top of page
Search

Is RAM the only bubble that can burst in the AI market?

Writer: Erwin SOTIRI
Erwin SOTIRI
21 hours ago
7 min read

Updated: 2 hours ago

Cartoon advertisement: a single computer memory stick displayed like crown jewels on a velvet cushion in a guarded glass showcase, with a certificate and a hanging price tag, admired by two well-dressed customers through the glass. Headline: The guards are new! Caption: Make your AI memorable. Ask about our payment plans.

Is the RAM price the only bubble that could burst in the emerging market of AI adoption?

A Jurisconsul analysis, 11 November 2026. Erwin Sotiri, Founding Partner. General information, not investment or legal advice.

Every discussion of the AI economy eventually runs into the same arithmetic. Training a frontier model costs a fortune; running one costs less than the subscription price you pay for it. But the second half of that sentence is now doing less work than it used to. Local inference, the kind that runs on a machine in your office, used to be a hobbyist's argument. In 2026 it has become a pricing argument, and the reason is not software. It is the price of memory.

1. The number that broke the model

Contract prices for a 16-gigabit DDR5 chip, the standard unit of desktop memory, went from $8.70 in October 2025 to $48 in September 2026. That is not a rounding error or a bad quarter. That is roughly a fivefold increase in twelve months, and it is the single most useful number in this debate.

It does not stop at the chip. A 32GB DDR5 kit that cost $95 to $250 in mid-2025 now runs $550 to $650. Server RDIMM spot prices broke $2,000. Counterpoint Research found that memory now represents 54% of the bill of materials for a low-end smartphone, up from 20% a year earlier. Intel's chief executive, speaking in September 2026, put memory at 70–80% of the cost of budget smartphones and laptops, with prices up five- to sevenfold depending on the category. Gartner projects PC prices up 17% and smartphone prices up 13% by the end of 2026, and expects PC shipments to fall 10.4%.

This is what a memory supercycle looks like from the inside. It is also, and this is the point, a tax that lands hardest on the devices that ordinary people and small firms actually buy.

2. The parity squeeze

The familiar argument is that AI generation costs much more than subscription prices, so the subscription model is where the value sits. That argument was built on data-centre economics: tokens are cheap because the hardware is amortised across millions of users, and the marginal cost of one more request is close to zero.

Local inference was supposed to lose that fight, and for a while it did. A 2026 total-cost-of-ownership study put a single RTX 4090 running an 8B model at roughly $58.60 a month — about $50 in amortised hardware plus electricity — against a budget API doing the same volume for around $6. The break-even only arrived at very high, steady volume, or if you already owned the machine and it would otherwise sit idle.

Then the hardware got expensive. The mid-range and high-memory machines that make local inference practical moved with the memory market, not with silicon progress. A 128GB unified-memory mini PC — the category that finally made running a useful model at home credible — launched at $1,999 for one configuration in mid-2025 and was listed at $3,449 by late September 2026. The same specification roughly doubled in about a year, and the reason given by the vendors, consistently, is LPDDR5X cost.

So the parity the subscription model relies on is now partly an artefact of a supply shock. If that shock reversed, the arithmetic would move.

3. The RAMageddon side effect

The most interesting observation in this whole story is not about AI at all. It is that Apple, a company that has spent forty years defending premium pricing, has spent the last year being described as the value option for local AI.

That is a strange place for Apple to be, and it is not a strategy. Apple's unified memory architecture — one pool shared by CPU and GPU — happens to be exactly the shape that large language models want. When the Mac mini was discovered to run agentic AI locally, it sold out for months. Demand was not the cause of the shortage: the shortage came from AI data centres absorbing every chip in the supply chain, and Apple itself had trouble getting the memory it needed. But the consequence is real. Apple raised the Mac mini's entry price from $599 to $799 in June 2026, then to $899 with the new generation, and removed higher memory ceilings earlier in the year as costs climbed.

The result is a genuinely odd market position: Apple as the practical, near-affordable way to run a model locally, while the PC side is defined by graphics cards whose street prices have detached from their labels. The RTX 5090 still carries a $1,999 launch price, but in the United States the cheapest in-stock model was reported at $9,400 in October 2026 — about 370% above that figure. Nvidia has repriced the GPU-and-memory kit it sells to board partners three times this year, by a compounded 32% to 49%, while leaving the official prices untouched. Memory is not the only cause of that gap, but it is the headline cause.

4. The chicken-and-egg problem

Which brings us to the harder question in the opening premise. Is the cost of inference genuinely high, or is inference pushing its own price up?

The honest answer is that both directions are visible in the same data. The costs are real: AI infrastructure spending reached $121.4 billion in the second quarter of 2026, up 53% year over year, and IDC raised its full-year figure to $521 billion. Deloitte expects hyperscaler capital spending to exceed $1 trillion in 2026, with roughly 30% of that going to memory, rising to 36% in 2027. J.P. Morgan projects global AI capital expenditure approaching $1 trillion in 2026, and around $5.5 trillion by 2030. Those are committed sums, and they have to be recovered.

At the same time, the concentration is doing some of the pricing. The three memory makers have redirected capacity towards high-bandwidth memory and enterprise DRAM, because the margins are several times better than a consumer module's. AI server demand is met through long-term agreements: SK Hynix has signed agreements covering more than half of its 2027–28 capacity, Micron says more than 75% of its 2027 output is committed, and Samsung's next-year supply is sold out. The effect is that buyers without leverage — the small studio, the regional accounting firm, the laptop buyer — end up paying spot-market rates that ran 30% to 50% above contract levels.

So the frontier labs do dominate the scarce hardware, and the scarcity is part of what makes their position durable. Revenue follows the compute; compute follows the contracts. Whether that dominance would survive universal access to 256–512GB of private memory is the interesting counterfactual, and it deserves a straight answer.

5. Would universal 256–512GB change the picture?

Partly, and less than the premise implies.

What it would change. Price is the binding constraint on local inference, not capability. If a capable box were available at €1,000–1,500, every firm with a genuine confidentiality reason to keep client files off someone else's servers — law firms, doctors, accountants, HR departments — would have a credible local option. That is not a niche. It is precisely the market that regulation, professional secrecy and the GDPR keep pushing towards private deployment. At €1,000, local inference stops being a hobby and starts being a procurement decision.

What it would not change. Fixing memory prices does not move frontier capability into the room. A 128GB box runs a good open model; it does not run the frontier model you are subscribing to. The largest models will still live in data centres, because the training runs and the inference clusters require capital and power no single customer can assemble. Universal access to large private memory would fragment the inference market, not abolish the training market.

What is actually happening. The hardware is arriving, slowly and at the wrong price. AMD's unified-memory mini PCs ship in 128GB configurations, but at $3,449 to $4,999 rather than $1,000. The next generation, expected around 2027–2028, is reported to bring a memory pool capable of 256GB, though AMD has confirmed nothing at SKU level. Whether that arrives at a price that matters depends on the same variable as everything else here.

6. So is RAM the only bubble?

No, and it is worth being precise about what kind of bubble this is.

Memory prices are cyclical by nature. Deloitte notes that memory makers are cautious about overbuilding, that fabrication capacity takes three to five years to come online, and that the price surge mostly reflects higher selling prices rather than more bits shipped. Every forecast in this article points the same way: tightness through 2027, with meaningful relief no earlier than late 2027 or 2028. A fivefold price rise that reverses is not a bubble bursting; it is a cycle turning, and it will take several quarters of consumer pain with it.

The genuine bubble risk sits one level up, in the finance. Analysts have raised questions about the useful lives hyperscalers assign to GPU assets — commonly five to six years against a two-to-three-year product cycle — and about circular financing, where an investor funds the customer that buys its product. The asset class most exposed is not memory but the depreciation schedule and the debt behind the build-out. Forrester's November 2026 assessment is that a burst is not imminent, largely because most of the money comes from healthy cash flows rather than leverage, while warning that a correction could upend suppliers, contracts and prices.

For a business deciding how to buy AI capacity, the practical conclusion is narrow and useful. Do not treat today's hardware prices as the long-run cost of local inference, because they are not — they are the price of a shortage. Do not assume cloud prices stay still either, since they inherit the same memory market. And read the depreciation assumptions in any vendor's financials before you sign a three-year contract, because that is where the money and the risk actually sit.

The RAM price is the bubble most people can see. It is not the only one that can burst.

Sources

 
 
bottom of page