AI

Black Friday GPU Deals for Local AI: 5060 Ti, 5070 Ti, 5090, 3090

This post contains affiliate links. If you buy through them, we may earn a small commission at no extra cost to you. Learn more.

Black Friday GPU deals for local AI and LLMs are hard to judge this year, because every new GDDR7 card you want costs well above its launch price. The RTX 5060 Ti 16GB launched at $429 and sits at about $800 on Amazon today. The RTX 5070 Ti launched at $749 and is close to double that. The cheapest new RTX 5090 we found on Amazon was about $7,700, for a card NVIDIA still lists at $1,999.

Original content from computingforgeeks.com - post 179274

This page ranks those three and a used RTX 3090, plus an Intel 24GB card that beats them all on VRAM per dollar. Each pick has a verified Amazon ASIN and a buy-below price, so you can tell a real Black Friday cut from a relabelled list price. We pulled every price from Amazon’s Creators API on October 29, 2026, so the figures and stock counts are live readings, not MSRPs.

The Black Friday GPU picks at a glance

For local inference, VRAM decides which models you can load and bandwidth decides how fast they answer. We picked one card per budget and gave each a price to wait for:

  • Best value 16GB card: PNY RTX 5060 Ti 16GB OC, about $800 and sold by Amazon itself. Buy below $780.
  • Fastest 16GB card, wait for a cut: MSI RTX 5070 Ti 16G Ventus 3X OC Black, about $1,480 from a third-party seller. Overpriced today; buy below $1,190.
  • Best 24GB value: a used RTX 3090 from eBay at about $1,100, or an HP OEM RTX 3090 (Renewed) on Amazon at about $1,650 if you want a return window. Buy below $1,000 used.
  • Most VRAM per dollar: ASRock Intel Arc Pro B60 Creator 24GB, about $1,000, sold by ASRock USA. No CUDA, so read its section first.
  • Skip at today’s prices: RTX 5090. Buy one only as a Founders Edition at $1,999.99, or put the money into two used 3090s.

Prices on these listings move daily, and most third-party offers show stock counts between “only 1 left” and “only 13 left”. Check the live price on each link before you decide. The buy-below numbers come from our own price log, explained next.

How we priced and ranked these cards

We ranked on what you actually pay per gigabyte of VRAM today, not on MSRP, because nobody on Amazon sells these cards at MSRP this autumn. The prices and speeds come from four sources, kept apart so you can see which number is a live reading and which is history:

  • Amazon Creators API, October 29: the buy-box price, condition, seller and stock message for every linked ASIN. Used-condition listings were excluded, which removed several cheaper 5070 Ti and 5060 Ti offers.
  • Our October 6 snapshot: the same API run that feeds our GPU price tracker, which gives a three-week trend for the same cards.
  • eBay sold prices: for the used RTX 3090, from the Get PC Parts sold-listing average dated October 24, because there is no new 3090 to price.
  • Our own benchmarks: tokens per second from vast.ai rentals we ran in June and October 2026, linked in each section.

We did not run a new lab for this page, since a deals page has no install steps to test. The speed figures come from those earlier rented-GPU runs, with links to the full logs, and the power and memory specs come from the NVIDIA, Intel and AMD product pages. NVIDIA’s family pages omit bandwidth and now show placeholder prices, so the NVIDIA bandwidth figures and launch MSRPs come from our earlier benchmark articles and launch coverage. Market context (Newegg medians, B60 listings, retailer dates) is cited where it appears. We left out cards under 16GB because a 14B model at Q4 takes about 9GB and leaves a 12GB card only about 3GB for context. Our GPU for LLM guide already covers that 12GB tier.

VRAM per dollar across the shortlist

The table below divides the October 29 price by the card’s VRAM. Lower dollars per GB means more model for your money, and the bandwidth column tells you how quickly that model generates tokens once it fits:

Card (listing)VRAMBandwidthPrice Oct 29$ per GBvs launch MSRP
Arc Pro B60 (ASRock Creator, new)24GB GDDR6456 GB/s$999.99$41.67no US MSRP published
RTX 3090 (eBay sold average, used)24GB GDDR6X936 GB/s$1,093$45.54launched at $1,499
RTX 5060 Ti 16GB (PNY OC, new)16GB GDDR7448 GB/s$799.99$50.00+86% over $429
RTX 3090 (HP OEM, Amazon Renewed)24GB GDDR6X936 GB/s$1,649.99$68.75launched at $1,499
Radeon AI PRO R9700 (Gigabyte AI TOP, new)32GB GDDR6640 GB/s$2,349.95$73.44+81% over $1,299
RTX 3090 FE (Amazon Renewed)24GB GDDR6X936 GB/s$1,929.97$80.42launched at $1,499
RTX 5070 Ti (MSI Ventus 3X OC Black, new)16GB GDDR7896 GB/s$1,483.49$92.72+98% over $749
RTX 5090 (ASUS ROG Astral, new)32GB GDDR71,792 GB/s$7,680.96$240.03+284% over $1,999

The used 3090 and the Arc Pro B60 both land in the low-to-mid forties per gigabyte with 24GB, the size that holds a 32B model at Q4 with room for context. The RTX 5090 costs more than five times as much per gigabyte as either. Its bandwidth lead, about twice a 3090, does not justify that. In practice the 16GB cards only make sense if your models fit in 16GB or you need Blackwell features such as FP4.

RTX 5060 Ti 16GB: the cheapest 16GB CUDA card

The PNY RTX 5060 Ti 16GB OC is the card we would buy first on a budget under $1,000. Amazon sells and ships it, and the cheapest new 5060 Ti 16GB has moved only $11 in three weeks.

PNY GeForce RTX 5060 Ti 16GB OC dual-fan GPU, a Black Friday deal pick for local LLM inference
PNY RTX 5060 Ti 16GB OC: 16GB GDDR7 on a 128-bit bus, 180 W. Image: PNY.

NVIDIA rates the 5060 Ti 16GB at 180 W with a 600 W system recommendation and a single 8-pin connector, so it drops into almost any desktop without a PSU upgrade. The weak spot is the 128-bit bus. Its 448 GB/s is half what the 5070 Ti gets, and token generation follows bandwidth closely. In our June vast.ai runs the 5060 Ti produced 75.1 tokens per second on Llama 3.1 8B at Q4_K_M under Ollama, against 125.0 on the 5070 Ti. On Qwen2.5 14B it was 42.3 against 73.2. The full method is in our RTX 5070 Ti vs 5060 Ti benchmark.

  • Who it’s for: 7B to 14B models, coding assistants and embeddings on one card, in a PC with a modest power supply.
  • Skip it if: you want 32B models or long contexts. Qwen3 32B at Q4_K_M is a 19.8GB file, so it spills into system RAM. When a 70B model spilled out of our 24GB 3090, generation fell to 3.44 tokens per second.
  • Interface: PCIe 5.0 x8. On an older PCIe 3.0 board it runs at a quarter of its link bandwidth, which slows model loading but not generation once the weights are in VRAM.

Our October 6 snapshot logged the cheapest new 5060 Ti 16GB at $788.99, and today’s PNY listing reads $799.99 (check the live price). Newegg’s median was $569.99 in June, before the summer retail rises. A Black Friday listing under $780 beats every Amazon reading we have logged this month, and anything near $600 is a real deal to buy on the spot.

RTX 5070 Ti: faster, but not 16GB cheaper

The RTX 5070 Ti gives you the same 16GB as the 5060 Ti on a 256-bit bus. That doubled bandwidth is the only reason to pay its premium for local AI.

MSI GeForce RTX 5070 Ti 16G Ventus 3X OC Black GPU for local AI, a Black Friday deal pick
MSI RTX 5070 Ti 16G Ventus 3X OC Black: 16GB GDDR7 on a 256-bit bus, 300 W. Image: B&H Photo.

Our benchmark put the 5070 Ti at 1.66 times the 5060 Ti’s token rate on Llama 3.1 8B (125.0 against 75.1 tokens per second), but today it costs 1.85 times as much. That works out to $11.87 per token-per-second for the MSI card against $10.65 for the PNY 5060 Ti. The faster card is now the slightly worse buy per unit of speed, not only per gigabyte. NVIDIA rates it at 300 W with a 750 W system recommendation, which matters if you are adding it to an existing build.

  • Who it’s for: agent loops and coding tools where response latency matters more than model size, on models that fit in 16GB.
  • Skip it if: you’re paying above $1,400. At that price a used 3090 gives you 8GB more VRAM and similar bandwidth for about $300 less.
  • Listing note: the MSI card is sold by Empowered PC, a third-party seller, with 13 in stock on October 29. The same seller’s Gigabyte WINDFORCE OC SFF read $1,449.00 on our first pull and $1,524.00 when we re-checked it less than an hour later, so these listings move fast.

Our October 6 snapshot logged a new 5070 Ti at $1,189.99, and Newegg’s median sat at $1,099.99 through the summer, so $1,190 is a fair line to wait for. Watch the condition line too. Several Amazon 5070 Ti listings at $1,299.99 today are marked Used in the API data, even though the titles read like new cards.

Used RTX 3090: still the 24GB value play

A used RTX 3090 is the cheapest way to get 24GB on CUDA. Bought at its eBay sold average, it undercuts every new card here on price per gigabyte except the Arc Pro B60. The Renewed Founders Edition linked below is the Amazon option with a return window, at a steep premium.

NVIDIA RTX 3090 Founders Edition 24GB used GPU for running LLMs locally, Black Friday renewed deal
RTX 3090 Founders Edition: 24GB GDDR6X on a 384-bit bus, 350 W, three slots. Image: NVIDIA.

The 3090 earns its place because 24GB holds Qwen3 32B at Q4_K_M entirely on the card, which no 16GB card can do. In our October llama.cpp runs a single 3090 generated 38.91 tokens per second on that model and 147.09 on Llama 3.1 8B. That is roughly half the 5090’s speed for about a seventh of its Amazon price. The numbers and the power logs are in our RTX 3090 vs RTX 5090 comparison. Two 3090s also run a 70B model: our dual RTX 3090 build reached 31.93 tokens per second on Llama 3.3 70B Q4_K_M with both cards capped at 290 W.

Get PC Parts put the average eBay sold price at $1,093 on October 24, with a 90-day average of $1,040 and $909 at the end of January. Amazon is far dearer: the Founders Edition (Renewed) reads $1,929.97 with one left, and an HP OEM RTX 3090 (Renewed) reads $1,649.99. You pay the Amazon premium for the Renewed return policy and nothing else. Before buying a used card from anyone, check these:

  • Photos of the actual card: the serial label, the 12-pin or 8-pin connectors and the fan blades, not stock images.
  • Memory temperature history: 3090s keep GDDR6X on both sides of the board, and ex-mining cards with dried pads run hot under long inference jobs. Budget for new thermal pads.
  • Power and space: 350 W per card and a 750 W system minimum for one; the Founders Edition takes three slots. Our PSU guide for multi-GPU builds covers the two-card case.
  • Buyer protection: pay through eBay or a marketplace with a return path, never a direct transfer.

We would wait for a used 3090 under $1,000, which is below the 90-day average. Skip the Amazon Renewed listings unless the return window is worth the extra $557 to $837 to you.

RTX 5090: the card to skip at Amazon prices

The RTX 5090 is the best single consumer card for local LLMs. At the prices Amazon shows today, it is also the worst buy in this guide.

NVIDIA GeForce RTX 5090 Founders Edition 32GB GDDR7 GPU for large language models
RTX 5090 Founders Edition: 32GB GDDR7 on a 512-bit bus, 575 W, two slots. Image: NVIDIA.

On paper it is everything the other picks are not: 32GB, 1,792 GB/s and FP4 support. In our runs it produced 277.78 tokens per second on Llama 3.1 8B. It also ran Llama 3.3 70B at IQ3_XXS at 52.83 with every layer on the card, a model a single 3090 cannot hold. NVIDIA’s RTX 5090 product page lists a 575 W board power and a 1,000 W system recommendation. The problem is price. The cheapest new 5090 on Amazon on October 29 was an ASUS ROG Astral at $7,680.96, sold by Amazon’s German store with two left. That makes it an import, so shipping time, import fees and the warranty region can differ from a US listing. Our October 6 snapshot logged $7,199.99, so the price has gone up heading into November.

NVIDIA’s US MSRP is still $1,999. HotHardware reported on October 19 that Best Buy lists the Founders Edition at $1,999.99 but rarely has stock, and that NVIDIA’s own US marketplace no longer lists the card. That is the only 5090 deal worth chasing this Black Friday. If no Founders Edition drop lands, the same $7,680 buys two used 3090s with 48GB and leaves over $5,000. It also buys a Radeon AI PRO R9700 32GB at about $2,350 if you can work with ROCm. Our used datacenter GPU guide covers the other route to 24GB and up.

Intel Arc Pro B60: the most VRAM per dollar

The ASRock Arc Pro B60 Creator is the cheapest new 24GB card on Amazon at about $1,000, sold and shipped by ASRock USA. That makes it the VRAM-per-dollar winner this Black Friday.

ASRock Intel Arc Pro B60 Creator 24GB blower GPU, best VRAM per dollar Black Friday pick for local AI
ASRock Arc Pro B60 Creator: 24GB GDDR6 on a 192-bit bus, 200 W blower. Image: ASRock.

Intel’s specification page lists 24GB of GDDR6 at 456 GB/s, a 200 W board power and a PCIe 5.0 x8 link, so it has 3090-class capacity with 5060 Ti-class bandwidth. That trade suits large models served at modest speed. A 32B Q4 model fits, but generation is bound by roughly the bandwidth of a 5060 Ti, so do not expect 3090 speeds. The blower cooler exhausts out of the back of the case, which suits a workstation with two or more cards side by side.

  • Who it’s for: builders who want 24GB or 48GB per box on a new-card warranty and are comfortable with llama.cpp’s SYCL or Vulkan backends.
  • Skip it if: your stack assumes CUDA. Tools that only ship CUDA builds, and most fine-tuning recipes, will not run on it.
  • Software support: llama.cpp runs on Arc through its SYCL and Vulkan backends; check your own runtime’s Intel GPU support before you buy.

A US store listed the Arc Pro B60 at $599 in September 2025, as Wccftech reported, so today’s $999.99 for the ASRock card is well above where the B60 started. Newegg lists other B60 boards between about $800 and $1,000, most shipping from overseas. Under $900 on Amazon is a fair Black Friday line, and anything near $600 is back to launch pricing. We have not benchmarked the B60 ourselves yet, which is why it sits below the CUDA cards in the picks list despite winning the table.

Why GPU prices are this high before Black Friday

GPU prices are high because memory got expensive, not because NVIDIA changed its list prices. Board partners have been paying more for GDDR7 since spring, and those costs went straight into street prices:

  • Partner cost rise, May 2026: VideoCardz reported NVIDIA raised RTX 5090 board prices to partners by about $300, citing GDDR7 costs.
  • Module pricing, August 2026: Tom’s Hardware reported that 2GB GDDR7 modules had risen to about $20 each, alongside TSMC wafer price increases.
  • US retail, June to August: Newegg’s median RTX 5060 Ti 16GB price rose about 40% (Tom’s table lists 39%), from $569.99 to $804.99, while the 5070 Ti held at $1,099.99, according to Tom’s Hardware’s August tracking.

GDDR7 cards took the hit, and they are the cards local AI buyers want. That is also why an older 24GB card on GDDR6X and Intel’s GDDR6 board look so good per gigabyte: neither depends on GDDR7 supply. Expect Black Friday discounts to be measured against inflated list prices. A strikethrough showing 20% off a $1,800 5070 Ti still lands above the summer median.

When the Black Friday GPU sales start

Black Friday falls on November 27, 2026. As of October 28 Amazon had not announced its sale dates, and its Black Friday Week started on November 20 last year. Other retailers have published these dates so far:

  • Walmart: Early Black Friday November 12 to 15, Black Friday November 23 to 29, Cyber Monday November 30 to December 2.
  • Best Buy: Black Friday listed for November 27. Best Buy is also where a $1,999.99 RTX 5090 Founders Edition would show up.
  • Amazon: not announced; sellers are being prepared for a Black Friday Week and Cyber Monday event.

We will re-read every listing on this page once Amazon’s event starts and update the prices and stock counts. If you are shopping for the rest of a homelab too, our Black Friday mini PC deals page tracks the Strix Halo boxes that compete with a GPU for running 70B models.

The cost per gigabyte at each budget

Most readers arrive with a budget rather than a card in mind. The table maps each budget to the card it buys at October 29 prices, with the largest model class it holds without spilling into system RAM:

BudgetBuyTotal VRAMPrice per GBLargest model it holds
About $800One RTX 5060 Ti 16GB16GB$50.0014B at Q4 with room for context
About $1,000One Arc Pro B60 (new)24GB$41.6732B at Q4, no CUDA
About $1,100One used RTX 309024GB$45.5432B at Q4 on CUDA
About $1,500One RTX 5070 Ti16GB$92.7214B at Q4 with room for context, faster
About $2,200Two used RTX 3090s48GB$45.5470B at Q4 split across two cards
About $7,700One RTX 5090 (Amazon)32GB$240.0370B at IQ3 on one card

Read down the price-per-GB column and the pattern is plain. The 24GB tier is where your money goes furthest, and the two-3090 row gives you 48GB for less than a third of one Amazon 5090. We’d spend a $1,100 budget on a used 3090 and an $800 budget on the 5060 Ti. Any 5090 money waits until a Founders Edition appears at $1,999.99. The 5070 Ti only wins when speed on small models matters more than capacity, so wait for it to fall under $1,190 before you treat it as a deal.

Keep reading

Ollama Models Cheat Sheet: Best Models by VRAM, Tested (PDF) AI Ollama Models Cheat Sheet: Best Models by VRAM, Tested (PDF) Open Source LLM Comparison Table (2026) AI Open Source LLM Comparison Table (2026) OpenCode CLI Cheat Sheet: Commands, Sessions and Fixes (PDF) AI OpenCode CLI Cheat Sheet: Commands, Sessions and Fixes (PDF) Best Laptops for Local LLMs: RTX 50 vs 128GB Unified Memory AI Best Laptops for Local LLMs: RTX 50 vs 128GB Unified Memory Black Friday Mini PC Deals: Proxmox, Local AI and Home Assistant AI Black Friday Mini PC Deals: Proxmox, Local AI and Home Assistant Best Deep Learning and PyTorch Books: 9 Picks Checked for PyTorch 2 AI Best Deep Learning and PyTorch Books: 9 Picks Checked for PyTorch 2

Leave a Comment

Press ESC to close