Speculative Growth and the AI "Bubble" [pdf]
economics.mit.edu
economics.mit.edu
We've seen speculative over-growth with a good legacy at least three times in the last three decades. First was the dot-com boom. Overpromotion made it necessary for every business to have a web site. That wasn't pre-ordained. The Web could have maxed out as a distribution system for catalogs, data sheets, academic papers, and similar business to business info. Overpromotion created the business to consumer web, which turned out to be useful.
The second overbuild was long-haul fiber optics. Look up Global Crossing. So much fiber was put into the ground and water that intercontinental spam is not a problem. That didn't have to happen. If traffic was billed, it wouldn't have happened. It turned out to be useful, but was not pre-ordained from the economics.
A third overbuild was the solar panel industry, especially in China. So much money was thrown at solar panel manufacturing that the price became very, very low. Solar deployment accelerated and started to take over, after decades of panels costing too much. Now China has a solar panel glut. They're dealing with it intelligently - minimum efficiency standards are coming into effect, and pollution controls on panel manufacturing are being tightened.
Strong disagree, demand for the internet was insatiable, all one had to do to see the future of the internet in the 90s was observe just one school age person using AIM or MSN Messenger.
thing is, a few of us (mea culpa, sadly) figured out how to use it, even in 94, to sell stuff. the technology even back then was adequate for this.
So in that world, we would have fiber, but not (much) international spam.
Really cheap fiber backbones are make a pure datagram Internet work. We still can't handle congestion in the middle of the network. Congestion has to be forced out to the edges. There was a period in the late 1990s when MAE-EAST and MAE-WEST, the big peering points, routinely went into congestion collapse and only about 30% of packets got through
[1] https://en.wikipedia.org/wiki/Connectionless-mode_Network_Se...
This burned a lot of effort in the late 1980s and 1990s. Much incompatibility. Attempts to use intermediate machines to help with protocol conversion. Attempts to build protocol converters. A need for O(N^2) different protocol converters. Some computer vendors had to be dragged, kicking and screaming, away from their proprietary networks. There was a lot of pressure from DoD to standardize.
Down at the wire protocol level, there were alternatives. Too many of them. IBM had Token Ring. Apollo had a different token ring. Datapoint had ARCnet. Industrial automation people liked RS422, which is still around. DEC had a triaxial coax system. ARPANET IMPs had their own strange parallel interface. Ethernet came in two coax forms before un-shielded twisted pair took over, plus a form of Ethernet which connected via a 15-pin D-sub connector. Sometimes you just skipped the Ethernet cable and plugged all the 15-pin cables into a multi-port repeater. Plus there were a bunch of photonic links.
Finally, IP won as the only long-haul packet format, RJ45 jacks handled most electrical interconnects, and most things just plugged in and worked.
If enough capital has been installed before learning removes the wedge, the economy lands in the high-capital state,
edit: as I keep reading the paper, I keep noticing some common sentence construction patterns, stylistic choices, and other little tics that I find frustrating because they are the very same things that I've been working on a few "writing-style skills" to get rid of
A few notes:
1. This assumes that there is notable ROI on 'AI labor'. That is still up for debate.
2. This assumes that the interests are currently falling, unless I misread the paper.
3. This affirms that we are in an over valuated, speculative bubble which will inevitably correct; but it needs to "correct" at the exact right time defined by multiple factors.
First, "correction" can be an euphemism for a disastrous financial crisis. It could take years and years for most people to see the end of the tunnel. I don't know if the end justify the means.
Do we really need to engineer a financial crisis to build more energy facilities? And will they be built the 'right way', using renewable energy for example? What if we invested half of those trillions directly in socially impactful measures, instead of having the money flow through a speculative bubble first?
Finally, I am not an economist, but I wonder how accurate a mathematical model is to the real world - i.e. what happens to the model when Donald keep changing the opening hours of the Hormuz?
It does feel a bit like trying to read tea leaves to me. This reminds me of Hari Seldon's psychohistory:
> In Foundation (1951), famed mathematician and psychologist Hari Seldon has developed the science of psychohistory, which uses sophisticated mathematics and statistical analysis to predict future trends on a galactic scale. He has predicted the unavoidable and relatively imminent fall of the Galactic Empire, and intends to establish the Foundation, "a repository of crucial, civilization-preserving knowledge" that will enable society to revive itself more quickly and efficiently [...] [1]
---
[1](https://en.wikipedia.org/wiki/Foundation_universe#Psychohist...)
There’s no real “we” in this case. The money is coming from private coffers, people looking for ROI on their hard-earned money. The money isn’t coming from a central planning process.
Yet, "we" will suffer the potential consequences.
GHC '07 can be directly traced to a (in retrospect) failure of regulations AKA the government AKA "we".
Our democratically elected government has the ability to strike, what they think, is the right balance.
The companies and investors cannot be expected to take the broad view, so it is the WE who are responsible in a way, are we not? In a way, "the people" have, on balance, decided what level of regulations.
We would be tempted to complain that the lag and fidelity of choices by "the people" cannot be faithfully represented by the government. However, there's no better solution than what we currently have that I can think of, but I would love to hear about a scalable approach that caters to more individuals' specific wishes on a very very long list of policy decisions and that has short turnaround time.
Truly dismal science of an Economics professor at MIT.
However, GPUs and memory chips are some of the fastest-depreciating capital investments one can make. Overinvesting in this generation's GPU model will either be wasted (because the chips are worthless in a few years) or result in a lot of underinvestment in future years (because instead of replacing those GPUs with newer models you keep using them, unwilling to admit you bought several times as much as you should have).
If we have reached the point in the cycle where the boom's proponents are trying to argue that even if it was all a mistake, maybe it's ok, then one suspects were might be late in the boom part of the cycle.
Interest rates in fiat currency, these days, are not due to scarcity of lendable funds. Saving rates by consumers actually are just a low-risk or high-risk form of rent seek8ng., depending on the economy.
There's no world these days where a vast amount of collected wealth by a few makes credit suddenly cheap for all consumers.
All wage earners will still be able to drive up inflation if they demand goods that are not able to keep up easily... therefore pressure to save (to not take out debt) will raise interest rates fine by itself.
Unless we are talking about fiscal-based government spending (and interest on fiscal debt), no increased access to credit will become "better for all". Businesses can absorb all the credit offered during good times and consumers may see none.
Business failures will suddenly dump some assets on the market at fire sale prices, but a lost job makes those fire sales of little solace.
I am looking to test if there are several second order effects of rising oil prices and supply chain issues that can exacerbate financial contagion from an AI bubble (assuming we are in one) and, to the best of my understanding, it depends on what kind of mechanisms fail to contain the fall out.
I don't think it is reasonable to assume doom, but I would imagine there needs to be considerations from a much broader perspective as to discuss the possibility of 'a larger capital stock, higher wages, and a lower interest rate' that is paper is asserting.
I could be wrong (I am still trying to assess my hypothesis), but I am skeptical that the increased value/productivity from AI can overcome a rising cost of living if the war is sustained.
"workers operate with a larger conventional capital stock and wages rise even as the worker share falls."
Liberal Democratic capitalism splits power into two primary buckets: political and economic.
Marx provided the critique of consolidated economic power. The Soviet union proved the dangers of consolidated political power and Hayek made the mechanism explicit.
The "election tampering" BS is their attempt to try to undermine plutocratic political power. This looks like an attempt to justify the insane concentration of economic power that clearly goes against Hayek's description of free markets as a mechanism to discover preferences. Whose preferences?
If worker share is falling, workers are losing their share of the economic voting mechanism. Whether or not the emerging capital ownership class chooses to keep rents and subscriptions affordable to the new working subclass if and when they accomplish this power grab is immaterial, no matter how much math they try to wrap the propaganda in.
So we're trying to prove "mathematically" that the K-shaped economy is "rational", and trying to circumvent the fact that the mathematical rules of economics depend on the social, political, and behavioural substrate by which those rules derive their efficacy.
Which is fine if nobody has agency, politics is irrelevant, and we just accept everything we're told at face value if it's framed in sufficiently mathematical language.
So you have it exactly backwards: the K-shaped economy refers to the continued rise of asset valuations as workers real wages decrease.
But I'm not going to keep arguing with you since you seem to be afflicted by "Math brain" (defined here by the belief in homo-economicus where humans can be accurately modelled as a selfish calculator - and please don't correct me on the calculator metaphor, I'm being hyperbolic).
Let's take what we do know from AI and cognitive science seriously: if we accept that humans are some sort of biological prediction machine, that doesn't give us a completely predictive model. There's a sort of "measurement problem" when it comes to self-prediction: while the goal of environmental prediction is clearly accuracy, the goal of self-prediction cannot be accuracy since self-prediction is inherently accurate by virtue of the actualisation of self-predictions. What then discriminates within the self-predictive model between multiple actualizable self-predictions?
If Math-brained believers in homo-economicus are to be believed, then human behaviour is completely determined by what basically amounts to greed. Certainly that's what generally selects for those who occupy positions of power and prestige in our society, and thus based on current political norms has turned into a somewhat self-fulfilling prophecy.
In this since Nietzsche was right: "God is dead." He was killed by enlightenment thinking and replaced with rational self-interested agents governed by predictable macro-economic models... which keep being wrong, because humans are not universally greedy and behaviour, politics, and psychology actually matter a great deal. (Nor is the new god of "rational" selfishly motivated acquisition, expansion, and power-seeking really that new of an invention - he is sometimes called "Moloch", and is in fact perhaps the oldest God, and humans keep discovering he's a pretty shitty God to worship; I have no doubt that he will prove equally destructive and unkind to us if we insist on continuing to and/or reverting to worshipping him dressed up in "rational" mathematical economic language).
And the K in K shaped economy came from turning the K 90 degrees so that the left arm (representing the low wage service sector) and the right arm (representing rising asset prices) were elevated while the middle was depressed. It got retconned to mean "increasing inequality" by those who couldn't bear to admit that something had gotten better for those lower on the wage scale (which objectively happened back then).
As much as you may want to move from the low to high equilibrium, it would be pretty hard in an ancient society. In the modern world, with enough exuberance, it seems much easier.
Because we definitely seem to be in a world where GDP inflation is getting more and more divorced from incentives to actually produce things.
With respect to AI and the paper being discussed specifically, all evidence I've seen is that AI's primary applications are within military and intelligence and that the escalating arms race in those domains is actively undermining things that people would generally like to see being produced (food and housing specifically are illustrative here).
Notably, one of the reasons that the military industrial complex can be hard to unwind is that it works well for capital and workers. Consider all the programs you've heard of which boast about providing jobs in every single state. Whether these different equilibria are better or worse for labour (or fractions of labour giving you your diverging K-shape) is a separate question and potentially changes over time.
One historical event you could model as a change in equilibria was the industrial revolution. Instead of elites consuming economic surpluses, the economy change into a mode where those surpluses could be reinvested in productive capital, new technology, and manufacturing in a self-reinforcing cycle (of course these avenues didn't really exist before). Famously, within this new equilibria, labour did not do well at first; awful conditions in factories with terrible life expectancies. But those workers' grandchildren lived lives of relative prosperity as a result.
All of which is to say, the existence and description of multiple equilibria in economies is a different question from how labour and capital fare in those equilibria.
A temporary overvaluation can build enough real capital that the economy lands in a permanently higher-capital equilibrium, even after the inflated valuations correct. The future for AI companies may look rather iffy, but the whole economy may not be as screwed as some fear.
If the "higher capital" that results from an AI boom consists of massively parallel computational resources that currently can only be fully utilised by AI and crypto, and if those things turn out to be a bust, the "higher capital" only has value if we find something else to do with it.
Maybe we will...
The model in my head is more like DotCom telecom. The massive overbuild in fiber was eventually used and even used for the purpose that it was imagined for during the boom. It's just that the companies that built it mostly went under and new owners acquired it at a profit-supporting price.
Data centers and electrical infrastructure has a similar long term value, but most of the AI investment is in compute/manufacturing capacity for current nodes which doesn’t age nearly as well.
I mean, compute depreciates, but I think there is zero chance that the value of inference or training is going to fall to zero. Market discovery will find the right price provided the market has the right degree of freedom. Given the type of market it is, I don't see how that won't be the case.
That implies to me that in the future we'll have models as good or perhaps better than the state of the art at the moment, but on hardware chips that can be put in places where you can't currently locate a datacenter, and operating at hundreds of times better power efficiency, which sounds pretty great.
This will lead to a superabundance of power-hungry compute power in the hyperscalers, and it's not entirely clear what can be done to consume it all and still run at a profit unless they manage to make ever greater gains for ever more compute-hungry models that cannot be run on consumer devices, unless they refresh their hardware at ever faster and more expensive rates.
The joke about data centers used to be that their core business was selling power at a loss; this may end up being true of the hyperscalers next.
Some workloads may make sense running for a few hours a day during cheap solar prices on outdated hardware, but in less than a decade the value is very much hitting zero.
Some people thought that it was misguided when they extended the depreciation cycle of the current AI build out year(s).
In terms of raw performance, there is still some headroom (maybe) but those gains are going to be marginal when you look at the amount of compute per watt (if its more than 5 percent I will be shocked). And that push is going to create a whole other set of problems (cooling is going to be an issue, it already is).
It is fairly likely that this hardware buildout has more legs than one might suspect based on history.
Most compute isn’t going to be new at a potential crash and recovery takes time.
It's like the tulip bubble of the 17th century [1]. Having a bunch of money tied up in useless tulip bulbs didn't do anything productive after the collapse.
And beyond physical infrastructure there are the intangible assets: the learning and the process innovation across multiple fields.
The upfront price for all that may end up steep, or fair, or even cheap… the truth is no one knows yet
It's like comparing a railway line from a mine to a smelter with a city's road network.
Utility scale power stations have outputs measured in GWs. Data centres are measured in MWs, although people are trying to build GW scale data centres at the moment. But even then a data centre will want a proper grid connection, otherwise they have a massive single point of failure in the form of the directly connected power station.
It’s also very unlikely that purpose built power station is capable of offering cheaper than grid power anyway, except in the very special situations like Hydro. So if you’re gonna build a datacentre, you will want a proper grid connection capable of providing all you needs. Even if you’re running on dirty gas turbines in car park initially while waiting the grid hardening happen. In the long term, that grid connection is always going to be the cheapest, most reliable source of power, ignoring it completely would be foolish.
Furthermore, to elaborate on my point above: building new power plants in tandem with data centres that demand GWs of additional power does nothing to address the needs of the grid itself. The grid is not built to handle all that new power, and the consumer electricity demand is not there anyway.
Building new power plants immediately isn’t really required to serve data centre, they’re huge amounts of spare capacity in most grids that only exists for extreme situations. The hard part is making sure any existing capacity can actually be delivered to the new data centre. New power plants will get built to rebuild spare capacity over time, or more likely, solar and wind will get built to reinforce that spare capacity over time.
Datacentres are useful grid loads because they all have backup power on site capable of powering the entire site as an island. Which means grid operators can ask them to disconnect quickly if they need to extra power for short periods of time to handle power spikes, or local grid faults. So it’s possible to connect quite a lot of extra data centre type load before generation capacity becomes a serious issue.
Really, now? Yeah, they can ask, but do data center operators just go "oh grid needs our power! Let's switch to backup!"
That seems unlikely. Why would a data center (or any large consumer of electrical power) be forced into a backup/contingency plan, when no crisis exists at all? Other than drilling/practice runs, I cannot see them readily agreeing to this. Is it actually a routine occurrence?
Of course they do. The grid is going to load shed regardless, better to disconnect in an orderly fashion (especially as the grid will pay you to do so) than wait for your local substation to disconnect you.
> That seems unlikely. Why would a data center (or any large consumer of electrical power) be forced into a backup/contingency plan, when no crisis exists at all?
Because they get paid to do it. This is pretty basic demand side management that’s been the bread and butter of grid management for decades now. Grids will offer discounts to large loads that allow themselves to be disconnected at short notice, and will then pay those loads for the periods they’re disconnected. There’s usually multiple tiers of discount and payment depending on how fast a load can disconnect, and how often the grid expects to call upon that facility.
Good for the economy, what about the value of the labor that it's currently screwing over?
I don't give a single damn if "the economy" grows if it means my skills become worthless and I become basically unemployable anywhere near my previous earning ability
Edit: even if the value of "the economy" does strongly in the future, is the value of "my labor" ever going to recover?
If no, then fuck it. Why should I care?