Jevons paradox explains why cheaper, open AI models could raise total spending on chips, data centers, and apps rather than shrink the market.
What is Jevons paradox, and why does it matter for AI? #
Jevons paradox says that when a resource becomes more efficient to use, and therefore cheaper, total consumption of that resource goes up rather than down. It was first observed with coal in 19th century England: more efficient steam engines used less coal per unit of output, but total coal consumption rose because cheap, efficient power made coal useful for far more things. Applied to AI, the same logic suggests that as open-source models drive down the cost of intelligence per token, total spending across the AI industry (chips, power, data centers, applications) could rise, not fall, because cheaper intelligence gets used in more places, by more people, for more tasks.
This matters right now because the AI industry is in the middle of a real fight over whether the most capable models should be open and freely modifiable, or closed and controlled by a handful of companies. Nvidia’s CEO Jensen Huang recently published a letter backing open-source AI, and most major tech CEOs signed on. Anthropic was the notable holdout. Whatever the ideological merits, the economic question underneath it is simpler: if open weights become the norm, who actually captures the money?
TL;DR #
Jevons paradox suggests cheaper, more efficient open-source models won’t shrink AI spending, they’ll expand total usage enough that overall demand for compute and infrastructure grows.Open source and closed source AI are built the same way; the only real difference is the business model, since open weights are given away rather than sold as a metered API.** Nvidia benefits either way**, because both open and closed models run on GPUs, and Nvidia has committed $20 billion toward open-source AI development specifically to grow the whole ecosystem.The application layer and infrastructure layer stand to gain the most if open models commoditize intelligence, since the value shifts from “who owns the smartest model” to “who builds the best product or serves it most efficiently.”Closed-source labs like OpenAI and Anthropic hold their lead today mostly because they scaled first and reinvested the resulting revenue, not because of any inherent technical advantage open models can’t close.China currently leads in open-source model quality(Moonshot AI’s Kimi K2 ranks close to top closed models on independent benchmarks) but is constrained by weaker access to advanced chips.Lower prices for intelligence tend to trigger more usage, not less, which is exactly the mechanism that could keep chip and data center demand climbing even as per-token costs fall.
Remy doesn't build the plumbing. It inherits it. #
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
How does the AI stack actually break down? #
Think of the AI industry as five stacked layers:
Chips. Nvidia, AMD, and specialized players like Cerebras and Groq build the GPUs and CPUs that train and run models.Energy and data centers. Hyperscalers like Google Cloud, Microsoft Azure, and AWS provide the power and physical infrastructure.Model providers. This includes both closed labs (OpenAI, Anthropic) and open-weight developers (Meta, Moonshot AI).Software infrastructure. Tooling for training and fine-tuning models, plus tooling that helps developers build applications on top of them.Applications. The products people actually touch: ChatGPT, Cursor, Replit, Lovable, and thousands of narrower tools.
Right now, closed-source labs capture an outsized share of attention and revenue because they got to the frontier first. OpenAI and Anthropic scaled early, generated massive revenue from selling access to their models, and reinvested that money into more compute, more researchers, and better subsequent models. That compounding advantage, not some inherent superiority of closed development, is why closed models still edge out open ones on most leaderboards today.
Is open-source AI actually behind closed models? #
Somewhat, but the gap is narrower than most people assume. On the Artificial Analysis leaderboard, the top intelligence rankings are usually held by Claude and ChatGPT, but Moonshot AI’s Kimi K2, an open-weight model out of China, sits right behind them. There’s no law of physics that keeps open models permanently behind. The lag exists mostly because of business model economics, not technical ceiling.
Building a frontier model requires enormous upfront capital. If a company gives that model away for free, it has to make money either by building applications on top of it or by running efficient infrastructure to serve it at scale, and both of those also require significant capital. A closed-source company only has to solve one problem: sell the intelligence directly. That’s a structural reason open-source AI has struggled to keep pace, even though nothing stops an open model from matching a closed one on raw capability.
Why does Nvidia win no matter which side wins? #
Nvidia doesn’t care whether the winning models are open or closed, because both run on its GPUs. Huang has committed $20 billion specifically to open-source AI development. That’s not pure altruism: growing the open ecosystem expands the total pool of AI usage, and more usage means more GPU demand regardless of who owns the model weights. This is Jevons paradox in practice. If open models make intelligence cheaper and more accessible, more companies build more products, and more products mean more inference requests, which means more chips.
That’s also why Nvidia was eager to get other major tech CEOs to co-sign its pro-open-source letter. A world with many competing open models, all needing to be trained and served, is a world with more total chip demand than a world with two or three closed labs rationing access.
Who else wins if open-source AI becomes dominant? #
Following the Jevons logic through the stack:
Application builders win big. If the underlying model is a commodity anyone can download and run, the differentiation shifts entirely to product experience, workflow integration, and distribution. Companies like Cursor or Replit compete on what they build around the model, not on owning it.Infrastructure and data center operators win. Someone still has to serve these open models at scale efficiently. That becomes its own business, separate from having built the model in the first place.Consumers and businesses win on price. More competing options plus more scrutiny of the code (since anyone can inspect and improve open weights) tends to push efficiency up and prices down.Chip makers win on volume. Even at lower per-token prices, the sheer expansion in total AI usage tends to outpace the price decline, echoing what happened with coal after steam engines got more efficient.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
The clearest loser, at least in relative terms, is the idea that owning a single closed frontier model is a durable moat. If open alternatives get consistently close enough in quality, the premium closed labs can charge for raw intelligence shrinks, pushing them to compete more directly at the application layer too.
Is betting on open-source AI a safe economic bet? #
It depends on where in the stack a company sits. For chip makers and infrastructure providers, betting on open source is close to a free option: they benefit from higher total usage almost regardless of who wins the model layer. For application builders, an open, competitive model layer is good news because it commoditizes the input they depend on and lets them compete on their own product rather than being at the mercy of a single closed provider’s pricing or policy changes.
For a company trying to build and monetize an open model itself, the picture is harder. The capital requirements to train a frontier-class model are enormous, and giving the result away means recouping that investment indirectly, through services, infrastructure, or downstream products. That’s a much less proven business model in AI than it has been in software generally, and it’s part of why only a small number of players (primarily those with other massive revenue streams, like Meta or China’s state-supported labs) have been willing to fund open frontier models at scale.
Frequently Asked Questions #
What is the difference between open-source and closed-source AI models?
Technically, very little. Both are built, trained, and fine-tuned using largely similar methods. The difference is distribution: open-source models release their weights publicly so anyone can run, inspect, or modify them, while closed-source models are only accessible through a company’s paid API or product.
Why would a company give away an AI model for free?
Because the value shifts elsewhere in the stack. Companies that release open models often make money by selling the infrastructure to run them efficiently, by building paid applications on top of them, or by growing an ecosystem that benefits their other business lines, as Nvidia does by selling more chips.
Does Jevons paradox mean AI prices won’t actually fall?
Per-unit prices for intelligence (cost per token) can and likely will keep falling. Jevons paradox is about total spending, not unit price. Cheaper intelligence tends to get used far more broadly, so total industry spending on compute and infrastructure can rise even as the cost of any single query drops.
Is China ahead in open-source AI?
China currently produces some of the strongest open-weight models, including Moonshot AI’s Kimi K2, which ranks close to top closed models on independent benchmarks. China’s constraint isn’t research talent or engineering, it’s access to the most advanced AI chips, where U.S. companies like Nvidia and AMD still hold an edge.
Who benefits most if open-source models become as good as closed ones?
Application developers and infrastructure providers benefit most directly, since commoditized intelligence pushes competition toward product quality and efficient serving rather than model ownership. Chip makers benefit indirectly through higher total usage, consistent with Jevons paradox.