GPU Prices Are Up 50%+: What to Buy for Local AI Right Now

Short answer: GPU prices are up 50%+ over MSRP because of a memory shortage, not scalpers. IDC forecasts AI datacenters may consume up to ~70% of world memory output in 2026. An RTX 5090 lists at $3,695+, the RTX 50 Super is delayed indefinitely, and 128GB unified-memory boxes (ASUS Ascent GX10, $3,099) are the best value for big local models.

Add AMD's new $3,999 Ryzen AI Halo box to that list and you have the whole market in one paragraph. The rest of this page is the detail: what things actually cost as of mid-July 2026, which numbers are confirmed versus leaked, and exactly what to buy depending on the models you want to run. If you came here planning to buy a 5090 at street price, read the unified-memory section first — it will probably save you money.

---

Why Prices Are Up: The Memory Supercycle {#why-prices-are-up}

Every GPU is mostly a memory purchase now. The industry calls it the memory supercycle: AI datacenter buildouts are absorbing DRAM and NAND at a scale consumer hardware has never competed with. IDC's forecast is the headline number — AI datacenters may consume up to ~70% of world memory output in 2026.

The knock-on effects are not subtle:

• GPU street prices are ~50%+ above launch MSRP across tiers, not just at the flagship.
• NVIDIA officially raised the DGX Spark Founders Edition MSRP from $3,999 to $4,699 (+18%) in late February 2026, explicitly citing the DRAM/NAND shortage. Same hardware, higher price. When a vendor raises MSRP mid-cycle and says the quiet part out loud, believe the shortage is real.
• The RTX 50 Super refresh is reportedly on hold over 3GB GDDR7 memory pricing (more on that below).

Nothing in the reporting suggests this clears in a quarter — memory contracts are the bottleneck, and the AI buildout writing those contracts shows no reported sign of slowing.

---

RTX 5090: The 2x-MSRP Trap {#rtx-5090-prices}

Here is what the flagship actually costs, as of mid-July 2026 (street prices; they move weekly):

One market report tracked the 5090's street price up 79% in three months, versus +35% for the RTX 5080 over the same period. The flagship is inflating fastest, consistent with AI buyers targeting the top card.

Our advice is blunt: do not pay 2x MSRP for a 5090 unless you are VRAM-bound today — meaning a model you need right now will not fit in the card you have. The 5090 is still the fastest thing you can put in a consumer PC, and for models that fit in 32GB it beats every unified-memory box on speed. But at $4,000+, you are paying datacenter-shortage prices for consumer silicon, with a leaked Super refresh sitting in the wings that could reset the mid-range.

If you need CUDA VRAM on a budget instead, the used market still matters: see our RTX 4090 vs 3090 comparison and the RTX 3090 local AI guide — a used 24GB card plus our 24GB VRAM model picks covers a lot of real workloads without flagship pricing.

---

The RTX 50 Super Delay {#rtx-50-super-delay}

The refresh that was supposed to fix the mid-range VRAM problem is stuck — and it is stuck on the same memory shortage.

What the supply-chain reporting says (this is reporting, not an official NVIDIA statement): cards reached at least one board partner, but the launch is on hold over 3GB GDDR7 memory pricing. The refresh is delayed indefinitely — not cancelled — and CES 2027 (January) is the most-cited realistic window.

The leaked specs, labeled as leaks:

Why this matters for buyers: a 24GB RTX 5080 Super would gut the case for paying $3,700+ for a 5090 just to get VRAM. That leaked lineup is the one genuine reason current prices could normalize. But the timing is unknowable, it hinges on the exact memory pricing that caused the delay, and waiting indefinitely for a leak is not a plan. Decide based on what you need to run this quarter.

---

Unified-Memory Boxes: The Math That Changed {#unified-memory-boxes}

While discrete GPUs inflated, a new category quietly won the price-per-gigabyte fight: small boxes with 128GB+ of unified LPDDR5X that CPU and GPU share. For big models, they now beat stacking consumer GPUs outright.

ASUS Ascent GX10 — $3,099.99 (the value pick)

The GX10 carries the same NVIDIA GB10 Grace Blackwell chip and 128GB LPDDR5X as the DGX Spark, and it is in stock at Amazon, Newegg, and the ASUS eShop at $3,099.99 (1TB) or $4,149.99 (4TB). That is ~$1,600 cheaper than the DGX Spark Founders Edition ($4,699 after the February hike) for the same silicon. If you want the NVIDIA stack in this category, this is the box.

The mid-2026 review consensus on the Spark/GB10 platform is worth internalizing: its value is memory capacity, not speed. A single RTX 5090 is faster for anything that fits in 32GB. The GB10 niche is 70B-200B-class models at usable speeds — the models a consumer card simply cannot hold. The software story is maturing fast: vLLM published an official DGX Spark guide on June 1, 2026. NVIDIA's CES claim of "up to 2.5x inference gains" from software is a vendor claim — treat it as such.

AMD Ryzen AI Halo Developer Platform — $3,999

Launched July 10, 2026, this is AMD's own-brand Strix Halo mini-PC: Ryzen AI Max+ 395 (16 Zen 5 cores, Radeon 8060S with 40 CUs), 128GB LPDDR5X-8000, 2TB SSD, 10GbE, Wi-Fi 7. It costs $3,999, is a Micro Center exclusive (in-store pickup only), and ships with ROCm, LM Studio, and ComfyUI preconfigured — Windows 11 Pro or Linux, same price. AMD positions it directly against the DGX Spark for up-to-200B-parameter local models. You pay $900 more than a GX10 but get 2TB storage and 10GbE in the box; you trade CUDA for ROCm. For the chip itself, our Strix Halo / AI Max+ 395 guide has the full picture.

Strix Halo minis — cheapest 128GB, if you shop carefully

The mini-PC market caught the same inflation as GPUs. Street prices are volatile — treat these as approximate:

At ~$1,999, the GTR9 Pro is the cheapest ticket to 128GB — roughly half the price of a street-price 5090, with four times the memory. More options in our best mini PC for Ollama roundup.

Apple — the laptop path

The MacBook Pro M5 Pro/M5 Max line (announced March 3, 2026, available March 11) reaches 128GB unified memory at 614GB/s on the M5 Max (M5 Pro: up to 64GB at 307GB/s), with a Neural Accelerator per GPU core, per Apple. It is the only way to carry 128GB of model memory in a backpack. Details and configs in our Apple Silicon AI buying guide.

And the ceiling, for context

The NVIDIA DGX Station (GB300, 784GB-class coherent memory) is orderable via OEMs — ASUS, Dell, GIGABYTE, MSI, Supermicro — at ~$97K-$125K with 4-12 week lead times (an MSI XpertStation WS300 was listed at $96,995.99 on CDW). Not a consumer product; it is here so you know where the category tops out.

---

What to Buy, By Model Size {#what-to-buy}

Match the box to the models you actually run. That is the whole game in 2026:

Two honest caveats. First, unified-memory boxes are slower per token than a 5090 for anything that fits in 32GB — buy them for capacity, not speed. Second, what does 128GB actually unlock? Frontier-class open-weight models like GLM-5.2 are exactly the class of model this hardware exists for. Browse the full picture on our hardware hub and the best GPUs for AI ranking.

---

What Is Coming Next {#whats-next}

Gorgon Halo (Ryzen AI Max PRO 400 series) — announced ~May 20, 2026, this is the confirmed next step: Max+ PRO 495 (16-core/40CU with the new Radeon 8065S), PRO 490 (12-core/32CU), PRO 485 (8-core/32CU), with up to 192GB LPDDR5X-8533 and up to 160GB allocatable as GPU memory (~273GB/s, +7% bandwidth). AMD claims one SoC can load a 300B-parameter FP4 model — "a first for any single SoC that is not a Mac Studio" — which is a vendor claim until third parties test it. OEM systems from ASUS, Lenovo, and HP are expected Q3 2026. If you can wait a quarter and want 192GB, this is the one dated thing worth waiting for.

Medusa Halo (Ryzen AI MAX 500) — Zen 6 + RDNA 5, LPDDR6, ~460GB/s. Rumor-stage only, realistically 2027-2028. Do not plan purchases around it.

RTX 50 Super — see above: delayed indefinitely, CES 2027 the most-cited window, all per supply-chain reports.

---

Verdict {#verdict}

The summer 2026 market rewards buyers who know exactly which models they want to run — and punishes everyone paying street price out of habit.

1. Do not pay 2x MSRP for an RTX 5090. $3,695 FE pricing is already +85%; $5,000+ AIB pricing is indefensible unless a model you need today will not fit in your current VRAM.
2. The RTX 50 Super delay is real but unofficial. Leaked 24GB mid-tier cards are the one credible force that could normalize prices — timing unknowable, CES 2027 most-cited. Waiting on a leak is not a strategy.
3. For big models, unified memory won. GX10 at $3,099 (same chip as the $4,699 DGX Spark), Ryzen AI Halo at $3,999, Strix minis from ~$1,999 — these beat stacking consumer GPUs on price-per-GB of model, full stop.
4. For models that fit in 32GB, a fast GPU still wins on speed. A used 24GB card remains the best value in raw tokens per second per dollar.

Buy capacity for the models you cannot fit. Buy speed for the ones you can. And in a memory supercycle, never buy either at panic prices.

---

Sources {#sources}

• Tom's Hardware — GPU street-price tracking and memory-shortage reporting
• TechPowerUp — GPU specifications database and market coverage
• VideoCardz, ServeTheHome, Wccftech — RTX 50 Super supply-chain reporting and Strix Halo mini-PC coverage (leak-grade where noted)
• NVIDIA developer forum, Micro Center, ASUS eShop — first-party pricing for DGX Spark, Ryzen AI Halo, and Ascent GX10

---

FAQ {#faq}