Technical Context
I went carefully through the Unconventional AI article, and the thesis there is not decorative but very grounded: stop searching for one perfect memory for everything. For real AI implementation, this is significant because inference and training have long been limited not just by math, but by the cost of data movement.
The authors propose to "smooth the triangle" of speed, density, and data retention. Not to push every cell to near-perfect accuracy like in classical computing, but to accept that AI workloads tolerate more flexible trade-offs.
I would break down their idea into two distinct memories. The first, for model weights: here density and retention matter because weights are hardly ever rewritten during inference. The second, for working state and KV cache: you need very fast memory right next to compute, even if retention is shorter and the architecture must compensate.
Here's where it gets really interesting. As candidates they promote gain cells, eDRAM, PCM, and 3D-stacked HBM on top of logic. Not as "one winner," but as a set of technologies in a single stack, each covering its own class of data.
I especially liked the emphasis on locality. If reading from external memory truly eats a large share of accelerator energy, then the conversation is no longer about elegant circuitry, but about the cost per token. And yes, the idea of keeping most of the model and state on-die or as close as possible looks not like a fantasy, but the next necessary step.
The article is fresh, July 2026, so this is not a retrospective but a very current signal toward the next generation of AI infrastructure.
What This Changes for Business and Automation
I see three direct consequences here. First: the winners will be those building inference-first hardware and services where cost per response matters, not just peak benchmark. Second: HBM is no longer the only sacred answer, which will open a market for cheaper and more specialized configurations.
Losers will be those who continue to design AI architecture with the old logic: "everything universal, everything maximally reliable, then we'll deal with energy." With that approach, automation with AI will quickly hit economic walls, especially on long context and large volumes of requests.
I constantly see the same problem with clients: everyone talks about the model, but underrates memory, networking, and the cost of each pipeline step. And that's exactly where it's often decided whether AI automation will fly in production or remain an expensive demo.
If you're already calculating how to make inference cheaper, package a private environment, or design AI integration without excessive hardware overhead, this is precisely the moment to rebuild the architecture from scratch. At Nahornyi AI Lab, we help with exactly that: we don't draw fancy diagrams for their own sake, but build working AI solutions for business under real constraints of cost, speed, and scale.