9For two years the dominant story has been that AI hardware is racing toward free: more capability per chip, smaller boxes, falling prices. That story is true as far as it goes. The problem is that it stops right before the part that actually decides your firm’s budget. The component that gates how much capable AI you can run in private is not the processor. It is the memory. And memory is getting more expensive, fast.
Every professional who deferred the sovereign transition told themselves some version of the same thing: “I’ll wait; the hardware will only get cheaper.” That assumption has quietly inverted. The processors fell in price, but the memory required to hold a capable model — and a deep working context on top of it — climbed sharply in the space of a year. A standard system with 64 GB of memory rose from well under two hundred dollars’ worth of memory many months ago into the four-figure range today. A workstation-class accelerator with a large unified memory store has jumped in the same window. This is not a rounding error. It is a supply squeeze.
Here is the structural reason, stated plainly: the chips that do the math are now abundant, but the memory that feeds them is not. A small handful of manufacturers worldwide produce the high-bandwidth memory that serious local models need, and nearly every AI builder is competing for the same limited output. When a scarce component sits at the leverage point of an entire industry, its price stops tracking the curve of everything else on the desk. That is exactly what is happening now.
What this means for a firm in practice:
1. Capability is gated by memory, not by raw speed. A model that fits in resident memory runs fluidly; one that doesn’t must be swapped to far slower storage, where it stalls. The practical ceiling on what a firm can run privately today is a memory question — how much can sit resident while several professionals work at once — not a question of peak processor power. Buying more raw compute while skimping on memory buys very little.
2. The price of admission is rising, not falling. The honest forecast for the memory-hungry part of the stack is not “wait and it gets cheaper.” It is the opposite. Deferral now carries a literal dollar cost: the same sovereign system will cost more later. The polite “let’s look again next year” is no longer free.
3. Locking in capacity is the new form of control. For a firm that needs private, on-premise capability, memory is its reserve. Acquiring it while it is still affordable — for the full engagement, not just a pilot — converts a commodity into an asset the firm owns outright, insulated from any future price revision set by a supply chain it does not control.
The strategic upshot is a timing argument for sovereignty. The privacy case for keeping client data in-house was always strong. What is new is the economics: the window to build that capability at today’s pricing is real, and it narrows. Firms that commit now buy control and a hedge in a single move. Firms that wait will pay a growing premium for the same outcome — while continuing to route privileged work through systems that transfer their terms by default.
The reputable, defensible posture is the one that owns its capacity early and keeps the client’s privilege inside the firm’s own walls, where it has always belonged.
The first step is not a hardware order — it is knowing your exposure before you commit a dollar. Run the Confidential AI Exposure Checklist at 9tailsadvisory.com/checklist, a five-minute, no-obligation read that shows exactly where your firm stands, so you size the investment from evidence rather than from a vendor’s pitch.