TL;DR: NVIDIA told its board partners to expect a ~$300 cost increase on RTX 5090 units (effective ~May 13, 2026), driven by GDDR7 memory costs that have risen roughly 40% since late 2025. The 5090 was already selling at $3,500–$4,000 in the wild; this hike pushes expected street prices toward $4,500–$5,000. For local AI builders, the 5090 was already a poor value proposition — this makes it worse, but the ripple effects on the RTX 5080 and 5070 Ti matter more.

	RTX 5090	RTX 5080	Used RTX 4090
Best for	70B+ models unquantized, serious fine-tuning	13B–34B at full quality, strong inference throughput	34B–70B quantized, max VRAM per dollar
VRAM / BW	32GB GDDR7 / 1,792 GB/s	16GB GDDR7 / 960 GB/s	24GB GDDR6X / 1,008 GB/s
Street price (May 2026)	$3,500–$4,000+ (rising)	$1,150–$1,400	~$2,450
The catch	GDDR7 hike pushes to $4,500–$5,000 soon	16GB ceiling limits 34B+ models at full precision	Used market also inflated; older memory tech

Honest take: Unless you need 32GB VRAM and 1,792 GB/s bandwidth specifically — and almost no home AI user does — buy an RTX 5080 at $1,150–$1,400 and skip the flagship pricing spiral entirely.

---

What actually happened

On or around May 13, 2026, NVIDIA reportedly notified its add-in card (AIC) partners — ASUS, GIGABYTE, MSI, EVGA, Zotac, and others — that the cost of GeForce RTX 5090 and RTX 5090D V2 boards was increasing by approximately $300 (around 2,000 RMB for the Chinese market). NVIDIA’s stated reason: rising GDDR7 memory costs that NVIDIA could no longer absorb.

NVIDIA has not changed the official MSRP on its website, which still reads $1,999 for the Founders Edition. That number has been a fiction in the retail market for months. The RTX 5090 has been selling well above $3,500 at major US retailers since supply normalized from its January 2026 launch crunch, with premium board partner models regularly listed above $4,000.

The $300 partner cost increase is a wholesale number — what NVIDIA charges the board partners who manufacture the cards. How much of that passes through to retail depends on each partner’s margin strategy, but the standard industry multiplier suggests retail prices will move $300–$450 higher from current levels. That puts the high end of 5090 pricing firmly in the $4,500–$5,000 range.

Why GDDR7 is getting expensive

GDDR7 memory is made by three companies: Samsung, SK Hynix, and Micron. All three are currently capacity-constrained, and all three have strong incentives to prioritize a different product over GDDR7: High Bandwidth Memory (HBM).

HBM is the memory stack used in data center AI accelerators — the H100s, B100s, and H200s that cloud providers are buying in enormous quantities. HBM commands a significantly higher margin than GDDR7. When wafer capacity is limited and a manufacturer has to choose between filling orders for HBM3e or GDDR7, the math is obvious.

Micron has stated it can currently meet only around 55–60% of core customer demand for memory products. Samsung warned in early 2026 that memory shortages would drive industry-wide price surges through the year. New fab capacity from Micron’s U.S.-based ID1 facility isn’t expected to come online until 2027.

The RTX 5090’s 32GB GDDR7 configuration uses a 512-bit memory bus — the widest in any consumer GPU. That 512-bit interface requires 16 GDDR7 chips to populate fully. At a 40% increase in per-chip cost, the memory stack on a 5090 has gotten substantially more expensive to build.

The RTX 4090 used GDDR6X memory on a 384-bit bus. GDDR6X supply pressures are lower; the 4090 is out of production and GDDR6X is more broadly available. This is part of why the RTX 5080 and 5070 Ti have seen proportionally larger retail price increases than mid-range cards that use less GDDR7 or lower-width buses.

What this means for RTX 5090 street prices

Before this hike: RTX 5090 street price was $3,500–$4,000, already roughly double its $1,999 MSRP.

After the partner cost increase propagates through distribution — typically 2–6 weeks from the effective date — expect the floor on RTX 5090 retail listings to drift toward $3,800–$4,000 for board partner budget cards and $4,500–$5,000 for premium/liquid-cooled models.

If you’re waiting for a “deal” on the RTX 5090, stop waiting. The trajectory is up, not down, and nothing in the GDDR7 supply picture suggests a correction before at least late 2026. The fact that NVIDIA maintained $1,999 as the official MSRP while passing real costs to partners suggests the company knows the official price is a marketing number — not a real-world procurement target.

For the specific case of local AI, there’s a second problem beyond raw cost: the RTX 5090 32GB at 1,792 GB/s was barely justifiable at $2,000 for most home lab setups. At $4,500, the value math collapses entirely.

The ripple: RTX 5080 and 5070 Ti

The 5090 is the headline, but the more relevant number for most local AI builders is what’s happening to the RTX 5080 and RTX 5070 Ti.

Both cards also use GDDR7 memory. The RTX 5080 uses 16GB on a 320-bit bus with ~960 GB/s bandwidth, MSRP $999. The RTX 5070 Ti uses 16GB GDDR7 on a 256-bit bus, MSRP $749.

Since November 2025, global GPU prices have increased nearly 15% on average across the RTX 50 series. The RTX 5080 and RTX 5070 Ti saw the largest individual increases in the lineup outside the 5090 itself — both up approximately 25%. A card with a $999 MSRP now streets at $1,150–$1,400. The RTX 5070 Ti, which launched at $749, has been selling at RTX 5080 MSRP or higher for weeks.

The pattern: higher GDDR7 content equals higher price pressure. The 5090 (32GB, 512-bit) gets hit hardest. The 5080 (16GB, 320-bit) gets hit moderately. The 5070 (12GB, 256-bit) saw only a ~$70 street price increase by comparison.

None of these increases have been formally announced by NVIDIA as MSRP changes. They’re supply-side pressure expressed in retail listings.

The RTX 5090 as a local AI tool: who it actually makes sense for

The RTX 5090’s case for local AI was always narrow, even at $2,000. The 32GB of GDDR7 memory matters for specific use cases:

• Running Llama 3.3 70B at Q4–Q5 quantization without CPU offload (fits in 32GB comfortably; the RTX 4090’s 24GB requires Q3 or aggressive offload)
• Fine-tuning 13B–34B models with QLoRA without gradient checkpointing slowdown
• Running two 30B-class models simultaneously for test-time compute workflows
• Image generation with SDXL + ControlNet + LoRA stacked at full precision

If none of those describe your actual workflow, the extra 8GB of VRAM over a used RTX 4090 (24GB) is dead money.

At $4,500+, the only home AI users who can rationally justify the RTX 5090 are those running sustained professional workloads that would otherwise require cloud GPU rental. If you’re regularly billing $400–$600/month to RunPod or Lambda Labs for 70B inference or fine-tuning work, the 5090 still pencils out as a 12–18 month ROI on hardware. For everyone else, it doesn’t.

What to actually buy right now (May 2026)

Given current pricing and the GDDR7 situation:

RTX 5080 at $1,150–$1,400: The correct flagship choice for most serious local AI users. 16GB GDDR7 at 960 GB/s handles 13B models at full Q8, 34B models at Q4, and Stable Diffusion / Flux workflows comfortably. The RTX 5070 Ti vs RTX 5080 comparison covers the 16GB ceiling in detail — the short version is that 16GB is constraining only if you regularly run 70B-class models. Buy the RTX 5080 if you can get it at $1,150–$1,250. Hold at $1,400+.

Used RTX 4090 at ~$2,450: The 24GB GDDR6X gives you meaningful headroom over the 5080’s 16GB, and the 1,008 GB/s bandwidth is only marginally behind the 5080’s 960 GB/s (within ~5%, effectively a wash for typical inference). The catch: used 4090 prices are also elevated from where they were six months ago, following general AI hardware demand. If you find a clean used 4090 from a gaming rig in the $2,000–$2,200 range, that’s a legitimate buy. At $2,450 market average, you’re paying a premium for VRAM that may or may not matter for your specific model tier.

Used RTX 3090 at $800–$1,100: The 24GB GDDR6X option for budget-constrained builders. Memory bandwidth (936 GB/s) is lower than the 4090, but VRAM capacity is identical. We covered the 3090’s case in detail in the used RTX 3090 value guide — the short version is that 24GB at sub-$1,000 still makes sense for inference-only workloads on Llama 70B models.

What to avoid: Buying a new RTX 5090 at $3,500–$5,000 for home AI use. The use case doesn’t justify the cost for most builders, and the pricing trajectory is not favorable.

What to watch over the next 90 days

The GDDR7 supply crunch won’t resolve quickly. Samsung won’t have meaningful new GDDR7 capacity online before late 2026 at the earliest. Micron’s U.S. fab (ID1) isn’t producing GDDR7 until 2027.

Two scenarios that could move prices lower: (1) AI data center demand for HBM cools, freeing fab capacity for GDDR7. (2) Samsung or SK Hynix announces a significant production ramp specifically targeting consumer GDDR7. Neither looks likely in the next quarter.

One scenario that could move prices sharply higher: NVIDIA announces an RTX 5090 Ti or refreshed flagship that uses even more GDDR7, creating new demand pressure. Given NVIDIA’s current focus on data center SKUs, this seems more likely in late 2026 or early 2027 than imminently — but it’s worth knowing the risk is there.

For local AI builders with a $1,500–$3,000 budget, the RTX 5080 at $1,150–$1,400 is the only card in the lineup where the current price-to-performance ratio still makes sense. The 5090 has become a collector’s item for home AI labs. The rest of the RTX 50 series is trading at premiums that erode their launch-era value propositions.

Keep checking retailer listings. Use Newegg, B&H Photo, and Amazon alerts rather than scalpers. And if the cloud vs. local math has shifted for you given these price increases, it’s worth re-running the numbers — RunPod renting an RTX 5090 at current cloud rates for a finite project may genuinely undercut buying one.

---

Frequently Asked Questions

Why is NVIDIA raising RTX 5090 prices if MSRP hasn’t changed? NVIDIA hasn’t changed the official MSRP ($1,999), but it has told the companies that actually build and sell the cards — ASUS, MSI, GIGABYTE, etc. — that the cost they pay NVIDIA per card is increasing by ~$300. Those companies then pass the increase to retailers, who pass it to consumers. The MSRP is effectively a marketing number; the real market price has been $3,500–$4,000 for months.

Does the GDDR7 price hike affect the RTX 5080 and RTX 5070 Ti too? Yes, though less severely than the 5090. The RTX 5080’s 16GB GDDR7 on a 320-bit bus contains half the memory chips of the 5090’s 32GB on a 512-bit bus. Both cards have already seen street price increases of roughly 25% since launch, with the RTX 5080 now trading at $1,150–$1,400 against its $999 MSRP.

Should I buy an RTX 5090 for local AI inference at current prices? Almost certainly not. At $4,000–$5,000, the RTX 5090’s 32GB GDDR7 and 1,792 GB/s bandwidth only make economic sense if you’re running sustained commercial workloads that would cost more to rent from a cloud provider. For personal or hobby use, the RTX 5080 or a used RTX 4090 delivers 80–90% of the real-world capability at 25–55% of the price.

When will RTX 5090 prices come down? Not before late 2026 at the earliest. The GDDR7 supply constraint is a structural issue — Samsung and Micron have allocated capacity to HBM for AI data centers, and new fab capacity won’t come online until 2027. Absent a significant cooling of data center demand or a GDDR7-specific production ramp, the supply/demand imbalance persists.

Is the used RTX 4090 a good alternative right now? At $2,000–$2,200 it’s excellent — 24GB GDDR6X at 1,008 GB/s handles most serious local AI workloads and has significantly more VRAM than the RTX 5080’s 16GB. At current eBay market average (~$2,450), it’s good but not great value. The 4090 used market is also elevated by the same AI hardware demand pressure affecting new cards.

---

Sources

• NVIDIA Reportedly Prepares RTX 5090 Price Hike Amid Rising GDDR7 Costs — TechPowerUp
• NVIDIA reportedly raises RTX 5090 and RTX 5090D V2 board prices due to GDDR7 cost increase — VideoCardz
• NVIDIA Passes Another $300 Hit to RTX 5090 Board Partners as VRAM Shortage Pushes Flagship GPU Toward $5,000 — WCCFTech
• NVIDIA is reportedly preparing an RTX 5090 price hike due to rising costs of GDDR7 memory — Tweaktown
• GeForce RTX 5090 Graphics Cards — NVIDIA Official
• GeForce RTX 5080 Graphics Cards — NVIDIA Official
• Global GPU Prices Grew By Nearly 15% Over Three Months — WCCFTech
• Samsung warns of memory shortages driving industry-wide price surge in 2026 — Network World
• RTX 4090 Price Tracker US — May 2026 — Best Value GPU
• Memory Shortages To Last Till At Least Q4 2027 — WCCFTech

Last updated May 29, 2026. Prices and specs change; verify current rates before purchasing.

Recommended Gear

• RTX 5080 (current generation, 16GB GDDR7)