TL;DR: The RTX 5090 and 5080 shortage is real and severe. Cards sold out in minutes at launch, scalper prices hit $3,500-5,750 (200%+ markup), and the situation won't meaningfully improve until Q3 2026. The root causes are TSMC capacity allocation and GDDR7 production constraints. For AI buyers, this affects workstation builds and creates unusual value in the RTX 4090 - if you can find one.
---
What's Happening
NVIDIA's GeForce RTX 5090 and 5080 launched on January 30, 2026. Within 5-15 minutes, every retailer was sold out. This isn't typical launch-day enthusiasm - it's a supply crisis that analysts are comparing to the cryptocurrency mining boom of 2021.
The numbers tell the story:
- RTX 5090 MSRP: $1,999 (Founders Edition)
- Actual retail prices: $3,000-4,000 where available
- Scalper prices: $3,500-5,750 on secondary markets
- Markup: 200%+ over MSRP
- RTX 5080: Similar situation, most listings above $1,300 vs $999 MSRP
A German retailer recently disclosed they cannot sell any RTX 5070 Ti, 5080, or 5090 cards due to distributor rationing. Supply is being allocated in single-digit quantities per store.
Why This Is Happening
Two interconnected factors are driving the shortage: manufacturing capacity and memory constraints.
TSMC Capacity Allocation
NVIDIA's consumer GPUs and datacenter accelerators compete for the same TSMC 4nm wafer capacity. The economics strongly favor datacenter products:
- H100/H200 margins: Significantly higher than consumer GPUs
- Capacity allocation: Only 20-30% of advanced node capacity goes to consumer GPUs
- Demand priority: Hyperscalers and enterprise customers get first allocation
NVIDIA isn't choosing to starve the consumer market - but when wafer capacity is limited, higher-margin datacenter products naturally take priority. The AI Memory Crisis we covered earlier is part of this same dynamic.
GDDR7 Production Constraints
The RTX 5090 uses GDDR7 memory - a new standard that SK Hynix and Samsung are still ramping. Production complexity and yield issues mean memory supply can't match GPU die production:
- New process: GDDR7 requires different manufacturing than GDDR6X
- Yield challenges: Early production runs have lower yields
- Competing demand: GDDR7 also needed for AMD's upcoming cards
NVIDIA has confirmed they're absorbing memory costs to protect MSRP pricing, but this doesn't solve the physical supply constraint.
Timeline: When Will Supply Improve?
Based on analyst reports and supply chain data:
| Period | Expected Situation |
|---|---|
| Now - March 2026 | Severe shortage, retail nearly impossible |
| April - June 2026 | Gradual improvement, still constrained |
| July - August 2026 | Meaningful supply increase |
| September+ 2026 | Expected normalization |
NVIDIA has stated all RTX 50-series SKUs will "continue to ship," but actual availability remains severely constrained. The RTX 5070 and 5070 Ti face similar issues despite lower price points.
What This Means for AI Workstation Buyers
The RTX 5090's 32GB GDDR7 and improved tensor cores make it attractive for local AI development. But the current market creates specific challenges:
If You Need Hardware Now
- RTX 5090: Expect to pay $3,000-3,500 if you can find one, or wait 4-6 months
- RTX 4090: Ironically becoming harder to find as 5090 buyers settle for previous gen
- Pre-built systems: Vendors like Maingear have limited RTX 5090 allocations in laptops
- Consider RTX 4090: Still 24GB VRAM, excellent for most local LLM work, and prices haven't spiked as dramatically
Some RTX 5090 cards are available through system integrators. Viperatech lists the MSI RTX 5090 Gaming Trio at $3,600 - a premium, but actually in stock. Central Computer has the ASUS ROG Astral RTX 5090 at $3,400.
For Local LLM Development
The VRAM jump from 24GB (RTX 4090) to 32GB (RTX 5090) matters for specific use cases:
| Model Size | RTX 4090 (24GB) | RTX 5090 (32GB) |
|---|---|---|
| 7B models | Comfortable | Comfortable |
| 13B models | Comfortable | Comfortable |
| 30B models | Tight, needs quantization | Workable |
| 70B models | Requires heavy quantization | Still needs quantization |
For most local development work, the RTX 4090 remains capable. The 5090 premium only makes sense if you're consistently pushing VRAM limits or need the improved FP8 tensor performance.
Multi-GPU Considerations
If building a multi-GPU workstation for larger models, consider that:
- 2x RTX 4090: 48GB total VRAM, likely easier to source
- 2x RTX 5090: 64GB total VRAM, but finding two cards is extremely difficult
- Professional cards: RTX 6000 Ada (48GB) or A6000 available without consumer market constraints
Bizon offers workstations configured with RTX 5090 or 4090 options - they have direct vendor relationships that sometimes yield better availability than retail channels.
The Transparent Pricing Problem
This shortage highlights an ongoing issue in the AI hardware market: pricing opacity. During supply crunches, some vendors:
- Remove prices entirely and switch to "contact for quote"
- Adjust prices daily without transparency
- Bundle unwanted components to obscure GPU pricing
We only list products with transparent, public pricing. During shortages like this, that means fewer RTX 5090 listings - but at least the prices shown are real. Vendors that hide pricing during supply constraints aren't providing the transparency buyers need to make informed decisions.
Recommendations by Scenario
"I need an AI workstation in the next 30 days"
Buy an RTX 4090 system or consider DGX Spark for a different approach. The 5090 premium isn't worth the hunt right now.
"I can wait 3-6 months"
Monitor availability. April-May should see improvement. Set up stock alerts with retailers.
"I need 5090 specifically for the 32GB VRAM"
Check system integrators like Bizon, Puget Systems, and BOXX who may have allocations. Expect to pay 50-80% over MSRP. Alternatively, look at professional cards with higher VRAM.
"I'm building for production AI workloads"
Consumer GPUs shouldn't be your primary path for production. The ASUS Ascent GX10 with GB10 (128GB unified memory) or multi-GPU server configurations provide more appropriate capacity.
---
The Bottom Line
The RTX 5090 shortage is real, caused by structural supply constraints, and won't resolve quickly. For AI development:
- RTX 4090 remains excellent for most local LLM work
- RTX 5090 premiums are currently 50-100% - decide if the wait is worth it
- Professional cards and workstation builders offer alternatives outside the consumer crunch
- Supply should meaningfully improve by Q3 2026
---
Related:
- The AI Memory Crisis: Why Your Next GPU Server Will Cost More
- Browse AI Accelerators and GPUs
- Browse AI Workstations
---
