I like knowing why. Not in a useful way — in a compulsive one. When a price moves I want the chain that moved it, all the way back, and I am not satisfied by "because AI."
So I drew it. Every link below is a real, checkable step between a chatbot most people tried in 2023 and a number you can see on Amazon this afternoon.
OneThe Two-Year Delay
The link was never a secret. The size of it was.
- The public AI moment was late 2023. Consumer memory prices did not move meaningfully until late 2025.
- Fabs take years. The decision to reallocate a production line shows up on a shelf long after it is made.
- By the time you feel it, the news cycle that caused it is three years stale and nobody is writing about it any more.
The same lag runs in reverse. When AI demand does soften, the relief will also take two to three years to reach retail — so the first "AI bubble popped" headline will not make your next upgrade cheaper.
Prediction logged 22 August 2026
TwoThe Cascade
Demand doesn't stop at the thing it wanted. It spills.
- It is all the same silicon capacity. Fabs now focus on making high-bandwidth memory, not memory for your PC.
- Substitution is what spreads it. Priced out of NVMe, people buy SATA — so SATA rises too. The escape route is the transmission mechanism.
- In February 2026 Micron retired the Crucial brand to concentrate on datacenter and HBM. Not a price rise — an exit.
Used and refurbished already stopped being an escape — that happened months ago. The retreat that's left isn't a cheaper part, it's no part at all: people simply stop upgrading.
So the thing to watch is manufacturers making that permanent. Expect soldered, non-upgradeable memory to spread from laptops into desktops and prebuilts — because if the amount of RAM is fixed at purchase, the maker buys it on a contract and you never get a second decision.
Prediction logged 22 August 2026
ThreeSame Money, Fifteen Months
This is the part where it stops being an industry story.
I bought two 2×32GB DDR5-6000 desktop kits in May 2025 — lucky me — for $220 plus tax each. That's 128GB for $440. About $170 US a kit at the time.
I did not time anything. I just happened to need memory before the wave arrived. The same 128GB today is $3,300.
Or put it the way that actually stings: $220 bought me 64GB last year. Today $219.99 buys one 8GB stick.
Same Money
Two stacks of DIMMs side by side, both wearing an identical $220 price tag. May 2025 shows four sticks. August 2026 shows one. No axis, no index, no conversion — just the same banknote buying one-eighth as much.
- My exact brand and capacity today: Corsair Vengeance DDR5 64GB (2×32GB) 6000MHz — $1,649.99, and the listing says only 8 left in stock. That is 7.5× what I paid. (Different model within the same 6000 line.)
- Not a cherry-pick: a Silicon Power 64GB 6000MT/s kit on the same page is $2,755.
- Amazon.ca's own price filter for "ddr5 6000" now runs $160 to $12,600+.
- The Crucial kits still listed read only 1 left in stock and only 2 left. You are watching the last consumer stock drain out in real time.
Here's the one that would genuinely surprise people: buying a finished computer becomes cheaper than building one.
Thirty years of received wisdom says building your own saves money. But the big makers buy memory on contract and you buy it at retail — and that gap is now wider than their whole assembly margin. If a prebuilt with 64GB undercuts the cost of the parts alone, the DIY logic every enthusiast grew up with quietly inverts.
Prediction logged 22 August 2026
FourWhy You Got Hit Hardest
Wholesale roughly doubled. You paid five to seven times.
Contract vs. Retail
Two lines from the same starting point. The contract line — what the big buyers pay — roughly doubles. The retail line goes near-vertical. The widening gap between them is labelled: this is you.
- Big buyers sign contracts months ahead at agreed prices. You buy whatever is left.
- Retail is the residual market — the leftover after allocation — and residual markets amplify every wobble in the main one.
- This is why "DRAM prices rose 80%" and "my RAM cost 7× more" are both true and not a contradiction.
Micron killing Crucial was not a one-off, it was the template. A second major brand exits consumer memory or storage before the end of 2027 — not with an announcement, just a product line that stops being restocked.
The gap closes from the wrong direction: not retail coming down, but the consumer tier quietly ceasing to be a business anyone wants.
Prediction logged 22 August 2026
FiveAnd The Thing It Was All For Hasn't Happened Yet
Human nature turned out to be the bottleneck, not compute.
Desire for AI-Based Productivity Improvement
Two overlaid normal curves. The upper one a thin dashed blue line — everybody, including the people who will never need this. The lower solid curve, the one we care about: the tech-curious, tech-first and tech-forced. X-axis: successful AI agent adoption. Y-axis: how many people. Four shaded bands beneath, and dark drops falling on the middle band — the failure stories, visibly pushing the peak to the right.
- 88% of enterprise agent pilots never ship. 80% of enterprise apps have an agent embedded; only 31% run one in production.
- McKinsey: roughly two-thirds of enterprises have experimented, fewer than 10% have scaled to real value — which is almost exactly the top band of the curve.
- Gartner forecasts over 40% of agentic projects cancelled by 2027, on unclear returns and weak controls.
- Story poisoning is the delay mechanism. One well-publicised agentic disaster gives permission to everyone who already disliked learning curves — and that permission lasts months.
Adoption will not be unlocked by a better model. It gets unlocked when ordinary people personally know someone who uses an agent daily — a social threshold, not a technical one. My guess: the second band starts moving properly in 2027, and the fat middle not before 2028.
Prediction logged 22 August 2026
SixHow It Started
The short version, at the scale it actually happened.
The Ignition Chain
Five linked blocks, left to right, each one's height scaled to its real dollar magnitude — so the escalation is the picture, not decoration. The 2023 moment is a sliver. What it set off towers over it.
- Nvidia was the first company ever worth $5 trillion, and set a record $5.5 trillion on 13 May 2026 — above Google and Apple.
- Hundreds of billions in datacenter commitments followed, and those commitments are contracts, not intentions.
There will be a quarter — and I think it lands in 2027 — where Nvidia reports record revenue and a slower growth rate at the same time. Both true, same press release.
The tell won't be the number. It'll be which one the market reacts to. If a record quarter sells off, that's the moment the story changed, and everything downstream in this page moves with it.
Prediction logged 22 August 2026
SevenThe Tug-of-War Nobody Wins
Fear of losing the AI race, against the plain wish to sell chips to everyone.
The Tug-of-War
Two ropes pulling opposite ways — export controls and national-security anxiety on one side, the commercial urge to sell on the other. The centre mark does not move. Whoever wins, the same company gets paid.
- Every serious attempt at a cheaper or alternative accelerator has so far left the incumbent's position intact.
- Restrict exports and you protect a lead while shrinking your own market. Sell freely and you fund a rival. There is no version where someone simply wins.
The dent, when it comes, will not come from a faster chip. It comes from somebody needing less compute for the same result — efficiency, not horsepower. That is the change worth watching, and it will be announced quietly.
Prediction logged 22 August 2026
EightThe Loop That Closes
The chain bites its own tail, and that is the part I find genuinely unsettling.
The Loop
A ring rather than a line, using the same four hardware pieces. Consumer hardware gets expensive, so people stop upgrading and rent AI in the cloud instead, so datacenter demand rises, so consumer hardware gets more expensive. The last link joins back to the first.
- Priced out of upgrading, the sensible move is to rent compute instead of owning it.
- Renting compute is demand for exactly the datacenters that took the parts.
- Every person who gives up on their own hardware makes the next person's hardware dearer.
This is the one I would most like to be wrong about. If it holds, the personal computer quietly stops being where computing happens for ordinary people — not by a decision anyone announced, but because owning became the expensive option. Check me on this in 2028.
Prediction logged 22 August 2026
What I actually think
None of this required a conspiracy, and none of it required anyone to behave badly. A lot of money found a use for a limited amount of factory capacity, and everything downstream of that capacity moved. The reason it feels mysterious is only the delay — effects arriving years after causes, in a different part of your life, with no label on them.
That is what I wanted the drawings for. Not to be angry about the price of memory. Just to be able to point at the chain and say: that, then that, then this.
Sources and how firm each one is
- My own receipt and a live Amazon.ca capture (22 Aug 2026) — the $220 purchase, the $1,649.99 Corsair 64GB kit, the $219.99 8GB stick, the $160–$12,600+ filter range, the Crucial stock levels. First-party, dated, re-checkable.
- Memory and storage price moves — Tom's Hardware RAM price index and reporting; RamRadar retail index (verified live, series from 21 Oct 2025). Percentages are from published tracking, not from my own basket.
- Nvidia valuation milestones — contemporaneous financial press, May 2026.
- Agent adoption figures — Gartner, IDC and McKinsey 2026 summaries. Enterprise surveys, so they describe large organisations, not individuals.
- Everything in an ochre box is a prediction, dated so it can be graded later. Predictions are not findings and are not sourced, because they haven't happened.