> Jevon’s Paradox

I've now seen this referenced two dozen times today which is well up from the 0 times I've seen it over the past year.

Is there some recent article referencing it that everyone is regurgitating?

Baader-Meinhof phenomenon, but also because everyone is writing about GPU demand and Jevon's paradox is an easy way to express the idea in a trite keyword.

Satya Nadella posted about it on Twitter.

https://x.com/satyanadella/status/1883753899255046301

Which seems pretty self-serving, and at the very least, premature.

Most people know Jevon’s Paradox there is just rarely an opportunity to bring it up, similar to Poe’s Law.

I’d bet most people don’t know what a paradox is.

I never knew there was an actual term for this, but I knew of the concept in my professional work because this situation often plays out when the government widens roads here in the States. Ostensibly the road widening is intended to lower congestion, but instead it often just causes more people to live there and use it, thereby increasing congestion.

Probably a decent amount of professions have some variation of this, so it probably is accurate to say most people know OF Jevon’s Paradox because it’s pretty easy to dig up examples of it. But probably much fewer know it’s actual name, or even that it has a name

IMHO it happens as long as you can find use cases that were previously unfeasible due cost or availability constraints.

At some point the thing no longer brings any benefits because other costs or limitations overtake. for example, even faster broadband is no longer that big of a deal because your experience on most websites is now limited by their servers ability to process your request. However maybe in the future the costs and speeds will be so amazing that all the user devices will become thin clients and no one will care about their devices processing power, therefore one more increase in demand can happen.

The increase in efficiency is usually accompanied with the process of commoditization as stuff get cheaper to develop, which is very bad news for nvidia.

If you dont need the super high end chips than Nvidia loses it's biggest moat and ability to monopolize the tech, CUDA isn't enough.

Anthropic doesn’t train on Nvidia as far as I know. Google doesn’t either.

I don’t see how DeepSeek changes things.

By the way, DeepSeek trains on Nvidia.

What do the others train on?

Anthropic on Amazon trainium chips

And TPU, and normal GPUs. They like to have diversity in their stack.

Nvidia.

"Anthropic currently uses systems containing chips from Nvidia as well as those with AWS’ Trainium and Inferentia chips to train its models." https://www.cio.com/article/3602879/anthropic-caught-up-in-a...

> Nvidia loses it's biggest moat and ability to monopolize the tech, CUDA isn't enough

CUDA is plenty for right now. AMD can't/won't get their act together with GPU software and drivers. Intel isn't in much better of a position than AMD and has a host of other problems. It's also unlikely the "let's just glue a thousand ARM cores together" hardware will work as planned and still needs the software layer.

CUDA won't be an Nvidia moat forever but it's a decent moat for the next five years. If a company wants to build GPU compute resources it will be hard to go wrong buying Nvidia kit. At least from a platform point of view.

CUDA will still be a moat for the near future and nobody is saying that Nvidia will die, but the thing is that Nvidia margins will drop like crazy and so will it's valuation. It will go back down to being a "medium tech" company.

Basically training got way cheaper, and for inference you don't really need nvidia, so even if there's an increase for cheaper chips there's no way the volume makes up for the loss of margin.

No, Nvidia's margins won't drop at all and the proof for this is Apple.

The units of AI accelerators will explode, the market will explode.

At the end of the day, Nvidia will have 20-30% of the unit share in AI HW and 70-80% of the profit share in the AI HW market. Just like Apple makes 3x the money compared to the rest of the smartphone market.

Jensen has considered Nvidia a premium vendor for 2 decades and track record of Nvidia's margins show this.

And while Nvidia remains a high premium AI infrastructure vendor, they will also add lots of great SW frameworks to make even more profit.

Omniverse has literally no competition. That digital world simulation combines all of Nvidia's expertise (AI HW, Graphics HW, Physics HW, Networking, SW) into one huge product. And it will be a revolution because it's the first time we will be able to finally digitalize the analog world. And Nvidia will earn tons of money because Omniverse itself is licensed, it needs OVX systems (visual part) and it needs DGX systems (AI part).

Don't worry, Nvidia's margins will be totally fine. I would even expect them to be higher in 10 years than they are today. Nobody believes that but that's Jensen's goal.

There is a reason why Nvidia has always been the company with the highest P/S ratio and anyone who understands why, will see the quality management immediately.

This. Well said.

People are blind to the llm gpt 3.5/r1 paradigm, and fail to see the other domains nvda is quietly setting up

Someone commented somewhere above that deepseek avoided using CUDA. So it means you can achieve very good results without Nvidia's CUDA.

"They skipped CUDA and instead used PTX which is a lower level instruction set"

CUDA in this case is the ecosystem, not the programming language.

Nobody ever got fired for buying Nvidia

Intel seems to be getting its act together. Battlemage is a decent mid-range GPU, and Gaudi 3 seems to be a fairly decent AI attempt.

> I say DeepSeek should increase Nvidia’s demand due to Jevon’s Paradox.

If their claims were true, DeepSeek would increase the demand for GPU. It's so obvious that I don't know why we even need a name to describe this scenario (I guess Jeven's Paradox just sounds cool).

The only issue is that whether it would make a competitor to Nvidia viable. My bet is no, but the market seems to have betted yes.

Long term: Likely true. Short-to-medium term: Nope.

Why not short term?

Demand isn’t there yet and most AI products are still not fully mature. It will take time.

Will you post on HN in year describing how much money you earned by buying the stock today?

I did buy:-) although the amount bought isn't impressive because I have other temporary cash flow constraints for a few months.

Jevon's paradox ony applies under certain conditions. It remains to be seen if it will hold in this case.

What do you understand those conditions to be?

Output quantity consumed (almost) always increases with falling inputs (ie, costs, whether in dollars or GPUs). But for Jevon's paradox to hold, the slope of quantity-consumption-increase-per-falling-costs must exceed a certain threshold. Otherwise, the result is just that quantity consumed increases while quantity of inputs consumed decreases.

Applied to AI and NVIDIA, the result of an increase in the AI-per-GPU on demand for GPUs depends on the demand curve for AI. If the quantity of AI consumed is completely independent of its price, then the result of better efficiency is cheaper AI, no change in AI quantity consumed, and a decrease in the number of GPUs needed. Of course, that's not a realistic scenario.

(I'm using "consumed" as shorthand; we both know that training AIs does not consume GPUs and AIs are also not consumed like apples. I'm using "consumed" rather than the term "demand" because demand has multiple meanings, referring both to a quantity demanded and a bid price, and this would confuse the conversation).

But a scenario that is potentially realistic is that as the efficiency of training/serving AI drops by 90%, the quantity of AI consumed increases by a factor of 5, and the end result is the economy still only needs half as many GPUs as it needed before.

For Jevons paradox to hold, if the efficiency of converting GPUs to AI increases by X, resulting in a decrease in price by 1/X, the quantity of AI consumed must increase by a factor of more than X as a result of that price decrease. That's certainly possible, but it's not guaranteed; we basically have to wait to observe it empirically.

There's also another complication: as the efficiency of producing AI improves, substitutes for datacenter GPUs may become viable. It may be that the total amount of compute hardware required to train and run all this new AI does increase, but big-iron datacenter investments could still be obsoleted by this change because demand shifts to alternative providers that weren't viable when efficiency was low. For example, training or running AIs on smaller clusters or even on mobile devices.

If tech CEOs really believe in Jevons Paradox, it means that last month when they decided to invest $500 billion in GPUs, then this month after learning of DeepSeek, they now realize $500 billion is not enough and they'll need to buy even more GPUs, and pay even more each one. And, well, maybe that's the case. There's no doubt that demand for AI is going to keep growing. But at some point, investment in more GPUs trades off against other investments that are also needed, and the thing the economy is most urgently lacking ceases to be AI.

Thanks for that.

There's a Jevons Paradox article up now and I'll put most of my thoughts there.

If you care to respond though, my first question would be what examples of falling input prices not subject to the Jevons Paradox are. Several of the more notorious ones involve energy, and that was Jevons's principle topic of study (The Coal Question most notably).

I've got my own theory of how technological mechanisms function, with an ontology of nine elements. Fuels are one of those, information is another. See prior comments: <https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...>

As might be pertinent to AI and LLM, whilst fuels and power applications seem to scale linearly against input (constant slope, if not 1:1 relation), information processing delivers far more variable returns, often with critical thresholds. Network effects and Metcalfe's Law are the best known of these (if highly inaccurate themselves, see Tilly-Odlyzko's refutation), but another is the limited returns of predictive and targeting applications.

For the latter, the 18 order of magnitude increase in computing power from 1965--2025 (60 years, about 20--30 Moore's Law cycles) has roughly doubled the length of accurate long-term weather forecasting from roughly 5 days to 10. It's made possible fully-resuable first-stage boosters for orbital spaceflight, which is visually impressive, but has only resulted in a five-fold reduction ($1,400/kg vs. $5,400/kg) in low-Earth orbit (LEO) launch costs (Falcon Heavy vs. Saturn V). AdTech makes possible far more targeted advertising, but to heavily diminishing returns, the core result has been an abandonment of non-targetable media by advertisers, notably print and broadcast, as well as an arms-race between the browser (for a very small fraction of the market) and advertisers (the largest of which also has the largest browser marketshare).

Which makes me wonder what applications AI LLMs might practically be put to. Advertising, manipulation, fraud, and propaganda certainly seem to be benefiting.

A minor quibble: It’s Jevons Paradox - the last name of the economist who formulated it is “Jevons”

Shouldn't it be Jevons' Paradox?

It's "the Jevons Paradox".

I’m also baffled by the reaction. Even with the ability to do more with less, the nature of the race still encourages everyone to do more with more.

Why should they invest in Nvidia now instead of investing companies which can capitalize on the applications of AI.

Also, why not invest in AMD or Intel bur Nvidia till now: Because Nvidia had the moat and there was a race to buy as much GPU as possible at the moment. Now momentarily Nvidia sales would go down.

For long term investers who are investing in a future, not now, Nvidia was way overpriced. They will start buying when the price is right, but at the moment it's still way too high. Nvidia is worth 20-30 billion or so in reality.

I was under the impression too that this would bump the retail customers demand for the 50 series given the extra AI and cuda cores, add to that the relatively low cost of the hardware. But I know nothing of the sentiments around wallstreet.

I don't feel like upgrading my 4090 that said. Maybe wallstreet believes that the larger company deals that have driven the price up for so long might slow down?

Or I'm completely wrong on the impact of the hardware upgrades.

Part of NVidias valuation was due to the perception that AI companies would need lots and lots of GPUs, which is still true. But I think the main problem causing the selloff was that another part of the popular perception was that NVidia was the only company who could make powerful enough GPUs. Now it has been shown that you might not need the latest and greatest to compete, who knows how many other companies might start to compete for some of that market. NVidia just went from a perceived monopolist to "merely" a leading player in the AI supplier market and the expected future profits have been adjusted accordingly.

> DeepSeek should increase Nvidia’s demand due to Jevon’s Paradox.

How exactly? From what I’ve read the full model can run on MacBook M1 sort of hardware just fine. And this is their first release, I’d expect it to get more efficient and maybe domain specific models can be run on much lower grade hardware sort of raspberry pi sort.

>How exactly? From what I’ve read the full model can run on MacBook M1 sort of hardware just fine.

No you can't. Unless you run their 1.5b model quantized. Almost useless.

Their full model, Deepseek V3 has 671B parameters. Not even remotely close to being able to run well on consumer hardware.

Only the distilled models can run on everyday machines, not the full model

With all the rate limits on Sonnet 3.5, people only want more of these models. We haven't even seen wide adoption yet.

No, nvidia's demand and importance might reduce in the long term.

We are forgetting that China has a whole hardware ecosystem. Now we learn that building SOTA models does not need SOTA hardware in massive quanties from nvidia. So the crash in the market implicitly could mean that the (hardware) monopoly of American companies is not going to be more than a few years. The hardware moat is not as deep as the West thought.

Once China brings scale like it did to batteries, EVs, solar, infrastructure, drones (etc) they will be able to run and train their models on their own hardware. Probably some time away but less time than what Wall Street thought.

This is actually more about nvidia than about OpenAI. OpenAI owns the end interface and it will be generally safe (maybe at a smaller valuation). In the long term nvidia is more replaceable than you think it is. Inference is going to dominate the market -- its going to be cerebras, groq, amd, intel, nvidia, google TPUs, chinese TPUs etc.

On the training side, there will be less demand for nvidia GPUs as meta, google, microsoft etc. extract efficiencies with the GPUs they already have given the embarrasing success of DeepSeek. Now, China might have been another insatiable market for nvidia but the export controls have ensured that it wont be.

>On the training side, there will be less demand for nvidia GPUs as meta, google, microsoft etc. extract efficiencies with the GPUs they already have given the embarrasing success of DeepSeek. Now, China might have been another insatiable market for nvidia but the export controls have ensured that it wont be.

Why? If DeepSeek made training 10x more efficient, just train a 10x bigger model. The end goal is AGI.

You are assuming that a 10x bigger model will be 10x better or will bring us close to AGI. It might be too unweildy to do inference on. Or the gain in performance maybe minor and more scientific thought needs to go into the model before it can reap the reward with more training. Scientific breakthroughts sometimes take time.

I’m not assuming 10x bigger will yield 10x better. We have scaling laws that can tell you more.

But I find it bizarre that you made the conclusion that AI has stopped scaling because DeepSeek optimized the heck out of the sanctioned GPUs they had. Weird.

I have not said that. I simply said that you now know that you can get more juice for the amount you spend. If you’ve just learnt this you would now first ask your engineers to improve your model to scale it rather than place any further orders with nvidia to scale it. Only once you think you have got the most out of the existing GPUs you would buy more. DeepSeek have made people wonder if their engineers have missed some more stuff and maybe they should just pause spending to make sure before sinking in more billions. It breaks the hegemony of the spend more to dominate attitude that was gripping the industry e.g $500 billion planned spend by openAI consortium etc

It doesn’t break the attitude. The number one problem DeepSeek’s CEO stated in an interview is they don’t have access to more advanced GPUs. They’re GPU starved.

There’s no reason why American companies can’t use DeepSeek’s techniques to improve their efficiency but continue the GPU arms race to AGI.

DeepSeek’s impact does not change any attitude.

I agree but in the short/medium term, I think it will slow down because companies now will prefer to invest in research to optimize (training) costs rather than those very expensive GPUs. Only when the scientific community will reach the edge of what is possible in terms of optimization that it will be back at pumping GPUs like today. (Although small actors will continue to pump GPUs since they do not have the best talents to compete).