The Staged-Upgrade AI PC: A Reference Architecture

This is a reference architecture, not a tested build. Components are specified from published manufacturer documentation and are compatible on paper. Nobody at AI Gear Stack has assembled this machine or carried it through the upgrades it is designed to absorb.

The seven architectures before it all assume a single purchase. You decide what the machine is for, you buy it, you are finished. This one assumes the opposite and is specified around the constraint that follows: the machine has to absorb upgrades over a year or two without being rebuilt.

That is a constraint about time rather than about money, and it changes which components you are allowed to economise on.

The honest opening: this costs more

Buying in stages is more expensive in total than buying once. You pay for capability you will not use for a year, and you pay it before you have the thing that would use it.

What you get for that premium is optionality — the ability to start now, at a tier you can afford, without the first purchase becoming an obstacle to the second. That is worth real money if you genuinely cannot buy at once. It is worth nothing if you can, in which case the $2,000 machine is simply a better use of the same total.

Be honest with yourself about which of those you are before reading on.

You cannot future-proof compute

“Future-proof” is mostly a marketing word, and it is worth saying exactly where it fails before specifying anything.

You cannot buy readiness for things that change. Sockets are replaced. Memory standards turn over. Power connectors change — the industry moved from 8-pin to 12VHPWR to 12V-2×6 inside three years. Accelerator architectures gain numeric formats nobody offered before. No amount of headroom bought today makes a socket accept a CPU designed after it.

You can buy headroom in things that have not changed in a decade. Four of them, and they are the entire specification:

So the thesis of this architecture is narrow and defensible: you cannot future-proof compute, only the chassis around it. Spend on the four things above, economise on everything else, and accept that the accelerator is a consumable.

Buy the GPU last

This is the recommendation that surprises people, and it follows directly from what the rest of this site says.

VRAM capacity is what gates which models you can run — every guide here says so. It is also the component that depreciates fastest and improves fastest, and the one whose price moves most. A card bought twelve months early is twelve months of depreciation spent on capability you were not yet using.

Meanwhile the parts that do not improve much — the supply, the case, the board — are the ones whose replacement means disassembling the entire machine. Replacing a power supply is not an upgrade, it is a rebuild with an upgrade in it.

That inverts the usual order into something like this:

Stage 2 is a deliberate placeholder. A 16 GB card is enough to run 13–14B models, learn the tooling and discover what you actually need — which is a much better basis for the stage 4 decision than a specification sheet is.

The four headroom decisions

Each of these costs more now and saves a rebuild later. Each is specified against the endgame, not against stage 2.

The power supply

Specify for the largest accelerator you can imagine buying, not the one you are starting with. A 600 W card plus a system needs a very different unit from a 180 W card plus the same system, and a 750 W unit replaced in a year means you bought two supplies.

The old objection — that a large supply is inefficient at low load — is much weaker than it was. Modern 80 PLUS Gold and Platinum units hold high efficiency down to around 20% load, so a 1,000 W unit running a 200 W machine is not wasteful in any way you would notice on a bill.

Specify ATX 3.1 with native 12V-2×6, because that is the connector current high-power cards use and adapters are one more thing to seat imperfectly. And size for transients rather than averages, for the reasons the $2,000 machine sets out — GPU current spikes are what trip protection, not steady draw.

The case

Buy for the longest, thickest card in the range you might end up with, and check the clearance with front fans fitted, which is routinely 20–30 mm less than the headline figure.

Two specifics worth over-buying: eight expansion slots rather than seven, and genuine front-to-back airflow with fans actually included. Both are cheap now and impossible to add later without moving every component into a different box.

The motherboard

This is where the temptation to economise is strongest and the consequence is worst, because the board is the component whose replacement means removing everything.

Specify, even though you will not use them for a year:

• Documented ×8/×8 bifurcation, if a second card is conceivable. The dual-GPU machine explains why the manual is the only authority here and why the second full-length slot on a cheap board is not the same thing.
• Two M.2 slots, so the second drive is an afternoon rather than a migration.
• Four DIMM slots, populated with two.
• A socket with a stated multi-generation life, which is the one place a manufacturer’s roadmap is worth paying attention to — though treat it as a probability, not a promise.

The memory

Two modules, not four — which this site already recommends for a different reason, and the two reasons happen to agree.

The usual argument is speed: populating all four slots on a consumer board typically drops DDR5 well below its rated frequency. The staged-upgrade argument is capacity headroom: 2 × 32 GB now leaves two slots free, so 128 GB later is an addition rather than a replacement. Buying 4 × 16 GB is slightly cheaper today, slower immediately, and a dead end.

The 64 versus 128 GB comparison explains when the extra capacity actually matters — mostly fine-tuning, where offloaded optimiser state alone can be around 56 GB for a 7B model.

The upgrade that usually is not one

Adding a second graphics card is often worse than replacing the first, and this is the most common expensive mistake in a staged plan.

Two cards give you the sum of their memory. They give you the sum of their bandwidth only under tensor parallelism, which means only with specific serving software — under Ollama or LM Studio you get capacity at roughly single-card speed. They also bring slot spacing, bifurcation, connector count, mains circuit capacity and thermal stacking, all of which the dual-GPU machine works through and none of which is pleasant to discover mid-upgrade.

Plan for it at stage 1 or accept that you are not doing it. A board without documented bifurcation and a case with seven slots is a decision against a second card, made unknowingly, a year before it bites.

If the goal is simply more VRAM, replacing a 16 GB card with a 24 GB one is usually the better move: one card, one power connector, one thermal problem, and the old card has resale value that offsets some of the cost.

The component specification

Read that as a specification rather than a shopping list. The requirement — headroom in power, volume, lanes and slots — holds; a named product is one example that satisfies it.

Before you order: verification checklist

The unusual thing here is that you are verifying capability you will not use for a year, at the only moment it is cheap to check.

1. Confirm ×8/×8 bifurcation in the motherboard manual, not from the presence of two full-length slots. If a second card is conceivable, this is the check that preserves the option.
2. Count expansion slots in the case. Eight, not seven.
3. Measure case GPU clearance with front fans fitted, against the longest card you might buy, not the one you are buying.
4. Count PSU connectors for the endgame card, not the starter card. Confirm native cables rather than adapters.
5. Confirm two DIMM slots are left empty, and that the modules you bought are on the board’s QVL in a two-module configuration.
6. Confirm the second M.2 slot does not disable something else when populated. Boards frequently share lanes between M.2 and SATA, and it is always in a footnote.
7. Check the socket’s published roadmap before buying the CPU — and treat it as a likelihood rather than a commitment.
8. Write down the endgame — the card, the memory total, the number of drives — and keep it with the build notes. A year later you will not remember what the headroom was for.

When not to build this

• If you can buy it all now. Staged costs more in total. The premium buys optionality you do not need, and the $2,000 machine is the better use of the same money.
• If you already own a machine. Start by diagnosing it rather than planning a new one — the machine you already own works out whether the chassis you have can host an upgrade at all, which is the question this plan exists to make unnecessary next time.
• If you are not sure you will keep going. The headroom is only worth it if stage 4 actually happens. If this is an experiment, buy a cheap machine or rent, and decide later.
• If your endgame is a different machine entirely. A plan that ends in a 70B dual-GPU build, an always-on server or a fine-tuning workstation should be specified against that architecture from stage 1, not discovered at stage 4.
• If unified memory would suit you better. A Strix Halo or Apple silicon machine is not upgradeable at all — the memory is soldered and the accelerator is not replaceable. That is a real disadvantage here and it is worth naming, because for some readers the quiet, low-power, high-capacity trade still wins.