On January 27, 2025, a Chinese startup called DeepSeek released an AI model that reportedly matched the performance of OpenAI’s best, trained for a fraction of the cost using fewer, less powerful chips. According to CNN, $NVDA shed nearly $600 billion in market value in a single session — the largest single-day market cap loss by any company in history. The Nasdaq dropped 3.1%.
On March 25, 2026, Google Research published a compression algorithm called TurboQuant that claimed to cut the memory requirements of running large AI models by six times. Per CNBC, SK Hynix fell roughly 6%, Samsung dropped nearly 5%, Kioxia declined around 6%, and Micron extended a five-day losing streak that wiped over 17% from its recent high. Cloudflare’s CEO immediately called it “Google’s DeepSeek.”
Same fear. Same script. Fourteen months apart.
If you were caught off guard by one of these, that’s understandable. If you were caught off guard by both, there’s a pattern worth understanding — because it will almost certainly happen again.
The Script, Every Time
Each of these events follows a near-identical three-act structure.
Act one: a credible player announces a genuine efficiency breakthrough in AI — a cheaper model, a compression algorithm, a new architecture that does more with less. The announcement is real and the technical claims hold up. This is not hype.
Act two: the market concludes that if AI can do more with less, the companies selling the “less” — chips, memory, data centre hardware — will sell fewer of them. Memory and GPU stocks sell off sharply. Financial media runs headlines about the AI hardware boom hitting a wall. A famous investor or CEO calls it a paradigm shift.
Act three: a 19th century economic theory called the Jevons Paradox gets invoked. Stocks stabilise and recover. Infrastructure spending commitments from Microsoft, Google, Amazon and Meta confirm they are not slowing down. The thesis that “cheaper AI = less demand” gets quietly shelved until the next time.
After DeepSeek, Microsoft CEO Satya Nadella was among the first to invoke the Jevons Paradox on X, posting “As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.” As reported by Reuters, fund managers at BlackRock, J. Safra Sarasin, Nordea Asset Management and Thematics Asset Management all cited the Jevons Paradox as their reason for staying long AI infrastructure after the DeepSeek selloff. European tech stocks hit new highs within weeks.
After TurboQuant, Wells Fargo made the same argument in a note to clients. Quilter Cheviot’s head of technology research told CNBC directly that the move was “evolutionary, not revolutionary” and that “it does not alter the industry’s long-term demand picture.” Retail sentiment on Micron on Stocktwits climbed to “extremely bullish” even while the stock was still falling.
The counterargument won both times. But here’s the question nobody seems to be asking: is the Jevons Paradox genuinely the right framework here — or has it just become a convenient comfort blanket that stops investors from doing harder thinking?
What the Jevons Paradox Actually Says — And Where It Breaks Down
William Stanley Jevons observed in 1865 that as coal-burning engines became more fuel-efficient, total coal consumption in Britain went up, not down. The reason was simple: cheaper, more efficient coal-burning unlocked a massive wave of new uses that hadn’t existed before. Efficiency didn’t kill demand. It created it.
Applied to AI: if running a model costs six times less, then applications that were previously too expensive to build suddenly become viable. More products, more queries, more inference workloads, more chips needed in total. The math works — as long as the pool of potential new uses is large enough to absorb the efficiency gain.
And for most of the AI buildout cycle so far, it clearly has been. The shift from training-compute to inference-compute is already well documented. According to Deloitte’s 2026 technology predictions, inference now accounts for roughly two-thirds of all AI compute, up from one-third in 2023. That explosion in inference demand has absorbed every efficiency gain the industry has produced, and then some.
But here’s where it gets interesting for investors: the Jevons Paradox is not a law of physics. It’s a historical observation that holds when two conditions are met. First, the new use cases unlocked by efficiency gains must be large enough to replace the volume lost through reduced per-unit consumption. Second, there has to be latent demand waiting to be unleashed — customers who wanted to use the technology but couldn’t afford it.
In 2025, both conditions were clearly true. AI was still in early adoption. Enterprises and developers were desperate to build AI products but constrained by cost. Every efficiency improvement immediately unlocked a wave of new deployments. The Jevons Paradox applied cleanly.
In late 2026 and beyond, those conditions may start to look different. AI adoption is no longer early-stage. The enterprises most motivated to deploy AI have already deployed it. The developers most excited about building AI products have already built them. The latent demand pool that made the Jevons Paradox so reliable as a recovery mechanism is gradually shrinking.
At some point — nobody knows exactly when — an efficiency breakthrough will arrive and the Jevons Paradox won’t fire cleanly. Not because the technology isn’t real, but because the marginal new use case it unlocks won’t be big enough to replace the demand that compression took away.
That’s the inflection point that matters for chip investors, and it hasn’t happened yet. But the fact that it hasn’t happened yet is not the same as it being structurally impossible.
The Real Risk Nobody Is Discussing
Here’s the uncomfortable question buried under all the Jevons Paradox reassurance: what if the pattern doesn’t break through one big event, but through a slow accumulation of small ones?
TurboQuant is not the first compression technique. It joins a long list of efficiency improvements that have been compounding quietly for years — quantization, speculative decoding, model distillation, mixture-of-experts architectures, continuous batching. Each one individually is “evolutionary, not revolutionary.” Each one individually gets dismissed as not altering the long-term demand picture.
But they’re not happening individually. They’re happening simultaneously and continuously. The cost of running a GPT-4 class model has fallen from roughly $20 per million tokens in late 2022 to around $0.40 per million tokens in early 2026, according to analysis published by GPUnex — a 50x cost reduction in three years. The Jevons Paradox has absorbed all of that. But each incremental efficiency gain is asking the paradox to do a little more heavy lifting.
What this means for investors is less about any single event like TurboQuant and more about the structural position of the different chip categories.
$NVDA sits in the most defensible position. GPUs are not purely a memory play — they’re the processing layer that runs the models, and more efficient software running on GPUs often means more GPUs get adopted, not fewer. TurboQuant, notably, was benchmarked on Nvidia H100s and makes them run faster. The company that makes the hardware more efficient is not the enemy of the hardware.
$MU, Samsung, and SK Hynix are in a more nuanced position. HBM — the high-end memory that powers AI data centres — is where the demand concentration is, and HBM is relatively insulated from software compression because it’s constrained by the physical memory needed to hold model weights, not just the KV cache that TurboQuant targets. Micron’s own fiscal Q1 2026 earnings release showed the company had already contracted its entire 2026 HBM supply before TurboQuant existed, with management projecting an HBM market growing to $100 billion by 2028. Standard DRAM — the commodity memory used in general servers — is more exposed to efficiency improvements, and that’s where the longer-term risk actually sits.
The companies whose moats are thinnest are the commodity hardware players who’ve been riding the AI wave without deep integration into the AI software stack. When the efficiency gains start compounding faster than new use cases can absorb them, it’s those players who face the real structural question.
What Investors Should Actually Do With This
The recurring pattern of AI efficiency panics followed by Jevons Paradox recoveries is itself investment-relevant information — specifically, it tells you that the market has a consistent tendency to over-price near-term efficiency threats and under-price the long-term demand expansion they enable.
That means the selloffs triggered by events like DeepSeek and TurboQuant have, historically, been buying opportunities in the core AI infrastructure stack. Micron recovered from the DeepSeek period. Nvidia recovered from it dramatically. The pattern has been consistent enough that dismissing the panic as noise has been the right call every time so far.
But the investor with a longer time horizon should hold two thoughts simultaneously. The Jevons Paradox is real, it applies to AI, and it has been the right framework for the past three years. And the conditions that make it reliable — abundant latent demand, early adoption curve, large new use cases waiting to be unlocked — are finite. They don’t disappear overnight, but they erode over time.
The third or fourth version of this panic will look exactly like the first two. The question is whether it will resolve the same way. The honest answer is: probably yes, for now. Definitely not forever.
Tracking these events as they happen is straightforward. Google’s research blog at research.google is where the next technical breakthrough will surface first. Micron’s investor relations at investors.micron.com publishes the HBM supply and demand data that tells you whether the hardware thesis is holding. And the Philadelphia Semiconductor Index is the cleanest barometer of how the market is pricing the sector in real time.
The pattern will repeat. Now you know what to look for when it does.
