty-five percent profit margin. That is about twelve billion dollars a year, and roughly eighty years to earn back the price. But the labs are not priced to stay flat. If revenue grew eightfold to around three hundred seventy-five billion, with the same margin, yearly profit would be about ninety-four billion, and the payback falls to about ten years.

That would be cheaper than Nvidia today. Everything depends on that growth arriving. That is the real difference. In my view, Nvidia's price assumes its profits are near a peak and may not last.

The labs' prices assume their profits mostly have not happened yet. One is priced for a cycle, the other for a future.

5:35 Why might the shovel seller deserve its discount?

There are real reasons for the low multiple. At the end of July, five direct customers accounted for about seventy percent of the money owed to Nvidia. When buyers are that concentrated, a few budget decisions can move a lot of revenue. Nvidia is also helping to finance its own demand.

It put thirty billion dollars into OpenAI's March round. In August, it agreed to guarantee up to one hundred five billion dollars of lease and power payments for a data center campus in Ohio, on behalf of an OpenAI affiliate, in exchange for the site using Nvidia hardware. Its biggest customers also have alternatives. Anthropic says it trains and runs its models on four chip families: Amazon's Trainium, Google's TPUs, AMD's GPUs, and Nvidia's GPUs.

Every workload that moves to another chip is revenue that never reaches Nvidia's seventy-five percent margin. And the labs still have to pay for all of this. In September, reports said OpenAI's own projections showed it burning about two hundred seventy-eight billion dollars of cash from twenty twenty-six to twenty thirty. If funding gets harder, chip orders could be among the first things to slow.

6:54 What did the shovel seller of 2000 teach us?

We have seen a shovel seller priced for the future before. [Sentence removed on this site: price target or valuation scenario. See Methodology.] The internet did arrive, and Cisco kept selling equipment for decades. But by October two thousand two, its shares had fallen almost ninety percent from the peak.

The business was real. The price had simply assumed too much growth, too fast. Here is the twist. In two thousand, the market priced the shovel seller as if the boom would never end.

Today it prices Nvidia well below its own average multiple, as if the boom might end soon. The optimism has moved from the shovel seller to the gold miners.

7:46 To be fair: why might the labs be worth it?

To be fair, the case for the labs is strong, and it deserves a full hearing. Anthropic's public filing shows revenue of about one point two billion dollars in the first half of twenty twenty-five, and about sixteen point three billion in the first half of twenty twenty-six. That is more than thirteen times as much in one year. Its losses shrank while revenue exploded.

The operating loss fell from about two point one billion dollars to about six hundred ninety-three million. It also generated about three point nine billion dollars of cash from operations in the first half. Its reported net loss was far larger, about eighty billion, but most of that came from a non-cash accounting charge. In the second quarter, its revenue grew by nearly fourteen hundred percent from a year earlier, while its compute and infrastructure costs grew about three hundred percent.

If revenue keeps growing faster than the cost of compute, margins can widen quickly. OpenAI's run rate was reported to have grown about seventy-seven percent in just three months. Software businesses that win their market can end up with very high margins, and the labs may be priced for exactly that outcome. Soon there will be a public test.

In late October, Anthropic filed to list its shares on the Nasdaq. Once it trades, a lab will be priced every day by the market, just like Nvidia.

9:16 What should a long-term investor take from this?

Here is the zoom-out lesson. A low multiple is not automatically cheap, and a high one is not automatically expensive. A multiple is the market's guess about how long profits will last and how fast they will grow. Your job is to decide which guess is wrong.

So when you compare them, check three things. First, who pays whom: how much of the chip seller's demand is funded by the buyers' fundraising. Second, how much a lab spends on compute for each dollar of revenue. And third, whether the next funding round comes in higher or lower than the last one.

Remember those nine hundred sixty-five billion dollars for a lab whose profits have barely begun. The market is not saying the labs are better businesses than Nvidia today. It is betting they will be tomorrow. [Personal holdings statement omitted on this site. The author may hold shares in companies discussed. See Disclosure.]

Not financial advice. Which would you rather own for a decade, the shovel seller or the gold miners? Tell me in the comments. If you missed it, watch the last episode on why unfinished data centers look so profitable on paper.

Think in decades, not quarters, and subscribe.