I posted two takes on Mastodon a few months apart that I want to stitch together. One was about LLMs. One was about RAM prices. They turn out to be the same story.
Twelve months of plateau
Hot take: It’s been 12 months since LLMs actually got more intelligent. We’ve only seen speed and application diversity increase since. We put the sum of all human knowledge into a super computer and it gave us this.
I wrote that in August 2025. I stand by it now, five months later, and the data has only firmed up.
Watch what’s been announced since mid-2024:
- Bigger context windows.
- Faster inference.
- Cheaper per-token pricing.
- More modalities — video, audio, longer code.
- Better tool use and agentic scaffolding.
Notice what isn’t on that list: the underlying model getting materially smarter at hard reasoning tasks. The benchmarks tell the same story when you read them honestly. Saturation on the easy ones, marginal gains on the hard ones, and a steady drumbeat of new benchmarks invented to make the curve look fresh.
The applications got dramatically better. The thing underneath didn’t.
”AI 2027 my ass”
That was my one-line reply to my own thread. The AI 2027 scenario — explosive capability gains, recursive self-improvement, AGI by decade’s end — is built on extrapolating the 2020–2024 curve. If that curve is bending, the scenario doesn’t survive contact with the new data.
I’m not saying capability is done. I am saying the trajectory has clearly changed and the industry’s spending plans haven’t caught up.
Which brings us to RAM
In January 2026 I wrote:
RAM prices exploded because a huge amount was bought with non-existent money for GPUs that have not been produced, for data centers that have not been built, needing infrastructure that may never appear, to satisfy demand that does not exist to attract investors who won’t go near this time bomb.
This is the bubble in one sentence. Let me unpack each clause.
Non-existent money. A lot of the AI infrastructure capex is circular. Chipmaker invests in cloud provider, cloud provider commits to buy chipmaker’s chips, chipmaker books revenue. Strip out the related-party transactions and the actual cash flowing in from external customers is a fraction of the headline numbers.
GPUs that have not been produced. Lead times on the latest accelerators are still measured in quarters. Companies are pre-paying for capacity slots they haven’t received and won’t receive for a year.
Data centers that have not been built. A modern AI data center takes 18-24 months from groundbreaking. The power has to come from somewhere. The cooling has to come from somewhere. Both are constrained in most regions where these are being built.
Infrastructure that may never appear. Transmission lines, sub- stations, water rights. The boring stuff that turns out to be the binding constraint. Permitting alone can kill a project.
Demand that does not exist. This is the load-bearing one. AI revenue at the application layer is real but small relative to the infrastructure being built. The bet is that demand scales 100x. The plateau says it might not.
Investors who won’t go near this time bomb. Notice how much of this is being funded with debt, vendor financing, and prepayments rather than fresh equity. Public-market investors have been pricing the risk in for a while, even as private valuations stay frothy.
The RAM part specifically
Why does this hit consumer RAM prices? Because HBM (the memory on the GPUs) and DDR5 (the memory in your laptop) share manufacturing capacity at Samsung, SK Hynix, and Micron. When the AI buyers swept the HBM pipeline, the fabs reallocated wafers. Consumer DDR5 supply dropped. Prices doubled. Servers running normal workloads — the ones that pay actual bills — got squeezed.
This is the part that hurts non-AI businesses. A small studio buying a workstation in 2024 paid one number. In 2026 they pay double. Their economics didn’t change. They just got conscripted into someone else’s capacity plan.
What I think happens next
I want to be careful here. I am not predicting a crash next week and I am not telling anyone to short anything. But the structure looks like every other bubble I’ve seen:
- A real technology with real value.
- Extrapolated into a future where it eats everything.
- Funded by capacity-first, demand-later commitments.
- Pricing pressure on adjacent markets that have nothing to do with the original thesis.
- A re-rating when the demand doesn’t materialize on the predicted schedule.
The technology stays. The pricing collapses. The capacity overhang ends up cheap for whoever wants to use it for something other than chasing AGI.
If you build software, the practical advice is: use the applications, don’t bet your business on the trajectory. The useful stuff that exists today will keep working. The “this will all get much better in 12 months” promise is the part I’d discount.