spot price
The AI buildout is not only financed by investors. Part of it is financed at the checkout, by people who never bought a GPU and never will.
An Echo Dot was $49.99. On Monday it became $79.99.
Amazon raised prices across its device line on 25 August — the Fire TV Stick 4K Max from $59.99 to $84.99, the base Kindle from $109.99 to $149.99, eero routers up as well — and attributed the increases to the cost of memory and storage. That is a sixty percent jump on the speaker. On a product category whose entire commercial logic was that the hardware is cheap enough to be an impulse buy, because the money is downstream.
The explanation is real. It is also the most interesting thing that happened in tech last week, and almost nobody covered it as a story about AI.
the number under the number
Memory has gone vertical. DRAM contract prices rose something like eighty to ninety percent in the first quarter of this year alone. On 7 August the mainstream DDR4 1Gx8 3200 chip printed $42.45 on the spot market — per the tracker Capital and Compute, reading DRAMeXchange, the highest that part has ever been logged since the series began in June 2016, and more than double where it sat in May.
DDR4. Not the exotic stuff. The commodity part, in a supposedly obsolete generation, at an all-time high, ten years into its own decline.
That inversion is the whole story. DDR4 is expensive because it is old. Samsung, SK hynix, and Micron have moved wafer capacity to high-bandwidth memory and DDR5, where the AI datacenter buyers are, and where the margins are. What is left making DDR4 in volume is Nanya and Winbond, at a scale set for a market that was supposed to be shrinking. Except the installed base did not shrink. Every desktop, every older server, every SSD controller, every router and set-top box and smart speaker still wants the part, and now there are three fewer serious suppliers making it.
IDC expects this to persist well into 2027.
this is an auction, not a tax
The instinct is to call it a tax, and that is the wrong shape. A tax is levied. This was bid.
There is one pool of wafer starts. Hyperscalers arrived with capital allocated at a scale that has no historical comparison and bid for the substrate. They did not outcompete the Echo Dot in the market for speakers. They outbid it in the market for silicon, which is upstream of the market for speakers, and the Echo Dot’s manufacturer discovered the result as a line item.
That is worth naming precisely, because the mechanism is common and the vocabulary for it is thin. Ivan Illich called it a radical monopoly: not one firm dominating a market, but one mode of doing things consuming the conditions that alternatives require. Cars do not beat walking by being better at walking. They beat it by eating the street — by making distances that assume a car, and then a built environment where the walk is no longer possible for anyone, driver or not. The monopoly is radical because it operates on the substrate rather than the product, and because you cannot opt out by declining to buy the thing.
Inference is doing this to memory. You cannot decline the price of DDR4 by not using AI. There is no version of the Kindle that routes around the auction. The commons being consumed here is fab capacity, and fab capacity is about as inelastic as a physical input gets — a new fab is three to five years and tens of billions, which means the supply curve for the next thirty-six months is essentially a vertical line, and every unit of demand added to a vertical supply curve is pure price.
who is actually paying
Follow the incidence.
The week of the price increases was also the week Nvidia reported. Second-quarter earnings of $2.22 a share against a $2.09 consensus, a robust forward number, the stock up 8.7 percent on the day and the Nasdaq up 1.6 percent behind it. The market read the print as confirmation that AI demand is real and durable, and it was right to.
Both things are the same event. The demand that made the earnings beat is the demand that made the speaker cost thirty dollars more, and the two facts were reported in different sections of the same newspapers on the same days by people who had no particular reason to connect them.
The distributional shape of that is not subtle. The buyer of a $49 smart speaker and the buyer of a $109 Kindle are not, as a population, the people holding the semiconductor complex in their brokerage accounts. The AI buildout’s cost is being partly socialized into the cheapest tier of consumer electronics, and the return is being capitalized in equities. Nobody designed this. That is characteristic of infrastructure transitions and does not make it less true — the incidence of a cost does not care about intent.
The category that gets hit worst is the one with the thinnest margin and the least ability to substitute. Which is precisely the category built on the promise that the device is nearly free.
the cheap-device model was always a loan
Here is the part I think matters longer than this cycle.
For fifteen years the consumer hardware bargain has been: sell the object at or below cost, recover it later through the service, the store, the subscription, the ads, the lock-in. That bargain has a hidden dependency, which is that component costs fall monotonically. The whole model is a bet on the cost curve. Sell at a loss today, and Moore’s Law plus commodity competition pays you back before the customer churns.
The cost curve just broke, in one direction, on one component, for a reason that has nothing to do with the products it prices. And the response was immediate and visible: Amazon did not eat it. A sixty percent increase means the subsidy is gone, and the subsidy going means the business model’s premise stopped holding.
If DDR4 stays where it is into 2027, the sub-$50 connected device does not come back at $50. It comes back with less memory, fewer features, more aggressive cloud offload — which costs inference, which costs memory, so that door is narrower than it looks — or it does not come back. The floor of the consumer electronics market gets raised by an industry that does not sell to that market and does not notice it exists.
There is a version of this in every commodity boom. Copper does it to housing. Shipping does it to everything. What is unusual here is the reflexivity: the thing consuming the memory is being sold to consumers as the reason to buy new devices, on devices whose price is rising because of it. You are being asked to upgrade to get the AI features, at a price set by the datacenters running the AI features.
Jevons would recognize the shape and would not be surprised by the outcome. Efficiency in the use of a resource does not reduce consumption of it; it makes the resource more useful and consumption climbs. Every improvement in inference cost per token has been immediately spent on more tokens. There is no equilibrium in that loop that arrives on its own.
The correction, when it comes, will not come from restraint. It comes from fabs, in about three years, or from an AI capex pullback, and I would not want to be holding a hardware roadmap that depends on which.
Meanwhile the speaker is eighty dollars.
Sources: 24/7 Wall St ↗ · Cryptopolitan ↗ · Capital and Compute DRAM tracker ↗ · TrendForce spot price insights ↗ · Yahoo Finance, market wrap ↗ · Illich, Tools for Conviviality ↗