What Your Local AI Rig Will Be Worth in Two Years

Bottom Line

• A used RTX 3090 retains roughly 67-87% of its launch price after six years. It launched at $1,499 on 24 September 2020 and sells at $1,000-1,300 as of August 2026. Nothing else in a PC does this.
• A used RTX 4090 trades above its own MSRP. Launch price $1,599; used street ~$2,270-2,600 in August 2026.
• Buy used, not new, if resale matters. A used card has already taken its depreciation step. A new RTX 5090 at $4,300-5,000 against a $1,999 MSRP is carrying a premium the market may reclaim.
• The strength is a shortage, not a property of the hardware. The same DRAM and GDDR7 squeeze that props up used prices is what made your purchase expensive. If supply normalises in 2027, you paid the peak and would sell into the dip.
• Apple Silicon holds value well but not better. Resale dealers report roughly 40-60% of retail after two years, versus 25-40% for comparable Windows laptops. Dealer-sourced, so indicative only.
• Non-GPU parts depreciate normally. Your PSU, case, and board are worth little at resale. Model the GPU separately from the rest of the build.
• The honest planning number is a range, not a point. Model a good case and a bad case, and check that the purchase still makes sense in the bad case.

Why This Number Decides the Whole Purchase

Our electricity break-even guide sets out the full cost formula:

Monthly cost of local
  = (hardware price − expected resale) / months you keep it
  + power
  + whatever you still pay a hosted provider

The first term is usually the largest, and resale is the only variable in it that most people guess at. Guess badly and the whole answer flips. A $1,150 card that returns $900 in two years costs about $10 a month in depreciation. The same card returning $300 costs about $35 a month. That gap is larger than the electricity bill.

So this page does one thing: it puts real August 2026 market numbers into that term, and it is honest about how fragile they are.

The Retention Table

Launch prices are manufacturer MSRPs at announcement. Used prices are US street ranges as of August 2026, drawn from eBay-derived trackers and price aggregators. Treat aggregator figures as indicative: they are scraped, not audited.

Read the last row before you read the first one. The Tesla P40 is what a datacentre GPU is worth once it stops running current software. It has 24GB, the same as a 3090, and it sells for a fifth of the price, because it lacks the numeric formats modern inference stacks expect. Retention is not about VRAM. It is about whether the card still runs what people want to run.

What Actually Drives Retention

Four things, in order of weight:

1. Does it run today’s models at today’s precision? This is the P40 lesson. A card that drops out of the software stack loses most of its value in a single year, no matter what is printed on the box.
2. VRAM per dollar against the current new market. A 24GB used card competes against a new 16GB card at a similar price. As long as model sizes keep pushing past 16GB, the used 24GB card wins that comparison and holds its price.
3. New supply. This is the 2026 story. NVIDIA has run three GeForce price increases in 2026, and memory is reported at over 80% of the bill of materials on some high-end cards. Constrained new supply lifts the whole used market.
4. Whether it is a whole computer. A Mac mini or Mac Studio sells to anyone. A bare GPU sells only to people who already have a machine to put it in. The wider buyer pool is part of why Apple hardware liquidates quickly.

Notice that only the first item is about the card itself. The other three are market conditions, and market conditions change.

The Risk Nobody Prices In

Here is the part other resale posts leave out.

The retention figures above and the prices you are paying are the same phenomenon. GDDR7 supply is the binding constraint on high-VRAM cards. PC DRAM contract prices rose 105-110% quarter over quarter. A 32GB DDR5 kit went from roughly $100 to $375-589. That shortage is exactly why a six-year-old 3090 still fetches four figures, and it is also why a 5090 costs $4,300-5,000 instead of $1,999.

So if you buy new today and sell in 2028:

• If the shortage persists, you keep most of your money. The retention table roughly holds.
• If memory supply normalises, new-card prices fall toward MSRP, used prices follow, and you paid a peak price to sell into a normal market. That is the double hit.

Nobody knows which happens. What you can do is choose the position with less downside. Buying used at $1,150 risks $1,150 against a floor the market has already tested. Buying new at $4,600 risks the same floor plus a $2,600 premium. That asymmetry, not raw performance, is the strongest argument for the used 24GB card in 2026.

Worked Example: Two-Year Cost of Ownership

Stated assumptions: 24/7 operation at 300W average, US average residential electricity of $0.18/kWh as of August 2026, 24 months held. The “pessimistic resale” column assumes memory supply normalises and used prices fall by about 40%.

Add power on top. At 300W the GPU builds cost about $39 a month in electricity; the Mac mini is a fraction of that, because it never draws anywhere near 300W. Our quiet 24/7 build guide covers why that gap matters for more than money.

The Mac row is the interesting one. Apple hardware depreciates more, but it depreciates predictably. The spread between the good case and the bad case is $3 a month, against $58 a month for the new 5090. If you dislike uncertainty more than you dislike cost, that is a real argument.

How to Actually Preserve Resale Value

Practical, in order of impact:

1. Keep the box and the accessories. Complete-in-box listings clear faster and higher on used marketplaces. This is free.
2. Do not run it at stock power 24/7. Cap the power limit. On an RTX 3090, inference at a 275W cap runs at the same speed as 350W, because the workload is memory-bandwidth bound. Lower sustained temperatures protect the card and cost nothing in throughput. See our undervolting guide.
3. Never disclose mining history, because you should not have any. Buyers discount heavily for it. Inference at a capped power limit is a gentler duty cycle than gaming.
4. Sell before the next generation ships, not after. The step down happens on announcement, not on availability.
5. Keep the original cooler. A card with an aftermarket block and no original shroud loses a large share of its buyer pool.

Common Mistakes

1. Using straight-line depreciation. GPU value moves in steps at product announcements, not on a smooth curve. A 20%-per-year model gives the wrong answer at every point.
2. Assuming last-gen gets cheaper. In 2026 several last-generation cards trade above their own launch MSRP. The RTX 4070 Ti Super launched at $799 and sells around $1,100-1,200.
3. Modelling resale on today’s prices for a sale two years out. Use a range. Check the decision still holds at the low end.
4. Counting the whole build’s resale. The GPU holds value. The PSU, case, cooler, and board largely do not. Model them at close to zero.
5. Forgetting selling costs. Marketplace fees and shipping take a real percentage. Subtract them before you compare against a rental bill.
6. Buying the top card because “it holds value.” It holds value in this market. The premium above MSRP is the part most exposed if the market changes.

See Also

• Local LLM electricity cost and break-even math — the full cost formula this page fills in
• Cloud GPU vs local GPU cost — the rent-or-own test, where resale is the deciding term
• RTX 5090 vs 4090 vs used 3090 — which of the three to actually buy
• Is a $5K local AI rig worth it? — the same question at a bigger sticker price
• NVIDIA RTX price hike: what to buy now — why the 2026 market looks like this
• Undervolting your GPU for a 24/7 local agent — the measured power sweep behind tip 2
• Keeping a 24/7 local AI box quiet — the other thing that decides whether you keep the rig

Sources

• Nvidia announces Ampere: RTX 3090 for $1,499 (Tom’s Hardware)
• RTX 3090 new and used price history (BestValueGPU) — scraped tracker, indicative
• MacBook resale value by year (ValueSnap) — resale-dealer data, indicative
• Which used Apple Mac retains the most value (Gadget Salvation) — resale-dealer data, indicative
• RTX 3090 power limit sweet spot for local LLM inference (Jean Brito)
• RTX A6000 48GB unveiled at $4,650, launched 5 October 2020 (Wccftech) — exact figure $4,649
• Tesla P40 specifications and recommended launch price, 13 September 2016 (Technical City) — $5,699 list

Used-market ranges for the 3090, 4090, A6000, and P40 are our own August 2026 price reference, compiled from eBay-derived trackers and GPU price aggregators. Aggregator sourcing is scraped rather than audited; we treat single-source figures as indicative and quote ranges rather than point prices.