Bubble-double-plus: how the AI boom props up the economy with debt, data centers, faith, and valuations without rational expectations of durable and monetizable productivity growth…
"We will have a truly ungodly huge amount of datacenters and GPUs and (largely natural gas) power plants when this build-out is done. For example, only the first of many is StarGate 1 in Abilene, Texas: 500,000 square feet and 200 MW per building; 8 buildings planned; a campus roughly 1 square mile.
These concentrations of computer power will be useful for something."
What's the rate of deprecation on a data center? GPUs burn out pretty fast right?
Not my expertise at all, but I've read that chips are amortized for 5-6 years, it used to be 1 year less. They can easily last 5 years, but under constant heavy use it is more like 1.5 years. As for data centers, all the ones built 5 years ago are obsolete for AI because they don't have the cooling infrastructure.
If you compare earnings vs cash flow for Meta, GOOGL, etc you see the yawning difference due to depreciation.
Apparently, high end NVIDIA chips burn out in two to three years. If they are used continuously, they last about 54 days. It helps to remember that a silicon atom has a 110pm radius, so 3nm traces are less than 15 atoms wide, so they can be fragile in the face of the electron wind. I'll recommend Cory Doctorow's "The AI that we'll have after AI" for some references.
There's also technological drift. In theory, NVIDIA or some competitor is going to come out with something better in three to five years. If you don't write down your assets, you'll have a big bill due when a competitor can build a bigger, more powerful datacenter or one just as powerful for a lot less money and using a lot less power and cooling. This comports with the old saying in Silicon Valley that the pioneers are the ones with arrows in their backs.
An even bigger risk is algorithmic. LLM models keep getting better but they're improving asymptotically. That means that good enough, optimized models like Deepseek may have a bigger impact on productivity than larger more expensive ones. Apple is banking on being able to run useful AI on its RISC processors rather than having to build expensive datacenters much as it banked on distributed personal computers back in the 1970s. There's even a hobbyist movement to run models on processors that cost $5-$10.
P.S. If they want to prove ad technology, maybe they should use a text game language like Inform 7 and try: After someone buys something, someone is no longer shopping for something. Stupid text games seem to do better user modeling than whatever Google and its ilk are using.
"3 nm traces are less than 15 atoms wide" is an awesome and terrifying idea. It is clear that natural-language interface is a very important and useful use case for MAMLMs. But it is not clear to me how large a neural network is needed to achieve what is essentially a front-end to a FAQ list. And it is not clear to me what other enormously economically valuable use cases are going to emerge. Most of this is that while I understand training a neural network to emulate a Typical Internet S***poster, I do not understand what all the other steps of building a modern GPT LLM are and do:
> Kaleberg ; Apparently, high end NVIDIA chips burn out in two to three years. If they are used continuously, they last about 54 days. It helps to remember that a silicon atom has a 110pm radius, so 3nm traces are less than 15 atoms wide, so they can be fragile in the face of the electron wind. I'll recommend Cory Doctorow's "The AI that we'll have after AI" for some references.
> There's also technological drift. In theory, NVIDIA or some competitor is going to come out with something better in three to five years. If you don't write down your assets, you'll have a big bill due when a competitor can build a bigger, more powerful datacenter or one just as powerful for a lot less money and using a lot less power and cooling. This comports with the old saying in Silicon Valley that the pioneers are the ones with arrows in their backs.
> An even bigger risk is algorithmic. LLM models keep getting better but they're improving asymptotically. That means that good enough, optimized models like Deepseek may have a bigger impact on productivity than larger more expensive ones. Apple is banking on being able to run useful AI on its RISC processors rather than having to build expensive datacenters much as it banked on distributed personal computers back in the 1970s. There's even a hobbyist movement to run models on processors that cost $5-$10.
> P.S. If they want to prove ad technology, maybe they should use a text game language like Inform 7 and try: After someone buys something, someone is no longer shopping for something. Stupid text games seem to do better user modeling than whatever Google and its ilk are using...
Agreed. Something I've read, but am unsure about the veracity, is how much data center processing is model estimation versus application. If a large percentage is going to application, then the operating costs are too high and this iteration doesn't make economic sense, they are just running a Ponzi scheme of investments. Which isn't to say that AI won't be useful, just not this iteration.
The incentive is to issue a press release with $ billion investments and gigawatts, that's what makes stock prices go up.
As one might expect, ChatGPT has one answer and everyone else seems to have some other. Altman supposedly claimed about 2 watt hours per query, but ChatGPT claims 0.3 to 0.34 watt hours as opposed to a Google search taking about 0.00003 watt hours. That's quite a difference. There's a discussion about the methodologies at:
I am not going to crawl down this rathole as tempting as it might be. Even the lower power estimate is three orders of magnitude higher than the cost of a search. I'm guessing this could be brought down to two orders of magnitude with optimization, and future hardware might drive it even lower. It still isn't clear how this is going to make money.
Railroads move goods, radios move audio, the Internet moves data, houses provide shelter, and AI ??? We're forcing AI to fit into social media and search engines because these have high profit margins for a few investors.
What if the prevalent use of AI isn't to attract eyeballs or replace white collar labor, but to automate manual labor? That's a boring abstraction in Silicon Valley. Imagine that China creates dexterous, sensing mechanical hands that can quickly assemble everything from I-phones to T-shirts. That wouldn't use giant data centers, the latest Internet data, LLM's, or produce video. I'm not predicting that, but I am suggesting that smaller, specialized models built for specific applications may be more revolutionary than giant solutions investments in search of a problem.
If you think of AI at the small scale, it could improve productivity. It would be like the way putting small electric motors everywhere boosted productivity in the early 20th century. One problem is that general purpose LLMs do a crummy job. They produce mediocre text and make myriad mistakes. Different electric motors were developed to solve different problems, so different AI engines could be applied similarly. We're already seeing some of this. A farmer doesn't need an LLM that can make up "facts" about Eudora Welty to figure out where a drone needs to spray pesticide or herbicide.
Who bears the cost of malinvestment? As costs skyrocket, there's a lot of load shifting. I'm not sure we know.
The fate of AI and the business cycle may not be synchronized. CEO investment decisions respond to stock price. If CEO's are rewarded when they announce layoffs, or are penalized when they expand AI investments (looking at Meta this morning), then we get a recession now even if AI turns out great in 15 years.
"Build it and they will come." So we cleared mesquite trees and built a giant data center by a methane pipeline near Abilene. Who are we expecting to come, paying customers, or equity investors? The incentives are screwed up.
What's the biggest % gainer in the S&P 500, YTD? Not NVDA. It is Robinhood. #2 & 3 are Western Digital and Seagate, the old time makers of memory storage, because AI video uses a lot.
In sillier news, 'Weasel' Huang was photographed in a Korean fast food chicken place, dining w/ the heads of Hyundai and Samsung. This caused the price of every Korean stock associated with chicken to rise by 20-30% on the day.
If you haven't watched it, I recommend this discussion "AI and the Future of Work" at CUNY - with Danielle Li, Daron Acemoglu, Zeynep Tufekci, and Paul Krugman. A very interesting mix of expertise in different domains and thoughts. I thought Tufekci had some very interesting approaches that are very different from the sort of mainstream technology and economics issues that I read. Quite eye-opening, in some cases.
Most Americans believe we are already in a recession. They've cut back their spending. Those better off are still spending vigorously, but a lot of that is based on the market valuations of their portfolios. If the values of those portfolios fall, they will be spending less like everyone else.
Still this bubble is different. Usually, an investment bubble means money spent on labor. There's even a song about it that was popular in the 1930s, Brother, Can You Spare a Dime? Our current investment bubbles in AI and crypto are interesting in that they, unlike building railroads and making them run against time, don't create a lot of jobs. If the stock market and wealth effect weren't involved, they could crash and burn and put maybe thousands, not hundreds of thousands, out on the street.
However, the stock market and wealth effect are involved. I suppose we'll see. Either that or we'll see folks out on the street with "will compile economic statistics for cash" signs.
"This [AI stock] is now like buying Amazon when it was at $1.67." Riiiight. I see adverts like this to suck in the small investor, from Uber drivers down to paper delivery kids. 1929 Mark II is on its way. We just don't know which year we are in..1928, 1927, 1925...
As the sages say, "You can go bankrupt trying to short the market, guessing when the crash will happen."
Best case: all top heavy finance falls into the Gulf of Whatever and washes away Phony Crypto - time to take a deep breath (not snort) and ride the tide of cooperative capitalism
Revenue from energy generation for data centers can be an important factor in rebuilding disrupted community economies.
Using methane from centralized municipal wastewater treatment, together with the natural gas infrastructure (3 million miles), pyrolysis of the methane produces pure carbon and clean hydrogen.
Hydrogen powers solid oxide fuel cells (sofc) to produce electricity and heat for water treatment and additional pyrolysis. Electrical generation is scaled to supply both data centers and variable loads for communities (the most expensive kilowatts). Pure carbon is shaped into blocks for efficient thermal storage of heat from sofc and data centers. Hydrogen is also available for transportation fuel as direct combustion or use in proton exchange membrane (pem) fuel cells to produce electricity for EVs. Renewables with electrical storage can generate additional inexpensive kilowatts and produce Joule heating in carbon blocks.
Most important for a growing number of communities is water scarcity. In many areas groundwater recharge is falling consistently behind growing normal water use.
Connecting efficient community water treatment with potential treatment of produced water (from production of oil and gas), and other contaminated sources can all create potable water.
Composite treatment technologies, inexpensive hydrogen, and thermal storage in proximity create synergy for potable water, support for variable electrical loads, alternative transportation fuel, and economic redevelopment. All major components are available 'off the shelf' with standard assembly by existing trades.
Operations are separate but accountable to funding by continuous risk assessment.
Financial Cooperation funding includes all stakeholders as stockholders, enabling equitable distribution of returns, and funding for additional critical community infrastructure.
"We will have a truly ungodly huge amount of datacenters and GPUs and (largely natural gas) power plants when this build-out is done. For example, only the first of many is StarGate 1 in Abilene, Texas: 500,000 square feet and 200 MW per building; 8 buildings planned; a campus roughly 1 square mile.
These concentrations of computer power will be useful for something."
What's the rate of deprecation on a data center? GPUs burn out pretty fast right?
Not my expertise at all, but I've read that chips are amortized for 5-6 years, it used to be 1 year less. They can easily last 5 years, but under constant heavy use it is more like 1.5 years. As for data centers, all the ones built 5 years ago are obsolete for AI because they don't have the cooling infrastructure.
If you compare earnings vs cash flow for Meta, GOOGL, etc you see the yawning difference due to depreciation.
Apparently, high end NVIDIA chips burn out in two to three years. If they are used continuously, they last about 54 days. It helps to remember that a silicon atom has a 110pm radius, so 3nm traces are less than 15 atoms wide, so they can be fragile in the face of the electron wind. I'll recommend Cory Doctorow's "The AI that we'll have after AI" for some references.
There's also technological drift. In theory, NVIDIA or some competitor is going to come out with something better in three to five years. If you don't write down your assets, you'll have a big bill due when a competitor can build a bigger, more powerful datacenter or one just as powerful for a lot less money and using a lot less power and cooling. This comports with the old saying in Silicon Valley that the pioneers are the ones with arrows in their backs.
An even bigger risk is algorithmic. LLM models keep getting better but they're improving asymptotically. That means that good enough, optimized models like Deepseek may have a bigger impact on productivity than larger more expensive ones. Apple is banking on being able to run useful AI on its RISC processors rather than having to build expensive datacenters much as it banked on distributed personal computers back in the 1970s. There's even a hobbyist movement to run models on processors that cost $5-$10.
P.S. If they want to prove ad technology, maybe they should use a text game language like Inform 7 and try: After someone buys something, someone is no longer shopping for something. Stupid text games seem to do better user modeling than whatever Google and its ilk are using.
"3 nm traces are less than 15 atoms wide" is an awesome and terrifying idea. It is clear that natural-language interface is a very important and useful use case for MAMLMs. But it is not clear to me how large a neural network is needed to achieve what is essentially a front-end to a FAQ list. And it is not clear to me what other enormously economically valuable use cases are going to emerge. Most of this is that while I understand training a neural network to emulate a Typical Internet S***poster, I do not understand what all the other steps of building a modern GPT LLM are and do:
> Kaleberg ; Apparently, high end NVIDIA chips burn out in two to three years. If they are used continuously, they last about 54 days. It helps to remember that a silicon atom has a 110pm radius, so 3nm traces are less than 15 atoms wide, so they can be fragile in the face of the electron wind. I'll recommend Cory Doctorow's "The AI that we'll have after AI" for some references.
> There's also technological drift. In theory, NVIDIA or some competitor is going to come out with something better in three to five years. If you don't write down your assets, you'll have a big bill due when a competitor can build a bigger, more powerful datacenter or one just as powerful for a lot less money and using a lot less power and cooling. This comports with the old saying in Silicon Valley that the pioneers are the ones with arrows in their backs.
> An even bigger risk is algorithmic. LLM models keep getting better but they're improving asymptotically. That means that good enough, optimized models like Deepseek may have a bigger impact on productivity than larger more expensive ones. Apple is banking on being able to run useful AI on its RISC processors rather than having to build expensive datacenters much as it banked on distributed personal computers back in the 1970s. There's even a hobbyist movement to run models on processors that cost $5-$10.
> P.S. If they want to prove ad technology, maybe they should use a text game language like Inform 7 and try: After someone buys something, someone is no longer shopping for something. Stupid text games seem to do better user modeling than whatever Google and its ilk are using...
Agreed. Something I've read, but am unsure about the veracity, is how much data center processing is model estimation versus application. If a large percentage is going to application, then the operating costs are too high and this iteration doesn't make economic sense, they are just running a Ponzi scheme of investments. Which isn't to say that AI won't be useful, just not this iteration.
The incentive is to issue a press release with $ billion investments and gigawatts, that's what makes stock prices go up.
As one might expect, ChatGPT has one answer and everyone else seems to have some other. Altman supposedly claimed about 2 watt hours per query, but ChatGPT claims 0.3 to 0.34 watt hours as opposed to a Google search taking about 0.00003 watt hours. That's quite a difference. There's a discussion about the methodologies at:
https://surfshark.com/research/chart/chatbots-energy-consumption
I am not going to crawl down this rathole as tempting as it might be. Even the lower power estimate is three orders of magnitude higher than the cost of a search. I'm guessing this could be brought down to two orders of magnitude with optimization, and future hardware might drive it even lower. It still isn't clear how this is going to make money.
Railroads move goods, radios move audio, the Internet moves data, houses provide shelter, and AI ??? We're forcing AI to fit into social media and search engines because these have high profit margins for a few investors.
What if the prevalent use of AI isn't to attract eyeballs or replace white collar labor, but to automate manual labor? That's a boring abstraction in Silicon Valley. Imagine that China creates dexterous, sensing mechanical hands that can quickly assemble everything from I-phones to T-shirts. That wouldn't use giant data centers, the latest Internet data, LLM's, or produce video. I'm not predicting that, but I am suggesting that smaller, specialized models built for specific applications may be more revolutionary than giant solutions investments in search of a problem.
If you think of AI at the small scale, it could improve productivity. It would be like the way putting small electric motors everywhere boosted productivity in the early 20th century. One problem is that general purpose LLMs do a crummy job. They produce mediocre text and make myriad mistakes. Different electric motors were developed to solve different problems, so different AI engines could be applied similarly. We're already seeing some of this. A farmer doesn't need an LLM that can make up "facts" about Eudora Welty to figure out where a drone needs to spray pesticide or herbicide.
So, this time it might really be different? Is that a fair conclusion?
We all hope for the Plus
Who bears the cost of malinvestment? As costs skyrocket, there's a lot of load shifting. I'm not sure we know.
The fate of AI and the business cycle may not be synchronized. CEO investment decisions respond to stock price. If CEO's are rewarded when they announce layoffs, or are penalized when they expand AI investments (looking at Meta this morning), then we get a recession now even if AI turns out great in 15 years.
"Build it and they will come." So we cleared mesquite trees and built a giant data center by a methane pipeline near Abilene. Who are we expecting to come, paying customers, or equity investors? The incentives are screwed up.
We are regressing as a country due to red areas routinely voting against their best (economic) interests.
Any predictions of AI helping with that?
What's the biggest % gainer in the S&P 500, YTD? Not NVDA. It is Robinhood. #2 & 3 are Western Digital and Seagate, the old time makers of memory storage, because AI video uses a lot.
In sillier news, 'Weasel' Huang was photographed in a Korean fast food chicken place, dining w/ the heads of Hyundai and Samsung. This caused the price of every Korean stock associated with chicken to rise by 20-30% on the day.
If you haven't watched it, I recommend this discussion "AI and the Future of Work" at CUNY - with Danielle Li, Daron Acemoglu, Zeynep Tufekci, and Paul Krugman. A very interesting mix of expertise in different domains and thoughts. I thought Tufekci had some very interesting approaches that are very different from the sort of mainstream technology and economics issues that I read. Quite eye-opening, in some cases.
https://www.youtube.com/watch?v=McYBgZrORi4
Most Americans believe we are already in a recession. They've cut back their spending. Those better off are still spending vigorously, but a lot of that is based on the market valuations of their portfolios. If the values of those portfolios fall, they will be spending less like everyone else.
Still this bubble is different. Usually, an investment bubble means money spent on labor. There's even a song about it that was popular in the 1930s, Brother, Can You Spare a Dime? Our current investment bubbles in AI and crypto are interesting in that they, unlike building railroads and making them run against time, don't create a lot of jobs. If the stock market and wealth effect weren't involved, they could crash and burn and put maybe thousands, not hundreds of thousands, out on the street.
However, the stock market and wealth effect are involved. I suppose we'll see. Either that or we'll see folks out on the street with "will compile economic statistics for cash" signs.
"This [AI stock] is now like buying Amazon when it was at $1.67." Riiiight. I see adverts like this to suck in the small investor, from Uber drivers down to paper delivery kids. 1929 Mark II is on its way. We just don't know which year we are in..1928, 1927, 1925...
As the sages say, "You can go bankrupt trying to short the market, guessing when the crash will happen."
Best case: all top heavy finance falls into the Gulf of Whatever and washes away Phony Crypto - time to take a deep breath (not snort) and ride the tide of cooperative capitalism
Revenue from energy generation for data centers can be an important factor in rebuilding disrupted community economies.
Using methane from centralized municipal wastewater treatment, together with the natural gas infrastructure (3 million miles), pyrolysis of the methane produces pure carbon and clean hydrogen.
Hydrogen powers solid oxide fuel cells (sofc) to produce electricity and heat for water treatment and additional pyrolysis. Electrical generation is scaled to supply both data centers and variable loads for communities (the most expensive kilowatts). Pure carbon is shaped into blocks for efficient thermal storage of heat from sofc and data centers. Hydrogen is also available for transportation fuel as direct combustion or use in proton exchange membrane (pem) fuel cells to produce electricity for EVs. Renewables with electrical storage can generate additional inexpensive kilowatts and produce Joule heating in carbon blocks.
Most important for a growing number of communities is water scarcity. In many areas groundwater recharge is falling consistently behind growing normal water use.
Connecting efficient community water treatment with potential treatment of produced water (from production of oil and gas), and other contaminated sources can all create potable water.
Composite treatment technologies, inexpensive hydrogen, and thermal storage in proximity create synergy for potable water, support for variable electrical loads, alternative transportation fuel, and economic redevelopment. All major components are available 'off the shelf' with standard assembly by existing trades.
Operations are separate but accountable to funding by continuous risk assessment.
Financial Cooperation funding includes all stakeholders as stockholders, enabling equitable distribution of returns, and funding for additional critical community infrastructure.
Del Llano Resources - Transition Microgrids
delllanoresources.com
Craig Collier – Project Director/Principal
ccollierenergy@gmail.com