As the author says: this capex isn’t a railroad, it’s a very expensive immediately depreciating asset.

We should be so lucky

We won't. Well lay off engineers to balance the books and destroy unneeded capacity.

A competent government might rent these company’s GPUs to cure major diseases once they’re sitting idle.

Bitcoin still needs a-minin’.

yearly AI capex is more than NASA and NIH combined

Very cheap game consoles and VR headsets. Unironically that could really help world peace and QOL: less news and doomscrolling, and people would have an outlet for stress, anger, and boredom.

If you have a GPU which you've used for AI training, but that's no longer valuable, you could sell that GPU; but then you'd incur taxable revenue.

If you destroy the GPU, you can write it off as a loss, which reduces your taxable income.

Its possible you could come out ahead by selling everything off, but then you'd have to pay expensive people to manage the sell off, logistics, etc. What a mess. Easier to just destroy everything and take the write-off.

On the other hand, drug discovery sounds like it's a candidate for really benefitting from AI. To fuel AI model development, there maybe has to be all the garbage that comes with AI.

The rest of the world has not caught up to current LLM capabilities. If it all stopped tomorrow and we couldn't build anything more intelligent than what we have now: there would be years of work automating away toil across various industries.

Re hype: Why is it that so many people are completely obsessed with replacing all developers and any other white-collar job? They seem to be totally convinced that this will happen. 100%

To me, this all sounds like an “end-of-the-world” nihilistic wet dream, and I don’t buy the hype.

Is it’s just me?

Aren't advancements in AI actually helping drug discovery?

This implies that a human life is valuable but what they spend it doing is not.

I'm always surprised by the number of people posting here that are dismissive of AI and the obvious unstoppable progress.

Just looking at what happened with chess, go, strategy games, protein folding etc, it's obvious that pretty much any field/problem that can be formalised and cheaply verified - e.g. mathematics, algorithms etc - will be solved, and that it's only a matter of time before we have domain-specific ASI.

I strongly encourage everyone to read about the bitter lesson [0] and verifier's law [1].

[0] http://www.incompleteideas.net/IncIdeas/BitterLesson.html

[1] https://www.jasonwei.net/blog/asymmetry-of-verification-and-...

Cloud gaming 2.0

During the 1990s dotcom boom we massively overbuilt fiber networks. It was indiscriminate and most of the fiber was never lit.

After the dotcom crash, much of this infrastructure became distressed assets that could be picked up for peanuts. This fueled a large number of new startups in the aftermath that built business models figuring out how to effectively leverage all of this dead fiber when you don't have to pay the huge capital costs of building it out yourself. At the time, you could essentially build a nationwide fiber network for a few million dollars if you were clever, and people did.

These new data centers will find a use, even if it ends up being by some startup who picks it up for nothing after a crash. This has been a pattern in US tech for a long time. The carcass of the previous boom's whale becomes cheap fuel for the next generation of companies.

I live next to an abandoned building from the Spanish property boom. It's now occupied illegally. Hype's over yet the consequence is staring at me every day. I am sure it'll eventually be knocked down or repurposed yet it'd be better had the misallocation never happened.

It seems like it's specifically based on nvidia's sales, which I assume almost entirely for deep learning. "Regular" data centers don't need many GPUs.

> Because everyone and their mother are labeling regular data centers as AI data centers.

Companies (generally) build things with an expectation for a return on their investment: what "regular" data centre usage would necessitate these kind of build-outs?

To sell more Postgres or WordPress VMs/instances? Is that being used to justify the spending in shareholder conference calls and regulatory filings?

I wish the hardware itself were renewable too.

Data centers want firm power to avoid underutilizing expensive assets. Solar and wind are intermittent. New gas has a years-long lead out. 12+ hours of batteries (solar in winter) are, in fact, not free or de minimis.

These companies would all LOVE to use small modular nuclear reactors in their datacenters... if the NRC ever got around to approving a license for one.

Many of these companies are very interested in small nuclear tech as a means to power these facilities. The major bottleneck now for most of them is finding sites where grid capacity exists to power them.

Talking to anyone in the space for more than 30 minutes and nuclear with come up.

I very much hope the hype cycle lasts long enough for some of this capital raining down from the sky to get these reactors deployed in the field, because those will be a lasting positive from this hype cycle - much like laying railroad infrastructure and fiber optic cables came from other hype cycles.

I've often said that the robber barons sucked, but at least they left massive amounts of physical infrastructure for the average person to benefit from. The hype cycles of late leave very little in comparison.

Why renewables? These guys don't have an energy problem. The states have made sure give them energy credits and other assurances on that. The problem they do have though is siphoning the water from everyone around them.

Microsoft actually has a design for mini datacenters that stay cool in the ocean and collect tidal energy. But it's way more fun to have states trying to court you into building datacenters cause it'll bring some jobs.

At least in the US, it doesn't need to be forced. In 2024 94% of new US powerplant capacity in the US were renewables or battery storage, we're on track for 93% for 2025, and based on announced plans the next few years will see very similar numbers. What few fossil fuel power plants are being built are exclusively natural gas, and a decent number of them are conversions of former coal plants to natural gas. Planned additional natural gas plant capacity is lower than at any point since the shale revolution started. Renewables have won.

> I just wish we forced every new data center to be built with renewables or something.

Already the case in Europe. And in the US, most of the biggest player are doing this: Google, Microsoft, Meta, AWS. By now those 4 are the largest buyers of renewable purchase agreement in the entire world. MS alone invested something like 20B in renewable purchase.

But the issue is that installation of renewable in the US is not bottlenecked by lack of demand, it's bottlenecked by permitting, zoning issue etc. The queue for power deployment right now is something like 100GW (i.e. how much production is paid to be built, but not yet built), that is around 10% of the current total US power capacity. So it's not really clear to me if buying more renewables helps making it's deployment faster through economies of scale, or if the purchase order is just sitting in a queue for years and years.

One notable exception is xAI/Grok, who has one of the biggest cluster, is powering it 100% with gas and afaik did not offset it by buying the equivalent in renewable. Having built the cluster in a what was a washing machine factory that does not have adequate power supply or cooling tech, they have been rolling in 35 mobile gas turbines (large trailer trucks that you connect to gas pipes) and 50+ refrigeration trucks. IMHO, it should be illegal to build such an energy consuming system with such a poor efficiency, but well.

There were some major network upgrades during COVID times dueto traffic growth induced by work from home.

Hence the title, "Honey, AI Capex is Eating the Economy."

> if this is a net positive for our society

Maybe that's something that can only be determined looking back. There are so many unknown unknowns.

Roughly 5% of energy in the US is dedicated towards AI datacenters. The current usage will double to 70–90 TWh/year by 2026-2027. For software heavy tech businesses it makes sense to host your own AI data center so that you could train it on your codebase and have developers build faster. Not sure if this benefits humanity that much..

Is Argentina a net positive for our society? There's the grilled beef, but every country has some kind of barbecue. There's a few soccer players I guess? Is Argentina worth the energy expenditure though?

The only sign you'll ever get is the pop, no?

What future evidence would convince you it's not a bubble?

We are only 2 years in! 1.2% of GDP is enormous! The fact that we can even make any of these comparisons is stunning.

- The birth of the space age, and more realistically the birth of the ICBM and satellite age. Both key to national security, and in the context of a cold war.

- 40% of long-distance ton miles travel by rail in the US. This represents a VAST part of the economic activity within the country.

- A literal plague, and the cessation of much economic activity, with the goal of avoiding a total collapse.

- ...Come on.

So we're comparing these earth-shaking changes and reactions to crisis with "AI"? Other than the people selling pickaxes and hookers to the prospectors, who is getting rich here exactly? What essential economic activity is AI crucial to? What war is it fighting? It mostly seems to be a toy that costs FAR more than it could ever hope to make, subsidized by some obscenely wealthy speculators, executives fantasizing about savings that don't materialize, and a product looking for a purpose commensurate to the resources it eats.

As a % of GDP it doesn’t seem very large, but that’s because our GDP is so massive. This is still an entire Norway’s worth of GDP.

Like 1.2% isn’t a big percentage, but neither is 3.4% - our total military expenditures this year.

Yeah that's my first reaction to. 1.2% doesn't sound much. It's just people making headlines out of thin air. If it lists the water and energy consumption I might be more concerned.

Slightly off-topic, but ~9% of GDP is generated by "financial services" in the US. Personally I think it's a more alarming data point.

You're looking at dots on a graph when you should be looking at lines and curves (and slopes of curves). The author makes this argument:

* Movement of capital from other fields to "AI". * Duration of asset value (eg, AI in months/years vs railroad in decades/centuries). * "Without AI datacenter investment, Q1 GDP contraction could have been closer to –2.1%".

Covid stimulus isn't directly comparable to the other three programs you listed. It was more like tax legislation than an industry.

1/100th of the US economy sounds massive to me.

The crux of the article is asking whether such a large investment is justified; downplaying the article saying it's only X% of the GDP compared to Y doesn't address the issue.

Those were all mega events though.

> Apollo program: 4%

More than a decade long. The technology and industry here was broadly shared. They did things like highjacked bra manufacturers to make space suits.

> Railroads: 6% (mentioned by the author)

We're still using this investment today.

> Covid stimulus: 27%

The virus that was killing us the fizzled is probably not the best example... Only arguments will ensue if I even attempt to point thing out in this one.

> WW2 defense: 40%

I mean Russia made its last lend lease payment in 2006. It lead to America dominance of the globe. It looks like an investment that paid it self off.

How much of the hardware spend on AI is going to be usable in 5years?

There are some deep fundamental questions one should be asking if they pay attention to the hardware space. Who is going to solve the power density problem? Does their solution mean we're moving to exotic cooling (hint: yes)? Have we hit another Moores law style wall (IPC is flat and we dont have a lot of growth left in clock, back to that pesky power and cooling problem). If a lot of it is usable in 5 years thats great, but then the industry isnt going to get any help from the hardware side and thats a bad omen for "scaling".

Meanwhile capex does not include power, data, constables or people. It may include training, but we know that can't be amortized. (how long does a trained system last before you need another, or before you need a continuation/update).

Everyone is going after AI under the assumption that they can market capture, or build some sort of moat, or... The problem is that no one has come up with the killer app where the tech will pay for itself. And many in the industry are smart enough not to try to build their product on someone else's platform (cause rug pulls are a thing).

"AI" could go the way of 3d tv's, VR, metaverse, where the hype never meets up with hope. That doesn't mean we wont get a bunch of great tooling out of it. It is going to need less academics and more engineering (and for that hardware costs have to drop...)

Total global military expenditure as a percent of global GDP is about 2.5%.

But it's not about "wow, how much!", it's about "wow, it even registers!"

Where is that improvement coming from? Hardware is already here to compute gemm as fast as possible.

We're already seeing it with DeepSeek's and other optimizations - like that law with highways - the wider the highway the even more usage of it. Dropping by 90% would open even more use cases.

For white-collar jobs replacement - we can always evolve up the knowledge/skills/value chain. It is the blue-collar jobs where bloodbath with all the robotics is coming.

This is where I do wish we had more people working on the theoretical CS side of things in this space.

Once you recognize that all ML techniques, including LLMs, are fundamentally compression techniques you should be able to come up with some estimates of the minimum feasible size of an LLM based on: information that can be encoded in a given parameter size, relationship between loss of information and model performance, and information contained in the original data set.

I simultaneously believe LLMs are bigger than the need to be, but suspect they need to be larger than most people think given that you are trying to store a fantastically large amount of information. Even given lossy compression (which ironically is what makes LLMs "generalize"), we're still talking about an enormous corpus of data we're trying to represent.

We do already have ASICs, see Google's TPU to get some cost estimates.

HBM is also very expensive.

I wouldn't wait. fpgas weren't design to serve this model architecture. yes they are very power efficient but the layout/p+r overhead, the memory requirement (very few on-the-market fpgas have hbm), slower clock speed, and just an unpleasant developer experience makes it a hard sell.

My chip friends constantly complain about Qualcomm owning a bunch of FPGA related patents that stifle any meaningful FPGA progress.

For a first approximation, yes? Their earnings have been growing steadily for years ($4-5B growth in each of the last 8 quarters), with no seasonal effects.

Q1*4 is highly likely to be a better estimate of their eventual 2025 calendar revenue than their current trailing 12 months revenue would be. Probably still a bit conservative, but easier to justify than projecting that growth continues at exactly the same pace.

> "... builds on the assumption that the economy is a zero sum game when it clearly is not"

Be cautious making assessments as to compounding effects; while it remains the critical attribute, the compounding nature of a system is not always obvious. For example, the author is correct that financing for AI CapEx is starving other fields of investment at least in the short term.

In addition, the massive investment may be overbuilt, but then likely will end up being useful long-term. This is just as the initial buildout of the internet to power pets.com was a bust at the time, but then led to Amazon, YouTube, Zoom, etc and in fact led to us being able to weather Covid better than you'd have expected.

The modern internet came from folks getting connected over-exuberantly based on near-term returns (with a lot of investors losing their shirts) but then humans figured out what the actual best use of the technology.

Highly recommend this book for more, Carlota Perez is very insightful: https://en.wikipedia.org/wiki/Technological_Revolutions_and_...