The Physical Bet

Five American technology companies will spend somewhere around seven hundred billion dollars this year on the physical infrastructure that AI runs on.

For scale, the entire United States defense budget is roughly a trillion.

That number is the reason I think the fad question is settled, and I want to walk through what the money is actually buying, because the answer is not what most people picture when they hear the word AI.

What seven hundred billion dollars buys

It buys concrete, copper, silicon and electricity.

A modern AI chip isn't really a chip. It's a processor die sitting on a silicon base layer with stacks of memory bonded millimeters away, close enough that data can move between them fast enough to matter. That assembly step is its own industrial process, and for about two years it was the thing limiting how many of these existed. Not the chip design. Not the wafers. The packaging line. There was plenty of everything else and not enough of the machine that puts the pieces together, which is the most ordinary manufacturing problem there is.

The memory is a similar story. Three companies on earth can make it at the required grade, they have been sold out a year forward for most of the last three years, and a commodity industry famous for brutal price cycles now finds itself holding an allocation queue.

Then the racks. The unit of production is no longer a chip, it's a cabinet drawing more than a hundred kilowatts and costing two to four million dollars, sitting inside a building that takes eighteen months to two years to put up.

And underneath all of it, the thing that actually decides the pace now, is power.

You cannot fake a substation

This is the part that convinced me, and it has nothing to do with anybody's forecast.

The queue to connect new generation to the American grid holds something over two thousand gigawatts of projects. The median time from filing a request to producing power is more than five years. Data centers used about four percent of American electricity in 2023 and the credible projections have that heading toward somewhere between seven and twelve percent by the end of this decade.

Now look at the lead times on the equipment. Large power transformers run about two and a half years from order to delivery. Gas turbines are worse. The order book at the largest manufacturer stretches past a hundred gigawatts and a new heavy-duty order placed today delivers around 2031. Microsoft signed a twenty-year power purchase agreement to restart a nuclear reactor at Three Mile Island. That reactor is expected back online in 2027, three years after the deal was signed.

Nobody signs a 2031 turbine order for a fad.

That's the whole argument in one sentence. Software commitments are cheap and reversible. A company that changes its mind about a software strategy cancels a contract and eats a quarter of pain. A company that changes its mind after committing to a substation, a transformer order and a twenty-year power contract is stuck with physical objects in the ground that are useless for anything else.

The lead times are the honest signal, because they are the one thing that can't be walked back for a press release.

It might still be a bubble

Now the other side, because I don't want to sell you something.

There is a real case that the financial structure around this is unsound and the people making it are not cranks.

Start with utilization. Reported numbers for how hard these clusters actually run are not impressive. Typical inference deployments average somewhere around a fifth of theoretical capacity, while studies of production training jobs put effective utilization around a quarter. Any plant manager looking at those figures would tell you the factory is beautiful and mostly idle.

Then pricing. Renting an H100, the workhorse chip of the last cycle, cost around eight dollars an hour at the 2023 peak. It now goes for roughly two. That is what a rental market looks like when supply arrives faster than demand.

Then the accounting fight, which is the one worth understanding. These companies depreciate their servers over five or six years. If the real useful life is closer to three, then profits across the sector are overstated by a large amount. There are serious investors who argue exactly that in public and have put money behind it. Amazon quietly moved some of its server lives back down from six years to five, citing the pace of AI development, which tells you the question is live inside the companies too.

And then the financing. A chip maker invests in a model company. The model company signs enormous cloud commitments. Those commitments book as backlog at the supplier. The same dollar shows up in several places on several balance sheets. This year the largest spenders started borrowing to fund capex rather than paying out of cash flow, which is a genuine change from two years ago.

None of that is nothing. If you are deciding where to put your retirement money, all of it matters enormously.

If you are deciding whether to learn to use these tools, none of it matters at all.

The part that already happened twice

America has run this experiment before and both times the answer came out the same way.

In the 1870s and again in the 1890s the country wired itself together with railroads, financed with enormous debt and spectacular optimism. The Panic of 1873 started when the bank underwriting the Northern Pacific failed, the New York Stock Exchange closed for ten days and the country entered the longest contraction ever recorded. Twenty years later it happened again, worse. In the year after the Panic of 1893, roughly a quarter of all American rail mileage went into receivership. The Northern Pacific failed. So did the Union Pacific, the Santa Fe and the Reading.

The track stayed exactly where it was.

Receivership reorganizes who owns a thing. It does not pull up the rails. New owners bought the same network at a fraction of the price and ran it, and over the last three decades of that century freight rates fell by something like sixty percent. That is a way of saying the cost of moving anything anywhere in America collapsed. The investors were destroyed. The country got a continental transportation network.

Then we did it again with glass. Between 1996 and 2001 American telecom companies spent something on the order of half a trillion dollars laying fiber optic cable. When the bust came, analysts estimated that the overwhelming majority of the installed fiber was dark, meaning it had been laid and never lit. Global Crossing went bankrupt in January 2002. WorldCom followed six months later, at the time the largest corporate bankruptcy in American history.

Then, about seven years later, video streaming and smartphones arrived and quietly filled all of it.

The fiber you are using to read this was paid for by people who lost their money.

The one place the analogy breaks

I want to be careful here, because this comparison gets used lazily and it has a real weakness.

Fiber doesn't rot. A strand of glass in a conduit is just as good in 2010 as it was in 2000, which is why waiting seven years for demand was survivable. Rails are the same, only more so. The physical asset outlived the people who financed it by decades. That gap is the entire reason the story ends well.

Graphics chips are not like that. They lose value fast, they run hot, and the useful-life question above is not academic. If demand for AI compute arrived seven years late, an enormous share of what has been bought would be worthless before it got used. That is a genuine difference and anyone telling you the fiber story without it is selling something.

So the honest way to think about this spend is to split it in two.

There is a melting half, which is the silicon. And there is a durable half, which is the buildings, the substations, the transmission lines, the transformers and the generation. Those are thirty-to-fifty-year assets. They are useful for any computing that happens for the rest of our working lives and for a great deal that has nothing to do with AI at all.

If this thing pops, the melting half takes the loss. The durable half is still there and somebody buys it cheap, exactly the way Level 3 bought fiber and exactly the way the reorganized railroads bought track.

What this means if you are not an investor

Here is where I actually land, and it's the reason I wanted to write this one down.

Almost every conversation about whether AI is a bubble is a conversation about capital allocation, held between people deciding where to put large amounts of money. That's a legitimate question and I have no special insight into it.

But most of us are not in that conversation. Most of us are trying to work out something much simpler, which is whether it's worth reorganizing how we do our jobs around a technology that a lot of loud people say is overhyped.

Those are different questions and they have different answers.

The financing might fail. The capability will not. The power is being built, the buildings are going up, the chips exist and the models are already in the hands of about forty percent of American workers whether their employers have noticed or not. Nothing about a market correction unlearns any of that. The people who learned to build things on the internet during the years when internet companies were going bankrupt did not turn out to have wasted their time.

That's the confidence I'd want a working person to have. Not confidence in the valuations. Confidence that the ground under your feet has actually moved and is going to stay moved.

In the last essay I argued that the companies getting value out of this are the ones redrawing how they work rather than bolting AI onto what they already had.

This one is the reason I think that's worth the effort. The capability is not going anywhere. It has been poured into concrete and contracted through 2031.

Which leaves the question I care about most, and it's the one I'll spend the last essay on. If this is permanent, and if the people who thrive in it are the ones who can see a system and build the fix, then we have to ask what we are teaching our kids.