The previous article: AI’s current state saves “the labor cost per unit of software.” The saved labor cost goes into company profit. But AI’s saving is not from nothing. It needs compute, energy, memory, and electricity. These physical costs do not vanish; they move.

This article asks: where does the cost AI saves go? This question is worth answering because the “AI saves money” narrative only talks about the saving side, not where the saved cost goes. Where it goes decides who actually pays.

The main answer: the cost moves to the price of shared memory components, then to consumer electronics — phones 13% more expensive and PCs 17% more expensive in 2026. This is a regressive transfer: those who don’t get AI’s cognitive benefit pay most for AI’s side effect.

1. Question: where does AI’s saving go

The conclusion of the previous article: AI’s current state saves “the labor cost per unit of software.” The saved labor cost goes into company profit, not into the engineer’s pocket.

But for AI to run, it needs compute, energy, memory, and electricity. These are physical resources with cost. The “labor cost” AI saves is exchanged for “compute/energy cost.” The cost moves; it does not vanish.

The scale of this cost depends on the speed of AI infrastructure expansion. The faster and more AI data centers are built, the more shared resources they take, and the more cost moves to consumer products. This cost is proportional to AI’s “saving”: the more AI saves, the larger the infrastructure, and the more cost moves to consumer products. This article asks: where does this cost go? Who pays?

2. Evaluation framework: cost relocation depends on whether the resource is shared or dedicated

The path of cost relocation depends on whether the resource AI uses is “shared” or “dedicated.”

Dedicated resources (AI-specific chips, data center land): the cost stays within the AI industry and does not spread to other industries.

Shared resources (memory, electricity): AI and consumer electronics use the same batch. AI takes the supply, the price rises, and it spreads to all industries that use the resource.

Memory is a shared resource, and the most critical one. The DRAM/NAND that AI data centers use is the same batch that phones and laptops use. More critically: the three memory makers’ HBM (AI-specific high-bandwidth memory) and conventional DRAM use the same wafer capacity. HBM is a high-price, high-margin product; when the makers use the same wafer capacity to make HBM, they make less of the cheaper conventional DRAM. So the stronger AI demand is, the more conventional DRAM supply contracts, and the more the price rises.

This explains why AI’s growth pushes up consumer prices — not because AI directly raises prices, but because AI takes shared resources. AI demand has no substitute, so memory makers can raise prices without losing volume; this round they have pricing power.

So: AI’s cost moves to consumer electronics depending on the supply and demand of shared resources like memory. Electricity is also a shared resource, but its path to consumer prices is longer; this article only handles memory.

3. Comparison: 2025–2026 memory price increases and consumer product pass-through

From 2025, the three memory makers (Samsung, SK Hynix, Micron) shifted capacity to make AI-specific high-price HBM, so the supply of conventional DRAM/NAND contracted and prices surged. The mechanism is a chain: AI is useful → companies invest in AI servers → servers need massive memory → memory makers shift to HBM → conventional DRAM/NAND supply contracts → prices rise → phone/laptop BOM gets more expensive → sale price rises. The table lists the key indicators.

MetricValueSource
AI data centers’ share of global memory70%tech-insider[3]
DRAM price 2024 start → 2026 end+400%J.P. Morgan[1]
Memory price 2026 end (further rise)+130%Gartner[2]
DDR5 contract price within 2025$7 → $19.5 (2.8×)tech-insider[3]
A high-speed memory segment since early 2025+700–800%CSIS[4]
DRAM inventoryabout 8 weeksIntuitionLabs[5]
2026 PC price (vs 2025)+17%Gartner[2]
2026 phone price (vs 2025)+13%Gartner[2]
2026 DRAM/NAND supply growth16%/17% YoYIDC[6]

Three observations. First, supply is taken: AI data centers consume 70% of memory, so conventional DRAM/NAND supply contracts. Second, prices surge: DRAM rises 400% in a year, a high-speed segment rises 700–800%, inventory drops to 8 weeks. Third, it passes to consumer products: in 2026, PCs are 17% more expensive and phones 13% more expensive.

SK Hynix warns the shortage may last beyond 2030[3]. IDC estimates 2026 DRAM/NAND supply growth is only 16%/17% YoY, well below history[6]. Supply cannot keep up with demand, so the price increase is structural, not a short-term fluctuation. Supply growth is well below demand growth, so the gap will keep widening.

So “consumer products get more expensive” is not a deduction; it is the price happening in 2026. A phone that sold for $15000 in 2025 is 13% more expensive in 2026, i.e. $16950, $1950 more. Part of the phone or laptop you buy this year is paying for AI data centers taking memory.

4. Judgment: who can afford it (regressive transfer)

The cost lands on “those who can least afford it and least enjoy AI’s cognitive benefit.”

Concretely: a person crushed by rent and food is not currently crushed by mental load. The AI relief of “thinking faster” is unrelated to him; he does not need AI to write code or reports. But he must buy a phone and a laptop, and these get more expensive because AI data centers take memory. In 2026 his new phone is 13% more expensive; part of that increase is the result of AI data centers taking memory. He does not get AI’s cognitive benefit, yet he pays for AI’s physical substrate. For him, AI is not a money-saving tool; it is a price-increase cause.

So AI’s cost is a regressive transfer: the benefit moves up (companies and engineers who use AI), the cost spreads down (all consumers of electronics, including those who can least afford it). This is the same structure as the previous article’s “saved labor cost goes into company profit”: cost socialized, benefit concentrated.

This is not an isolated case. Everyone who buys consumer electronics pays this cost, just in different amounts.

Who can afford it? Those who don’t get AI’s benefit pay most for AI’s side effect.

Applicable conditions: supply-demand imbalance in shared resources (memory).

Limitations:

  • The pass-through magnitude is disputed. Memory is part of a phone’s BOM, not a 1:1 pass-through to price. Gartner’s +13% is a forecast, not a measurement.
  • It may reverse long-term. If AI makes compute/energy cheaper, memory demand may fall. But short-term (2026), the price increase is structural.

Ending

Back to the question: where does the cost AI saves go? The answer: it moves to the price of shared memory components, then to consumer electronics (2026 phones +13%, PCs +17%). This is a regressive transfer.

This cost is not an “externality” of AI; it is the direct price of AI’s growth — the more AI infrastructure expands, the higher the memory cost of consumer products. For consumers, this cost is hidden: you will not see an “AI memory surcharge” on the invoice, but it is embedded in the price.

Applicable conditions: supply-demand imbalance in shared resources (memory). Limitations: the pass-through magnitude is disputed; it may reverse long-term.

Next article: who collects this price increase — the memory oligopoly, and whether the whole AI investment loop can sustain.