Dual RTX 5060 Ti vs Used RTX 3090 for Local LLMs: This Build Died in the Shortage

Short answer

The dual RTX 5060 Ti build is dead as a budget recommendation. As of August 2026:

• one 5060 Ti 16GB medians about $805, so two cost ~$1,600
• a used RTX 3090 runs $1,000–$1,500, with fair eBay asks at $1,200–$1,300
• NVIDIA has ended 5060 Ti 16GB production

The build only made sense when two new cards with warranties undercut one used 3090. That condition no longer exists. Buy the 3090.

The build the internet still recommends

Through 2025, the pitch was strong: two 5060 Ti 16GB cards at $430–$500 each gave you 32GB of VRAM for around $900, new, with warranties, versus a used 3090 of unknown mining history. A Hacker News front-page thread (item 44198805) is exactly this debate, and old Reddit and YouTube guides still repeat the $450 number.

Every part of that pitch depended on the price. The price is gone.

What the shortage did, as of August 2026

Three confirmed changes killed the build:

1. The card nearly doubled. Price trackers at Digital Citizen and TweakTown put the median US 5060 Ti 16GB at about $805 in August 2026, up 39% from June and roughly 88% over MSRP.
2. NVIDIA ended 16GB production. VideoCardz reports ASUS confirming the 16GB variant is EOL with production halted, and TrendForce reports NVIDIA reallocating supply to 8GB variants. The card is evaporating, not recovering.
3. The replacement is on hold. The rumored RTX 50 SUPER cards with 24GB are delayed indefinitely over 3GB GDDR7 chip costs, roughly $60–70 per chip versus $20 for 2GB parts, per eTeknix and Notebookcheck. Possibly CES 2027. There is no cheap 24GB cavalry coming.

Used 3090s rose too, from the old $700–$1,000 range to $1,000–$1,500 per resale trackers. But they rose less, and that flipped the ranking.

The bandwidth argument was always there

Even at 2025 prices, the dual build had a technical catch. NVIDIA lists the 3090 at 936 GB/s of GDDR6X bandwidth. The 5060 Ti has 448 GB/s, so the 3090 is about 2.1x one card. In split-GPU inference, a model sharded across two 5060 Tis still decodes each layer at single-card bandwidth. As HN commenters put it, two 5060 Tis roughly equal one 3090 combined.

Add the dual-GPU tax: a board with two usable slots, a bigger PSU, tensor-parallel configuration, and runtimes that behave worse across two cards than one. The 3090 gives you 24GB of simple, fast VRAM in one slot.

What to buy instead

24GB on a 3090 covers most 20B–32B quantized models, which remains the practical local coding-agent tier. See best local LLM for RTX 5090 if you go the new-card route.

When would the dual build come back?

Watch two numbers. If the 5060 Ti 16GB ever sells near $500 again, and used 3090s stay above $1,200, the math partially returns. With production ended, that combination is unlikely. Treat any guide quoting $450 per card as a historical document.

Final recommendation

Buy the used RTX 3090 at $1,200–$1,300. It has more bandwidth than the dual build delivers per GPU, no multi-GPU complexity, and it now costs less than two 5060 Tis as of August 2026.

The dual 5060 Ti build was a good idea at a price that no longer exists. Do not pay 2026 prices for a 2025 recommendation.

Next steps

• Run the Local LLM Fit and Speed Estimator
• AMD R9700 vs RTX 3090 for local LLMs
• RTX 5090 vs 4090 vs used 3090 for local LLMs

Sources

• https://news.ycombinator.com/item?id=44198805
• https://www.digitalcitizen.life/
• https://www.tweaktown.com/
• https://videocardz.com/
• https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/