0xResearch returns to its analyst roots. Daily alpha and insights now coming directly from our research team. Investment research, market commentary, and unfiltered data-driven takes in your inbox.

Friday charts: AI needs a bubble

A short stretch of 1880s-era train tracks in Omaha, NE, is a living, working legacy of a speculative frenzy for railroad investment. 

What remains of the original Omaha Belt Line — a branch of track serving industrial warehouses — is a rare example of a bubble-era investment still earning a return an incredible 150 years later.

Similarly, there’s a good chance this newsletter was delivered to you over infrastructure built during the dotcom bubble — fiber optic cables have a useful lifetime of 30+ years.

The GPUs that helped me write it, however, will likely be obsolete in just two or three.

GPUs are the defining feature of the AI investment boom that Ben Thompson says has “obviously crossed the line into bubble territory.” 

With data center investments already worth hundreds of billions, it’s likely to be an even bigger bubble than either the railroad or dotcom ones.

But will it be as useful?

“The question now is whether or not this is a productive bubble,” Thompson continues.

Artificial intelligence is expected to be far more consequential than either railroads or the internet, but the half-life of a GPU is so fleeting, it’s uncertain how consequential the current investment boom will prove to be.

Thompson thinks the best chance to build the equivalent of railroad tracks or fiber optic cables is “a massive buildout in power generation, which would be a benefit for the next half a century.”

It will have to be massive indeed.

Just the Stargate data center in Abilene, TX, is expected to draw an unfathomable 1.2 gigawatts of power. 

If that number sounds familiar to you, it’s probably because it’s the same amount of power Doc Brown estimated would be required to send Marty McFly back to the future.

In the year 1955, where Marty found himself stranded, the only conceivable source of that much power was a bolt of lightning. (Fortuitously predicted for the next week.)

In the year Marty was trying to get back to — 1985 — the idea of that much power was still so unimaginable that people hadn’t even agreed on how to say the word: “gigawatt” (as we do now) or “jigawatt” (as they do in the movie).

In the year 2025, however, Stargate is building that capacity to check for typos in this newsletter.

To be fair, it also helped explain to me what the hell a gigawatt actually is.

Watts are a measure of how quickly energy is used — a single watt defined as the amount of energy used each second to push with one “newton” of force over one meter of distance.

The lightbulb in the stylish lamp behind you might have forty of them, meaning that it can draw forty watts of power from an electrical grid at any given moment.

The Stargate data center in Texas has 1.2 billion of them: enough to simultaneously run thirty million lightbulbs, power a million households, charge hundreds of thousands of electric vehicles, check this newsletter for typos and punctuation, or travel back to the future in a vintage DeLorean. 

Most of those watts are being provided by a natural gas power plant built onsite for the sole purpose of running the data center.

The plant should have a useful life of about 25 years. So, if Stargate isn’t needed for that long, it could be connected to the Texas power grid to power something else.

Many, many more power plants will have to be built.  

Bain estimates that by 2030, AI data centers will need 200 gigawatts of power (100 of which would be in the US).

Providing it would require some combination of 170 nuclear power plants, 300 natural gas plants, or 4,000 solar farms (as estimated by a ChatGPT query that briefly drew ~0.0000005 gigawatts of power).

That scale of construction won’t happen through careful planning alone. There’s a good chance it’ll take another round of collective financial overexcitement to make it happen. 

Sam Altman seems to be counting on it: “People will overinvest and lose money.”

But if they leave us with 200 gigawatts of power to use for a few decades after, that could prove very productive.

Let’s check the charts.

What we have so far:

US data centers currently have an estimated capacity of 45 gigawatts, of which 17.9 is dedicated to AI.

The forecast:

2,000 terawatt-hours (a measure of energy) is roughly equivalent to 238 gigawatts (a measure of power). Energy is the quantity of work possible, while power is how fast that work happens (or so AI tells me).

Moore’s Law isn’t fast enough:

More efficient GPUs would lower the need for new power, but Bain estimates that the demand for compute will grow twice as fast as Moore’s law.

It could go either way:

The red line above is what GDP should do if AI achieves singularity and is also benign, according to research from the Dallas Fed. The purple line is what it does if “the machines become malevolent, and this eventually leads to human extinction.” 

When Federal Reserve economists are including “extinction” in their models, you know it’s different this time.

So far, it’s benign:

The latest report on the impact of AI on labor from the Yale Budget Lab is packed full of charts with lines that go dead sideways — a reassuring sign that AI is not taking our jobs yet.

If we build it…

Ramp estimates that 44% of US businesses now pay for AI services, with an average contract value for AI products of $530K (expected to exceed $1 million in 2026). One-year retention rate is above 80%.

How it’s being financed so far:

The bear case in a picture: Bloomberg details the circular financing of AI capex that has people remembering the circular financing of the dotcom boom.

Different this time:

The free cash flow yield for the median stock in the S&P 500 is almost triple what it was at the top of the dotcom bubble.

The biggest AI winner (per capita): 

The .ai domain extension was introduced in 1995 as a country code for Anguilla. The British overseas territory (population 19,000) gets nearly half of its revenue from selling .ai web addresses to AI startups.

The railroad boom left us with tracks. The dotcom boom left us with fiber.

Let’s hope the AI boom leaves us with power.

Have a great weekend, productive readers.

— Byron Gilliam

_{Brought to you by:}

Katana was built by answering a core question: What if a chain contributed revenue back into the ecosystem to drive growth and yield?

We direct revenue back to DeFi participants for consistently higher yields.

Katana is pioneering concepts like Productive TVL (the portion of assets are actually doing work), Chain Owned Liquidity (permanent liquidity owned by Katana to maintain stability), and VaultBridge (putting bridged assets to work generating extra yield for active participants).

Stablecoins are having an institutional moment. How will that play out?

Leaders in the institutional world will hash it out live onstage at DAS London.

Get your tickets today with promo code: BREAKDOWNNL for £100 off.

📅 October 13-15 | London