Does Jevon's Paradox mean more demand for Nvidia and TSMC with DeepSeek?

Jevons’ Paradox happens when making something more efficient actually leads to using more of it instead of less.

The Expectation: Efficiency means using less of a resource (like fuel with efficient cars).

The Reality: Efficiency makes the resource cheaper to use, so people use it more, increasing overall consumption.

DeepSeek has made AI models cheaper, so demand for Nvidia and TSMC chips should go up, not down. With cheaper models, consumption grows, leading to more chips being bought to power AI. Better models help Nvidia and TSMC since their hardware is still needed to make them. I would only worry if DeepSeek used chips from Chinese manufacturers, but they trained their models on Nvidia GPUs made by TSMC.

Also, companies like OpenAI and Anthropic will push even harder to stay ahead by investing more in data centers. Even DeepSeek’s paper says moving forward requires more computing power:

“Distilling more powerful models into smaller ones is effective, but advancing intelligence further needs stronger base models and more large-scale reinforcement learning.”

I think you’re right, but this doesn’t mean the stock price will go up right away. Long-term, cheaper AI helps chip makers, but right now prices are inflated, and if that bubble pops, it could take time for a recovery.

@Quinlan
Exactly. DeepSeek is about optimizing models for now. When that optimization reaches its limit, then companies will start buying more chips again. Nvidia won’t break records on chip sales unless it’s already tied to existing deals. This is just how cycles in the semiconductor market work.

@Quinlan
How is TSMC inflated? They’re growing fast and only have a 28 P/E ratio.

@Quinlan
Some might argue TSMC’s current wafer prices are high because of the insane AI demand. If that demand drops, prices might follow.

@Quinlan
Why would demand drop when AI chips get more efficient? Jevon’s Paradox suggests efficiency will actually increase demand.

@Quinlan
Not always. Think about if we had a lightbulb that was 100x more efficient. Sure, we’d light more places, but 100x more? Probably not. With AI, companies might decide they don’t need as much hardware if models can be trained for less.

@Quinlan
But DeepSeek didn’t make AI 100x more efficient. Their method distills existing models to add reasoning, which still requires powerful foundational models for distillation. It doesn’t replace the need for those foundational models.

@Quinlan
You’re missing the latest update from DeepSeek V3.

@Quinlan
You’re oversimplifying the point LOE made.

@Quinlan
What about my argument is oversimplified?

So everyone’s suddenly a Jevon’s Paradox expert? Pretty funny to see.

Val said:
So everyone’s suddenly a Jevon’s Paradox expert? Pretty funny to see.

Never heard of it before, but I thought this would help companies using AI. If it makes things cheaper, my company is definitely going to benefit.

Val said:
So everyone’s suddenly a Jevon’s Paradox expert? Pretty funny to see.

Haha, sounds like Occam’s Razor all over again.

Consumption will go drastically up. When consumption goes up, more Nvidia and TSMC chips will be purchased to power them.

I don’t think this guarantees anything for AI. Right now, we don’t really know what global demand for AI looks like. A lot of what we’re seeing comes from companies and startups betting on the future of AI. If that future doesn’t happen as expected, it could backfire.

Counterpoint: Nvidia was already overpriced. This news is just the market correcting itself. The bubble was going to pop eventually.

Reese said:
Counterpoint: Nvidia was already overpriced. This news is just the market correcting itself. The bubble was going to pop eventually.

Why call this a disruption? It actually supports Nvidia’s growth.

Reese said:
Counterpoint: Nvidia was already overpriced. This news is just the market correcting itself. The bubble was going to pop eventually.

It’s seen as a disruption because it makes people question the dominance of U.S. tech. That’s enough to cause panic selling.

Think about when multi-core CPUs first came out. Everyone bought them, but software couldn’t use all the cores yet. Then, software caught up, and now multi-core is standard. DeepSeek could be a similar case. It’s making software catch up to hardware, which could shift demand but still lead to more overall growth.

I’m confused about your point. If training models is cheaper, wouldn’t companies order fewer chips to save money?