The Jevons Paradox Meets Chip Iteration: Revisiting AI Compute Investment

Jevons' Paradox Meets Chip Iteration: Revisiting AI Compute Investment

Source: Semiconductor Horizons Published: 2026-11-03 15:50 AI chips GPUs generate posters AI demand is highly likely to continue growing beyond expectations, but a considerable portion of the first wave of infrastructure investment will still inevitably suffer asset impairment.

Gartner's View vs. Jevons' Paradox

Gartner released a forecast report that has triggered industry concerns about AI infrastructure risk. In 2026, AI service providers' investment in infrastructure will approach $1 trillion; between 2027 and 2030, they will invest another $6 trillion. To cover this expenditure, bear the interest on borrowing, and simultaneously achieve a 12% return rate, by 2030 AI data center operators will need to obtain approximately $8.5 trillion in revenue. But Gartner expects that the AI-related revenue ultimately flowing into these data centers will be at most only $3.6 trillion. On this basis, Gartner calculates that by 2030, more than $3.5 trillion of this investment will be at risk of asset impairment, and more than 80% of emerging cloud vendors may be unable to survive as independent companies. This is a risk assessment Gartner provides based on its assumptions; it does not mean that this entire amount will be written off.

But in my view, the premise of this calculation is that AI's deployment scenarios and business models will maintain their current form. However, the AI adoption curve may only just be beginning. The well-known Jevons' Paradox points out that the lower the barrier to using a technology, the more people will discover new ways to apply it. Therefore, even if the cost of technology declines, the total revenue generated by all applications may instead rise. For example: after automobile fuel efficiency improves and the cost per kilometer of fuel consumption decreases, people will drive more and travel farther, ultimately increasing total fuel consumption. The same applies when oil prices fall: when oil is cheap, people travel more; when oil prices rise, many people switch to public transportation.

Back to AI compute. According to Epoch AI statistics, since 2023, under the benchmarks it tracks, the compute cost required to reach an equivalent benchmark score has fallen by roughly 13 times per year (this conclusion is affected by the rules for selecting benchmarks).

According to Jevons' Paradox, lower barriers to use and continuously falling costs will spawn new applications and drive up overall usage. This will be a long-term trend, but it does not mean that demand growth will necessarily outpace the rate of price decline.

From this perspective, Gartner's judgment is rather pessimistic; it is based on current application scenarios and existing demand, while the scale of demand in the coming years is very likely to change. Even so, there is another key variable: the form of gas stations remains unchanged year after year, but the chips installed in data centers (which can also be called AI factories) are different every year, whether through generational replacement or architectural innovation. New-generation chips require additional investment and may also cause impairment of old investments. This further increases the uncertainty of Gartner's forecast.

Signals Released by the Chip Rental Market

At present, the value-retention cycle of "old chips" is longer than expected. Recently a16z published a chart based on Ornn rental data (as of October 19), comparing the current rents and expected rents of various generations of Nvidia GPUs. The expected calculation method is simple: take the initial rental price of a chip and assume that the rent declines linearly to zero over six years, roughly consistent with the server depreciation pace in the financial reports of large cloud vendors. The starting rent of each generation of chips is different, so the starting point of the depreciation curve is also at different stages of the product life cycle. At present, the actual rents of all models are above this expected line: A100 is 9% higher, H100 is 54% higher, B200 is 111% higher, and H200 is 143% higher.

GPU rental market conditions Data source: a16z

a16z also mentioned that most of the rental premium gap opened up after April, when Token demand exploded, which is consistent with the effect of Jevons' Paradox, although this chart alone cannot prove causality.

But two points need to be noted: for most of their life cycle, the rents of these chips are actually below this baseline; in addition, high rent does not equal profitability — even if the chips are rented out, after deducting electricity, cooling, networking, data center space, and financing costs, operators may still lose money.

Low-Cost Oil Wells and High-Cost Oil Wells

Recently some have compared compute to new oil, and this metaphor is apt in many respects. Tokens are equivalent to barrels of crude oil, the commodity customers actually purchase; each generation of GPU corresponds to an oil field. Some oil wells are easy to exploit and low-cost, produce crude oil of better quality, and have a very low cost per barrel; others are difficult to exploit and produce crude oil of average quality, with a higher cost per barrel. Applied to GPUs, the cost of generating one Token mainly depends on the power consumption of that generation of chips and whether it can handle the corresponding tasks.

When oil prices rise, all oil wells will be put into production, even high-cost oil wells; when oil prices fall, the highest-cost oil wells will be shut down first. The chip market is the same. When demand was booming since April this year, even the A100 was in short supply, and the rents of all models were above the expected line; a16z's chart precisely reflects this phenomenon. Once new-generation chips are supplied at scale, they become the new low-cost oil wells; compute prices then decline, and the oldest batch of chips is the first to be eliminated. For example, Microsoft plans to deploy its self-developed Maia accelerator at scale starting in 2027, claiming a 40% increase in the number of Tokens produced per megawatt of compute, a figure that has not yet been verified by third parties.

Tokens Are Not Equivalent

The oil metaphor has a limitation: it assumes that every barrel of crude oil is the same, and this is also where much current analysis oversimplifies the problem. Different Tokens and the chips that generate Tokens are not the same. Interactive and agentic tasks have very high latency requirements and need the latest, high-speed memory; batch tasks, such as overnight document processing, synthetic data generation, and model evaluation, care more about cost and are almost insensitive to latency. For a chip that has already been fully depreciated, the remaining cost is mainly electricity, cooling, and operations and maintenance, and it is often the lowest-cost choice for running such batch tasks. Therefore, the compute market is not a unified market, but is divided by task type into multiple segmented markets.

But there is another feature here: the incremental demand brought by the Jevons paradox is itself highly sensitive to price, and many applications have commercial value only when AI compute is cheap enough. Some of this belongs to high-value business, but a large amount of the new demand is batch tasks, and older chips are most competitive in such scenarios. An interesting recent example is the model Jev, launched by TypeSafe AI. It outputs structured decision results rather than text, and the developer claims its cost is hundreds of times lower than that of large language models. Such models can offload a large number of agentic decision tasks from expensive large models. In my view, it is very well suited to old hardware, although TypeSafe has not disclosed what it runs on.

At the other end, frontier large models continue to grow larger and increasingly adopt sparse architectures, that is, mixture-of-experts (MoE) architectures: when generating a single Token, only a small subset of the model's modules is activated. For ultra-large-scale sparse models, the bottleneck is often not raw compute, but memory capacity and high-speed inter-chip interconnect. Thus the basic unit of compute shifts from the chip to the entire rack. Old GPUs are not unable to run; rather, they have insufficient memory capacity and limited interconnect capability. In the future, new architectures better suited to sparse models may also emerge.

Beyond these two types of scenarios, there are also various models such as recommendation, image generation, and scientific simulation, whose hardware requirements differ from one another.

Taken together, the market presents a barbell shape: giant frontier models gravitate toward the latest rack-scale systems; lightweight, specialized models run stably on old chips. My judgment is that the pressure is concentrated in the middle range: chips that are neither frontier nor yet fully depreciated, and whose book residual value is still high. If this judgment holds, a large amount of the investment in 2024 and 2025 happens to fall in this range.

Heterogeneous Data Centers

This adds another layer of complexity: data centers in the past also mixed multiple types of hardware, but the categories were relatively limited. A decade ago, cloud data centers carried different CPUs, memory, storage, and network equipment, and only needed to match tasks to a relatively predictable hardware list. The depreciation curve of servers was also gentle; a new generation of CPUs offered limited improvement over the previous generation, and old servers could reliably generate revenue for many years, similar to car rental: last year's models are not much different from the new ones.

Today the variety of hardware is much richer. In addition to CPUs, there are GPUs, TPUs, and various AI accelerators; inference tasks are also split into different stages, and each stage is most efficient on different chips. Nvidia launched the Rubin CPX chip specifically for the Prefill stage in 2025, then temporarily shelved it and shifted to a rack solution based on Groq licensed technology for decoding tasks. In addition, high-performance computing (HPC) scenarios often require high-precision arithmetic, while many AI chips do not prioritize support for such capabilities. Matching tasks to the right chip has become much harder, and each generation of chips depreciates much faster.

The leading big players can withstand this diversity of hardware. When scale is large enough, they can keep all types of hardware fully utilized and schedule tasks to the lowest-cost available chips. Small and medium-sized data centers find this hard to do, and general-purpose GPUs remain the safest choice for them: as they age, they can shift from training to inference and then take on batch tasks. Specialized hardware has a lower per-task cost but higher asset risk, because it is hard for it to take on other types of business.

Self-built data centers are another matter. Enterprises running fixed, mature businesses can choose highly tailored hardware solutions and do not need to chase every new generation. For such workloads, top chips are mainly used to train in-house models or for a small number of GPU-accelerated solvers.

The Silicon Roadmap Determines the Pace of Depreciation

This is also what I see as the core of the problem: many big companies spread server costs over six years in their financial reports, but this is only an accounting rule and does not represent the actual economic cycle in which chips can still generate revenue. The profit cycle of a chip is determined by the designers of the next generation of chips. Nvidia's iteration pace, the improvement in Tokens per watt for each generation, new memory standards, data formats, and the software ecosystem that keeps old chips usable together determine the competitive life cycle of a GPU. In this sense, the silicon technology roadmap determines the real depreciation timetable for investment.

Developing chips in-house is equivalent to controlling your own timeline. Microsoft Maia, Google TPU, and Amazon Trainium can all independently regulate the retirement pace of compute clusters, although demand and competitors will also impose constraints. For emerging cloud providers that started out by renting Nvidia GPUs, the pace is in the hands of suppliers and customers.

Nvidia sits at both ends of the contradiction: on the one hand, it keeps iterating new products, accelerating the depreciation of customers' existing compute assets; on the other hand, according to The Economist, Nvidia has also promised that if customers cannot repay loans taken out to purchase chips, Nvidia may assume up to $300 billion of the related debt. Faster iteration can sell more new chips, but it also dilutes the value of the loan collateral (old GPUs). Jensen Huang wrote in August that the A100 will still have useful value in 2029; that is a technical judgment, but he also needs to weigh this contradiction, even as Nvidia continues to launch new generations of products.

The automobile-and-chip analogy

Back to Gartner's core question: once the market turns, who will take on these chips? The car-rental industry is a good reference. Rental companies mostly borrow to buy cars in bulk, with the vehicles serving as loan collateral, and ultimate profit or loss depends on the disposal price of used cars. In early 2024, Hertz sold off about 20,000 electric vehicles and recorded a loss of about $245 million. Hertz mainly attributed this to high repair and collision costs, but used-car depreciation was also a contributing factor: Tesla cut new-car prices, directly pushing down the used EV market. Suppliers' pricing decisions eroded the value of customers' assets. GPUs carry higher risk: the generational improvement in chips is far greater than in car model updates; the used-chip market is smaller and more niche; and the power resources supporting compute are not as abundant as parking spaces.

New signals from earnings reports

Earnings season has brought new clues. Meta reported earnings and narrowed its 2026 capital expenditure to $135–145 billion, citing continued rising infrastructure costs and spending on externally purchased AI token services. Microsoft, Alphabet, and Amazon then released earnings in succession. All three mentioned that AI compute demand remains higher than newly added capacity supply, and Microsoft and Google are even renting out compute externally to fill the gap. Together, the three plan to invest nearly $600 billion in 2026; both Alphabet and Amazon said capital expenditure will increase significantly further in 2027.

Several details in the earnings reports are worth noting: Microsoft said part of Azure's growth was due to "continued efficiency gains in CPU and GPU clusters"; just like rental car companies, making existing assets more productive is itself part of the business. Amazon said short-term rental prices for compute on the previous generation of Trainium chips showed a premium compared with the first half of the year, confirming the pattern reflected in the a16z chart—this time for Amazon's in-house chips. Amazon also mentioned that inference revenue already accounts for more than half of AWS AI revenue, exactly the kind of price-sensitive business suited to older chips and specialized chips. Google now directly sells complete TPU systems that can be deployed in customers' own data centers; Amazon plans to sell Trainium compute externally outside AWS starting at the end of 2027. In-house silicon roadmaps are becoming products that the big companies sell externally, and the pace of more and more market participants will follow these big companies' roadmaps rather than being led only by Nvidia.

Investors have also realized once again: rising usage does not equal rising revenue in lockstep. Alphabet lost about $500 billion in market value in a single day, triggered by Google's admission that advertising revenue from its AI search mode is lower than from traditional search. Users use AI more, but Google's monetization revenue per search has instead declined—at least at this stage. And advertising revenue is precisely the source of funding that supports large-scale compute buildout, which is crucial for all companies purchasing chips.

More and more investment needs to rely on borrowing. Meta, Alphabet, and Amazon all reported negative free cash flow; Microsoft, which still maintains positive net cash inflow, also issued $25 billion in new bonds the day after releasing its earnings.

Back to Jevons

We can revisit the origin of this paradox. In 1865, economist William Stanley Jevons found in The Coal Question that more efficient steam engines did not reduce Britain's coal consumption but instead pushed it higher—after costs fell, coal found more use cases. This conclusion has stood the test of time: as steam engines kept improving, British coal consumption continued to rise, although efficiency was not the only driving factor. But this does not mean all mines can survive; as the industry expands, the least efficient mine owners will still go bankrupt. I predict the AI market will move toward a similar situation: AI demand is highly likely to continue growing beyond expectations, but a considerable portion of the first wave of infrastructure investment will still inevitably suffer asset impairment. The two outcomes share the same source; they simply view the same silicon roadmap from the perspectives of buyers and compute operators, respectively.

This article is reproduced from media reports or online platforms and represents the author's personal position or views. Our republication is solely for sharing and does not mean we approve of or endorse it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly; as a neutral platform service provider, we will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

The more agents there are, the better Intel looks 2026-11-03 16:35

Global smartphone shipments plunge 19%, revenue rises 3% against the trend 2026-11-03 16:28

Apple's robot is not on a humanoid path 2026-11-03 16:24