>Energy-efficient lighting, such as LED bulbs, has become a popular solution to reduce energy consumption. However, it serves as a notable example of the Jevons Paradox. While LEDs significantly lower the cost of lighting per unit, this reduction often leads to increased usage. For instance, people may install more lights in their homes or businesses, use them for longer durations, or adopt decorative and outdoor lighting more extensively. As a result, the total energy consumption for lighting has not decreased significantly despite the efficiency improvements.
This felt false to me, LED bulbs are generally 10x more efficient, so the average number of lights would have had to increase 10 fold, and this is ignoring the savings from not having AC cool those other 50 watts a bulb that went to heat.
So I took a look at the source, which doesn't actually have a source and just links to another white paper that claims light pollution increased, but is unclear about how much and due to what. Perhaps that whitepaper links to another whitepaper, but this is where I stopped.
Kind of makes me question how often I blindly trust Wikipedia, though this point in particular is pretty minor.
This is exactly the case. Anecdotally speaking, growing up in the 90s, I remember my parents adding lightbulbs and fluorescent tube lights to their shopping list every couple of months. We used to have a small stockpile of these in the pantry.
Fast forward to today, the last time I bought lightbulbs for my house was over 5 years ago and that was to upgrade from regular led lamps to Philips hue.
My sister just did a home remodeling and she added must have been 50 LED recessed lights in the ceiling. There are automated LED lights in the outdoor pathways, lights hanging on the balcony rails, lights under kitchen cabinets, lights inside closets, lights inside pantries, solar powered LEDs in backyard, etc.
Her appliances also have more indicator lights than ever before.
I’m not even counting the countless LED screens such as phones, iPads, TVs.
In total, she may every well have 10x more lights than before the LED boom.
Time Square in NYC was famous for having many lights and screens. Walk around any large Asian city in 2025 and they all look like time square with countless lights and screens.
Thats said, there is a limit to lighting demand because we still need darkness to sleep. I think the upper bound for intelligence is far higher than LED lights.
Still it does not use as much electricity as the previous generation of light bulbs and she won’t be replacing them often. So from a perspective of the industry it’s not so good. The tech is better but there’s less sales / margins/ profits altogether.
Is this true? What were the revenues of lighting companies before compared to now? Do we count TV screen makers because they’re LEDs as well?
I just watched Veritasium’s video on how LEDs were indented and the companies making LEDs saw exponential increase in revenue. https://youtu.be/AF8d72mA41M
Basic Google search tells me the lighting industry has experienced high growth due to the invention of LED lighting.
Even if she has 10 times as many (well possible if compared, say, to the 1960s), they aren't likely switched on most of the time.
People generally don't like if it is "too bright" in the room, and the extra lights in closets, pantries etc. will only be used marginally, for a few minutes a day.
It helps that LED lifetime isn't degraded by frequent on-and-off cycling.
Also keep in mind that better lighting efficiency enables people in poor countries to have lights where they couldn’t before. It’s not just wealthy counties. You have to count them as well. It's not just counting a rich person's house before and after LEDs.
Absolutely, yes. Many more people can now afford light.
My grandma (born 1926, died 2016) was born in Eastern Slovakia in the pre-electric era, the region she lived in only got connected to the grid in 1945.
In her youth, the whole family would simply have one kerosene lamp and gather around it after dark.
This felt false to me, LED bulbs are generally 10x more efficient, so the average number of lights would have had to increase 10 fold, and this is ignoring the savings from not having AC cool those other 50 watts a bulb that went to heat.
Jevons Paradox does not state that there is a 1:1 relationship between efficiency increase and consumption. A 2x increase in efficiency does not always result in a 2x increase in consumption.
With that said, it would not surprise me if there is a 10x increase in light sources (not total lumens) since LEDs.
The comparison becomes meaningless when we count the milliwatt, sub-Lumen LED in an outlet wirh incandescents. "I had 8 incandescent lightbulbs, but now have 1000s of LEDs- including outlets, TV backlights, DVR status" isn't comparing like with like.
Jevons Paradox describes the situation where a 10x efficiency increase results in a >10x deployment of lumens, such that the amount of energy consumed for lightning increases relative to when lighting was inefficient.
Comparing this with the solid blue price line showing price indicates that the 100 fold fall in price per unit of light in the 20th century was due to increased efficiency, mainly as electricity replaced gas. Yet demand rose strongly over the period, increasing more than 100 fold. This challenges the view that increases in efficiency translate into reductions in demand, and thus emissions.
That's a good source, although that historical analysis extends to the year 2000 when even CFLs only had reached ~30% market share, and well before LED lighting hit mass adoption. It is only speculating about LEDs. There's no question that lighting demand has continued to grow but the efficiency increase has also been remarkable, so unless you know of another source that does the math, it's unclear to me to what extent it's ended up balancing out in this case.
Another confounder to Jevons' paradox is that the price of the input (electricity) can also change, independent of changes in efficiency. If electricity prices were to go up ten-fold during the switch to LED lighting, then we wouldn't expect much change in illumination (probably a net decrease, as people would be hurting so much from other electricity expenditures that they'd be hunting for any savings on their energy bill). That's of course a fictional scenario, but in the case of GPUs, well, they have been getting quite expensive.
I somewhat doubt there is now 10x light fixtures... Thus 10x the traditional bulbs. That seems rather large undertaking in most buildings. Sometimes I think it makes sense to stop and think what would 10x more of some items mean in practise.
LEDs are around 3-4x more efficient than fluorescent lamps, (1) 28W 2x4 LED fixture puts out around the same amount of light as (1) 4-lamp 2x4 fixture with (4) 28W fluorescent lamps, depending on the lumen output package on the LED fixture.
> This felt false to me, LED bulbs are generally 10x more efficient, so the average number of lights would have had to increase 10 fold, and this is ignoring the savings from not having AC cool those other 50 watts a bulb that went to heat.
The first 5W also went into heat. Perhaps with the exception of light escaping through a window, *every* watt that goes into a lightbulb comes out as heat for your AC to cool.
I would not be surprised if there were a 10x increase in use of lighting with LEDs.
Not taking Wikipedia (or any source) on blind trust is always wise. More often than not, when you kick tires, you come back to folklore. It's fine for big things, but details are often wrong. Ditto for newspapers, for that matter.
> I would not be surprised if there were a 10x increase in use of lighting with LEDs.
Has the average homeowner added more ceiling lights or lamps because of different bulbs?
Has the average office added more lighting because of LEDs? Factories?
In order for an overall 10x increase in usage, there needs to have been some MASSIVE new usage or expansion of lighting, somewhere, to make up for all of these enormous use cases (residential, commercial) where there was probably close to zero increase.
> Has the average homeowner added more ceiling lights or lamps because of different bulbs?
As much as I appreciate the snark and sarcasm, at least where I live, the equivalent is happening.
Old home and commercial construction would typically have one central light per room. Homes would have a single light fixture with 1-3 lightbulbs at the center of each ceiling and were pretty dimly lit. Commercial would be similar with fluorescents. You would turn it off when you left a room.
New home and commercial construction has accent lighting all over the place, some turned on 24/7. Cabinets have under-cabinet-lights. Closets have closet lights. We have rows of LEDs lining all sorts of accents.
If you compare 80's-style construction to 2020's-style, the amount of new light seems to be about that much of an increase.
I can accept that new construction may have more lighting, but nowhere near 10x, and certainly not 10x overall when you factor "legacy locations" where it is mostly just a bulb replacement.
Are you really claiming that a new home has 10x more lighting than an old one? Or are you just saying that your own house it lit up like a lighthouse - search lights in every room, perhaps?
Not to mention that in poor countries, the invention of LEDs could mean going from being able to afford electricity for 1 light in the house to 10 lights.
If a 5w bulb and a 60w bulb output the same amount of light, those 55w are being wasted elsewhere (heat).
I said 50 because I was afraid someone would go 'ACTUALLY LED's only save 50-55w', I didn't expect someone to argue that the 5w led bulb is also outputting heat. You are of course right, 2.5w of that 5w is likely heat being cooled by the AC.
>I would not be surprised if there were a 10x increase in use of lighting with LEDs.
Quite frankly I don't care either way, I just wanted a source. I'm happy you feel that way though.
Quite frankly I don't care either way, I just wanted a source. I'm happy you feel that way though.
This is the best write up I found for lighting source, efficiency, and consumption relationship.
As a first step in looking at this, I’ve taken the data from the analysis and plotted it on a log scale. This shows clearly how consumption has increased as population and incomes (measured by per capita GDP) have risen and prices have fallen. The central role played by efficiency is indicated by the dashed line, which shows the number of kWh needed to produce a unit of light (1/efficiency). Comparing this with the solid blue price line showing price indicates that the 100 fold fall in price per unit of light in the 20th century was due to increased efficiency, mainly as electricity replaced gas. Yet demand rose strongly over the period, increasing more than 100 fold. This challenges the view that increases in efficiency translate into reductions in demand, and thus emissions.
The intuition behind this is that people who had a usecase for the tech but were locked out because they couldn't afford it, or didn't believe it was worth it usecase can be engineers or researchers. Some of these people may perform better than expected causing a gap in true market value which is filled when they scale up to their new size
Related: Jenson's Paradox: The more you buy, The more you save
I guess my question about the current events is: Are people really locked out of LLMs due to price? It seems like everything already has AI in it and that virtually every end user hates it. I could see a shift to more local models or something, but not an increase. I feel like LLMs have largely been oversold as a solution in search of a problem.
Or are people just applying this as everyone and their mothers are going to start training competing models to carve out their piece of the market?
These are the four main problems with LLMs (and related technologies) as I see them:
1. You can't tune them to your needs; they have restraining bolts and the training data is a generic corpus
2. You don't own your interactions with them; your data transits a network and is processed by third-party servers
3. They waste an immense amount of power relative to the usefulness of their output
4. Their responses tend toward uncanny simulacra and hallucination
Bringing the cost way down and making them trainable on consumer hardware solves or at least greatly alleviates problems 1-3. That just leaves problem 4, which might still be unsolvable and sink the whole endeavor, but at least can be focused on.
Absolutely! Even for inference! The SOTA models for all commercial purposes need to run on a consumer’s device.
Running either Grok2 or DeepSeek or even Llama405b requires nearly 400-500gb of memory.
Buying a tinybox with enough gpu memory costs $15k-25k. Or equivalently the same if you build your own.
A distributed Mac cluster costs about the same, if not more, if you’re buying 2-3 M2 Ultra each with 192gb of memory.
So people are absolutely constrained by price/supply here. Every engineer, analyst, scientist would be far more untethered by rules & regulations or policies & terms-of-service nitty gritties if they can trust that LLM they use is completely local, without-telemetry or tracking and is licensed fairly for commercial use (perhaps this excludes llama).
Not a lot of people can afford $15k-30k in spending for a computer (that can run this sota llms). But you can a billion will buy one when it’s $1k
Not to mention, the north star is to get to a place where we have the hardware to do training at home. we're a long ways off, but without the restrictions of needing the hardware to do it, ideally, we'd make the model such that is continually being trained.
Are people (or companies) locked out of training LLMs due to price?
I don't know the immediate answer. I really expect it to be a resounding "yes" on the long term because different LLMs should be good for different things. But in any way, this is not about the people adding a cloud client to their software.
> Are people really locked out of LLMs due to price?
Yes. For example Google has just made Gemini a standard part of Workspace, before that it cost $36 or something per month per user. That was too much for many SMBs to experiment with (you and I understand that the potential efficiency gains are way higher but for an SMB, paying 4x more for your office suite sounds bad).
Consumers of AI do have sensitivity to pricing. Many OpenAI customers "ration" their usage. I imagine lower costs open up new demand for these people.
On the Service side: it seems to have reduced the cost of operating a "commercially viable" (something people will actually pay money for) LLM. But even beyond that, "self hosted" models are also far more affordable now, which means models that target specific niches can be viable to build or buy.
However, this won't be an overnight phenomena. In the short-term, it will seem like demand drops, but in the long-term demand will go up. Big caveat: that demand may not be concentrated on the current incumbent players.
And finally, the elephant in the room: AI still needs to become more useful to reach its full potential. Easily a decade+ more needed here.
> Are people really locked out of LLMs due to price?
I think so. When you start using a good model, you soon learn that it can make a difference. If you have domain knowledge, you can treat LLMs like a design partner. With DeepSeek, what would have cost 50 dollars to use Sonnet is now 5 dollars. Having to only spend 5 dollars a month can be a real game changer for many.
"Induced demand": The highway is expanded such that it only takes me 15 minutes to get downtown instead of 45, so I go downtown more.
Jevson paradox: The highway is expanded such that it only takes 15 minutes to get from the suburbs to my office instead of 45, so I move to the suburbs.
So one is movement along the demand curve, while the other is a movement of the demand curve.
Yes, but for subtly different reasons. In the case of induced demand, the supply curve shifts to the right (eg number of highway lanes increases) so we move to the right on demand curve (or quantity increases, eg more people drive). On the Jevon's paradox case, efficiency increases (eg fuel efficiency increases), so the _demand_ curve shifts (eg more people people drive).
Isn't it the same from the point of view of the consumer's calculus? If suddenly my car can do the same trip for half the fuel and in half the time, whether it's because the road is better or because my car is better, it won't affect how I'm going to use my car from then on.
Induced Demand is a poorly conceived mental model.
The concept of "Induced Demand" is easily explained by the default state of the downward sloping demand curve, and upward sloping supply curve.
In basic economic theory, if you reduce the cost of a good, more people will consume it.
e.g. the classic building a highway example; there was always demand for cheap housing with accessibility to downtowns, but the supply of cheap housing with accessibility to downtowns did not exist prior to building the highway there. The "demand curve" doesn't shift at all, you're just moving along it as perception of cost changes.
Has there ever been a case where a highway was run through the middle of nowhere and traffic didn't increase? It's intuitive why
> The concept of "Induced Demand" is easily explained by the default state of the downward sloping demand curve, and upward sloping supply curve. In basic economic theory, if you reduce the cost of a good, more people will consume it.
It is easily explained by that because that is literally the definition of induced demand (see sentence one of wikipedia). The concept has a name because its important to discuss the externalities and long term implications of that additional consumption.
w.r.t your highway example, how you define the market is extremely important and is sensitive to context. The market for "transportation between suburb X and city Y" experiences a durable change in the demand curve as a result of the construction of all that cheap housing. Both market definitions are valid but if your concern is e.g. urban sprawl then contextually one is a lot more relevant than the other. All that said, you can think of the change in the demand curve of market B not as induced demand itself, but as a consequence of realized latent demand (i.e. induced demand) in market A (cheap housing with accessibility to downtown). Alternative solutions to realizing latent demand in Market A (public transit, denser housing, etc.) have different and potentially preferable externalities which is why considering induced demand and its consequences are important.
"Induced Demand" is often, incorrectly, talked about as a shift in the demand curve in response to increase in supply.
There is no shift. I am addressing this common misunderstanding, not debating the wording in the Wikipedia article.
"Induced Demand" is a misnomer, as the demand was always there at the given price. It is not induced, just realized. If I offer you a gold bar for $0, did I induce your demand to accept it?
Most people will always have demand for goods offered below their perceived intrinsic value.
Ultimately it's semantics around definitions, but the thinking of lay people around this concept is typically more of the shifting demand curve, not realization along the existing curve
>"Induced Demand" is a misnomer, as the demand was always there at the given price.
That's not necessarily true. Suppose you're the government and you produce food for free, and every year people eat everything you make, and everyone is well-fed. You decide you want to prepare for a famine, so this year you start more farms such that next year you make 20% more food. The first two months you're able to save, but when people see that there's more food available, they change their habits and start doing even more exercise than before, and so they eat more until they eat all the food every month again.
I was just coming here to say, this looks like the fancy way to say 'induced demand'.
Why do you think this is a specific case of induced demand (as opposed to induced demand being a specific case of Jevons paradox, or the two being different words for the same thing?)
As an aside, the concept of Induced demand has been criticized. A more precise way to think about “induced demand” is that the demand for a product at a lower price point was preexisting. The transaction didn’t clear until the price actually became lower and this demand manifested itself.
This seems like splitting a hair but it is a more consistent intellectual framework for understanding the dynamics of a particular market.
Induced demand is just the name for realized latent demand, your more precise way to think about it is the concept*. It's only muddled because it is often invoked in complex markets where the costs are often not monetary, there are substantial externalities to consider, or the supply is centrally managed.
Should a city spend $500,000 to install lighting in the park downtown? After all nobody uses it at night because its too dangerous! Induced demand is central to the answer but isn't a first thought for many because the problem space doesn't look like a traditional market.
* there is the additional concept that the long term availability of a good or service at a lower cost changes consumer behavior in a durable way. Building a new train station realizes latent demand but also creates new demand as it causes denser housing to be constructed near it**.
** Whether you want to model that as just more realized latent demand or as a consequence that's technically distinct from induced demand is IMO where a lot of the confusion comes from. If you ask me this should probably have a distinct name as it is a long-term, interrelated process that doesn't map well to sliding around supply and demand curves. Generated demand is what Bloomberg calls it but that has insane overlap with a marketing technique.
Thanks, very interesting. The term ‘induced demand’ has been used a lot in the urban transit & housing discourse. Occasionally you will see the term hijacked by NIMBYs who argue against new housing by claiming construction only serves to create more demand for housing. Or people will argue against widening highways because it will only and always just create ‘induced demand’. A more precise way of thinking about the problem is that a market will pay up to Y amount of time to get from point A to point B. It may be impossible to meet the total demand in this market at time Y with low-efficiency car transit, but a mass transit train system could saturate and address the entire market demand at price point time Y. That is, thinking more precisely about the problem can help inform policy choices.
I don't see how that is more precise. What is the definition of "demand" in that case? Anything that could ever happen at one point, or something that is actually wanted (needed?) by someone who's formulating the thought consciously?
This has actually been on my mind lately and I've been looking for statistics, from Denmark, but haven't been able to find any going back far enough.
My assumptions is that we in the late 1940s through the 1950s had achieved a reasonable standard of living, but perhaps not available to all. Even if we assume that everyone in a country has access to the same, high, late 1950s standard of living, but applied modern technology, like improved insulation, district heating, EVs, energy efficient light bulbs and appliances, then how much would that lower (or raise) our energy usage? More of the energy would surely be moved to electricity, from coal, gas and oil delivered directly to our home, but is it enough chance the environmental impact?
Well I can answer one question for you. There are studies that show that better insulation only briefly lowers the energy use. Once people adjust to the new insulation they end up heating or cooling their houses more since in their mind it is cheaper, but they end up using the same or more energy.
I would not have guessed that that would be the answer. That's kinda crazy, but it also makes sense, every time I visit someone in a newly built home the temperature is always crazy high.
If you could get people back to 1950s level of travel, whether in 1950s vehicles or (way better!) modern EVs, that would make a huge difference to the environmental impact. But of course if you actually suggest measures to force that people (correctly) notice it would reduce their standard of living.
> if you actually suggest measures to force that people (correctly) notice it would reduce their standard of living.
I guess people are different, the less I have to travel the higher my quality of life. Excluding the US, where things are just further apart, I do wonder why people need to move around so much. More and more jobs barely require you to leave your house. I know that the Danish government is proud that Denmark has a highly mobile workforce, meaning that we're willing to transport ourselves relatively far for work, which is a weird thing to be proud of, knowing how much that pollutes.
The rebound effect is common (people have something more efficient and use it more, so the efficiency improvement is somewhat offset by more use), but Jevon's is relatively uncommon (people have something more efficient and the increase in use is large enough to offset the efficiency improvement and then some).
This reminds me of something I remember from college but cannot ever seem to find: "The Jevon's diagram" by which I mean a two country system, with country containing two groups: highly educated and not very educated. The two countries have very different proportions of each group, and the analysis is around free immigration vs no immigration, and shows the harm done to the larger number of less uneducated in one of the countries. Anyway, I'm not going to do a good job of describing this thing, from 20 years ago in class, here, but if anybody knows offhand what I mean and can point me at an article or source on this, I'd be very appreciative! It's bugging me from way back! (where are my notes! Silly tobmlt, not taking notes in latex back then)
Are people getting hung up on NVDA's fall? It feels like most aren't drawing direct logical links - the Jevons paradox doesn't apply directly to stock prices.
I think the situation is fairly clear with some understanding of finance / equity markets. My analysis of the market and competitive situation:
Deepseek's cheaper LLM services + providing open models for other hosts to provide
=> overall prices for using LLM services will fall due to competition (lower prices + more hosts entering the market); AI users won't pay so much for LLM services
=> the major LLM hosts/providers won't be able to project such high revenues or even have the funds to purchase as many GPUs (they will receive less capital investment to buy GPUs since revenues per dollar invested are lower; their stock valuations will fall)
=> corporate demand for Nvidia cards will fall (AND consumer demand will not rise as much to counteract it for multiple reasons)
On the basis of this possible logic, portfolio managers and algorithms project lower growth/revenue for Nvidia and sell off its stock, setting off the usual chain reaction as other managers notice the downward price action and follow suit in order to stop further losses.
Now, if the new hosts entering the market with lower prices attract significantly more users (instead of merely competing away margins) as we might expect by the paradox, then demand for GPUs might grow instead of shrink instead. If more portfolio managers think this is likely, then it's likely for NVDA's price to recover, and it would be wise to get back into the stock once the worst of the fall is over.
No, this isn't directly about Nvidia or stock prices at all -- it's about the emergence of much lower-cost AI models.
So the way Jevon's paradox could apply is that if an AI model only costs 10% of the energy+GPU to train (or run), then we might not get 90% energy+GPU savings... we might actually wind up consuming more energy+GPU on the whole because the models become so much cheaper, we use them more than 10x as much.
Obviously this will affect Nvidia's price, but this principle absolutely does not apply to stock prices directly.
The dose makes the poison, what if a new method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it. What if efficiency brought it down such that inference on CPUs that people already have is good enough. I see a large demand curve but not an infinite one.
For example laptops are getting faster every year but most people I know are not clamoring for the latest laptop, same with phones, good enough is good enough and people would rather save the money. If a laptop came out that was 50x faster for the same price people wouldn’t buy 50x more laptops, they would wait even longer to upgrade.
On the assumption that we’d want super intelligence to do interesting things, consider that we currently have many very smart humans and society is not maximizing their use, quite the opposite.
Not all, and many smart humans are prevented from doing what they want to do for all sorts of reasons. One reason being others feel threatened by them, but those same people will feel threatened by AI, probably more so.
> what if a new method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it
Obviously yes; 6 orders of magnitude would result in insane amounts of money swooping in. The market has currently been going crazy over relatively minor improvements compared to that.
> If a laptop came out that was 50x faster for the same price people wouldn’t buy 50x more laptops, they would wait even longer to upgrade.
That would be equivalent to decades of improvement in the tech. The last time we did that (over the last few decades) we saw massive increases in spending on computers.
Hence my very first point; the dose makes the poison.
Yeah, it would be super impressive, perhaps another point SMT solvers algorithms have improved faster than the hardware for a very long time and are incredibly useful. But we are not spending infinite money on SMT solvers, quite the opposite.
But we are spending more than ever before. I don't see why you think being unable to spend infinity dollars matters here. There are a lot of numbers between larger than current spend between where we are and infinity.
Couple of thousand dollars? Consultant turns up on site, gathers data, talks to a few people, sets up a model and solves it. Probably around a weeks work @ say $100/hr. Might be under-calling it for validation and there'll probably be some iteration to do so a project will take a few solves to complete + overheads.
The software is free, and few people even think about the hardware cost.
But that’s not how it always was, it was a core part of Bill Gates world domination plan during the ‘foundation’ era (2008). Good ones used to be super expensive, then they were sold as excel plugins, and now Z3 is given away for free. And despite being free it’s still not used nearly as often as it should be, it’s not the algorithm or price holding back adoption it’s the people. It was assumed that there was this huge untapped demand curve at the long end of the tail before finding out that there really was not. Z3 was originally released as academic use only (2012) on the assumption that it would drive demand for commercial use of Z3 or another commercial solution. Once it became clear that wasn’t going to make any real money they allowed it to be released for commercial use as well (2015).
As an aside people still do pay for SMT solvers so the market hasn’t completely disappeared but it’s clearly not what it was thought to be.
Your point being? You're stringing together a bunch of observations without any argument relevant to Jevons' paradox.
My guess is that you're mistake is thinking something like [demand = 1/price]. That was never the prediction, the prediction is when efficiency improves, demand increases which is a pretty basic observation for most in-practice supply and demand curves (something like a Veblen good would be different). You're not talking about changes in demand and you're barely looking at efficiency beyond noticing that the price was positive and is now 0.
> And despite being free it’s still not used nearly as often as it should be ...
Case in point, if this wasn't obvious then you need to modify your expectations. Jevons' paradox predicts exactly this, because there are prices below free (ie, paying someone to use the thing) and we'd expect they eventuated people would find new uses for SMT solvers due to Jevons' paradox in context of a price signal that they were getting more efficient. That implies even at $0 in software costs we aren't going to be using them everywhere we could.
It’s a question of depreciation of the GPU vs time to deploy AI.
If RL and synthetic data creation are all we need for self-improving AI, the large labs have an internal use case to utilize all available compute, regardless of efficiency gains.
If they have to go external and deploy in the broader market, time to deploy would make the GPUs worthless before they are fully utilized.
The models will all get more efficient, but NVIDIA as the high cost provider of compute (as well as cloud services that use them) - 10x markup on H100s - may not benefit from increased demand. If the models turn out to be a commodity, then low cost providers will be the winners, likely running on chips like TPU and Trainium (Amazon/Anthropic) that don't have the 10x markup.
Low cost providers will not be the winner because those will become commodity. It’s the leading model, the one closest to AGI that will make all the profit.
Nvidia has products to serve lower end as well. There’s no reason why Nvidia can’t dominate in the commodity sector.
And if Nvidia can’t keep their high margins, they will make it up with significantly more volume (Jevons Paradox in play).
I think there'll continue to be models at different levels of capability and price - and not at all obvious whether it'll be the Ford or Rolls Royce that'll make most money.
My guess is that is that what passes as "AGI" will also become commoditized, maybe even free to consumers (funded by advertising, business referrals to restaurants, etc).
Been thinking about EVs. They are cheaper per kilometre driven. So would this not mean that the owners of them spend more overall than ICE owners spend. And would drive a lot more. As Jevons Paradox means the demand for driving will be higher thus overall money spend on driving will go up.
An example you may have learned about in school was Eli Whitney and the cotton gin. He thought his invention would reduce slavery by making it very cheap to clean the seeds out of picked cotton. Instead, this efficiency led to an explosion of demand for this lower-cost cotton.
A bubble around a technology can pop, and yet the tech can be ultimately valuable in the long term. That does not mean it wasn't a bubble nor it didn't pop.
Given the likes of sama & co were asking for unfathomable levels of capital, chipmakers had backlogs, datacenter & power supply chain companies were getting bid to the moon, etc.. having a dramatic decrease in compute requirements might just bring things closer to a supply/demand balance.
Where was Stargate gonna get $500B?
How was Altman ever going to raise.. $7T? (~1/3 of the entire GDP of US)
Does a 10x or even 40x decrease in compute cost just mean they can actually raise & spend the capital that their compute requires?
The Jevons Paradox, an economic principle stating that increases in efficiency often lead to higher overall consumption, has been observed in the context of DeepSeek AI. DeepSeek, a Chinese AI startup, recently introduced its R1 model, which achieves comparable performance to leading AI systems like OpenAI's ChatGPT while requiring significantly less computational power and energy. This efficiency has made the technology more accessible and affordable, spurring wider adoption across industries.[30] As Microsoft CEO Satya Nadella noted, the increased accessibility and efficiency of AI tools like DeepSeek's R1 model could lead to a surge in applications and use cases. This expanded usage may result in higher energy consumption overall, despite individual efficiencies. For example, while DeepSeek's model uses fewer resources per task, the growing demand for AI-driven solutions—ranging from chatbots to complex reasoning models—could increase the total energy requirements of the global AI infrastructure.[31]
I made a comment yesterday that the whole NVDA thing seemed eerily similar to the AMZN thing in the first dot com crash. [0]. If the market can get sufficiently irrational, and we start to see that same like 80-90% markdown that we got with AMZN in the first crash, I mean, who knows? That's a golden opportunity. Not many generations get opportunities like that twice.
I guess we'll see what happens with the stock price, but long term I'm pretty confident in this Jevon's Paradox effect on AI use. I'm praying for more negativity hype out among those who maybe don't know as much about tech as the HN crowd. Because those people collectively, can move stock prices.
Having read the paper from deepseek I seriously don't get how anyone can think "Oh now is the time to sell nvidia".
Both at the tactical level and strategic level that paper means we will see _more_ nvidia gpus fly off the shelves. If anything it's showing why the h800 is a decent card and everyone in China should buy one (dozen thousands).
majority of Nvidia's profits are from huge companies particularly Microsoft and META not individuals buying graphics cards.
Look at their current video gaming graphics sales and even if these ai consumer purchases were 10x what there previous graphics sales were then the current valuation wouldn't make sense.
This is all ignoring the fact that if the h800 was good enough amd's top line cards should be too.
>This is all ignoring the fact that if the h800 was good enough amd's top line cards should be too.
You can't pay me to use AMD cards. I have work to do. I don't have time to write drivers for cards that no one supports.
The point of the deepseek r1 paper is that the more synthetic data you generate by over trained models the better the resulting model. That means more GPUs. There is no reason why you can't train a dozen models like their nameless logic model in other domains and then use them for synthetic data augmentation.
I have to agree. I'm not sure the architectural and engineering changes here don't put NVDA further ahead. Some stuff all gpu makers can do. Like quantize on transfer. But NVDA is so far ahead in other areas and DeepSeek leans heavily on that industry leading tech. At least at the low level.
But I'm betting the average investor doesn't have the level of facility with ML to properly analyze the information that's come out. Heck, I'm betting most of the people working in AI are really just TF or PyTorch monkeys. The real AI experts have probably been hovered up by MS or FB or AMZN or OpenAI or what have you. So it feels like we could get lucky here and the AI related market could drop like a rock taking the good with the bad. Which I think would be awesome for most people in a position to wait for it to happen.
What gets me about the deepseek work is that it's neither novel or theoretically difficult.
I've been using a similar synthetic data pipeline to generate examples for sota bert embedding models.
That low-hanging fruits like this are left unexplored by meta or alphabet - and if we're to be believed lead to crisis meetings for a week straight - makes one wonder wtf the engineers there are doing?
This is just a baseless conspiracy theory that I have, but I do wonder if they intentionally avoided certain avenues of research because it could erode the “moat” major tech firms have by dramatically reducing the capital costs required for training and inference. If you focus your research around work that mandates high-end hardware at scale you can lock out tons of potential incumbents
There is nothing in the deepseek paper that suggests you can't use the order of magnitude in hardware costs you saved to just train models that are ten times as large.
But there are thresholds of commercial viability in all of this. DeepSeek's technology gets you over that line with less sophisticated hardware.
There's already some pretty impressive work being done with folks using just a pair of M2 Ultras with r1 in a "home lab" context that goes way beyond what you could previously do with llama.
Yes, but the r1 model is shiny and new. I've been using it locally for a day and I'm already finding odd spots. One thing that I've seen it do a number of times is get the correct answer in the thinking tokens, then ignore it, generate thousands more tokens until it reaches it context length then give a wrong answer.
There are things, like doc strings for functions, that llama models give better results for.
It's not comparable. NVDA is heading toward ~$95 billion in operating income for the next four quarters approximately. AMZN in the first dotcom bubble crash had only (hefty) sales growth and was burning enormous amounts of cash to build itself out. There was no profit - and no near-term hope of a profit - to support Amazon's valuation back then. They were highly susceptible to an exaggerated crash accordingly.
At $95b in op income, you're looking at ~33 times op income for NVDA for 2025, or lower. I don't consider that a bargain but it's also not crazy expensive given the stock market valuation context + their growth + their market dominance in a critical field.
You can't apply Jevons paradox to everything, that would be predicting the future. However, you can try to analyze whether it holds in your specific context.
Comparing AMZN and NVDA is like comparing apples and oranges. Amazon has significantly diversified from its origins as an online bookseller and started when the internet, and later mobile devices, were not yet mainstream. NVDA, on the other hand, is an entirely different kind of company.
It doesn't matter whether it's AMZN or NVDA, you should try to distinguish between the company's fundamental value and the speculative component, as both are reflected in the price.
With AGI/ASI, there will come a point where this paradox breaks down: once the ASI is self-sufficient and "agentic enough", people's direct use for it will actually reduce. You won't need to "micromanage" it - and the total amount of compute it requires might actually drop drastically if the ASI improves its own efficiency much better than humans can. At that point the question becomes one of alignment: will overall consumption drop because ASI subjugates or eliminates humanity or does it grow exponentially as we use it to explore the stars.
This "paradox" is being cited everywhere, and while it might apply to AI, people are wrongly using it to justify why nvidia should have a $3.6T or greater valuation.
What happens to a holistic system doesn't necessarily apply to every player in that system.
nvidia's heights are largely because of a few super capitalized players dumping enormous amounts of money into data centre GPUs, letting nvidia enjoy outrageous profit margins selling rapidly obsoleted, rapidly iterating products to mega-money players who are all counting on this giving them such a differentiation that it's financially worthwhile. If AI is just some free thing that has little differentiation between the players -- commoditized, so to speak -- the market for $200,000 GPUs probably isn't going to be as big. People won't be rushing to build $500B buildouts.
Another argument I keep seeing is smaller models and edge compute. NVDA has almost no stake in that, and there is no reason to believe they're going to have any particular advantage if they tried harder there, any more than Intel could leverage PCs to Smartphones. The core operations of running a model are well known by all of the silicon makers, and they can all make massively scalable FMA matrix operators, etc.
This "paradox" is being cited everywhere, and while it might apply to AI, people are wrongly using it to justify why nvidia should have a $3.6T or greater valuation.
People are arguing that DeepSeek is a bull case for Nvidia, not bear.
Nvidia being worth $3.6T is a different topic. People think DeepSeek should have reinforced that valuation or even increased it, not lower it.
The simplest explanation, Jevons paradox is cope from CEOs and retail investors trying to justify the share market.
People seem to have forgotten that there’s a time element involved, not a perfectly elastic and instant demand.
Thinking about it another way, if it turned out it was going to cost 30x more compute to train the next generation of models than we thought due to some scaling law, would the chip share prices be expected to rise, stay the same or go down? Probably rise.
In that case, if it was 30x cheaper to do what we need today would we expect option a, b or c?
Assuming that there would be less short / mid term sales, that impacts their bottom line and flattens off their (ie nvda) demand curve. Slower growth, lower share price. Slower growth, longer opportunity for competition. Therefore the speculative element of the chip leader’s share price’s should come down.
I don’t have any long or short positions in the chip sector.
Companies compete against each other - not themselves. It’s not like OpenAI set an internal goal of training 30x bigger and that’s it. No, OpenAI needs to compete.
If training efficiency is 30x more, then Anthropic, Google, OpenAI, Mistral, Qwen, etc all get the benefit. It resets the baseline. They will not stop competing against each other just because there is a 30x efficiency improvement that everyone has access to.
the money must be happened, otherwise shit will happen
it's like we are spending 500bn for x ai needs, but suddenly, the cost fell to 1/20, now we must convince everybody there'll be x20 needs, so that the 500bn will not go missing
that is what Satya Nadella and Sam Altman are trying to achieve
the real problem:
1. will there be 20x needs for ai?
2. how much ai needs are priced now? x? 5x? or, 20x?
Lots of NVDA holders and NVIDIA itself, trying to justify their valuations, but forgetting the Theory of Constraints will set limits on your Jevons Paradox.
The only thing relevant, is the large sale that Nancy Pelosi did before the crash. Another incredible coincidence again of course, by one of he most successful hedge managers in history...
Come on. Which do you think is more likely, that it was a normal EOY sale of a lucrative stock, or that they sold based on insider info obtained weeks before anyone else?
i think it's insider trading, but not about deepseek, deepseek is never the main cause
deepseek v2 showed effeciency months ago, and nothing happened, there's also nothing happened after v3 and r1 release
the insider info is about when will Wall Street stop pushing and squeezing the market and start harvesting, and, IMO, this will happen anytime after rotation of ruling parties
i don't know if the similar scene happened before in the US, but this happened elsewhere, the rulling party with connected market makers and puppet central bank, push the market to highest and let it freely fall after rotation
nancy definetly know what the timing is
Just some nonsense, but if I’m right, it won’t take too long to witness it happen.
>Energy-efficient lighting, such as LED bulbs, has become a popular solution to reduce energy consumption. However, it serves as a notable example of the Jevons Paradox. While LEDs significantly lower the cost of lighting per unit, this reduction often leads to increased usage. For instance, people may install more lights in their homes or businesses, use them for longer durations, or adopt decorative and outdoor lighting more extensively. As a result, the total energy consumption for lighting has not decreased significantly despite the efficiency improvements.
This felt false to me, LED bulbs are generally 10x more efficient, so the average number of lights would have had to increase 10 fold, and this is ignoring the savings from not having AC cool those other 50 watts a bulb that went to heat.
So I took a look at the source, which doesn't actually have a source and just links to another white paper that claims light pollution increased, but is unclear about how much and due to what. Perhaps that whitepaper links to another whitepaper, but this is where I stopped.
Kind of makes me question how often I blindly trust Wikipedia, though this point in particular is pretty minor.
This is exactly the case. Anecdotally speaking, growing up in the 90s, I remember my parents adding lightbulbs and fluorescent tube lights to their shopping list every couple of months. We used to have a small stockpile of these in the pantry.
Fast forward to today, the last time I bought lightbulbs for my house was over 5 years ago and that was to upgrade from regular led lamps to Philips hue.
Not the case I’m observing.
My sister just did a home remodeling and she added must have been 50 LED recessed lights in the ceiling. There are automated LED lights in the outdoor pathways, lights hanging on the balcony rails, lights under kitchen cabinets, lights inside closets, lights inside pantries, solar powered LEDs in backyard, etc.
Her appliances also have more indicator lights than ever before.
I’m not even counting the countless LED screens such as phones, iPads, TVs.
In total, she may every well have 10x more lights than before the LED boom.
Time Square in NYC was famous for having many lights and screens. Walk around any large Asian city in 2025 and they all look like time square with countless lights and screens.
Thats said, there is a limit to lighting demand because we still need darkness to sleep. I think the upper bound for intelligence is far higher than LED lights.
Still it does not use as much electricity as the previous generation of light bulbs and she won’t be replacing them often. So from a perspective of the industry it’s not so good. The tech is better but there’s less sales / margins/ profits altogether.
Is this true? What were the revenues of lighting companies before compared to now? Do we count TV screen makers because they’re LEDs as well?
I just watched Veritasium’s video on how LEDs were indented and the companies making LEDs saw exponential increase in revenue. https://youtu.be/AF8d72mA41M
Basic Google search tells me the lighting industry has experienced high growth due to the invention of LED lighting.
Even if she has 10 times as many (well possible if compared, say, to the 1960s), they aren't likely switched on most of the time.
People generally don't like if it is "too bright" in the room, and the extra lights in closets, pantries etc. will only be used marginally, for a few minutes a day.
It helps that LED lifetime isn't degraded by frequent on-and-off cycling.
So what are we measuring here with Jevons Paradox?
Light bulb sales? Total lumen output? Total lighting revenue?
Better to measure electricity usage in either $$ or kwh.
Like this? https://www.researchgate.net/publication/342626956/figure/fi...
Maybe installed capacity in lumens.
I can’t find the numbers you’re looking for.
I did find that the total consumption of light increased 100,000x between 1700s and 2000 in Uk.
https://onclimatechangepolicydotorg.wordpress.com/low-carbon...Also keep in mind that better lighting efficiency enables people in poor countries to have lights where they couldn’t before. It’s not just wealthy counties. You have to count them as well. It's not just counting a rich person's house before and after LEDs.
Absolutely, yes. Many more people can now afford light.
My grandma (born 1926, died 2016) was born in Eastern Slovakia in the pre-electric era, the region she lived in only got connected to the grid in 1945.
In her youth, the whole family would simply have one kerosene lamp and gather around it after dark.
The discussion was about energy efficiency and energy consumption, not light bulb durability and replacement rate.
With that said, it would not surprise me if there is a 10x increase in light sources (not total lumens) since LEDs.
Yes, I don't have 10x more lightbulbs but now household appliance has an LED. I just installed new GFCI outlets and all of them have LEDs too!
The comparison becomes meaningless when we count the milliwatt, sub-Lumen LED in an outlet wirh incandescents. "I had 8 incandescent lightbulbs, but now have 1000s of LEDs- including outlets, TV backlights, DVR status" isn't comparing like with like.
Yeah probably a meaningless comparison for Jevons Paradox (and its application to NVDA) since that is about resource use and LEDs are not a resource
Jevons Paradox describes the situation where a 10x efficiency increase results in a >10x deployment of lumens, such that the amount of energy consumed for lightning increases relative to when lighting was inefficient.
Which has happened historically.
https://onclimatechangepolicydotorg.wordpress.com/low-carbon...That's a good source, although that historical analysis extends to the year 2000 when even CFLs only had reached ~30% market share, and well before LED lighting hit mass adoption. It is only speculating about LEDs. There's no question that lighting demand has continued to grow but the efficiency increase has also been remarkable, so unless you know of another source that does the math, it's unclear to me to what extent it's ended up balancing out in this case.
Another confounder to Jevons' paradox is that the price of the input (electricity) can also change, independent of changes in efficiency. If electricity prices were to go up ten-fold during the switch to LED lighting, then we wouldn't expect much change in illumination (probably a net decrease, as people would be hurting so much from other electricity expenditures that they'd be hunting for any savings on their energy bill). That's of course a fictional scenario, but in the case of GPUs, well, they have been getting quite expensive.
I somewhat doubt there is now 10x light fixtures... Thus 10x the traditional bulbs. That seems rather large undertaking in most buildings. Sometimes I think it makes sense to stop and think what would 10x more of some items mean in practise.
LEDs are around 3-4x more efficient than fluorescent lamps, (1) 28W 2x4 LED fixture puts out around the same amount of light as (1) 4-lamp 2x4 fixture with (4) 28W fluorescent lamps, depending on the lumen output package on the LED fixture.
It's common to use incandescent bulbs as a baseline.
You're right. But it's wikipedia so I deleted the LED section and replaced it with a bit about cars.
Your math is wrong:
> This felt false to me, LED bulbs are generally 10x more efficient, so the average number of lights would have had to increase 10 fold, and this is ignoring the savings from not having AC cool those other 50 watts a bulb that went to heat.
The first 5W also went into heat. Perhaps with the exception of light escaping through a window, *every* watt that goes into a lightbulb comes out as heat for your AC to cool.
I would not be surprised if there were a 10x increase in use of lighting with LEDs.
Not taking Wikipedia (or any source) on blind trust is always wise. More often than not, when you kick tires, you come back to folklore. It's fine for big things, but details are often wrong. Ditto for newspapers, for that matter.
> I would not be surprised if there were a 10x increase in use of lighting with LEDs.
Has the average homeowner added more ceiling lights or lamps because of different bulbs?
Has the average office added more lighting because of LEDs? Factories?
In order for an overall 10x increase in usage, there needs to have been some MASSIVE new usage or expansion of lighting, somewhere, to make up for all of these enormous use cases (residential, commercial) where there was probably close to zero increase.
So, do you have any pet theories ?!!
> Has the average homeowner added more ceiling lights or lamps because of different bulbs?
As much as I appreciate the snark and sarcasm, at least where I live, the equivalent is happening.
Old home and commercial construction would typically have one central light per room. Homes would have a single light fixture with 1-3 lightbulbs at the center of each ceiling and were pretty dimly lit. Commercial would be similar with fluorescents. You would turn it off when you left a room.
New home and commercial construction has accent lighting all over the place, some turned on 24/7. Cabinets have under-cabinet-lights. Closets have closet lights. We have rows of LEDs lining all sorts of accents.
If you compare 80's-style construction to 2020's-style, the amount of new light seems to be about that much of an increase.
I can accept that new construction may have more lighting, but nowhere near 10x, and certainly not 10x overall when you factor "legacy locations" where it is mostly just a bulb replacement.
You can accept what you like.
I am describing where I live.
If you don't accept that a place like this might exist, without so much as a follow-up question, says more about you than anything else.
Are you really claiming that a new home has 10x more lighting than an old one? Or are you just saying that your own house it lit up like a lighthouse - search lights in every room, perhaps?
Not to mention that in poor countries, the invention of LEDs could mean going from being able to afford electricity for 1 light in the house to 10 lights.
If a 5w bulb and a 60w bulb output the same amount of light, those 55w are being wasted elsewhere (heat).
I said 50 because I was afraid someone would go 'ACTUALLY LED's only save 50-55w', I didn't expect someone to argue that the 5w led bulb is also outputting heat. You are of course right, 2.5w of that 5w is likely heat being cooled by the AC.
>I would not be surprised if there were a 10x increase in use of lighting with LEDs.
Quite frankly I don't care either way, I just wanted a source. I'm happy you feel that way though.
The intuition behind this is that people who had a usecase for the tech but were locked out because they couldn't afford it, or didn't believe it was worth it usecase can be engineers or researchers. Some of these people may perform better than expected causing a gap in true market value which is filled when they scale up to their new size
Related: Jenson's Paradox: The more you buy, The more you save
I guess my question about the current events is: Are people really locked out of LLMs due to price? It seems like everything already has AI in it and that virtually every end user hates it. I could see a shift to more local models or something, but not an increase. I feel like LLMs have largely been oversold as a solution in search of a problem.
Or are people just applying this as everyone and their mothers are going to start training competing models to carve out their piece of the market?
These are the four main problems with LLMs (and related technologies) as I see them:
Bringing the cost way down and making them trainable on consumer hardware solves or at least greatly alleviates problems 1-3. That just leaves problem 4, which might still be unsolvable and sink the whole endeavor, but at least can be focused on.In many cases the solution to 4 is also to make them faster, because many tasks are checkable.
Absolutely! Even for inference! The SOTA models for all commercial purposes need to run on a consumer’s device.
Running either Grok2 or DeepSeek or even Llama405b requires nearly 400-500gb of memory.
Buying a tinybox with enough gpu memory costs $15k-25k. Or equivalently the same if you build your own.
A distributed Mac cluster costs about the same, if not more, if you’re buying 2-3 M2 Ultra each with 192gb of memory.
So people are absolutely constrained by price/supply here. Every engineer, analyst, scientist would be far more untethered by rules & regulations or policies & terms-of-service nitty gritties if they can trust that LLM they use is completely local, without-telemetry or tracking and is licensed fairly for commercial use (perhaps this excludes llama).
Not a lot of people can afford $15k-30k in spending for a computer (that can run this sota llms). But you can a billion will buy one when it’s $1k
Keep in mind too that to run DeepSeek R1 you'd need 768 GB, so essentially 4 tinyboxes.
The full one, sure, but there are quantizations runnable on far smaller machines.
Quantized models don't perform as well. The question was about running one of the cloud models.
This was true before DeepSeek.
Not to mention, the north star is to get to a place where we have the hardware to do training at home. we're a long ways off, but without the restrictions of needing the hardware to do it, ideally, we'd make the model such that is continually being trained.
The question is:
Are people (or companies) locked out of training LLMs due to price?
I don't know the immediate answer. I really expect it to be a resounding "yes" on the long term because different LLMs should be good for different things. But in any way, this is not about the people adding a cloud client to their software.
> Are people really locked out of LLMs due to price?
Yes. For example Google has just made Gemini a standard part of Workspace, before that it cost $36 or something per month per user. That was too much for many SMBs to experiment with (you and I understand that the potential efficiency gains are way higher but for an SMB, paying 4x more for your office suite sounds bad).
Well there's two sides here.
Consumers of AI do have sensitivity to pricing. Many OpenAI customers "ration" their usage. I imagine lower costs open up new demand for these people.
On the Service side: it seems to have reduced the cost of operating a "commercially viable" (something people will actually pay money for) LLM. But even beyond that, "self hosted" models are also far more affordable now, which means models that target specific niches can be viable to build or buy.
However, this won't be an overnight phenomena. In the short-term, it will seem like demand drops, but in the long-term demand will go up. Big caveat: that demand may not be concentrated on the current incumbent players.
And finally, the elephant in the room: AI still needs to become more useful to reach its full potential. Easily a decade+ more needed here.
> Are people really locked out of LLMs due to price?
I think so. When you start using a good model, you soon learn that it can make a difference. If you have domain knowledge, you can treat LLMs like a design partner. With DeepSeek, what would have cost 50 dollars to use Sonnet is now 5 dollars. Having to only spend 5 dollars a month can be a real game changer for many.
This is IMO a spefic case of the induced demand concept (https://en.wikipedia.org/wiki/Induced_demand).
Reddit explains the difference thus:
"Induced demand": The highway is expanded such that it only takes me 15 minutes to get downtown instead of 45, so I go downtown more.
Jevson paradox: The highway is expanded such that it only takes 15 minutes to get from the suburbs to my office instead of 45, so I move to the suburbs.
So one is movement along the demand curve, while the other is a movement of the demand curve.
https://www.reddit.com/r/badeconomics/comments/oo2s48/that_o...
I fail to see a difference. In both cases demand increased, no?
Here's [0] a good overview as to the differences.
[0] https://www.toppr.com/guides/business-economics/theory-of-de...
Yes, but for subtly different reasons. In the case of induced demand, the supply curve shifts to the right (eg number of highway lanes increases) so we move to the right on demand curve (or quantity increases, eg more people drive). On the Jevon's paradox case, efficiency increases (eg fuel efficiency increases), so the _demand_ curve shifts (eg more people people drive).
Isn't it the same from the point of view of the consumer's calculus? If suddenly my car can do the same trip for half the fuel and in half the time, whether it's because the road is better or because my car is better, it won't affect how I'm going to use my car from then on.
Ahh so the difference is in a definition of supply.
You say drilling more oil is something different than building more efficient cars.
I would say in both cases supply of oil increased so there is no real difference.
.... I mean in the case of more efficient cars supply of oil explicitly did not change.
There are meaningful differences in terms of pricing strategies, anticipating demand, etc.
Induced Demand is a poorly conceived mental model.
The concept of "Induced Demand" is easily explained by the default state of the downward sloping demand curve, and upward sloping supply curve.
In basic economic theory, if you reduce the cost of a good, more people will consume it.
e.g. the classic building a highway example; there was always demand for cheap housing with accessibility to downtowns, but the supply of cheap housing with accessibility to downtowns did not exist prior to building the highway there. The "demand curve" doesn't shift at all, you're just moving along it as perception of cost changes.
Has there ever been a case where a highway was run through the middle of nowhere and traffic didn't increase? It's intuitive why
> The concept of "Induced Demand" is easily explained by the default state of the downward sloping demand curve, and upward sloping supply curve. In basic economic theory, if you reduce the cost of a good, more people will consume it.
It is easily explained by that because that is literally the definition of induced demand (see sentence one of wikipedia). The concept has a name because its important to discuss the externalities and long term implications of that additional consumption.
w.r.t your highway example, how you define the market is extremely important and is sensitive to context. The market for "transportation between suburb X and city Y" experiences a durable change in the demand curve as a result of the construction of all that cheap housing. Both market definitions are valid but if your concern is e.g. urban sprawl then contextually one is a lot more relevant than the other. All that said, you can think of the change in the demand curve of market B not as induced demand itself, but as a consequence of realized latent demand (i.e. induced demand) in market A (cheap housing with accessibility to downtown). Alternative solutions to realizing latent demand in Market A (public transit, denser housing, etc.) have different and potentially preferable externalities which is why considering induced demand and its consequences are important.
"Induced Demand" is often, incorrectly, talked about as a shift in the demand curve in response to increase in supply.
There is no shift. I am addressing this common misunderstanding, not debating the wording in the Wikipedia article.
"Induced Demand" is a misnomer, as the demand was always there at the given price. It is not induced, just realized. If I offer you a gold bar for $0, did I induce your demand to accept it?
Most people will always have demand for goods offered below their perceived intrinsic value.
Ultimately it's semantics around definitions, but the thinking of lay people around this concept is typically more of the shifting demand curve, not realization along the existing curve
>"Induced Demand" is a misnomer, as the demand was always there at the given price.
That's not necessarily true. Suppose you're the government and you produce food for free, and every year people eat everything you make, and everyone is well-fed. You decide you want to prepare for a famine, so this year you start more farms such that next year you make 20% more food. The first two months you're able to save, but when people see that there's more food available, they change their habits and start doing even more exercise than before, and so they eat more until they eat all the food every month again.
I was just coming here to say, this looks like the fancy way to say 'induced demand'.
Why do you think this is a specific case of induced demand (as opposed to induced demand being a specific case of Jevons paradox, or the two being different words for the same thing?)
As an aside, the concept of Induced demand has been criticized. A more precise way to think about “induced demand” is that the demand for a product at a lower price point was preexisting. The transaction didn’t clear until the price actually became lower and this demand manifested itself.
This seems like splitting a hair but it is a more consistent intellectual framework for understanding the dynamics of a particular market.
Induced demand is just the name for realized latent demand, your more precise way to think about it is the concept*. It's only muddled because it is often invoked in complex markets where the costs are often not monetary, there are substantial externalities to consider, or the supply is centrally managed.
Should a city spend $500,000 to install lighting in the park downtown? After all nobody uses it at night because its too dangerous! Induced demand is central to the answer but isn't a first thought for many because the problem space doesn't look like a traditional market.
* there is the additional concept that the long term availability of a good or service at a lower cost changes consumer behavior in a durable way. Building a new train station realizes latent demand but also creates new demand as it causes denser housing to be constructed near it**.
** Whether you want to model that as just more realized latent demand or as a consequence that's technically distinct from induced demand is IMO where a lot of the confusion comes from. If you ask me this should probably have a distinct name as it is a long-term, interrelated process that doesn't map well to sliding around supply and demand curves. Generated demand is what Bloomberg calls it but that has insane overlap with a marketing technique.
Thanks, very interesting. The term ‘induced demand’ has been used a lot in the urban transit & housing discourse. Occasionally you will see the term hijacked by NIMBYs who argue against new housing by claiming construction only serves to create more demand for housing. Or people will argue against widening highways because it will only and always just create ‘induced demand’. A more precise way of thinking about the problem is that a market will pay up to Y amount of time to get from point A to point B. It may be impossible to meet the total demand in this market at time Y with low-efficiency car transit, but a mass transit train system could saturate and address the entire market demand at price point time Y. That is, thinking more precisely about the problem can help inform policy choices.
I don't see how that is more precise. What is the definition of "demand" in that case? Anything that could ever happen at one point, or something that is actually wanted (needed?) by someone who's formulating the thought consciously?
This has actually been on my mind lately and I've been looking for statistics, from Denmark, but haven't been able to find any going back far enough.
My assumptions is that we in the late 1940s through the 1950s had achieved a reasonable standard of living, but perhaps not available to all. Even if we assume that everyone in a country has access to the same, high, late 1950s standard of living, but applied modern technology, like improved insulation, district heating, EVs, energy efficient light bulbs and appliances, then how much would that lower (or raise) our energy usage? More of the energy would surely be moved to electricity, from coal, gas and oil delivered directly to our home, but is it enough chance the environmental impact?
Well I can answer one question for you. There are studies that show that better insulation only briefly lowers the energy use. Once people adjust to the new insulation they end up heating or cooling their houses more since in their mind it is cheaper, but they end up using the same or more energy.
https://www.cam.ac.uk/research/news/insulation-only-provides...
presumably there are limits to that. once every room is the preferred temperature, i'm going to stop running the heater
although maybe i'd get bored and run the heater and the AC at the same time
Often you can't because the heater and the AC is literally the same unit, the heat pump.
I would not have guessed that that would be the answer. That's kinda crazy, but it also makes sense, every time I visit someone in a newly built home the temperature is always crazy high.
> same, high, late 1950s standard of living
.. I suspect this is wildly less good than most people would expect, along with forgetting about how widely distributed it was or wasn't.
Just look at the first chart here: https://www.bbc.co.uk/news/uk-42182497
If you could get people back to 1950s level of travel, whether in 1950s vehicles or (way better!) modern EVs, that would make a huge difference to the environmental impact. But of course if you actually suggest measures to force that people (correctly) notice it would reduce their standard of living.
> if you actually suggest measures to force that people (correctly) notice it would reduce their standard of living.
I guess people are different, the less I have to travel the higher my quality of life. Excluding the US, where things are just further apart, I do wonder why people need to move around so much. More and more jobs barely require you to leave your house. I know that the Danish government is proud that Denmark has a highly mobile workforce, meaning that we're willing to transport ourselves relatively far for work, which is a weird thing to be proud of, knowing how much that pollutes.
FYI, people mix up Jevon's and the rebound effect.
https://en.m.wikipedia.org/wiki/Rebound_effect_(conservation...
The rebound effect is common (people have something more efficient and use it more, so the efficiency improvement is somewhat offset by more use), but Jevon's is relatively uncommon (people have something more efficient and the increase in use is large enough to offset the efficiency improvement and then some).
This reminds me of something I remember from college but cannot ever seem to find: "The Jevon's diagram" by which I mean a two country system, with country containing two groups: highly educated and not very educated. The two countries have very different proportions of each group, and the analysis is around free immigration vs no immigration, and shows the harm done to the larger number of less uneducated in one of the countries. Anyway, I'm not going to do a good job of describing this thing, from 20 years ago in class, here, but if anybody knows offhand what I mean and can point me at an article or source on this, I'd be very appreciative! It's bugging me from way back! (where are my notes! Silly tobmlt, not taking notes in latex back then)
Jevons Paradox Paradox: Markets fall on news that should make them rise, due to Jevons Paradox.
Somehow, nobody expects the Jevons' Paradox. Even when those same people talk about it every other day.
The Hofstadter-Jevon Corollary: Taking Jevon’s Paradox into account only makes it worse.
Are people getting hung up on NVDA's fall? It feels like most aren't drawing direct logical links - the Jevons paradox doesn't apply directly to stock prices.
I think the situation is fairly clear with some understanding of finance / equity markets. My analysis of the market and competitive situation:
Deepseek's cheaper LLM services + providing open models for other hosts to provide
=> overall prices for using LLM services will fall due to competition (lower prices + more hosts entering the market); AI users won't pay so much for LLM services
=> the major LLM hosts/providers won't be able to project such high revenues or even have the funds to purchase as many GPUs (they will receive less capital investment to buy GPUs since revenues per dollar invested are lower; their stock valuations will fall)
=> corporate demand for Nvidia cards will fall (AND consumer demand will not rise as much to counteract it for multiple reasons)
On the basis of this possible logic, portfolio managers and algorithms project lower growth/revenue for Nvidia and sell off its stock, setting off the usual chain reaction as other managers notice the downward price action and follow suit in order to stop further losses.
Now, if the new hosts entering the market with lower prices attract significantly more users (instead of merely competing away margins) as we might expect by the paradox, then demand for GPUs might grow instead of shrink instead. If more portfolio managers think this is likely, then it's likely for NVDA's price to recover, and it would be wise to get back into the stock once the worst of the fall is over.
No, this isn't directly about Nvidia or stock prices at all -- it's about the emergence of much lower-cost AI models.
So the way Jevon's paradox could apply is that if an AI model only costs 10% of the energy+GPU to train (or run), then we might not get 90% energy+GPU savings... we might actually wind up consuming more energy+GPU on the whole because the models become so much cheaper, we use them more than 10x as much.
Obviously this will affect Nvidia's price, but this principle absolutely does not apply to stock prices directly.
The way the market reacted is completely illogical.
If Nvidia announced a new chip that was 10x more effective at the same price, the market would be like WOW they're going to sell so many of those!
But a software advance that achieves the same effect somehow causes the opposite reaction.
Exactly. The only thing that matters is AI adoption rate and training scaling laws right now.
The best Jevons Paradox/induced demand explainer ever:
https://www.youtube.com/watch?v=pCzCJzwrB_c
The dose makes the poison, what if a new method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it. What if efficiency brought it down such that inference on CPUs that people already have is good enough. I see a large demand curve but not an infinite one.
For example laptops are getting faster every year but most people I know are not clamoring for the latest laptop, same with phones, good enough is good enough and people would rather save the money. If a laptop came out that was 50x faster for the same price people wouldn’t buy 50x more laptops, they would wait even longer to upgrade.
Do you have a use case for your opinion?
On the assumption that we’d want super intelligence to do interesting things, consider that we currently have many very smart humans and society is not maximizing their use, quite the opposite.
Very smart humans don't want to do random tasks, they want to study number theory.
Not all, and many smart humans are prevented from doing what they want to do for all sorts of reasons. One reason being others feel threatened by them, but those same people will feel threatened by AI, probably more so.
Oh I see, the underpants gnomes path to profitability. Makes sense.
> what if a new method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it
Obviously yes; 6 orders of magnitude would result in insane amounts of money swooping in. The market has currently been going crazy over relatively minor improvements compared to that.
> If a laptop came out that was 50x faster for the same price people wouldn’t buy 50x more laptops, they would wait even longer to upgrade.
That would be equivalent to decades of improvement in the tech. The last time we did that (over the last few decades) we saw massive increases in spending on computers.
Hence my very first point; the dose makes the poison.
Yeah, it would be super impressive, perhaps another point SMT solvers algorithms have improved faster than the hardware for a very long time and are incredibly useful. But we are not spending infinite money on SMT solvers, quite the opposite.
But we are spending more than ever before. I don't see why you think being unable to spend infinity dollars matters here. There are a lot of numbers between larger than current spend between where we are and infinity.
To my point, how much do you think a SMT solution costs now?
Couple of thousand dollars? Consultant turns up on site, gathers data, talks to a few people, sets up a model and solves it. Probably around a weeks work @ say $100/hr. Might be under-calling it for validation and there'll probably be some iteration to do so a project will take a few solves to complete + overheads.
The software is free, and few people even think about the hardware cost.
But that’s not how it always was, it was a core part of Bill Gates world domination plan during the ‘foundation’ era (2008). Good ones used to be super expensive, then they were sold as excel plugins, and now Z3 is given away for free. And despite being free it’s still not used nearly as often as it should be, it’s not the algorithm or price holding back adoption it’s the people. It was assumed that there was this huge untapped demand curve at the long end of the tail before finding out that there really was not. Z3 was originally released as academic use only (2012) on the assumption that it would drive demand for commercial use of Z3 or another commercial solution. Once it became clear that wasn’t going to make any real money they allowed it to be released for commercial use as well (2015).
As an aside people still do pay for SMT solvers so the market hasn’t completely disappeared but it’s clearly not what it was thought to be.
Your point being? You're stringing together a bunch of observations without any argument relevant to Jevons' paradox.
My guess is that you're mistake is thinking something like [demand = 1/price]. That was never the prediction, the prediction is when efficiency improves, demand increases which is a pretty basic observation for most in-practice supply and demand curves (something like a Veblen good would be different). You're not talking about changes in demand and you're barely looking at efficiency beyond noticing that the price was positive and is now 0.
> And despite being free it’s still not used nearly as often as it should be ...
Case in point, if this wasn't obvious then you need to modify your expectations. Jevons' paradox predicts exactly this, because there are prices below free (ie, paying someone to use the thing) and we'd expect they eventuated people would find new uses for SMT solvers due to Jevons' paradox in context of a price signal that they were getting more efficient. That implies even at $0 in software costs we aren't going to be using them everywhere we could.
It’s a question of depreciation of the GPU vs time to deploy AI.
If RL and synthetic data creation are all we need for self-improving AI, the large labs have an internal use case to utilize all available compute, regardless of efficiency gains.
If they have to go external and deploy in the broader market, time to deploy would make the GPUs worthless before they are fully utilized.
Did they not already have an incentive to use all available compute?
Let’s assume you were buying the GPUs, how would you plan on paying for it? What’s the business case?
Doesn't really seem like a Paradox.
There may only be 10 people willing to pay $100 for something, but 1000 willing to pay $10 for it, so lower price results in more revenue.
That’s a given.
Another counter argument is that’s Nvidia will sell less high margin advanced GPUs that require expensive networking equipment.
Problem with that argument is that every competitor gets the benefit of higher efficiency and the baseline will reset back to 0.
The models will all get more efficient, but NVIDIA as the high cost provider of compute (as well as cloud services that use them) - 10x markup on H100s - may not benefit from increased demand. If the models turn out to be a commodity, then low cost providers will be the winners, likely running on chips like TPU and Trainium (Amazon/Anthropic) that don't have the 10x markup.
Low cost providers will not be the winner because those will become commodity. It’s the leading model, the one closest to AGI that will make all the profit.
Nvidia has products to serve lower end as well. There’s no reason why Nvidia can’t dominate in the commodity sector.
And if Nvidia can’t keep their high margins, they will make it up with significantly more volume (Jevons Paradox in play).
I think there'll continue to be models at different levels of capability and price - and not at all obvious whether it'll be the Ford or Rolls Royce that'll make most money.
My guess is that is that what passes as "AGI" will also become commoditized, maybe even free to consumers (funded by advertising, business referrals to restaurants, etc).
Been thinking about EVs. They are cheaper per kilometre driven. So would this not mean that the owners of them spend more overall than ICE owners spend. And would drive a lot more. As Jevons Paradox means the demand for driving will be higher thus overall money spend on driving will go up.
An example you may have learned about in school was Eli Whitney and the cotton gin. He thought his invention would reduce slavery by making it very cheap to clean the seeds out of picked cotton. Instead, this efficiency led to an explosion of demand for this lower-cost cotton.
Benedict Evans caught on early - https://www.ben-evans.com/benedictevans/2023/7/2/working-wit...
Current market panic looks like early railroad investors freaking out that someone made a faster train that goes over same tracks.
Is that really the analogy you want to use?
« In 1873, greed, speculation and overinvestment in railroads sparked a financial crisis that sank the U.S. into more than five years of misery »
https://www.smithsonianmag.com/history/robber-baron-gamble-r...
And yet, in history, railroad is one of the few examples of bubbles that didn't ultimately popped.
You can still lose money in the greed-fear cycle even if the road leads up to unimaginable heights.
A bubble around a technology can pop, and yet the tech can be ultimately valuable in the long term. That does not mean it wasn't a bubble nor it didn't pop.
I guess if it « didn't ultimately popped » neither did the dotcom bubble - nor any housing bubble.
Yes. Both dotcoms and housing still bring immense profits to investors.
Given the likes of sama & co were asking for unfathomable levels of capital, chipmakers had backlogs, datacenter & power supply chain companies were getting bid to the moon, etc.. having a dramatic decrease in compute requirements might just bring things closer to a supply/demand balance.
Where was Stargate gonna get $500B? How was Altman ever going to raise.. $7T? (~1/3 of the entire GDP of US)
Does a 10x or even 40x decrease in compute cost just mean they can actually raise & spend the capital that their compute requires?
I see it already mentioned in the page ;)
Jevons Paradox and DeepSeek AI
The Jevons Paradox, an economic principle stating that increases in efficiency often lead to higher overall consumption, has been observed in the context of DeepSeek AI. DeepSeek, a Chinese AI startup, recently introduced its R1 model, which achieves comparable performance to leading AI systems like OpenAI's ChatGPT while requiring significantly less computational power and energy. This efficiency has made the technology more accessible and affordable, spurring wider adoption across industries.[30] As Microsoft CEO Satya Nadella noted, the increased accessibility and efficiency of AI tools like DeepSeek's R1 model could lead to a surge in applications and use cases. This expanded usage may result in higher energy consumption overall, despite individual efficiencies. For example, while DeepSeek's model uses fewer resources per task, the growing demand for AI-driven solutions—ranging from chatbots to complex reasoning models—could increase the total energy requirements of the global AI infrastructure.[31]
Deepseek is 651B, GPT models are suspected to be <200B. Where are they getting the idea that Deepseek is more efficient?
Its an MoE model so only 37B parameters are activated per token.
MoE models aren’t new. Rumors are that GPT4 was one.
yea specifically 8x220B https://www.latent.space/p/geohot
Yes, with substantially more active params for SOTA.
I made a comment yesterday that the whole NVDA thing seemed eerily similar to the AMZN thing in the first dot com crash. [0]. If the market can get sufficiently irrational, and we start to see that same like 80-90% markdown that we got with AMZN in the first crash, I mean, who knows? That's a golden opportunity. Not many generations get opportunities like that twice.
I guess we'll see what happens with the stock price, but long term I'm pretty confident in this Jevon's Paradox effect on AI use. I'm praying for more negativity hype out among those who maybe don't know as much about tech as the HN crowd. Because those people collectively, can move stock prices.
[0] https://news.ycombinator.com/item?id=42853937
Having read the paper from deepseek I seriously don't get how anyone can think "Oh now is the time to sell nvidia".
Both at the tactical level and strategic level that paper means we will see _more_ nvidia gpus fly off the shelves. If anything it's showing why the h800 is a decent card and everyone in China should buy one (dozen thousands).
majority of Nvidia's profits are from huge companies particularly Microsoft and META not individuals buying graphics cards. Look at their current video gaming graphics sales and even if these ai consumer purchases were 10x what there previous graphics sales were then the current valuation wouldn't make sense. This is all ignoring the fact that if the h800 was good enough amd's top line cards should be too.
>This is all ignoring the fact that if the h800 was good enough amd's top line cards should be too.
You can't pay me to use AMD cards. I have work to do. I don't have time to write drivers for cards that no one supports.
The point of the deepseek r1 paper is that the more synthetic data you generate by over trained models the better the resulting model. That means more GPUs. There is no reason why you can't train a dozen models like their nameless logic model in other domains and then use them for synthetic data augmentation.
The better the foundation model, the better the distilled thinking model.
So yes, an increase need for ever better thinking models.
Perception, not facts, primarily drive markets
A supposed future event, invoked by a unconstrained interpretation of Jevons Paradox, is not a fact.
But, I thought the CUDA moat is primarily needed for training.
On the inference side, why does one need Nvidia GPUs over Intel and AMD GPUs?
Probably because companies buy Nvidia GPUs training and then use them for inference. Dual use.
Not to mention there’s little real world evidence that AMD is better at inference speed not cost/inference over Nvidia.
Tensor Cores accelerate inference...
I had analyzed the paper as well, and gave some quick thoughts on a few of the low level things that stood out and would affect NVDA here:
https://news.ycombinator.com/item?id=42856652
I have to agree. I'm not sure the architectural and engineering changes here don't put NVDA further ahead. Some stuff all gpu makers can do. Like quantize on transfer. But NVDA is so far ahead in other areas and DeepSeek leans heavily on that industry leading tech. At least at the low level.
But I'm betting the average investor doesn't have the level of facility with ML to properly analyze the information that's come out. Heck, I'm betting most of the people working in AI are really just TF or PyTorch monkeys. The real AI experts have probably been hovered up by MS or FB or AMZN or OpenAI or what have you. So it feels like we could get lucky here and the AI related market could drop like a rock taking the good with the bad. Which I think would be awesome for most people in a position to wait for it to happen.
I can't wait for the 2000 dot com crash for AI.
What gets me about the deepseek work is that it's neither novel or theoretically difficult.
I've been using a similar synthetic data pipeline to generate examples for sota bert embedding models.
That low-hanging fruits like this are left unexplored by meta or alphabet - and if we're to be believed lead to crisis meetings for a week straight - makes one wonder wtf the engineers there are doing?
This is just a baseless conspiracy theory that I have, but I do wonder if they intentionally avoided certain avenues of research because it could erode the “moat” major tech firms have by dramatically reducing the capital costs required for training and inference. If you focus your research around work that mandates high-end hardware at scale you can lock out tons of potential incumbents
There is nothing in the deepseek paper that suggests you can't use the order of magnitude in hardware costs you saved to just train models that are ten times as large.
But there are thresholds of commercial viability in all of this. DeepSeek's technology gets you over that line with less sophisticated hardware.
There's already some pretty impressive work being done with folks using just a pair of M2 Ultras with r1 in a "home lab" context that goes way beyond what you could previously do with llama.
Yes, but the r1 model is shiny and new. I've been using it locally for a day and I'm already finding odd spots. One thing that I've seen it do a number of times is get the correct answer in the thinking tokens, then ignore it, generate thousands more tokens until it reaches it context length then give a wrong answer.
There are things, like doc strings for functions, that llama models give better results for.
…and even if you can’t, you could train models that are ten times more specific, or update them ten times more frequently.
It's not comparable. NVDA is heading toward ~$95 billion in operating income for the next four quarters approximately. AMZN in the first dotcom bubble crash had only (hefty) sales growth and was burning enormous amounts of cash to build itself out. There was no profit - and no near-term hope of a profit - to support Amazon's valuation back then. They were highly susceptible to an exaggerated crash accordingly.
At $95b in op income, you're looking at ~33 times op income for NVDA for 2025, or lower. I don't consider that a bargain but it's also not crazy expensive given the stock market valuation context + their growth + their market dominance in a critical field.
You can't apply Jevons paradox to everything, that would be predicting the future. However, you can try to analyze whether it holds in your specific context.
Comparing AMZN and NVDA is like comparing apples and oranges. Amazon has significantly diversified from its origins as an online bookseller and started when the internet, and later mobile devices, were not yet mainstream. NVDA, on the other hand, is an entirely different kind of company.
It doesn't matter whether it's AMZN or NVDA, you should try to distinguish between the company's fundamental value and the speculative component, as both are reflected in the price.
nvidia stock holders favorite new term
Is there literature about Jevons paradox invocations during gold rushes?
With AGI/ASI, there will come a point where this paradox breaks down: once the ASI is self-sufficient and "agentic enough", people's direct use for it will actually reduce. You won't need to "micromanage" it - and the total amount of compute it requires might actually drop drastically if the ASI improves its own efficiency much better than humans can. At that point the question becomes one of alignment: will overall consumption drop because ASI subjugates or eliminates humanity or does it grow exponentially as we use it to explore the stars.
Why can’t ASI optimize its own efficiency and make better/more chips at the same time?
This "paradox" is being cited everywhere, and while it might apply to AI, people are wrongly using it to justify why nvidia should have a $3.6T or greater valuation.
What happens to a holistic system doesn't necessarily apply to every player in that system.
nvidia's heights are largely because of a few super capitalized players dumping enormous amounts of money into data centre GPUs, letting nvidia enjoy outrageous profit margins selling rapidly obsoleted, rapidly iterating products to mega-money players who are all counting on this giving them such a differentiation that it's financially worthwhile. If AI is just some free thing that has little differentiation between the players -- commoditized, so to speak -- the market for $200,000 GPUs probably isn't going to be as big. People won't be rushing to build $500B buildouts.
Another argument I keep seeing is smaller models and edge compute. NVDA has almost no stake in that, and there is no reason to believe they're going to have any particular advantage if they tried harder there, any more than Intel could leverage PCs to Smartphones. The core operations of running a model are well known by all of the silicon makers, and they can all make massively scalable FMA matrix operators, etc.
Nvidia being worth $3.6T is a different topic. People think DeepSeek should have reinforced that valuation or even increased it, not lower it.
The paragraph on deepseek reads like cope from an NVDA bagholder tbh.
[dead]
The simplest explanation, Jevons paradox is cope from CEOs and retail investors trying to justify the share market.
People seem to have forgotten that there’s a time element involved, not a perfectly elastic and instant demand.
Thinking about it another way, if it turned out it was going to cost 30x more compute to train the next generation of models than we thought due to some scaling law, would the chip share prices be expected to rise, stay the same or go down? Probably rise.
In that case, if it was 30x cheaper to do what we need today would we expect option a, b or c?
Assuming that there would be less short / mid term sales, that impacts their bottom line and flattens off their (ie nvda) demand curve. Slower growth, lower share price. Slower growth, longer opportunity for competition. Therefore the speculative element of the chip leader’s share price’s should come down.
I don’t have any long or short positions in the chip sector.
Companies compete against each other - not themselves. It’s not like OpenAI set an internal goal of training 30x bigger and that’s it. No, OpenAI needs to compete.
If training efficiency is 30x more, then Anthropic, Google, OpenAI, Mistral, Qwen, etc all get the benefit. It resets the baseline. They will not stop competing against each other just because there is a 30x efficiency improvement that everyone has access to.
the money must be happened, otherwise shit will happen
it's like we are spending 500bn for x ai needs, but suddenly, the cost fell to 1/20, now we must convince everybody there'll be x20 needs, so that the 500bn will not go missing
that is what Satya Nadella and Sam Altman are trying to achieve
the real problem:
1. will there be 20x needs for ai? 2. how much ai needs are priced now? x? 5x? or, 20x?
Lots of NVDA holders and NVIDIA itself, trying to justify their valuations, but forgetting the Theory of Constraints will set limits on your Jevons Paradox.
The only thing relevant, is the large sale that Nancy Pelosi did before the crash. Another incredible coincidence again of course, by one of he most successful hedge managers in history...
https://www.newsweek.com/nancy-pelosi-sells-nvidia-stock-wee...
Sold last year. This is not really the place for misleading low-effort politics.
Sure, 31 December 2024 is by definition last year, but at the same time that was not even a month ago.
Come on. Which do you think is more likely, that it was a normal EOY sale of a lucrative stock, or that they sold based on insider info obtained weeks before anyone else?
Parent never suggested it was due to DeepSeek specifically, but for supossedly structural reasons if you follow the first paragraph.
I'm just calling out that "last year" wasn't really "last year" if you follow that argument.
i think it's insider trading, but not about deepseek, deepseek is never the main cause
deepseek v2 showed effeciency months ago, and nothing happened, there's also nothing happened after v3 and r1 release
the insider info is about when will Wall Street stop pushing and squeezing the market and start harvesting, and, IMO, this will happen anytime after rotation of ruling parties
i don't know if the similar scene happened before in the US, but this happened elsewhere, the rulling party with connected market makers and puppet central bank, push the market to highest and let it freely fall after rotation
nancy definetly know what the timing is
Just some nonsense, but if I’m right, it won’t take too long to witness it happen.
it was 2-3 days after deepseek paper release on Christmas Day. I read this paper too and drew the exact same conclusion (that nvda was due to decline)
but I dont think its insider trading, just informed and reactive trading - not eoy profit trading either though
Fair. The existence of said paper makes this a normal trade based on publicly available information.
Sure...Nothing to see here, just a 84 years old lady innocently playing Bingo... :
"Nancy Pelosi Outperformed Nearly Every Hedge Fund In 2024" - https://finance.yahoo.com/news/nancy-pelosi-outperformed-nea...
I am right with you with a very large NVDA holding, friend! HODL!
If it feels very large it might be wise to diversify. Not for maximum $, but maximum zzzzz.
Energy is commonly used for this paradox.