Jevons Effect Applies to DeepSeek, Not Nvidia

Many financial analysts and investors are concerned that lower-cost language models such as DeepSeek will reduce demand for high-end graphics processor units, servers and high-performance computing services. The drop in Nvidia market value by about $600 billion after the DeepSeek unveiling of its latest model, which, it is claimed, was developed at a fraction of the cost of other leading generative AI models, attests to that fear. 

Others do not believe a dip in need for high-end Nvidia servers will happen. The Jevons Paradox or Jevons Effect has been cited in this regard. The Jevons Effect, also known as the rebound effect, is the proposition that increases in efficiency in resource use can lead to an increase in the consumption of that resource, rather than a decrease.

So the reasoning is that cheaper AI models and cheaper inferences will lead to much more AI use,and therefore continued demand for Nvidia servers, even high-end machines. Some might point out that the claimed DeepSeek cost advantages should properly mean the Jevons Effect happens to DeepSeek and other models, not the servers running those models. 

So the reduction in demand for Nvidia servers would be indirect. That has not deterred many observers, who believe significantly-lower-cost model training and inference operations are possible (with or without DeepSeek contributions). We already have seen such effects.

source: a16z

 

In principle, some argue, such trends might mean model operators could reduce the need for high-end graphics processing units and other acceleration chips. That remains to be seen. 

Broadly speaking, the history of computing suggests the “lower costs lead to wider usage; more usage; more demand for servers” argument has legs. Computing resource demand always has increased as the cost drops. 

But there also is an argument that lower-cost models and inferences will increase demand for all those products. Paradoxically, lower training and inference costs will stimulate more demand, possibly lead to new use cases 

Also, keep in mind that generative AI is only one form of AI. Perhaps SeepSeek points the way to lower processing costs for GenAI. That does not necessarily mean improvements for general AI or even machine learning. 

One might argue it remains necessary to use high-end processors and accelerators for use cases moving towards artificial general intelligence. 

The Jevons Effect, or Jevons Paradox suggests that an increase in efficiency in resource use can lead to an increase in the consumption of that resource, rather than a decrease, as some might expect, over the long term. 

Industry/Resource	Efficiency Improvement	Observed Increase in Consumption
Coal (Historical Example)	More efficient steam engines in the 19th century	Increased coal consumption as steam engines became more widely used
Oil & Gas	Fuel-efficient car engines	More people drive longer distances, leading to sustained or increased fuel demand
Electricity	Energy-efficient lighting (LEDs, CFLs)	More lighting is used due to lower energy costs, offsetting savings
Computing Power	More efficient processors (Moore’s Law)	Increased use of computing applications and data-intensive services
Water Usage	Efficient irrigation systems	More land is cultivated, increasing total water use
Internet Bandwidth	Faster and cheaper data transmission	More streaming, gaming, and cloud computing, increasing total bandwidth consumption
Paper Consumption	Digitalization and efficient printing	More documents are printed due to lower per-page costs
Ride-Sharing & Public Transit	Efficient, affordable transport services (Uber, electric buses)	More trips taken, sometimes replacing walking or biking
Air Travel	Fuel-efficient aircraft	Lower costs lead to increased air travel demand
Food Production	High-yield farming techniques	More food is produced and consumed, increasing overall agricultural resource use

As applied to more-affordable generative artificial intelligence solutions, a drastic decrease in cost might-also lead to a decrease in buying of high-end AI servers, for example, even if model training can be done on lower-capability servers. (at least in some cases). 

The argument is that efficiency doesn’t reduce demand—it increases it. As AI becomes cheaper and more accessible, more businesses, startups, and individuals will adopt it and new higher-end use cases will emerge, driving more higher-end GPU sales. 

Perhaps the bigger immediate concern is that many contestants have essentially overpaid for their infrastructure platforms. 

That becomes a business issue to the extent that, in competitive markets, the lower-cost producers tend to win. 

On the other hand, the Jevons Effect works when the price of the inputs does not change. If the price of high-end Nvidia servers does drop, then supply and demand principles--and not the Jevons Effect--will operate. And lower prices for high-end Nvidia GPUs then sustains demand. 

The Jevons Effect suggests that improved product efficiency leads to greater overall resource consumption rather than reductions. That might apply to use of AI models, AI as a service or power consumption. 

But  many have speculated that AI models such as DeepSeerk would lessen the need for high-end graphics processor units, for example. That might hold only if the prices for such servers remains the same. 

And the general rules of computing economics suggest lower prices with scale and time. So the ultimate impact of DeepSeek and other possible model contenders on demand for high-end Nvidia servers might be more neutral than some fear.  

And, in any cases, one might argue any effect on high-end server demand might affect GenAI models more than broader and more complicated artificial general intelligence operations.

Will DeepSeek Prove Neutral for Nvidia and AI-as-a-Service Providers?

Many financial analysts and investors are concerned that lower-cost language models such as DeepSeek will reduce demand for high-end graphics processor units, servers and high-performance computing services. 

But there also is an argument that lower-cost models and inferences will increase demand for all those products. Paradoxically, lower training and inference costs will stimulate more demand, possibly lead to new use cases 

Also, keep in mind that generative AI is only one form of AI. Perhaps SeepSeek points the way to lower processing costs for GenAI. That does not necessarily mean improvements for general AI or even machine learning. 

One might argue it remains necessary to use high-end processors and accelerators for use cases moving towards artificial general intelligence. 

The Jevons Effect, or Jevons Paradox suggests that an increase in efficiency in resource use can lead to an increase in the consumption of that resource, rather than a decrease, as some might expect, over the long term. 

Industry/Resource	Efficiency Improvement	Observed Increase in Consumption
Coal (Historical Example)	More efficient steam engines in the 19th century	Increased coal consumption as steam engines became more widely used
Oil & Gas	Fuel-efficient car engines	More people drive longer distances, leading to sustained or increased fuel demand
Electricity	Energy-efficient lighting (LEDs, CFLs)	More lighting is used due to lower energy costs, offsetting savings
Computing Power	More efficient processors (Moore’s Law)	Increased use of computing applications and data-intensive services
Water Usage	Efficient irrigation systems	More land is cultivated, increasing total water use
Internet Bandwidth	Faster and cheaper data transmission	More streaming, gaming, and cloud computing, increasing total bandwidth consumption
Paper Consumption	Digitalization and efficient printing	More documents are printed due to lower per-page costs
Ride-Sharing & Public Transit	Efficient, affordable transport services (Uber, electric buses)	More trips taken, sometimes replacing walking or biking
Air Travel	Fuel-efficient aircraft	Lower costs lead to increased air travel demand
Food Production	High-yield farming techniques	More food is produced and consumed, increasing overall agricultural resource use

As applied to more-affordable generative artificial intelligence solutions, a drastic decrease in cost might-also lead to a decrease in buying of high-end AI servers, for example, even if model training can be done on lower-capability servers. (at least in some cases). 

The argument is that efficiency doesn’t reduce demand—it increases it. As AI becomes cheaper and more accessible, more businesses, startups, and individuals will adopt it and new higher-end use cases will emerge, driving more higher-end GPU sales. 

Perhaps the bigger immediate concern is that many contestants have essentially overpaid for their infrastructure platforms. 

That becomes a business issue to the extent that, in competitive markets, the lower-cost producers tend to win. 

On the other hand, the Jevons Effect works when the price of the inputs does not change. If the price of high-end Nvidia servers does drop, then supply and demand principles--and not the Jevons Effect--will operate. And lower prices for high-end Nvidia GPUs then sustains demand. 

The Jevons Effect suggests that improved product efficiency leads to greater overall resource consumption rather than reductions. That might apply to use of AI models, AI as a service or power consumption. 

But  many have speculated that AI models such as DeepSeerk would lessen the need for high-end graphics processor units, for example. That might hold only if the prices for such servers remains the same. 

And the general rules of computing economics suggest lower prices with scale and time. So the ultimate impact of DeepSeek and other possible contenders might be more neutral than some fear.  

AI Performance Improvement Will be "Stair Step" Rather than Continuous

Many observers now worry that artificial intelligence models are hitting an improvement wall, or that scaling the models will not bring the same level of improvements we have seen over the past few years. 

 That might be worrisome to some because of high levels of investment in the models themselves, before we actually get to useful applications that produce value and profits for businesses. 

Of course, some would note that, up to this point, large language model performance improvements have been based on the use of larger data sets or more processing power. 

And slowing rates of improvement suggest that further value using just those two inputs might be reaching its current limit.

 

 Of course, some of us might note that there is a sort of “stair step” pattern to computing improvements, including chipsets, hardware and most software. Moore's Law, where the doubling of transistor density on integrated circuits happens about every two years, is a prime example of stairstep progress. 

The expansion of internet bandwidth also tends to follow this pattern, as do capacity improvements on backbone and access networks, fixed and mobile. 

The evolution of operating systems, smartphones and productivity tools also often sees periods of rapid innovation followed by stabilization for a time, before the next round of upgrades.

So concern about maturing scaling laws, while apt, does not prevent us uncovering different architectures and methods for significant performance improvement. 

AI "Performance Plateau" is to be Expected

There is much talk now about generative artificial intelligence model improvement rates slowing. But such slowdowns are common for most--if not all--technologies. In fact, "hitting the performance plateau," is common. 

For generative AI, the “scaling” problem is at hand. The generative AI scaling problem refers to diminishing returns from increasing model size (number of parameters), the amount of training data, or computational resources.

In the context of generative AI, power laws describe how model performance scales with increases in resources such as model size, dataset size, or compute power. And power laws suggest performance gains will diminish as models grow larger or are trained on more data.

Power laws also mean that although model performance improves with larger training datasets, but the marginal utility of more data diminishes.

Likewise, the use of greater computational resources yields diminishing returns on performance gains.

But that is typical for virtually all technologies: performance gains diminish as additional inputs are increased. Eventually, however, workarounds are developed in other ways. Chipmakers facing a slowing of Moore’s Law rates of improvement got around those limits by creating multi-layer chips, using parallel processing or specialized architectures for example

Technology	Performance Plateau	Key Challenges	Breakthroughs or Workarounds
Steam Engines	Efficiency plateaued due to thermodynamic limits (Carnot cycle).	Material limitations and lack of advanced thermodynamics.	Development of internal combustion engines and electric motors.
Railroads	Speed and efficiency stagnated with steam locomotives.	Limited by steam engine performance and infrastructure capacity.	Introduction of diesel and electric trains.
Aviation	Propeller-driven planes hit speed and altitude limits (~400 mph).	Aerodynamic inefficiency and piston engine limitations.	Jet engines enabled supersonic and high-altitude flight.
Telecommunications	Copper wire networks reached data transmission capacity limits.	Signal attenuation and bandwidth limitations of copper cables.	Transition to fiber-optic technology and satellite communication.
Automotive Engines	Internal combustion engine efficiency (~30% thermal efficiency).	Heat losses and material constraints in engine design.	Adoption of hybrid and electric vehicle technologies.
Semiconductors (Moore's Law)	Scaling transistors beyond ~5 nm became increasingly difficult.	Quantum tunneling, heat dissipation, and fabrication costs.	Development of chiplets, 3D stacking, and quantum computing.
Renewable Energy (Solar)	Silicon solar cells plateaued at ~20–25% efficiency.	Shockley-Queisser limit and cost of advanced materials.	Emerging technologies like perovskite solar cells and tandem cells.
Battery Technology	Lithium-ion batteries plateaued at energy density (~300 Wh/kg).	Materials science constraints and safety issues.	Development of solid-state batteries and alternative chemistries.
Television Display Technology	LCD and OLED reached practical resolution and brightness limits.	Manufacturing cost and diminishing returns in visual quality.	Introduction of micro-LED and quantum dot technologies.

The limits of scaling laws for generative AI will eventually be overcome. But a plateau is not unexpected.