Every major technological revolution promises greater efficiency. Conventional wisdom says that should reduce demand and lower costs. Yet AI economics appears to be doing the opposite. A 160-year-old economic principle known as Jevons Paradox may explain why. As AI becomes cheaper, faster and more accessible, businesses aren’t using less of it; they’re finding entirely new ways to consume more.

Jevons Paradox, identified by William Stanley Jevons in 1865 while studying coal, holds that improving resource efficiency does not cut consumption; it lowers barriers, unlocks new use cases, and ultimately drives demand higher.

Microsoft CEO Satya Nadella has repeatedly pointed to Jevons Paradox to explain why advances in AI are driving soaring compute demand rather than reducing it. As models become faster, cheaper and more capable, businesses don’t simply use AI more efficiently; they find entirely new ways to deploy it.

Three companies at the forefront of AI hardware, infrastructure and enterprise software explain how this dynamic is reshaping the competitive landscape for startups and small businesses.

Efficiency That Unlocks New Demand In Hardware

Faster, cheaper hardware is removing bottlenecks and transforming AI economics. Tensordyne builds inference racks using 90% less power than alternatives, an advance that embodies Jevons’ insight: lower costs mean organizations run more AI, not less.

“Jevons Paradox is gaining momentum because the type of work AI is actually doing has become more complex and intensive,” says cofounder Gilles Backhus. “When ChatGPT launched a few years ago, it was often one input in, one output out. Now, a single developer's agent might be performing hundreds of tasks at a time, each requiring detailed reasoning that uses more compute.”

Customers Value Speed And Model Quality

The shift to agentic systems has already boosted compute use per user and task, but hardware has struggled to keep pace, leaving many facing rising costs even as they rush to adopt AI.

“To outrun Jevons Paradox, we need to make it much cheaper to run AI workloads, without compromising on what customers value: speed and model quality,” says Backhus. “Software will get more efficient, but the biggest needle-mover will be hardware: chips and racks.”

The Tensordyne Napier inference system uses less energy, less space, and less money per token, keeping costs down while still running 13 times higher throughput than Nvidia. Systems, and the AI models served by them, are typically either fast and expensive, or slow and cheap. The company plans to eliminate that compromise, making it less capital-intensive to run high-performance AI workloads.

“Startups have a limited supply of capital to burn, so each token is much more valuable,” says Backhus. “If chips and data centers are more than 10x cost-efficient, models will be cheaper to run, and the margins of financially constrained early-stage companies would transform overnight. It could be rocket fuel for growth.”

Lower costs could democratize AI R&D, unlocking use cases currently too expensive to test. For founders, the core lesson lies in the underlying economics. “If a state-of-the-art customer support model currently costs $5 a session, but becomes $0.50 because the hardware is cheaper, then we will see ten or even a hundred times the number of companies adopting it,” says Backhus.

Productivity That Creates New Possibilities In Professional Services

Lower-cost AI only matters if businesses can find new ways to use it. That’s exactly what companies like Orbital are seeing. In fields like law and real estate, where output has long been tied to hours worked, AI doesn’t just cut time; it enables work that was previously uneconomical or impossible. Orbital, a real estate AI platform supporting 200,000 annual transactions across the U.S. and U.K., sees this play out daily.

“Developers designing these AI systems aren't working fewer hours either, usually the same or sometimes more; they're just getting far more done,” says CTO Andrew Thompson. “Human capital plus token capital is now producing more software per head.”

Software Is Inherently Creative

The real question is what that extra capacity enables. “Definitely higher total output, but the more interesting point is what that output buys you,” says Thompson. “The single greatest predictor of a software team's success is the rate of shipping new things to customers. Software is inherently creative. Nothing is real until customers actually use a product.”

He also points out that building software is a discovery process. If AI supercharges that discovery loop, you can ship more and arrive at solutions faster. “Those are the efficiency gains that people aspire for when levelling up their work with AI,” he says.

The ‘10x engineer’, once a rarity, is now achievable for those who master AI, letting small teams deliver what once required entire departments. “Fewer people producing the same or more output means productivity per person is up, and compensation is rising for the individuals who've leveraged AI effectively,” says Thompson. “I think that’s why salary growth is spiking inside tech for various roles.”

He anticipates the Jevons effect creating a virtuous cycle of adoption, where greater utility drives ever-higher consumption of AI tools, but with some nuance.

“Back in the day, Intel kept producing CPUs faster, but Microsoft Word never sped up at the same rate,” he says. “That's because Word wasn't static, and as CPUs could handle more, Microsoft's developers built richer and richer functionality into Word that wasn't previously possible. The hardware gains were absorbed by new capability rather than speed.”

Due diligence in commercial real estate is also being rewritten, with firms able to review every document for every property and produce a full set of due diligence reports as if every property had been manually reviewed. “This is a new offering that law firms can sell to clients, and work that simply didn't exist before because it wasn't economically viable,” says Thompson.

Accessibility That Fuels the Flywheel Of Infrastructure

As more businesses adopt these AI-powered workflows, demand for compute doesn’t plateau; it accelerates. That’s where infrastructure providers such as Verda come in. Better hardware drives efficiency, and easier access to compute accelerates the Jevons effect further.

The AI infrastructure provider has grown revenue twentyfold in two years, topping a $100 million annual run rate and raising $117 million in April. That growth stems not just from easier access, but a reinforcing cycle, where improved hardware and more capable models advance together.

A big part of the growth story for Verda, and for the AI industry generally, is the models and their capabilities getting better, moving from almost-good-enough to genuinely production-ready. CTO Arturs Poli says: “Rather than thinking of rapid growth purely as a by-product of easier access to compute, I'd see it as hardware and models improving together: accelerators getting more capable, and the models running on them becoming more accurate. They go hand in hand, and it's a clear Jevons dynamic with very rapid, hard-to-predict growth.”

Shaping Value And Pricing

Democratizing access removes longstanding barriers favoring large enterprises. High-quality, open-weight models paired with more easily available compute have opened the door to lower-cost agentic workflows. It’s also easier for businesses and enterprises to deploy new services leveraging AI compute, as cost and availability become less of a barrier to entry.

This flywheel is already reshaping value and pricing. Poli expects the trend of increased usage to speed up, driven by more efficient and diverse ways to consume intelligence.

And sustainability is deeply tied to this expansion.

“For us, sustainability matters enormously, and we're hoping to pave the way with our own practices,” he says. “Concretely, we're thinking about everything from producing new clean energy to using excess heat to warm homes.”

If the companies building AI are right, Jevons Paradox may prove to be one of the defining economic forces shaping the next phase of AI economics. Rather than simply lowering costs, greater efficiency could expand AI into markets, products and services that would previously have been impossible to justify economically.