neat essay. do you think its relevant that the lifecycle of a GPU or much of the other components in AI datacenters depreciate fully within years, vs rail infra that lasts for decades+. is it really "capex" or "opex" when the facility needs a full refresh every couple years?
3 year depreciation is a short lifetime for expensive CPUs, memory and servers. Replacement will probably be constrained by supply availability. Regarding railroad equipment lifespans: "In 1900, steel rails lasted only 5 to 10 years due to brittle Bessemer steel and heavy steam trains, while untreated wood ties rotted in 4 to 6 years. In 2026, modern alloy steel rails last 30 to 50 years, and treated or concrete ties routinely last 30 to 50 years.Steel Rail LifespanIn 1900: The transition from soft wrought iron to Bessemer steel was taking place. However, this early steel was brittle and prone to internal defects, and axle loads quickly battered the rails.In 2026: Advanced metallurgy (such as head-hardened steel alloys) creates rails with immense tensile strength and wear resistance. The lifespan is largely dictated by tonnage rather than years; major freight routes average 40 to 50 years, while high-speed, heavy-haul lines may replace rails every 20 years.Railroad Tie (Sleeper) LifespanIn 1900: The vast majority of the 110 million ties used annually were untreated. Because they were highly vulnerable to moisture, fungi, and insect infestations, they rotted out in just 4 to 6 years.In 2026: Ties are treated with powerful preservatives (such as creosote and borates), extending their lifecycle to 30–40 years. In high-traffic and heavy-tonnage corridors, railroads frequently utilize prestressed concrete ties that reliably last 40 to 50 years. Additionally, high-durability composite (recycled plastic) ties designed to bypass rot in extreme environments are now hitting 50-year lifespans." Google AI
i cant tell if this is a human or not - the "google ai" signature seems a tell - but sure there are parts of rail networks that require faster cadence capex, but the biggest asset - by far - is the contiguous right of way and land networks, which does not need additional capital. Think about the replacement cost to build a coast to coast rail network today - with right of way access through each town it rolls through. If I understand correctly, that was a huge part of the actual costs, and i don't see a real parallel here in the datacenter build. Yes, they buy the land and water, but its peanuts compared to the assets that depreciate much quicker.
I once owned a bunch of railroad bearer bonds. I actually had to go to my safe deposit box, take out the bond, use a scissor to clip the coupons printed with the promissory note, stuff them into special envelopes with glassine windows and hand them with a deposit slip to a teller to get my interest payments. My favorites were Missouri-Pacific bonds that paid 5% and would have come due in 2045. My father explained that railroads had all kinds of weird bonds because they were always going broke and getting their debt restructured. These M-P bonds were extra weird in that they were sold flat. Usually, when you buy a bond, you have to pay the seller the accrued interest for the time they owned it between semi-annual interest payments. There was none of this for these bonds. This meant that if you bought the bonds right before the interest payment, you kept the whole thing. You might imagine that the price of the bonds would rise and fall to reflect this, but this was before modern computing and then as now the bond market had a bazillion bonds all thinly traded and each with its own quirks. No one wanted the headache. My father loved these bonds because he could buy a pile of them before interest was paid, grab the interest payment, and then sell them for roughly what he paid for them. It was free money which was as popular back then as it is now. I never got into that kind of trading. I just liked getting 10% interest because the bonds were sold at a big discount in the 1980s. I sold them decades ago when I bought my first house, but more recently I went looking for them. It took some serious searching, but the internet came through with a complex history of acquisitions, refinancing and related shenanigans followed by the most recent owner retiring the bonds well before 2045. I imagine the railroad is still running.
P.S. I read an article in a 19th century issue of Scribner's stating the the US government pissed off somewhere between one and three billion dollars on the railroads. They didn't used the term "pissed off", but it was used as an example of government waste and inefficiency.
The other place where I think the railroad industry has something to teach about economics and disruption is with the advent of the diesel locomotive. In the days of steam, there were a handful of companies that were major producers of steam engines. When diesel came about, they all tried their hand at the new technology. But in the end, it was two new companies (GE, and EMD - the Electro-Motive Division of General Motors), which never produced steam engines, that came out on top. (Alco - the American Locomotive Company - was the steam manufacturer that lasted the longest in the diesel era, but even they went out in the 1960s.)
That's right, a new technology came out and rendered obsolete the technology that the steam locomotive builders had perfected over decades. Tell me this same thing isn't happening right now with EV builders supplanting the "old" car manufacturers? Their core drivetrain technology is now obsolete. They're all trying to adapt to the new world - but is it possible that, in 20 years, none of them will still exist?
For stock market investors, there are a set of variables that, when aligned, have forecast periods of forward 12 month double digit market declines. These have involved :
a) the S&P 500 price residing below its 10 period moving average (monthly basis) value on "June 30th"or "July 31"
b) the YTD S&P500 return being negative into June 30th / July 31
c) variables a & b falling within "Presidential term years" 1, 3, or 4
d) the 3 month T bill yield being higher than the 10 year Treasury note yield (yield spread inversion) within 24 month proximity to variables a, b, & c
2nd or Mid term years are exempt from the process as their July - June returns have been predominately positive
Signaling record
July - June S&P500 return
7/1969 - 6/1970 -22.8%
7/1973 - 6/1974 -14.5%
8/1981 - 7/1982 -13.2%
7/2001 - 6/2002 -18.0%
7/2008 - 6/2009 -26.2%
This also applied to 7/1931 - 6/1932 ( -62% ) and 7/1873 - 6/1874 ( -10.4% )
Short term Treasury yield data for 1931 is sketchy and non existent for 1873. Yet for 1873, variables a, b, c could still be applied :
- the "stock index" price was below it's moving average on June 30th,
- the YTD return was -2.7%
- 1873 was a 1st Presidential year
So one can monitor for these variables for a possible negative return period within an AI driven market decline.
A very interesting essay. The question is, how can retail investors best position themselves to profit from AI? Given that individual companies go bust, it seems to me the answer is the S&P 500 index fund, or perhaps the tech-focused QQQ fund, which by definition always contains the biggest players.
(I say "perhaps" because sometimes the profit goes to ancillary players, e.g. not to the gold miners but to the people who sell them picks and shovels.)
But I'm thinking of the fact that the railroads prompted the invention of entirely new kinds of finance. Index funds themselves weren't invented until 1976, so a retail investor in the 1860s would have had a hard time devising a way to reliably capture the profits from rail.
So perhaps it'll be some other instrument that will best capture AI gains. I just listened to a fascinating and unsettling podcast episode in which Ezra Klein interviewed Yuval Noah Harari. Harari said the exact same thing: new infrastructures, like rail, triggered the development of new financial instruments. But Harari raised the possibility that AI itself will invent the new instruments, and human investors won't be able to understand them well enough to take advantage of them.
Maybe we need a new kind of index fund that buys into a bit of every new emerging financial infrastructure, like smart contracts, units of compute, and so on--whether humans understand them or not. Most would fail, but just as with stocks, the few that do succeed will realize enough gains to make up for the failures.
There are already a lot of sector funds that let one invest in particular industries.
I'm not sure investors really want financial contracts they can't understand. i imagine stock promoters would like them to bamboozle the naive and a rarified group of investors who can actually understand them if only by sheer dint of AI computing would find them useful.
External scenario: US companies accumulate HUGE costs for AI subscriptions and look to reduce their costs. China produces many alternatives to OpenAI, Anthropic, etc. Sure, the Chinese models are not as advanced as the top US ones, but they are constantly improving as China builds its own GPUs and CPUs. US companies opt to reduce their expenses and use the Chinese models, disguising that fact. US AI revenue collapses and an AI infrastructure crash happens. Financial repercussions shake the stock market as private equity/banks go belly up.....
neat essay. do you think its relevant that the lifecycle of a GPU or much of the other components in AI datacenters depreciate fully within years, vs rail infra that lasts for decades+. is it really "capex" or "opex" when the facility needs a full refresh every couple years?
Yes this is the key question!
3 year depreciation is a short lifetime for expensive CPUs, memory and servers. Replacement will probably be constrained by supply availability. Regarding railroad equipment lifespans: "In 1900, steel rails lasted only 5 to 10 years due to brittle Bessemer steel and heavy steam trains, while untreated wood ties rotted in 4 to 6 years. In 2026, modern alloy steel rails last 30 to 50 years, and treated or concrete ties routinely last 30 to 50 years.Steel Rail LifespanIn 1900: The transition from soft wrought iron to Bessemer steel was taking place. However, this early steel was brittle and prone to internal defects, and axle loads quickly battered the rails.In 2026: Advanced metallurgy (such as head-hardened steel alloys) creates rails with immense tensile strength and wear resistance. The lifespan is largely dictated by tonnage rather than years; major freight routes average 40 to 50 years, while high-speed, heavy-haul lines may replace rails every 20 years.Railroad Tie (Sleeper) LifespanIn 1900: The vast majority of the 110 million ties used annually were untreated. Because they were highly vulnerable to moisture, fungi, and insect infestations, they rotted out in just 4 to 6 years.In 2026: Ties are treated with powerful preservatives (such as creosote and borates), extending their lifecycle to 30–40 years. In high-traffic and heavy-tonnage corridors, railroads frequently utilize prestressed concrete ties that reliably last 40 to 50 years. Additionally, high-durability composite (recycled plastic) ties designed to bypass rot in extreme environments are now hitting 50-year lifespans." Google AI
i cant tell if this is a human or not - the "google ai" signature seems a tell - but sure there are parts of rail networks that require faster cadence capex, but the biggest asset - by far - is the contiguous right of way and land networks, which does not need additional capital. Think about the replacement cost to build a coast to coast rail network today - with right of way access through each town it rolls through. If I understand correctly, that was a huge part of the actual costs, and i don't see a real parallel here in the datacenter build. Yes, they buy the land and water, but its peanuts compared to the assets that depreciate much quicker.
I once owned a bunch of railroad bearer bonds. I actually had to go to my safe deposit box, take out the bond, use a scissor to clip the coupons printed with the promissory note, stuff them into special envelopes with glassine windows and hand them with a deposit slip to a teller to get my interest payments. My favorites were Missouri-Pacific bonds that paid 5% and would have come due in 2045. My father explained that railroads had all kinds of weird bonds because they were always going broke and getting their debt restructured. These M-P bonds were extra weird in that they were sold flat. Usually, when you buy a bond, you have to pay the seller the accrued interest for the time they owned it between semi-annual interest payments. There was none of this for these bonds. This meant that if you bought the bonds right before the interest payment, you kept the whole thing. You might imagine that the price of the bonds would rise and fall to reflect this, but this was before modern computing and then as now the bond market had a bazillion bonds all thinly traded and each with its own quirks. No one wanted the headache. My father loved these bonds because he could buy a pile of them before interest was paid, grab the interest payment, and then sell them for roughly what he paid for them. It was free money which was as popular back then as it is now. I never got into that kind of trading. I just liked getting 10% interest because the bonds were sold at a big discount in the 1980s. I sold them decades ago when I bought my first house, but more recently I went looking for them. It took some serious searching, but the internet came through with a complex history of acquisitions, refinancing and related shenanigans followed by the most recent owner retiring the bonds well before 2045. I imagine the railroad is still running.
P.S. I read an article in a 19th century issue of Scribner's stating the the US government pissed off somewhere between one and three billion dollars on the railroads. They didn't used the term "pissed off", but it was used as an example of government waste and inefficiency.
Great post, very interesting!
The other place where I think the railroad industry has something to teach about economics and disruption is with the advent of the diesel locomotive. In the days of steam, there were a handful of companies that were major producers of steam engines. When diesel came about, they all tried their hand at the new technology. But in the end, it was two new companies (GE, and EMD - the Electro-Motive Division of General Motors), which never produced steam engines, that came out on top. (Alco - the American Locomotive Company - was the steam manufacturer that lasted the longest in the diesel era, but even they went out in the 1960s.)
That's right, a new technology came out and rendered obsolete the technology that the steam locomotive builders had perfected over decades. Tell me this same thing isn't happening right now with EV builders supplanting the "old" car manufacturers? Their core drivetrain technology is now obsolete. They're all trying to adapt to the new world - but is it possible that, in 20 years, none of them will still exist?
For stock market investors, there are a set of variables that, when aligned, have forecast periods of forward 12 month double digit market declines. These have involved :
a) the S&P 500 price residing below its 10 period moving average (monthly basis) value on "June 30th"or "July 31"
b) the YTD S&P500 return being negative into June 30th / July 31
c) variables a & b falling within "Presidential term years" 1, 3, or 4
d) the 3 month T bill yield being higher than the 10 year Treasury note yield (yield spread inversion) within 24 month proximity to variables a, b, & c
2nd or Mid term years are exempt from the process as their July - June returns have been predominately positive
Signaling record
July - June S&P500 return
7/1969 - 6/1970 -22.8%
7/1973 - 6/1974 -14.5%
8/1981 - 7/1982 -13.2%
7/2001 - 6/2002 -18.0%
7/2008 - 6/2009 -26.2%
This also applied to 7/1931 - 6/1932 ( -62% ) and 7/1873 - 6/1874 ( -10.4% )
Short term Treasury yield data for 1931 is sketchy and non existent for 1873. Yet for 1873, variables a, b, c could still be applied :
- the "stock index" price was below it's moving average on June 30th,
- the YTD return was -2.7%
- 1873 was a 1st Presidential year
So one can monitor for these variables for a possible negative return period within an AI driven market decline.
A very interesting essay. The question is, how can retail investors best position themselves to profit from AI? Given that individual companies go bust, it seems to me the answer is the S&P 500 index fund, or perhaps the tech-focused QQQ fund, which by definition always contains the biggest players.
(I say "perhaps" because sometimes the profit goes to ancillary players, e.g. not to the gold miners but to the people who sell them picks and shovels.)
But I'm thinking of the fact that the railroads prompted the invention of entirely new kinds of finance. Index funds themselves weren't invented until 1976, so a retail investor in the 1860s would have had a hard time devising a way to reliably capture the profits from rail.
So perhaps it'll be some other instrument that will best capture AI gains. I just listened to a fascinating and unsettling podcast episode in which Ezra Klein interviewed Yuval Noah Harari. Harari said the exact same thing: new infrastructures, like rail, triggered the development of new financial instruments. But Harari raised the possibility that AI itself will invent the new instruments, and human investors won't be able to understand them well enough to take advantage of them.
Maybe we need a new kind of index fund that buys into a bit of every new emerging financial infrastructure, like smart contracts, units of compute, and so on--whether humans understand them or not. Most would fail, but just as with stocks, the few that do succeed will realize enough gains to make up for the failures.
There are already a lot of sector funds that let one invest in particular industries.
I'm not sure investors really want financial contracts they can't understand. i imagine stock promoters would like them to bamboozle the naive and a rarified group of investors who can actually understand them if only by sheer dint of AI computing would find them useful.
External scenario: US companies accumulate HUGE costs for AI subscriptions and look to reduce their costs. China produces many alternatives to OpenAI, Anthropic, etc. Sure, the Chinese models are not as advanced as the top US ones, but they are constantly improving as China builds its own GPUs and CPUs. US companies opt to reduce their expenses and use the Chinese models, disguising that fact. US AI revenue collapses and an AI infrastructure crash happens. Financial repercussions shake the stock market as private equity/banks go belly up.....
I read 1942.