Unified-Memory AI PCs Are Suddenly 75% More Expensive: What the EVO-X2 and RTX PRO 6000 Price Surge Means for Home Lab Buyers (Mid-2026)

By RunAIHome Team · 14 min read
hardware-pricesevo-x2strix-halortx-pro-6000rtx-3090ddr5buying-guidehome-lab2026

TL;DR: Memory-heavy AI hardware repriced hard in July 2026: the GMKtec EVO-X2 128GB that launched at $1,999 now debuts at $3,499, and NVIDIA’s RTX PRO 6000 went from $8,565 to $13,250. The more memory a machine carries, the harder it got hit. Used cards with old GDDR6X are now the least-bad value left.

Used RTX 3090 24GBEVO-X2 128GBRTX PRO 6000 96GB
Winter 2025/26 price~$966 (winter low avg)$1,999–$2,199$8,565 MSRP
July 2026 price$1,050 eBay floor / $1,252 avg$3,499–$3,649$12,099 Newegg / $13,250 NVIDIA
Changeup ~30%up ~75%up 41–55%
Cost per GB of memory~$52~$27~$138

Honest take: If you need 24GB of fast VRAM, buy a used RTX 3090 now — it’s up 24% and still the cheapest 936 GB/s on the market. If you were waiting to buy an EVO-X2 at the $1,999 we reviewed it at in June, that machine no longer exists; rent a RunPod A100 until LPDDR5X pricing stabilizes.

Five weeks ago, our EVO-X2 review called it “a sub-$2,000 mini PC that runs 235B models.” That headline is now wrong through no fault of the machine. On July 18, VideoCardz reported GMKtec’s newest 128GB configuration debuting at $3,499 — positioned as the cheap option, $150 below the existing 128GB/2TB model that now sells for $3,649. The exact same silicon, the same 256 GB/s of LPDDR5X-8000, at 59–75% above the $1,999–$2,199 street price we quoted four weeks earlier.

This isn’t a GMKtec story. It’s a memory story, and it has repriced every tier of home-lab AI hardware at once. Here’s the full damage report as of July 31, 2026, what’s driving it, and the updated buying table.

The July receipts, machine by machine

GMKtec EVO-X2: $1,999 → $3,499–$3,649

The EVO-X2 launched in May 2025 at $1,799 pre-order / $1,999 retail for the 128GB + 2TB configuration. When we reviewed it on June 23, 2026, street pricing was $1,999–$2,199 and it was the clear capacity-per-dollar champion: 128GB of unified memory that runs Qwen3-235B-A22B at ~11 tok/s, in a box that idles at 8–14W.

As of mid-July, the 128GB/2TB model lists at $3,649, and the “news” was GMKtec introducing a $3,499 128GB variant to serve as the budget entry. The 64GB/1TB configuration holds at $1,999.99 — which tells you exactly where the cost lives: the second 64GB of LPDDR5X now effectively costs $1,500 at retail.

The value proposition inverted in one month. At $1,999, the EVO-X2 was a no-brainer for anyone who wanted to run 100GB-class MoE models slowly but locally. At $3,499, you’re paying discrete-GPU-tower money for 256 GB/s of bandwidth — a quarter of what a used RTX 3090 delivers.

RTX PRO 6000 Blackwell: $8,565 → $13,250

Tom’s Hardware confirmed in late July that NVIDIA raised the RTX PRO 6000 Blackwell’s price on its own marketplace to $13,250 — a 55% increase over the $8,565 launch MSRP from March 2025. Newegg undercuts that at roughly $12,099, which now counts as the discount.

The card’s problem is structural: 96GB of GDDR7 in a clamshell layout makes it the single most memory-dense discrete card you can buy, which means it absorbs more of the GDDR7 shortage per unit than anything else in the catalog. When we analyzed it in June, the case for a home lab was already thin at ~$8,600. At $12,000+, the $50K four-card GLM-5.2 build we covered last week would cost roughly $14,000 more in GPUs alone if you started buying today.

Used RTX 3090: ~$1,010 → $1,252 average

The used market didn’t escape — it just moved slower. When we covered NVIDIA skipping its 2026 consumer refresh in June, BestValueGPU’s tracker had already climbed off a winter average low near $966. ResalePrices’ July data now shows a $1,252 market average across 336 listings, with a fair asking range of $1,200–$1,299; BestValueGPU’s July tracker shows an eBay floor around $1,050, with remaining new units at $1,488 on Amazon.

That’s a ~30% climb since winter for five-year-old silicon — and it’s still the recommendation. Nothing else sells 24GB at 936 GB/s for anywhere near $1,100. Our value-king analysis from May has aged in one direction: every alternative got more expensive faster.

RTX 5090 and the rest of the 50-series

The RTX 5090 sits around $4,329 on Amazon as of July 11 per TechTimes, with premium AIB variants past $5,000 — against a nominal $1,999 MSRP that exists only on paper. TechTimes separately reports NVIDIA has cut RTX 50-series supply to board partners by roughly 20%, with the RTX 5070 Ti effectively sidelined from shelves and global street prices up 19% since late 2025. Reports from December 2025 (TechPowerUp, VideoCardz) had NVIDIA planning 30–40% production cuts for H1 2026, hitting the RTX 5070 and 5060 Ti hardest. The details of the GDDR7 price-hike story we covered in May keep getting worse, not better.

DDR5, for anyone building a CPU-offload rig

Per Newegg’s own July 2026 memory explainer, 64GB DDR5-6000 kits now start around $875 with thin stock — the company pegs current kits at four to five times mid-2025 pricing. TweakTown’s framing: 64GB of DDR5-6000 now costs more than a PlayStation 5. If your plan was a 256GB–384GB DDR5 box for CPU inference on big MoE models, that plan roughly tripled in cost since our June price-surge report.

Why everything with memory repriced at once

One cause, several symptoms. AI datacenter demand for HBM is soaking up DRAM wafer capacity at Samsung, SK Hynix, and Micron; what’s left gets allocated to server DDR5 first, and consumer LPDDR5X, DDR5, and GDDR7 fight over the remainder.

TrendForce’s numbers track the squeeze: Q1 2026 saw record quarter-over-quarter contract price increases across every memory category (The Register summarized the outlook as DRAM prices “expected to double” in Q1), Q2 kept climbing as cloud providers locked in long-term agreements, and the July 3 forecast for Q3 2026 still shows conventional DRAM contract prices rising 13–18% quarter-over-quarter. The moderation isn’t relief — TrendForce attributes it to consumer buyers hitting the limit of what they’ll pay, not to supply recovering.

Unified-memory machines get hit hardest for a simple reason: memory is most of what they are. A Strix Halo box is 16 Zen 5 cores attached to 128GB of high-density LPDDR5X — the exact commodity being redirected to datacenters. NVIDIA’s production cuts compound it on the GPU side: fewer new cards at higher prices push buyers to the used market, which is why even the RTX 3090 is appreciating.

The updated buying table (July 31, 2026)

Here’s where our standard recommendation tiers land after the repricing:

BudgetBuyJuly 2026 priceWhat you get
~$700RX 9070 XT 16GB$670–$700 street640 GB/s, best under-$1k card, ~12% over MSRP
~$1,100–$1,300Used RTX 3090 24GB$1,050 floor / $1,252 avg936 GB/s, CUDA, still the value pick
~$1,300AMD R9700 32GB$1,299 list (if findable)32GB GDDR6 at 640 GB/s — our July review
~$2,000EVO-X2 64GB$1,999.99Capacity play, halved — 64GB ceiling
~$3,500EVO-X2 128GB$3,499–$3,649100GB+ MoE capacity; wait if you can
~$4,300+RTX 5090 32GB$4,300–$5,000+1,792 GB/s, fastest consumer card
~$12,000+RTX PRO 6000 96GB$12,099–$13,25096GB GDDR7; rent before you buy

Notice what’s stable: GPUs whose memory was bought and soldered years ago (used 3090) or that use last-generation GDDR6 (RX 9070 XT, R9700) moved least. Everything built on LPDDR5X, GDDR7, or current DDR5 moved most. The shortage is repricing memory, and the market is repricing machines in proportion to how much memory they contain.

The cloud math just changed underneath you

The standard defense of buying an EVO-X2 was the rental break-even. That math moved a lot. RunPod’s Secure Cloud A100 80GB runs $1.39/hr (PCIe) as of July 2026:

runpod_a100_hr = 1.39          # RunPod Secure Cloud, PCIe, Jul 2026
for price in (1999, 3499):
    hours = price / runpod_a100_hr
    months_at_4h = hours / 4 / 30.4
    print(f"${price}: {hours:,.0f} hours = {months_at_4h:.0f} months at 4 hr/day")

Expected output:

$1999: 1,438 hours = 12 months at 4 hr/day
$3499: 2,517 hours = 21 months at 4 hr/day

(Our arithmetic; it ignores electricity — which favors the 8–14W-idle EVO-X2 slightly — and RunPod storage fees, which cut the other way.)

At $1,999, the EVO-X2 paid for itself in about a year of steady use, comfortably inside the machine’s useful life. At $3,499, break-even stretches to 21 months at 4 hours a day — long enough that the calculus in our rent-vs-buy guide tips back toward renting for anyone who isn’t running inference daily, and long enough that LPDDR5X prices may normalize before you’d have broken even anyway.

”I was about to buy X — what do I do now?”

You were about to buy an EVO-X2 128GB at ~$2,000. The problem: that price is gone, and paying $3,499 for 256 GB/s of bandwidth is a bad trade against everything else on the table. The fix: if your target models fit in 24GB, buy a used RTX 3090 instead — it’s 3× the EVO-X2’s measured speed on models that fit (our review’s benchmarks: ~95 tok/s vs ~30–45 tok/s on 7B–13B). If you genuinely need 100GB+ capacity for big MoE models, rent the workload on RunPod monthly and wait. Nothing about the EVO-X2’s hardware got better at $3,499; don’t reward the reprice.

You need 24GB now. Buy the used 3090 this week, not next quarter. It’s already up ~24% since March, NVIDIA’s supply cuts keep funneling buyers into the used market, and there is no plausible source of downward pressure before the rumored 2028 RTX 60-series. The floor price has only moved one direction for eight months.

You own an EVO-X2 or a 3090 already. You’re up. The June-vintage EVO-X2 at $1,999 is now, functionally, an appreciated asset — and the same is true of any 24GB card bought before spring. Run them; don’t sell into hype unless you’re exiting local AI entirely.

You were eyeing the RTX PRO 6000. Rent one first. RunPod and its competitors will lease you Blackwell workstation silicon by the hour for less than the price appreciation the card gained per month this year. If 96GB is genuinely your working set, the four-card build math still works for businesses — but at $12,000+ per card, a home lab buyer is paying a 55% shortage tax on day one.

You just want to run good models on hardware you already own. This is the quietly correct answer for most people in a shortage. The best-model-per-GPU leaderboard is full of 2026 MoE releases that made 8–16GB cards dramatically more capable than they were a year ago — the software got better faster than the hardware got expensive. FOSS inference stacks that squeeze existing cards are covered well at aifoss.dev, and if your actual need is coding assistance, an API subscription through the tools covered at aicoderscope.com costs less per month than the price drift on any GPU in this article.

What would have to change for prices to fall

Three things to watch, none imminent. First, TrendForce’s Q3 read says buyer resistance — not new supply — is what’s slowing increases; actual new DRAM fab capacity lands 2027 at the earliest. Second, NVIDIA restoring consumer GPU production would require datacenter demand to blink first; every report through June says the opposite. Third, GMKtec’s $3,499 “debut” pricing is itself a tell — vendors are building the shortage into new SKUs rather than eating margin, which means they expect it to last.

Our working assumption for the rest of 2026: memory-dense hardware stays expensive, used last-gen cards keep drifting up, and the best value moves keep being either old GDDR6X or no purchase at all.

FAQ

Did the EVO-X2 get worse? No — same Ryzen AI Max+ 395, same 128GB LPDDR5X-8000 at 256 GB/s, same ~11 tok/s on Qwen3-235B-A22B. Only the price changed. At $1,999 it was the best capacity value in local AI; at $3,499 it’s an expensive way to buy bandwidth that a $1,100 used card quadruples.

Is the used RTX 3090 still worth it at $1,252 when it cost $1,010 in March? Yes, grudgingly. Price-per-GB/s it still embarrasses everything: ~$1.34 per GB/s of bandwidth versus ~$13.67 for the EVO-X2 128GB at $3,499. The risk isn’t overpaying today — it’s that five-year-old cards carry five-year-old fan bearings and thermal pads. Budget $30–50 for repadding.

Should I wait for prices to normalize? Only if your need is optional. TrendForce still projects 13–18% quarterly DRAM contract increases for Q3 2026, and new fab capacity is a 2027 story. “Wait” is a real position for the EVO-X2 128GB and the RTX PRO 6000; it’s a losing position for used 24GB cards, which have risen every month for eight months.

Does the 64GB EVO-X2 at $1,999 make sense as the fallback? Sometimes. 64GB of unified memory still runs 30B-class MoE models fast (70–100 tok/s in our testing) and mid-size models a discrete 16GB card can’t hold. But at the same $1,999 you can have a used RTX 3090 plus most of a supporting PC, and the 3090 is 3× faster on everything that fits in 24GB.

Is any of this priced-in panic rather than real cost? Some, probably — GMKtec adding a $3,499 SKU $150 under its own $3,649 model looks like anchoring. But the underlying inputs are verifiably real: record Q1 2026 DRAM contract increases, 13–18% more in Q3, NVIDIA’s 20% supply cut to partners, and $875 64GB DDR5 kits at retail. This is a genuine input-cost shock with some opportunism on top.

Sources

Last updated July 31, 2026. Prices and specs change; verify current rates before purchasing.

Products linked in this article:

Was this article helpful?