Why Local-AI Mini PC Prices Doubled in 2026 — And Whether to Buy Now or Wait

Quick answer: Local-AI mini PC prices roughly doubled between late 2025 and mid-2026. The GMKtec EVO-X2 (128GB Strix Halo) went from a pre-sale price of about $1,999 to roughly $3,399 on the street in July 2026, and NVIDIA raised the DGX Spark from $3,999 to $4,699 on 2026-02-27 — an 18% first-party hike — citing memory supply constraints. The cause is a DRAM shortage, not vendor opportunism: global DRAM contract prices rose roughly 180% in Q1 2026 versus Q4 2025. Waiting is the intuitive move and, for this category, the wrong one — TrendForce's 2026-07-30 outlook projects the DRAM supply-demand gap widening through 2027, with new fab capacity unlikely to land in volume before late 2027/2028. Verdict: if you truly need 128GB of unified memory, buy existing shelf stock now; if you don't, right-size down to a 32–64GB box and keep the four figures.

What actually happened to prices — the receipts

Every buyer's guide on the internet, including several of ours, was written against a price structure that no longer exists. So before any analysis, here is the damage, SKU by SKU, with dates attached. A price without a date is a liability in this market, and these are moving weekly.

Two of these rows are worth staring at. The DGX Spark row is first-party: this is not a scalper or a retailer padding a margin, it is NVIDIA — the company with more supply-chain leverage than anyone in this industry — publicly conceding it could not absorb its memory bill. And the EVO-X3 row is the tell for what comes next: a brand-new SKU, same memory, same storage, double the launch price. Nothing about the product changed. Only when its DRAM was bought.

The GMKtec EVO-X2 is the exhibit everyone recognizes, because it was the box the whole category was recommended on. If you read our EVO-X2 review or our best mini PC for local LLMs guide when they published, the hardware verdict in them still holds exactly — the silicon didn't change. The price did, and this post is the amendment.

Why memory, and only memory, is driving this

The market data is not subtle:

• Global DRAM contract prices rose roughly 180% in Q1 2026 versus Q4 2025.
• TrendForce projected a further ~63% rise in Q2 2026, on top of that Q1 jump.
• Consumer RAM is running 3–6× its 2024–25 lows.
• IDC expects average PC prices up to +8% in 2026 — and some vendors have resorted to shipping pre-builts without RAM installed.

Now the part specific to this niche, and the reason our category got hit roughly 2× while IDC's whole-market number is 8%. A unified-memory AI box is the most DRAM-exposed product in consumer computing.

On a normal mini PC, memory is a modest slice of the bill of materials — 16GB or 32GB of DDR5, maybe 10–15% of the price, and a vendor under cost pressure can quietly ship a 16GB SKU instead of 32GB. On a unified-memory box, memory is the product. The entire pitch of a Ryzen AI Max+ 395 machine is 128GB of LPDDR5X-8000 with up to 96GB allocatable to the GPU — that pool is the reason it can hold a 70B-class model that no consumer GPU can. Strip it to 64GB and you have not made a cheaper AI box; you have made something that no longer does the job people buy it for.

So three things compound:

1. The most expensive component is the one that spiked. 128GB of LPDDR5X-8000 is the single largest line item on these boards.
2. It's soldered. There's no configure-down option, no BYO-RAM escape hatch, no buying DIMMs later when prices normalize. The 128GB is bought at the factory, at whatever DRAM costs that week.
3. It's the premium grade. LPDDR5X-8000 (and -8533 on the HP Z2 Mini G1a) is not the commodity DDR5 tier; it competes for the same capacity that HBM and datacenter demand are consuming.

"NVIDIA raised the DGX Spark's official price from $3,999 to $4,699 on February 27, 2026, citing memory supply constraints — an 18% increase on a flagship product from the vendor with the most supply-chain leverage in the industry. When first-party pricing moves like that, the conclusion is not that vendors got greedy. It's that a 128GB unified-memory box is a DRAM product wearing a computer's chassis, and DRAM contract prices rose roughly 180% in a single quarter."

Will prices come down? The honest answer

This is where nearly every generic RAM-shortage article ends the same way — "prices are crazy right now, wait it out" — and where, for this category specifically, we think the primary sources say the opposite. Attribute the following to TrendForce, not to us: it is a forecast, and forecasts are wrong sometimes.

TrendForce, press release dated 2026-07-30 — three days before this post — is titled "Diverging Memory Market Outlook in 2027 as DRAM Supply Remains Tight While NAND Flash Supply Conditions Ease." The word doing the work is diverging. Their projection:

• The DRAM supply-demand gap widens further in 2027 — not narrows.
• NAND flash eases; DRAM does not. This matters, because it means the "memory prices are falling!" headlines you will see in 2027 may well be about the wrong kind of memory. Cheaper SSDs do nothing for a 128GB LPDDR5X box.
• New fab capacity is not expected in volume before late 2027 / 2028, with structural shortage running toward 2028.

Read plainly, that removes the thing that makes waiting a strategy: a defined end date. "Wait six months" is a plan. "Wait until an unspecified point after new fab capacity ramps, which is currently projected for late 2027 at the earliest, in a market whose supply gap is forecast to widen first" is not a plan — it's a hold with no exit condition, on hardware you wanted in order to do work.

We'll state the counter-case fairly, because it exists: forecasts have been wrong before, memory is a famously cyclical market, and a demand shock (an AI capex pullback, say) could loosen DRAM faster than anyone projects. If that happens, buying today looks like bad timing. That's the risk you're accepting. What you should not do is treat "prices will come down soon" as the safe default assumption — the most recent primary-source forecast in the market says the opposite, and it is three days old.

The one thing still working in your favor: contract-priced shelf stock

Here is the mechanism that makes "buy now" a rational call rather than a capitulation, and it is the part the generic explainers miss because they're written for DIY PC builders.

OEMs buy DRAM on forward contracts; you buy on spot. A mini PC vendor negotiates memory supply months to years ahead at contract prices. A person buying DIMMs on Amazon this afternoon pays today's spot price. That means the memory soldered into a box already sitting in a warehouse was very likely purchased at pre-spike contract pricing — you are, in effect, buying DRAM at last year's price with a computer wrapped around it.

This is a genuine 2026 inversion worth saying out loud: the pre-built unified-memory box is now the cheap side of the memory market, not the expensive side. For years the enthusiast reflex was "buy the barebones and add your own RAM." At 3–6× spot pricing, that reflex is now the expensive path.

And the evidence that this shield is expiring is exactly the two data points above:

• The EVO-X3, launched 2026-07-06 at $3,600 — same memory and storage as the EVO-X2, twice the launch price. New SKU, new contract, new price.
• Framework has publicly said its forthcoming 192GB Ryzen AI Max+ PRO 495 configuration will cost "much more" than the 128GB one, attributing it explicitly to LPDDR5X pricing.

Old inventory carries old memory costs. New SKUs carry new ones. That gap is the entire buying window, and it closes as stock turns over.

The Framework Desktop at $1,999 (catalog price, 2026-06-29 — verify at checkout) is the clearest surviving example. It is direct-only, so it never rode the Amazon price swings, and it is the cheapest credible 128GB box we track by a wide margin at that figure. Treat that price as perishable rather than permanent.

What we'd buy today, at today's prices

The thesis reshuffles the catalog. Below is the revised recommendation by route. Prices marked (catalog) are our 2026-06-29 seed figures and have not all been re-verified against this week's street pricing — check the live price before you commit. The two we have re-verified for this post are the EVO-X2 and the DGX Spark, and both moved.

Route 1 — You genuinely need 128GB

Buy shelf stock, and buy the cheapest credible one, because at these prices the chassis-and-I/O differences no longer justify big premiums. All Strix Halo boxes run identical silicon: same Ryzen AI Max+ 395, same ~256 GB/s theoretical (~215 GB/s real) bandwidth, same 96GB allocatable ceiling. You are choosing a case and an I/O panel, not performance — as our tokens-per-second deep-dive shows, peak throughput is a silicon fact, identical across all of them.

Note what the spike did to the ranking: the HP's premium used to be roughly $1,300 over an EVO-X2. Against a $3,399 EVO-X2 the gap is a few hundred dollars, and vPro plus a 3-year warranty starts to read like value rather than tax. Price moves change recommendations, not just budgets.

Route 2 — You need CUDA

The NVIDIA DGX Spark is now $4,699 officially, up from $3,999 (2026-02-27). The ASUS Ascent GX10 ($2,999–$4,100, catalog) is the same GB10 superchip, same 128GB, same 273 GB/s, same ConnectX-7 200GbE, same DGX OS — in an ASUS chassis with cheaper SSD tiers.

That gap was always there, but the repricing widened it in relative terms, and it is now the clearest arbitrage on this page: if you need the CUDA stack rather than the NVIDIA logo, price the GX10 first. Our DGX Spark vs Strix Halo comparison covers the platform fork itself; the only thing that changed is which side the money sits on. And be clear-eyed about what $4,699 buys: 273 GB/s of bandwidth, which is a capacity play and a software play, not a speed play.

Route 3 — Right-size down (the one most people should take)

Here's the genuine 2026 reversal, and it deserves to be said plainly: Apple's pricing held far better through the spike than the PC side did. For years Apple was the expensive unified-memory option and Strix Halo boxes were the value play. At a $3,399 EVO-X2, that has inverted for anyone who doesn't need the full 128GB.

• Mac Mini M4 Pro — $1,399–$1,599 (catalog). 64GB, 273 GB/s — more bandwidth than any Strix Halo box — silent, and comfortable for 30B-class models. The right-size hero.
• Mac Mini M4 — $499–$799 (catalog). 24GB / 120 GB/s; 7–14B models only, but the cheapest door into unified memory.
• Mac Studio M4 Max — $1,999–$5,999 (catalog). Up to 128GB at up to 546 GB/s, i.e. more real bandwidth than any Strix Halo or GB10 box.
• Mac Studio M3 Ultra — $3,999 (96GB). Still the bandwidth king at 819 GB/s, but the 256GB and 512GB configs were pulled during the shortage — the same crisis expressed as unavailability instead of price.

And below that, the tier the price spike quietly made the smartest buy in the catalog — because a $429 box's memory bill is a rounding error next to a 128GB one:

• Beelink SER8 — $449–$599 (catalog). Ryzen 7 8845HS, 32GB. The value pick for 13B inference.
• GMKtec M6 Ultra — $429–$549 (catalog). Zen 4, 32GB, real headroom for 13B and parallel agents.
• GMKtec M8 — $389–$459 (catalog). Dual 2.5GbE, 16GB — edge and cluster setups.
• MAGICNUC AS1 — $229–$299 (catalog). 16GB, Windows 11 Pro included. The cheapest always-on agent host we track.

The arithmetic is stark. The delta between a $3,399 EVO-X2 and a $449 SER8 is roughly $2,950 — and for a workload of 7–13B models and agents, both boxes do the job, one of them just holds models you'll never load. Before the spike that delta was about $1,550 and easy to hand-wave as future-proofing. At $2,950 it isn't.

Should you wait for the 192GB Ryzen AI Max+ 495?

Guardrail first: everything in this section concerns unreleased hardware. Treat every spec, price, and date below as "needs verification" — we are not asserting availability.

AMD's Gorgon Halo / Ryzen AI Max 400 series launched 2026-05-21, and a 192GB Ryzen AI Max+ PRO 495 configuration has been discussed for Q3 2026 — reported timing, not a commitment. Framework has already said publicly that its 192GB config will cost "much more" than the 128GB one, attributing it to LPDDR5X pricing.

Our read: this is more memory at a worse price-per-GB in the worst possible market. A 192GB config is a brand-new SKU, which by definition means brand-new memory purchased at post-spike contract prices — it is structurally the opposite of the shelf-stock advantage described above. You'd be waiting an unconfirmed amount of time to pay a premium, in a market where the vendor has pre-announced the premium.

The narrow case where waiting is right: you specifically need models that fit in 192GB but not 128GB, and nothing smaller will do. That is a real requirement for a small number of people. For everyone else, the extra 64GB is not worth an open-ended wait plus an explicitly pre-announced price increase. For background on the silicon family, see our Ryzen AI Max+ 395 explainer.

When buying is the wrong call entirely

This section costs us the flagship click, so take it as the honest one.

If you run 7–14B models, 128GB is dead money at any price — and at 2026 prices it's an expensive mistake. The reason the 128GB tier exists is capacity: it holds models no consumer GPU can. That is a real and valuable thing. It is also completely wasted if your actual workload is a coding assistant, a Home Assistant brain, a RAG backend over your notes, or a handful of always-on agents. Those run fine at 30–45 tok/s on a $429–$599 box.

The correct sequence — and the one the price spike now punishes people for skipping — is size the model first, then buy the box. Our guide to the best local LLM models for a 128GB mini PC is the place to do that. Run the models you care about on hardware you already own or rent, decide what you actually need resident in memory, and only then price a box.

Where 128GB does earn $3,399: large MoE models. GPT-OSS 120B runs at roughly 31 tok/s at ~120W on Strix Halo silicon — ServeTheHome's measurement on the Beelink GTR9 Pro, cross-checked by StorageReview running the same model on an HP Z2 Mini G1a with no discrete GPU. That's comfortably interactive, on a model no consumer GPU holds. If that is your workload, the box is justified at the new price. If it isn't, no price makes it a good buy. Our GPT-OSS 120B setup guide covers what that actually looks like in practice.

One last piece of the arithmetic: the purchase price is only half of what a box costs you. Our companion post on local-AI mini PC power consumption and 24/7 running cost covers the other half — roughly $15–$60/year in electricity depending on idle draw. Cost-to-buy plus cost-to-run is the number that should decide this, and the cost-to-buy side just moved a lot more than the cost-to-run side did.

Bottom line

• Prices roughly doubled, and the cause is DRAM: contract prices +~180% in Q1 2026, consumer RAM at 3–6× its 2024–25 lows. Unified-memory boxes are the most exposed product in consumer computing because memory is the product.
• Waiting has no defined exit. Per TrendForce (2026-07-30), the DRAM gap is projected to widen through 2027, NAND eases but DRAM doesn't, and new capacity isn't expected in volume before late 2027/2028. That's a forecast — but it's the most recent one, and it points the other way from the conventional advice.
• Shelf stock is the anomaly worth exploiting. Boxes already built carry pre-spike contract memory. The EVO-X3 at 2× and Framework's "much more" warning on 192GB are the proof that the window is closing.
• Need 128GB: Framework Desktop at $1,999 is the best buy on this page. Need CUDA: price the ASUS Ascent GX10 before the $4,699 DGX Spark. Don't need 128GB: the Mac Mini M4 Pro at 64GB / 273 GB/s, or a $429–$599 32GB box — and pocket the difference.

The uncomfortable summary is that the honest answer to "buy now or wait?" isn't really about timing at all. It's about whether you needed 128GB in the first place. If you did, the market is not going to hand you a better moment on any schedule anyone can name. If you didn't, the price spike did you a favor by making that obvious.