The Jevons Paradox: Why Cheaper AI Is Bullish for Memory, Not Bearish

I use AI every day for product research, marketing analysis, competitor monitoring, tasks that used to require hiring a real human, briefing them, managing them, and hoping they stayed and performed. Also dealing with attitudes sometimes.

Here is how I think about the cost. I pay for AI subscriptions the same way I pay salaries. When a model gets better and can do more for me, I upgrade my subscription without thought. Willingly. The cost feels the same because the value went up proportionally. I am not paying for tokens. I am paying for output. And when the output improves I pay more, not less.

That is Jevons Paradox in one paragraph. Efficiency improvements do not reduce spending. They shift the ceiling of what is worth spending on. The value of results are worth paying more for.

What DeepSeek Actually Proved

In January 2025 a Chinese AI lab dropped a model that matched GPT-4 level performance at a fraction of the training cost. Nvidia lost $600 billion in market cap in a single day. The narrative was immediate and loud, AI needs less compute than we thought, the GPU supercycle is over, sell everything.

I was not surprised. That is what Chinese engineers do. More with less. They have been doing it for decades across every industry they enter. Just hacking it.

But I also knew it was not an Nvidia killer. Not even close.

Here is why. DeepSeek proved you could train a frontier model more efficiently. It said nothing about what happens when you deploy that model to a billion users running agentic workflows. Training efficiency and inference demand are completely different conversations.

And the market figured this out within weeks. Nvidia is now reporting $81 billion in quarterly revenue. Memory stocks are near all time highs. The DeepSeek panic was one of the most expensive overreactions in recent market history for anyone who sold into it. If you had conviction and held, you would have been paid.

The Mechanism Most People Miss

Jevons Paradox has a specific mechanism that people skip over when they cite it.

William Stanley Jevons observed in 1865 that more efficient steam engines led to more coal consumption, not less. Not because each engine burned more coal. But because cheaper operation made steam power economically viable for applications that previously could not justify the cost.

The same thing happens every time AI gets more efficient.

When GPT-3.5 came out it was impressive but expensive for most business applications. When GPT-4 came out it was better and cheaper per token. Did businesses use fewer tokens? No. They built entirely new applications that did not exist before because the economics finally worked.

When DeepSeek proved frontier AI could be achieved at lower training cost, it did not reduce inference demand. It made frontier AI viable for applications and geographies and companies that previously could not afford it. Every new application is a new source of inference demand. Every new inference session needs memory.

My Own Business Is the Proof

I did not start using AI less when models got more capable. I started using it for things I previously would not have attempted.

Product research that used to take a junior employee three days now takes an AI agent three hours. I do not pocket the savings. I redirect the time to researching ten more products. Market analysis that previously required an expensive consultant now happens in the background while I am on a call. I do not stop at one analysis. I run five.

The tasks that AI does not replace are not even the interesting ones. The interesting development is the tasks that did not exist in my workflow before AI, things I never attempted because the time cost made them economically irrational. Those tasks now exist. And they all require compute and memory to run.

Every e-commerce operator who discovers what I discovered adds more sessions to the global inference load. Every session needs memory bandwidth. The efficiency gains do not compress demand. They expand the universe of what gets attempted. Think about that when other old industries join.

The Agentic Multiplier

There is a second layer to this that most Jevons discussions miss entirely.

The shift from chatbot AI to agentic AI is not just more usage of the same thing. It is a fundamentally different memory demand profile. A chatbot query is a brief spike in memory bandwidth. An agent running a 40 step browser task maintains persistent context across the entire session, screenshots, action history, intermediate results, error states, for minutes or hours at a time.

The efficiency improvements that make agentic AI cheaper to run also make it viable for businesses that could not previously afford continuous agent deployment. Cheaper per session means more sessions running simultaneously means more sustained memory bandwidth demand.

DeepSeek made AI more efficient. Agentic AI made each session more memory intensive. The net effect is not a wash. It is a multiplier.

The Counterargument Worth Taking Seriously

I want to be honest about where Jevons can fail.

If efficiency improvements arrive so dramatically and so fast that they outrun adoption, a genuine 20x efficiency gain arriving before new use cases can absorb the freed capacity. You get a temporary demand air pocket. The stock market prices that air pocket immediately even if the long term demand story remains intact.

That is the DeepSeek scenario that spooked Nvidia for one day. It was wrong about the direction but correct about the mechanism. A sufficiently dramatic efficiency breakthrough can cause short term pain even inside a long term bull story.

The signal I watch is VC deal flow into memory efficient inference companies. When Sequoia or a16z writes a $200 million check into a startup claiming 10x memory efficiency at scale, that is the market starting to price a disruption. That check has not been written yet at that scale. When it is, I will reassess.

Until then, every time someone tells me cheaper AI is bearish for memory I think about my payroll.

I keep paying it. I keep getting more done. And I keep wanting to do more.

That is Jevons Paradox. And it is the most underappreciated driver of why the memory supercycle has more room to run than the efficiency bears think.

Remember when mobile data was expensive enough that watching a video on your phone felt like a luxury? Teenagers today scroll short form video for hours without a second thought. They have never experienced data as something scarce. The efficiency gains did not reduce data consumption. They created a generation that consumes more data before breakfast than most people did in a month in 2005.

That is where AI is going. The generation entering the workforce today will not think of AI compute as something to be rationed. They will use it the way teenagers use data, constantly, casually, without thinking about the cost.

And every one of those sessions needs memory.

I hold a long position in DRAM ETF (ticker: DRAM). Not investment advice. Do your own research.