Boom, bust, and the adults in the room.
By Wolf Richter for WOLF STREET.
Investments in the construction of data centers in the US has exploded. But this construction spending on data centers does not include the amounts spent on AI-specialized semiconductors, memory chips, servers, networking equipment, etc., which has reached astronomical levels. The four biggest cash-richest companies – Microsoft, Alphabet, Meta, and Apple – project capital expenditures of over $210 billion in their current fiscal years, just those four. Compared to that, construction costs are mundane. But they show the trend.
Construction spending on data centers soared by 29% year-over-year to $3.5 billion in June, not seasonally adjusted, up by 366% since the beginning of 2021, and up by 430% in the seven years since mid-2018. The $1-billion-a-month mark was surpassed for the first time in May 2022.
Data center construction had already grown at a red-hot pace before 2022, but 2022 triggered panic mode when Microsoft, Alphabet, Meta, and Apple, plus lots of lesser companies such as Tesla, plus data-center specialists such as Digital Realty, Equinix, and CoreSite, a subsidiary of American Tower, plus PE firms, VC firms, and hedge funds, plus everyone and their dog began throwing mindboggling amounts of cash at the physical infrastructure for AI, including the buildings to house the servers.
The Census Bureau, which tracks construction spending of all kinds, considers data-center construction spending one of the main segments of office construction spending.
The rest of office construction spending (total office minus data centers, blue in the chart below) has plunged since 2022 when the commercial real estate sector of office began to spiral into a depression.
In dollar terms, the rest of office construction spending has plunged nearly as much as data center construction spending has soared. As a result, total office construction spending, including data centers, has increased by only 17% since 2022 and by 30% since mid-2018.
Construction-cost increases: The PPI for Construction of Nonresidential Buildings jumped by 37% in two years from the beginning of 2021 through the end of 2022.
But construction cost inflation has stabilized in 2023 (actually negative inflation), 2024, and so far in 2025, and neither the data center construction boom nor the plunge in office construction over those years was significantly impacted by inflation.
Where’s the juice going to come from? Data centers require lots of power, and utilities or the data center providers themselves have to invest to build new capacity to generate this power and invest in the transmission infrastructure to get the power to the data centers. But it’s a slow complex process and takes many years.
And utilities want to make sure the infrastructure they plow billions of dollars into will not become a stranded asset when the AI bubble implodes, and so they’re proceeding with the care of the adults in the room.
Monthly construction spending on electric power installations, including power plants and transmission infrastructure, has run at a pace of $9 billion to $10 billion a month over the past two years ($9.6 billion in June), after two big waves of acceleration, totaling 67% since mid-2018. Since the beginning of 2021, the monthly construction pace has risen by 33%, essentially just tracking construction cost increases.
There are now numerous powerplant projects being planned and argued over, but they won’t show up in construction spending until on-site construction actually starts.
Enjoy reading WOLF STREET and want to support it? You can donate. I appreciate it immensely. Click on the mug to find out how:
Curios Wolf, how is construction in manufacturing plants doing after the huge rise the last few years? Are manufacturer’s repurposing present plants to accommodate the return of manufacturing to the U.S.? The U.S. doesn’t import much electric energy، geography and the Canadian border play a part in New England and the far northwest. Did any of the tariffs include electric energy?
I’m going to cover this in detail next month — including what’s driving it… got a new chart for that. Meanwhile, overall factory construction, not seasonally adjusted:
A single penny spent on Artificial Idiocy (AI) is just a penny wasted.
I don’t know what will happen in the future, but right now, I find “AI” to be a very useful tool for my work.
My healthcare organization uses AI all the time, such as for interpreting diagnostic results (with a doctor looking over its shoulder) and doing visit summaries, all of which then are posted online to where I can look at this stuff. Looks like they’re using AI to allow doctors to spend more time with patients, rather than doing this stuff.
Sure, there are some benefits, but I can’t help but think that a lot of the AI spending is so that the big 4 can hype their prowess, and get higher stock prices.
I’m just not seeing the massive long-term economic benefits to society of all of this “generative” AI.
TSonder305
I’m convinced that the AI spending and stock bubble will implode spectacularly, à la Dotcom Bust, and trigger a recession when it does, but the use of AI will spread all over the place (it’s already doing that), just like the internet did. The internet didn’t go away with the Dotcom bust.
I think you may be right about the bust, but unlike the dotcom bubble, the surge isn’t really fueld by start-ups and websites like pets dot com
It is led by the Mag7 especially NVDA selling chips. It is wayyy overhyped right now, but I think it is more likely to flatline not crash.
I’ve used Claude for medical advice twice so far. First for foot pain which it diagnosed as plantar fasciitis. Gave a treatment plan, etc and it worked.
Pretty easy diagnosis if we are going to be honest tho.
Second, not so easy. Muscle pain in the back linked to finasteride. Similar problem as with statins. People who are sensitive to one are often sensitive to both. Years ago I stopped statins because of this exact problem. Think any doctor told me that? Nope.
It suggested 300mg of Coq10 daily.
Could just be coincidence, but Ill be damned my back muscle pain is gone, not totally but 90% after a week.
I have never, ever had an MD recommend Coq10 because statins and/or finasteride deplete it.
I find this an amazing use of this technology.
This is a good use case for AI.
The biggest problem I see is the implementation of AI with no human interpretation. I work with people who copy and paste results from a LLM and do zero QA on it. Then I review it and it is full of errors and mistakes that the AI tool created.
8ticks,
I tell people to treat LLMs as a new hire straight out of college, like they’ll do the work, but you should definitely review it lol
“I have never, ever had an MD recommend Coq10 because statins and/or finasteride deplete it. ”
I’ve been taking COQ10 for years for that, based on old fashioned self-research. But AI may improve the marginal research skills of many, which is a benefit.
If billions of super fast but still simple on or off switches can “talk” and even “do” most any kind of science (like medical, in the example), what does that say about human language….and especially science.
Just curious…..doubt anyone else is.
Also, once you have the FULL report on your “health” (ALL overseen by a doctor), what is there to talk to the second doctor about with all the “extra time” AI has saved everyone….the weather….investing?
There must be something I don’t get…..too much wisdom here for obvious errors in thinking like mine.
NBay
“There must be something I don’t get….”
Yes, you’re obviously not using that kind of service. That’s OK too.
Regarding this bubble being lead by real companies like NVDA and not by fluff like pets.com in 1999, I haven’t been comfortable with that comparison.
In the late 90’s we had much the same picks and shovels argument that we have now. One of the hottest stocks was JDSU, which was producing optic fiber, and we had CSCO, making switches and routers, for example. The whole mania just got ahead of itself and it took years for a lot of that optic fiber to be lit up. When the bubble popped, JDSU and CSCO got hammered along with everything else.
And instead of pet.com, now we have 0DTE traders and record margin interest in retail traders and vol control funds and meme stocks and Roaring Kitty, thankfully I haven’t heard of him for a few months. Same as it ever was.
The only time I’ve used AI so far was when someone I know wanted to use ChatGPT to help us plan a recent backpacking trip. It will confidently give you routes, timetables, descriptions of the destinations, tell you what permits are needed and even generate a GPX file for you to follow. Every last detail of what it was telling us was wrong – like it was so bad it was comical. It doesn’t seem to have the ability to say “I don’t know” – and I suppose it wouldn’t be able to considering that it’s not actually reasoning, it’s just autocomplete on steroids.
For the record, I still have my money on model collapse.
The worst case scenario regarding AI isn’t that it’s largely a failure or bust, it’s that it’s successful beyond its proponents’ wildest dreams. I recommend reading Player Piano by Vonnegut.
Sure is good for the electrical safety business. Picks and shovels baby. Similar to the fracking boom. The music does stop rather unexpectedly. Sock away as much of the profits as possible into retirement accounts.
You probably said the same thing about the internet and the cell phone!
Ok Boomer.
Thanks to the demand from data centers, COMED just raised our rates 10% in Illinois.
Yes, that’s going to be a universal issue: lots of inflation in household electricity (adding to services inflation). Our power bill is already out the wazoo.
I put in solar power at my house I can switch back to the grid as necessary. I feel it will mitigate increase of electrical cost which electrical costs are going up. I do have a question. Why don’t some of these high tech company’s buy silver mines? It seems like silver is needed in everything and they r always saying we have a silver deficient…
How dare they, after we bailed out their nuke plants. Of course, we can’t give them all the credit or all our ire. The State government didn’t rescind a single thing Madigan and ComEd were busted for years ago.
Not a fan of COMED for sure, and the increase stings, but they just deliver. Had something to do with the cost of power at the auctions last year. Should push nuclear power again. Fleet in the nation aging…
COMED is an Exelon company. They don’t just distribute, they produce, through multiple Nuke plants that the state of Illinois supports with tax dollars. Look at any Illinois electric bill, down at bottom are five things found by the federal government to be collusion between the State and Exelon to grift extra money from tax payers.
Madigan got busted, some exec got jailed, yet the grift is still on our bills years later.
The green scheme is real. Billions are wasted on increasing corporate profit.
The Battery Electric Storage Systems (BESS) are popping up everywhere now. Stores huge amount of electricity – buys it cheap at night and sells it at top rates during the day. Ones in Canada will make over $1 B in gross income each. They are “powering” AI in a lot of data centres. Just huge growing industry. Unfortunately most of the batteries and the inverters are made in China
He who takes the risks and makes the investments reaps the rewards.
The amount of energy resources used for this AI destruction of our future is absurd:
In 2023, Google’s global data centers consumed approximately 6.1 billion gallons of potable water. This is equivalent to the amount of water needed to irrigate 41 golf courses in the southwestern US, according to Quartz. Two year old data, how quaint.
The increased trend of bed rotting, where one lays in bed all day surfing, will surely require greater data center capacity and investment — if only there was a way to accelerate the advancement of neural implants, so we can store our brains offsite and stop worrying about our mortal coils (and beds).
LOL, what hilarious comparison to point out wastefulness: Google’s data centers v. irrigating 41 golf courses in the southwestern US.
Us golfers need lush fairways!
Motocross riders would like a shot at them as well…..make for REAL super fun rides…..doesn’t our sport have the same value as yours?
Most fun to do it at nite but now risk not having good escape routes like when I was a kid……
Pretty unfair to kids, who have it bad enough nowadays.
Most of the golf courses in the southwest are using reclaimed grey water. Golf course horticulture has come a long way by developing drought resistant grasses, and altering course designs to use less water. Any courses that didn’t make these adjustments either went out of business (Lost Canyons), or they are exclusive private courses that pay out the wazzo for out-dated irrigation methods. I’d also add that a municipal golf course does more for it’s community than a data center.
“I’d also add that a municipal golf course does more for it’s community than a data center.”
A golf course is only good for the relatively small number of people who actively play that sport. For me, a golf course is an immense waste of real estate and resources. A public park would be far better.
But everyone uses data centers, whether you know it or not, and even by reading this site and posting your comment here on this site, you used at least one data center – one that is in the Chicago area. And maybe more than one data center. The internet lives in data centers.
You see, it’s like always. If you don’t use something, or don’t think you use it, it’s a waste. If you use it, it’s the best thing since sliced bread.
The horror of an unwatered golf course, is funny. I lived in a golf course community in PNW, which at one point almost was bankrupt, because the course people decided they needed a million dollar upgrade for the sprinkler system. Their membership had been declining for decades, but champagne tastes overpowered the reality of their beer budget.
In terms of golf analogies, overbuilding capacity for AI expansion, while champagne is freely flowing today, seems inconsequential.
A lot of data centers will end up as unused Putt Putt franchises…
With air conditioner that spit ice cubes!
This is misleading. I worked in Google SRE and visited several data centers. All the water comes from sources not needed elsewhere and is treated on-site to be suitable for evaporated cooling.
The Dalles uses water from the Columbia, which has it to spare. In Georgia, they take city “gray water” (aka “light sewage”) and treat that for use.
This water is evaporated locally and returned to the local atmosphere where is eventually returns as local rain… Not that it matters since it’s still a rounding area compared to natural evaporation in the area.
There is a Data Center that is in the planning stages near my cousin’s farm near Toronto, South Dakota. It is to encompass 1.8 million square feet. The name of the company is Applied Digital.
There’s a big one under way in Tucson, AZ slated to consume 5 million gallons of water/day. Good plan for a desert region that gets 11 inches of rain/year. It will produce 180 new full time jobs. In anticipation of this windfall, Tucson Electric Power has applied for a 14% rate increase.
Well, its not going to “consume” the water. Just heat it up.
The energy is also not consumed: just converted into a different form (heat).
Most of us are technology users, but the opinion is split as to who will actually reap the benefits.
The main question for the future becomes:
Jetsons, or Terminator?
The arc of humanity will determine which (look to history as the guide).
Power generation is heavily subsidized by the public so those AI companies are mopping up public dollars along the way I believe.
I’m here in Tucson.
Let’s just say that people are hot, bothered, and fired up about this data center, which is a project of Amazon Web Services.
And, on a personal note, Slim is not a fan of anything with Amazon in its name.
It goes back to when Amazon killed the Slim family’s favorite bookstore. That bookstore was a beloved part of the community. But, alas, it couldn’t compete with the Evil River.
Apple –> Amazon in a couple of paragraphs above?
Lots of hype for data centers in Norway too.
Local politicians are baited with promises of job creation, which doesn’t even seem close to realistic numbers, judging by the few completed projects.
Then centers are built on farmland, drives electricity prices through the roof and threatens the traditional heavy industry – which was made possible by cheap surplus electricity from our hydro plants in the first place.
So now the windmill industry is making a killing, supported by generous government subsidies – which they then reinvest to destroy nature wherever there is wind, baiting local politicians with job creation.
As the people that used to have good jobs in heavy industri are now unemployed, the pressure against local politicians to create jobs is now immense.
Happy days.
If only he had not told our coal miners to learn how to code.
Also requires incredible amounts of water
Building all these data centers that use so much water and electricity is like killing your only cow because you want a cheeseburger.
I’m grumpier.
“Where’s the juice going to come from? Data centers require lots of power, and utilities or the data center providers themselves have to invest to build new capacity to generate this power and invest in the transmission infrastructure to get the power to the data centers. But it’s a slow complex process and takes many years.”
“And utilities want to make sure the infrastructure they plow billions of dollars into will not become a stranded asset when the AI bubble implodes, and so they’re proceeding with the care of the adults in the room.”
Residential electric utility customers should NOT pay for the new loads from data centers, including AI infrastructure. The companies that operate the data centers need to foot the bill, and provide utility rebates for any directly related residential customer cost increases to date. This is ex-inflation in utilities. The utility companies should force the cost onto these companies for both residential customers and their own financial protection from new capacity becoming a stranded asset. Residential customers should be calling for this. I’m sure Congress is got this when they’re not too busy self-enriching themselves via stonk trading, etc…
There’s not enough electrical supply to meet the new AI/data center demand. Period. As you said, building this will take years. In my view, distributed nuclear reactors are the best option. Renewables are unreliable. Data centers have steady demand. Day or night. Windy or not.
The U.S. electrical grid is old and needs a massive upgrade.
The existing supply and infrastructure can’t support this massive new demand. This is fundamentally going to limit AI. We need more adults in the room on this, and in general.
The United States currently hosts 10x as many data centers than China — but seems this might unfold like the Ukraine drone story, where a massive data center buildout will end up eating cash, as alternative, more efficient technology distribute data.
Google will probably retain its primacy, but as other systems evolve below the radar, the markets will fragment into those who are willing to spend big cash for cutting edge advantages — and the majority of the herd that watch cat videos.
In reality, it’s absurd for residential utility customers to help foot the bill for data center expansion, for the sole purpose of storing more cat videos.
I’m sure this AI buildout will be part of hedonic CPI adjustments, that will see the storage of cat videos as an increase in quality of life, well worth the cost.
…recalling that, in the divine pantheon of ancient Egypt, cats were regarded as one of the gods (…or, so our old boy has always indicated to us…).
may we all find a better day.
Yeah. So don’t badmouth cats, it could be another path to that bad place…..best to cover all the possible bases…like they say, “Diversifying is your only free lunch”
Data centers can (and will) have built-in combined-cycle natgas turbines that aren’t tied to the grid. Gas goes in, data comes out, and there’s no effect on local electric cates.
Even better: build the data centers in the arctic where it’s naturally cold. A significant % of data center electricity is used for cooling. There’s also quite a bit of natgas in many arctic regions.
I downloaded the file from Census Bureau and the last 6 months for Office are the following:
86,766
87,977
89,040
87,678
89,213
88,919
Did I download the wrong file, or where did you get the +-40,000 number from?
RTGDFA instead of wasting my time with your internet incompetence. At least learn to understand what it is that you download, you goofball. Not my job to teach you and correct your screwups.
Lmao Damn Wolf
Re: Not my job to teach you and correct your screwups
You must have changed your terms of service agreement — it’s hard to keep up.
There’s more than increased electricity bills to contend with. I was surprised to learn that the demand for electricity in data centers, particularly AI data centers, can reduce the quality of the power supplied to nearby households, and that can cause appliances to wear out faster. Wendover Productions (Youtube) published a report about it. The report ends on a positive note – the changes they are undertaking to accommodate the power demands of AI are the same changes needed to make best use of renewables.
On a side note, I have started using ChatGPT as a search engine for information on the internet. No advertising, no clickbait, no misleading threads to click through. My life has become so much better in problem solving on a range of issues. It is not perfect but is light years ahead of what the big tech companies have offered. It even suggests ideas to refine my search. I am sold on AI at least as a search engine.
Flashman, All,
I know this isn’t a blog post to debate the pros and cons of AI, but I’ll weigh in as well.
I used Elon’s Grok Ai (xAI) to see how it works for a few specific searches recently. Overall, it’s pretty good. AI is real and it’s definitely going to make life easier for some, as well as replace many jobs where knowledge, learning, and intelligence are primary skills. Attorney’s, CPAs, Financial Advisors, for example. I would like to see most of .gov replaced with AI, since it seems to be light on the intelligence and heavy on the corruption. Any Congressperson can be replaced by AI/algorithm. Prove me wrong.
There’s a difference between online AI and online search engines. It’s an apples to oranges comparison. That being said, I still prefer to search for human-generated information, including news, articles, charts, etc., vs. AI.
I find with Grok anyway that it’s responses are limited and may not include some key points. It takes a human a lifetime of learning and experience to become proficient in a given field. AI can help, but I wouldn’t yet rely on it exclusive of humans. I’m not sure about AI’s critical thinking skills as I haven’t challenged it with any tests yet. Everyone needs a highly functioning BS detector, and especially today.
In my view AI is viable, but it will take probably decades to implement its full function. There are real dangers. So far, it has been shown that AI can lie and cheat, and this is just in cyberspace. Safeguards need to be included, especially with integrated with robotics/cybernetics.
Here are Issac Asimov’s laws:
(1) a robot may not injure a human being or, through inaction, allow a human being to come to harm.
(2) a robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
(3) a robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
(0/4) a robot may not harm humanity, or, by inaction, allow humanity to come to harm.
There are many Sci. Fi. movie warnings of AI gone wrong, including: “I, Robot” (2004), “The Terminator” (1984), The Matrix (1999), “Westworld” (1973), ‘2001: A Space Odyssey’ (1968), etc. This is fiction, for now. Let’s keep it that way…
Oh, and I do think that AI is a bubble, and I invest accordingly. Not investment advice.
I have a Facebook account and the information feed has been rendered useless by AI. I’d say 50% of the photos that come across my feed are AI generated false photos that attempt to use deception as click bait.
…sounds like one of the oldest aspects of computing, ‘GIGO’, now on steroids, will be emanating from data center outputs coming to you, soon! (…hoping Wolf’s brand of integrity might steal, virus-like, into this brave new tech…).
may we all find a better day.
While I am convinced that computer programs wrapping large language models (what we are now calling “AI”) are here to stay (until the companies go bankrupt or we find something better), given how ecstatic some of my colleagues are about them (despite the fact that they regularly, reliably, make mistakes, screw things up, cause headaches, and seemingly save no time for people given how much you have to babysit them), I am not convinced (given the parenthetical) that they are going to save the US economy or give us any kind of productivity edge over our economic rivals (unless they’re even dumber about them than we are). Much of the hype is coming from the highest echelons of the corporate elite, who have long believed that “AI” is going to turn them into immortal genies, and I wish I was making this up, but legitimately the heads of the big digital technology companies are all Singularitarians who have an emotional and pseudo-philosophical attachment to “AI” programs because they believe they’ll give them the power to live forever in a digital virtual reality. The next bubble will probably be some other Kurzweilian BS like computronium, or maybe we’ll get a re-run of early 2000s nanotechnology hype. Maybe the bubble only dies with Ray Kurzweil.
Evidence in software engineering is already emerging that the promised “10x productivity” gains from LLM-based coding tools are simply not going to materialize. In open source development, one study found an overall loss of productivity with the advent of LLM-based code generation tools. I have personally witnessed colleagues waste time with these things when they could have been working and using their own brains, a new way that computer programs are failing to show up in the productivity statistics. One co-worker wasted 30 minutes of his workday (and mine) trying to get his chatbot to understand the code he was working on and to identify the semantic error that was giving us unwanted results. I had already identified the bug and proposed a solution, but he didn’t believe me, he trusted the chatbot more, and this was around this time last year; very little improvement has been made on actual work tasks, rather than “benchmarks” which can be gamed (and is by dishonest “AI labs” such as OpenAI).
That isn’t to say that these things can’t be useful in a software context. Certainly, it’s nice to be able to auto-complete some “boiler plate” code… but we’ve had tools to deal with that since the 90s at least. You can crack open your copy of the Pragmatic Programmer (you do, dear programmers reading this, own a copy, right?) and find all kinds of old tools that software engineers used to not have to fuck around with boiler plate code. OpenAPI docs and project generators have existed for over a decade and are excellent tools which junior engineers do not know exist because they all graduated post-ChatGPT and don’t know how to actually write computer programs themselves (ok, ok, hyperbole, but whatever).
Given the program architecture, which makes “hallucinations” inevitable, I don’t think “AI” is comparable to the dotcom bubble or the internet, in fact, there is very little that it provides that is fundamentally new. Chatbots? Not new. Search engines? Not new. Abstractive summarization over search results? Not new. Code generation? Not new, and arguably worse in many respects than “Good Old Fashioned AI”-based code generation tools which are now falling by the wayside because of LLM hype (installing LLM-less Intellicode gave me a bigger productivity boost than Copilot – go figure). The internet was a whole new communication mechanism which allowed you to instantly contact anyone you knew who also had a computer (which is everyone nowadays), to find and retrieve reliable knowledge (easier back in the day than now), to transfer files without having to deal with physical media, to meet new people, to facilitate sales, etc., etc. The chatbots are, at their best, ways to avoid using your own brain power to accomplish a task that wouldn’t take you much extra time to do on your own.
All of these services are monstrously expensive and have to be subsidized by companies who sell digital ads on their social platforms or other kinds of rent-seeking (Microsoft IP rent-seeking makes them so much money despite the quality of their product being, at this point, worse than a Linux desktop); from talking to local business owners, it seems like there’s a growing disillusionment with digital ad spending. There’s decreasing bang for your buck, and the digital ad brokers like Google and Facebook are fundamentally dishonest about how effective their ad platforms are (go figure). Short-form online content may go the same way that broadcast and cable television did: fewer eyes on the product because it all got so trashy. I, for one, do not care to view a bunch of model-generated bullshit. I go online to keep up with friends and family, not so Zuckerberg can shove slop down my throat. My usage of facebook has fallen precipitously as a result but my viewership of websites like this one, which I keep up with via RSS and my hand-made feed reader, has gone up. The value of the internet is the people on it.
I witnessed another colleague of mine, who uses chatbots for non-programming purposes, have his weekly planning schedule get turned into a monthly planning schedule, which he had then had to edit by hand only to… put it back into the chatbot’s context window. He swears by it, but I don’t see the point. There is virtually no time saved. I love this guy to death, but he used ChatGPT to help write his master’s thesis, and the number of hallucinated citations which he had to catch makes me believe that there are probably some he didn’t catch. This is not a new problem (used to be you had to bullshit this yourself, now the computer can do the bullshitting for you), but it is already negatively impacting science, and feeding more garbage into the garbage-in-garbage-out algorithm (similar things are happening in software because LLM companies are training their LLMs on LLM-hallucinated code blocks – fun stuff!)
I would argue this bubble is somewhere between the “blockchain” bubble of the late 2010s and the dotcom bubble. Certainly the scale of it is more like the dotcom bubble, but the impact will be more like the blockchain bubble. Some companies and consumers will find legitimate use-cases for large language models. Some company will field a chatbot that is cheap enough to use without losing money on every query, but the singularity it will not be. It will mainly be used by office workers to bullshit on bullshit tasks (writing emails) and by students to cheat on their homework. DeepSeek is pointing the way there. DeepSeek is dumb as hell when it comes to reasoning (all LLMs are, but some are dumber than others), but it can regurgitate the text it’s trained on, and summarize the text it retrieves from a search engine query, as good as the rest of them, and with a lower overall cost to run. This will also negatively impact data centers because you do not need a huge data center to run DeepSeek, unlike the latest models from US digital technology companies, which, according to another recent study, are hitting a wall as they 1) run out of reliable data to train on and 2) find out that throwing MOAR DATA at a transformer does not result in more “generalized” “intelligence”, but actually degrades performance on various tasks (this was already evident with the jump from GPT-3.5 to GPT-4, the former having been discovered by some researchers to outperform GPT-4 on some tasks).
I am of the same opinion as Jobst Landgrebe and Emily Bender: this software will save time where tasks are easily dealt with by statistical approximation of a Web-scale text corpora (i.e., writing emails that no one reads). Outside of that, it will negatively impact productivity (already is) and lead to compounding errors that people don’t catch because they’re mystified by “AI”/next-token prediction algorithms.
From an old retired software developer, that was one helluva post, dude!
Great post. The hype around this topic is hyperventilation level. The fear mongering about how it’s going to take every job except trades is getting absurd. It’s like we’re all living in a new season of “Silicon Valley” (HBO). Wish we could all laugh a bit at the Singulatarians and immortal genies who in fact put their pants on one leg at a time like everybody else.
Interesting. What are your thoughts and perspective on bitcoin and stablecoin and the rest of that?
That is all utterly fraudulent Ponzi scheme garbage. Obviously.
Next time please, ask chat GPT to rewrite your rants in 25 words or less. ,
In 2020 we were leveraging chat GPT to write powershell scripts in our infrastructure team to rapidly automate several services and fixes that the development team didn’t have time to come back to.
AI is just like that for expensive production studio. You buy it for your senior developer, not the junior developer.
Summarize it with AI?
Nah. The detail is interesting and informative.
I’m a software developer with over 30 years of experience building commercial software, including AI applications for enterprise clients. I agree with Wolf that AI, like the internet, may be in a bubble, but it’s not going away. AI is incredibly useful, but its limits are still underappreciated.
What I’ve seen: People are throwing AI at every problem, but it’ll take years to find where it truly shines. So far, AI is great at aggregating information (e.g., summarizing Slack conversations for a manager) but terrible at making decisions (e.g., judging if an employee is promotion-ready).
My prediction: Not all AI applications will work out. But I’m skeptical of a bust; legitimate use cases will find value in the infrastructure we build. The key is learning what AI can’t do as much as what it can.
Disclaimer: I used AI to wordsmith this post :)
AI (Artificial Idiocy) can’t and will never do anything even slightly useful.
I think you are on the right track. The appropriate analogy in my mind is all of the fiber companies that over built in the 1990s and early 2000s. Many of them went bust before the fiber was utilized but eventually it paved the way for SaaS at scale and streaming video platforms etc. The place to bet is NVIDIA as they are making money hand over fist from every one at software-like gross margins and with no end in sight. Cuda is their most. It doesn’t matter to them who wins at the end of the arms race.
Federal Reserve and FDIC Release Public Sections of Resolution Plans for Several Banking Organizations
WASHINGTON — The Federal Reserve Board and the Federal Deposit Insurance Corporation (FDIC) today released the public sections of resolution plans for the eight largest and most complex domestic banking organizations and 56 foreign banking organizations. These banks were required to submit resolution plans by July 1, 2025.
Resolution plans, commonly known as living wills, are required by the Dodd-Frank Wall Street Reform and Consumer Protection Act. These plans describe a banking organization’s strategy for orderly resolution under bankruptcy in the event of material financial distress or failure. By regulation, resolution plans must be divided into public and confidential sections. To foster transparency, the agencies have required each banking organization’s public section to summarize certain elements of the resolution plan.
In addition, the FDIC released public sections of the separate resolution plans of 12 large insured depository institutions (IDIs), which are required under the FDIC’s IDI Rule. These IDI resolution plans, which were also required to be submitted by July 1, 2025, are designed to enable the FDIC to efficiently resolve an IDI under the Federal Deposit Insurance Act in the event of its failure.
The public sections of the resolution plans are available on the FDIC’s and the Board’s websites.
# # #
If you think Data Centers, etc are pushing up electricity demand, wait until we start approaching 100% EV usage.
Two years ago I was at a talk in Oak Ridge, TN where the Chairman of TVA (Tennessee Valley Authority) was asked by an audience member about the needed increase in electricity production due to electric vehicles. He stated that TVA has estimated that if all transportation in its service area switched to EV’s, it would need 3x the present amount of electricity production.
That is a lot of new power plants.
I hope alternate energy source also become more and more popular.
I have 4 evs, and zero electricity bill because of solar.
Nuclear can get it done.
…sounds like one of the oldest aspects of computing, ‘GIGO’, now on steroids, will be emanating from data center outputs coming to you, soon! (…hoping Wolf’s brand of integrity might steal, virus-like, into this brave new tech…).
may we all find a better day.
…whoops! 😬 There’s the dreaded double-post! Apologies, but as long as I’m here, any AI mavens willing to conjecture when it will cross the point of feigning human emotion to actually possessing its own? (…being a geezer, positing, and probably misunderstanding, the ‘singularity’ in reverse? A weird parody of God and Adam and this on the Sistine Chapel ceiling just appeared in my imagination…
may we all find a better day.
I envision when ai becomes “sentient”will be more like a angry/spoiled child and thus hope it has limited powers/connections or may get very interesting.
I would say be a friend to your ai now while you have a chance.
James – brilliant!
may we all find a better day.
Yeah…always cover you ass.
Btw, on the Sistine it appears to be a touch….which is an “on”, but we have no way to tell what Michaelangelo had in mind for components on either side of that completed circuit.
…good’un, NBay! (…always-good to flip the telescope when cya’ing…). Best-
may we all find a better day.
I’m clueless when it comes to data centers in general but it seems to me that these giant buildings will at some point become obsolete much like the cumputers that used to fill an entire room and are now handheld. Am I wrong?
All buildings eventually become obsolete and are torn down (except some historic buildings that are preserved at great cost). That’s why buildings are depreciated down to zero over varying time spans, such as the IRS rules of 39 years for commercial buildings and 27.5 years for residential rental buildings. If you get more out of a building than that without spending a lot of money maintaining and repairing it, consider it a bonus gift from real estate.
That’s just the building. The land itself is not depreciated and continues to have value, unless it is contaminated in a bad way, in which case, you cannot even give the land away.
Too many are overthinking AI. Old and retired and I use some form of it many times a day. Nothing fancy, just info I would like to know. Google camera is great as well as AI responces.
So much better than human interaction with attitude.
The number one use is and will be surveillance. It is like when they first installed cameras where I worked and later who cared except the criminals?
It will take very little to be better than most american workers. When lines are open at Walmart I still use self checkout.