in short
Every measure of compute cost is collapsing — roughly 10x a year for equivalent AI performance. Yet total spend, hardware scarcity, and power demand keep climbing. The unit price falls; the bottleneck just migrates to whatever is newest and scarcest.
In late January 2025, a Chinese lab called DeepSeek released a model that matched frontier performance at a fraction of the training cost, and the market did something revealing: it sold off power and nuclear stocks. Constellation, Vistra, GE Vernova, and Talen all fell double digits in a day (AIProem). The logic seemed airtight — if AI just got dramatically cheaper to run, it will need less hardware and less electricity, so the companies selling power are worth less.
Microsoft’s CEO was among the first to say the market had it backwards, invoking a 160-year-old idea to explain why (AIProem). When a resource becomes more efficient to use, we tend to use so much more of it that total consumption goes up, not down. The reaction to DeepSeek wasn’t the end of the AI power trade. It was a textbook misreading of what falling prices actually do.
Every cluster, however advanced its chips or elegant its architecture, runs on electricity that has to come from somewhere. That question — where the power comes from — is now shaping how the next decade of AI infrastructure actually gets built. And it’s being answered in three very different places: on land, under the ocean, and in orbit.
Why Falling Compute Prices Don't Lower the Bill
The core idea here has a name — the Jevons paradox — and understanding it is the whole game.
What the Jevons Paradox Actually Says
In 1865, the economist William Stanley Jevons noticed that more efficient steam engines didn’t reduce Britain’s coal consumption; they increased it, because cheaper steam power made coal worth using for far more things. The efficiency gain expanded the market faster than it shrank the per-unit demand. A modern academic framing of this same rebound dynamic, applied directly to AI, treats it as a pair of linked paradoxes rather than a single curiosity (arXiv 2504.10503).
The formal condition is precise, and worth stating plainly because it’s where most commentary gets sloppy: falling prices raise total consumption only when demand is elastic — when a given percentage drop in price triggers a larger percentage rise in usage. If the cost of a task halves and users respond by running more than twice as many tasks, total spending rises even though each task is cheaper. If they run only 50% more, total spending falls and efficiency wins. Jevons is not a law of nature. It is a statement about elasticity, and it holds only while demand for the cheaper thing is nearly insatiable.
The Refined Claim: The Bottleneck Migrates, It Doesn't Vanish
That framing lets us sharpen the intuition that cheaper compute hasn’t made AI cheaper. The precise version is this: the unit cost of compute is falling relentlessly — and so, eventually, is the price of any specific chip. What doesn’t fall is the cost of the scarcest input at any given moment. Because demand for frontier AI capability runs ahead of supply, the bottleneck never disappears. It migrates — from last year’s chip to this year’s, and, increasingly, to the power and physical infrastructure underneath them. An investor who watches the falling unit price and concludes the constraint is easing is watching the wrong number.
The Evidence: Prices Down, Everything Else Up
Three data pictures, laid side by side, tell the whole story.
Unit Cost Is in Freefall
The price of a fixed level of AI performance is dropping faster than almost any technology in history. Andreessen Horowitz named the trend “LLMflation”: for a model of equivalent quality, inference cost has fallen roughly 10x per year, so that GPT-3-level output that cost about $60 per million tokens in 2021 sold for around $0.06 by late 2024 — a thousandfold decline in three years, faster than the cost curve of PC computing or dot-com bandwidth (a16z). Epoch AI, using a benchmark-anchored method, measures a median decline near 50x per year across tasks, accelerating past 200x per year since early 2024 (Epoch AI). Stanford’s AI Index clocked a 280-fold drop for GPT-3.5-equivalent quality in about eighteen months (VoxBooster synthesis). Some of this is better chips; a meaningful slice — roughly 3x per year — is pure algorithmic efficiency stacked on top (Gundlach, Lynch, Mertens & Thompson, MIT — “The Price of Progress,” arXiv 2511.23455). By every measure, compute is deflating.
Yet Total Spend and Volume Explode
Now the other side of the ledger. OpenAI’s inference bill for 2025 was reported near $8.4 billion — roughly four times the prior year — even as the per-token cost of everything it runs collapsed. Google reported processing about 3.2 quadrillion tokens per month by mid-2026, on the order of seven times the prior year’s volume (VoxBooster). This is the elasticity condition being met in real time: unit prices fell 10-to-50-fold, and usage rose fast enough to more than swallow the saving. The bill didn’t shrink with the price. It grew against it. As one analysis put it, the historical rhyme is the internal combustion engine — a roughly 100x efficiency leap that drove global oil demand up more than 200-fold (AIProem). Cheaper never meant less.
The Bottleneck Migrated to Frontier Hardware
Here is where the client’s instinct needs one correction, and the correction makes the thesis stronger. It is not true that “hardware doesn’t get cheaper.” Older hardware gets dramatically cheaper: a used H100 that traded near $40,000 in late 2023 changed hands for around $6,000 by mid-2026 — an 85% collapse — as newer chips arrived, and cloud rental rates for the same card fell 64–75% from their peaks (IntuitionLabs; Introl). The compute market is now two-speed.
What stays scarce and expensive is the frontier. Nvidia’s Blackwell generation (B200, GB200 NVL72) sells at roughly $30,000–$50,000 per GPU (IntuitionLabs) and has run persistently short: an early order backlog above $50 billion (Winzheng), a booking pipeline reported around half a trillion dollars, and a CEO describing demand as exceeding expectations “by 10x” (AI CERTs). The pricing power shows up in the financials — Nvidia’s most recent quarter put data-center revenue around $51 billion at roughly 73% gross margin (AI CERTs). Hyperscalers locked up most of the available frontier allocation through 2026 and into 2027 via forward contracts, crowding out everyone without a seat at the table, and sovereign buyers are now queuing behind them (Spheron). The unit cost of compute falls; the price of this quarter’s best compute does not, because that is where the scarcity always relocates.
And to Power
The same rebound is now visible in electricity. The energy used per query keeps falling, but aggregate demand keeps climbing. The International Energy Agency estimates global data-center electricity use rose about 17% in 2025 — with electricity at AI-focused facilities up roughly 50% — and projects the total will nearly double from about 485 terawatt-hours in 2025 to around 950 by 2030 (Brookings, citing IEA; Eurelectric, citing IEA “Energy and AI,” 2025). The physical density is the tell: a single Blackwell rack draws on the order of 120 kilowatts, against an industry-average rack of about 7.6 (ModulEdge). Efficiency per computation improved. Total power intensity went up anyway. The scarce input migrated again — from silicon to the grid.
The Challenge: Where the Paradox Could Break
Intellectual honesty requires naming the conditions under which this thesis fails, because they are real.
The Jevons dynamic holds only while demand elasticity stays above one — while each price cut unlocks more than enough new usage to outweigh it. That is an empirical fact about the present, not a guarantee about the future. If AI’s useful applications saturate, or if a broad, durable “killer” use case fails to keep materializing, demand growth could decelerate below the efficiency rate — at which point the paradox inverts and falling prices genuinely do shrink consumption. The rebound is a phase, not a permanent law.
There’s a measurement caveat, too. Nobody has cleanly estimated the price elasticity of AI compute; what we observe is that consumption and efficiency are rising together, which is consistent with a rebound but also with the simpler fact that enormous capital is flooding the sector for reasons unrelated to price. Rising spend is not, by itself, proof of Jevons — it’s proof of demand, whatever its cause. And the two-speed hardware market cuts both ways: frontier prices fall more slowly than commodity ones, but they do eventually fall (Data Storage). If model capability plateaus, the frontier premium compresses and the bottleneck stops migrating — it just eases. The whole thesis rests on the frontier staying ahead of supply, which is true today and is an assumption, not a certainty.
Conclusion: Price the Bottleneck, Not the Sticker
For an investor, the deflation in headline compute cost is a head-fake. It is real, it is fast, and it is almost irrelevant to where the money gets made or trapped. The durable value sits with the inputs that stay inelastic and supply-constrained through the cycle: frontier accelerators, the power to run them, and the physical infrastructure — cooling, transmission, land with a grid connection — that neither efficiency nor a cheaper chip can conjure on demand.
So the number to watch is not the price of a token, which will keep falling toward zero. It is whether demand elasticity stays above one — whether the world keeps finding uses for cheaper intelligence faster than intelligence gets cheaper (Divyam.ai; Medium). As long as it does, every efficiency gain lands not as a saving but as an invitation to consume more, and the bill, the silicon shortage, and the power crunch all keep rising together. Compute gets cheaper. AI does not. The bottleneck just moves — and the returns move with it.
source
covers
Foundational unit-cost-decline curve
Independent, benchmark-anchored decline rate
Cross-source synthesis; spend and volume figures
Algorithmic efficiency ~3x/yr, isolated from hardware gains
IEA data-center electricity figures (both cite IEA directly)
DeepSeek sell-off; Nadella’s Jevons framing
Academic framing of the rebound thesis
Used-hardware price collapse
Frontier-tier price levels
Rental price floor; spend-vs-price divergence
Frontier pricing power; revenue/margin
Order backlog + booking pipeline
Forward-contract allocation, crowd-out
Power-density escalation
Two-speed market; bottleneck migration
Accessible mechanism explainer
Practitioner confirmation of spend-up dynami
Commentary-tier framing only, not primary evidence
