The Soldered Memory Trap: Why 'Buy Less Now, Upgrade Later' Fails on Unified-Memory AI Boxes

Bottom Line (August 2026)

• All of them solder it. Mac (on-package), Ryzen AI Max+ 395 (soldered LPDDR5X, no SO-DIMM), DGX Spark (on-package LPDDR5X). Zero upgrade paths.
• It is a physics tradeoff, not just a margin grab. The 256-bit bus that delivers the bandwidth cannot run through a socket at these tolerances. The thing that makes the box good is the thing that locks it.
• The 2026 twist: the config you planned to upgrade to may be deleted. Apple removed the 64GB Mac mini M4 Pro, the base M4’s 32GB/64GB, and the Mac Studio’s 256GB and 512GB.
• So “buy small, upgrade the box later” has failed twice over — the memory is fixed, and the bigger machine you were saving for stopped existing.
• Buy for the model class you want in two years. On these platforms there is no second chance.

What “Soldered” Actually Costs You

On a normal desktop, RAM is a deferred decision. Buy 32GB, add 32GB more in a year. That flexibility is worth real money, and most people price it in without noticing.

Unified-memory AI boxes remove it entirely:

It is worth being fair about why. These platforms reach their bandwidth using a wide (256-bit) memory bus running LPDDR5X at very short trace lengths. A socketed DIMM arrangement at that bandwidth, in that thermal and physical envelope, is not something you can currently buy at any price. The soldering is what buys the bandwidth. A Strix Halo box with DIMM slots would be a slower Strix Halo box.

So the tradeoff is honest. The trap is not that vendors solder memory. The trap is how buyers respond to it.

The Trap: “I’ll Buy the Cheap One and Upgrade the Box Later”

Nearly everyone who gets caught knows the memory is soldered. Their plan accounts for it: buy the 24GB machine now, and in two years buy the 64GB machine.

That plan has a hidden dependency. It assumes the 64GB machine will still be for sale.

In 2026 that assumption broke, and this is the part almost nothing on the web has caught up with. During the memory shortage, vendors did not merely raise prices — they deleted configurations mid-generation:

Read that table as a buyer, not a spec-sheet reader. Someone who bought a 24GB Mac mini M4 Pro in early 2026 planning to step up to 64GB later cannot execute that plan. The rung they were climbing toward was removed from the ladder while they were saving for it.

The line nobody else prints: on a soldered-memory platform, your upgrade path is not a technical fact, it is a product-line fact — and product lines shrink. Every other buying guide treats the vendor’s config menu as a fixed backdrop. In 2026 it was the most volatile variable in the whole decision. Apple’s top two Mac Studio memory tiers vanished in ten weeks.

There is a second casualty here. A large volume of local-AI content was built on “run a 600B model on a 512GB Mac Studio.” That machine cannot be bought. If you are reading a guide that recommends it, the guide is describing a 2025 market.

How Much Should You Actually Buy?

Because there is no later, the sizing rule for unified memory is different from the one for a desktop. Buy for the model class you expect to want in about two years, not the one you run this week.

Rough guide for local LLM work, keeping in mind the OS and apps need several GB and you cannot allocate the whole pool to a model:

Bandwidth matters as much as capacity, and the two are not correlated across these platforms. A machine can hold a model and still decode it slowly. Capacity determines whether it runs; bandwidth determines how fast. Check both before buying — see what local LLM fits my machine.

The counter-argument, stated fairly

“Buy more than you need” is exactly what a vendor charging a premium for soldered memory wants you to hear, and this guide should not pretend otherwise. Memory upgrade tiers on these machines carry steep margins, and the 2026 shortage widened them further.

The honest version of the rule is narrower: buy up one tier from what you need today, not three. If the extra capacity serves only a hypothetical workload you cannot describe, that money is better spent on rented compute for the occasional big job. Buying a 128GB box to run a model you have never run is how people end up with an expensive machine running 8B models.

What to Buy in This Category Right Now

Two stock warnings, because in this category availability is a bigger risk than price. Strix Halo mini-PCs are chronically out of stock — list prices are real but frequently unbuyable, and marketplace stock trades well above them. And check the exact config before ordering any Mac: several previously standard tiers no longer exist, so a listing or guide describing them may be stale rather than a deal.

See Also

• Mac mini M4 Pro: 48GB or 64GB? — the config decision, now that 48GB is the ceiling
• Mac mini vs Mac Studio for Local LLMs — picking the chassis
• Strix Halo vs Mac Studio M4 Max 128GB — the two 128GB platforms compared
• Which Strix Halo Mini-PC to Buy — box-by-box, with stock notes
• Best Local LLMs for 128GB RAM — what the top tier actually runs
• What Local LLM Fits My Machine? — size before you buy
• Should I Wait for the RTX 60 Series? — the same timing question on the GPU side