TL;DR: If your current PC has a full-length PCIe x16 slot, a 750W+ power supply (or $110 to fix that), and 32GB+ of RAM, upgrading it with a used RTX 3090 costs roughly $1,050–$1,450 and gets you 95% of what a $2,500+ new build delivers. The DRAM crisis is the deciding factor: the 64GB of RAM a new build needs jumped 485% in a year. Buy new only if your case, PSU, or platform physically can’t take the card.

	Drop-in GPU upgrade	GPU + PSU upgrade	New rig (DIY or prebuilt)
Best for	Towers with 750W+ PSU and 32GB+ RAM	Most 2019–2024 mid-range towers	OEM boxes, laptops, multi-GPU plans
Total cost (Sep 2026)	$1,050–$1,343 (used RTX 3090)	~$1,160–$1,450	$2,200 prebuilt to $2,700+ DIY
Hidden line item	None — your RAM is already paid for	$110 Corsair RM850e	64GB DDR5 now averages $1,118
The catch	Transient spikes can trip a weak PSU	Case clearance: a 3090 runs ~313mm	You’re rebuying RAM at crisis prices

Honest take: Upgrade. A used RTX 3090 dropped into the tower you already own is the best price-to-tokens deal of 2026, because the components a new build forces you to rebuy — RAM above all — have never been more expensive relative to what they do for inference speed.

The usual version of this question is about gaming, where a five-year-old CPU genuinely drags down a new GPU. Local AI inference doesn’t work that way. Once a model’s weights are loaded into VRAM, token generation is bounded by the GPU’s own memory bandwidth — the CPU, the PCIe generation, and the rest of the platform are close to spectators. That one architectural fact, plus a memory market that has gone sideways, flips the default answer toward upgrading in 2026.

Before you price anything, run your target model and context length through the VRAM calculator to confirm which VRAM tier you actually need, and the local vs cloud cost calculator to sanity-check whether you should be buying hardware at all this quarter.

Why 2026 tilted the math toward upgrading

Two verified numbers frame the whole decision.

First, the GPU. A used RTX 3090 — still the value king for local AI — sold for around $1,050 on eBay in early September 2026 per BestValueGPU’s tracker, with active asking prices averaging $1,343 (range $1,287–$1,411) per ResalePrices. That card brings 24GB of VRAM at 936 GB/s and a measured 161 tok/s on GPT-OSS 20B in llama.cpp’s community benchmark thread (discussion #15396). Whether it goes into a 2020 tower or a 2026 tower, it produces the same tokens per second.

Second, the RAM. Tom’s Hardware’s price index put a 64GB DDR5-5600 kit at an average of $1,118 in August 2026 — up 485% from roughly $191 in August 2025 — and TrendForce forecasts conventional DRAM contract prices climbing another 13–18% in Q3. The AI datacenter buildout is eating the DRAM supply, and consumer kits are collateral damage. We covered the start of this in our DDR5 and SSD price surge breakdown; it has gotten worse since.

Put those together: the 32 or 64GB of RAM sitting in your current PC is an asset that would cost $500–$1,100 to replace at today’s prices, and it contributes exactly as much to inference speed in your old tower as it would in a new one. Every dollar a new build spends re-buying memory, storage, a case, and a PSU is a dollar that isn’t buying VRAM — and VRAM is the only thing the models care about. Our May $2,000 workstation parts list already flagged the memory crunch changing the math; that same build specced today lands closer to $2,700.

The five-minute audit of the PC you already own

Four checks decide whether your tower can take a 24GB card. No screwdriver needed for the first one.

1. PCIe slot and generation. Any full-length PCIe x16 slot from the PCIe 3.0 era onward is fine. If there’s already a discrete NVIDIA GPU in the machine:

$ nvidia-smi --query-gpu=name,pcie.link.gen.current,pcie.link.gen.max,power.limit --format=csv
name, pcie.link.gen.current, pcie.link.gen.max, power.limit
NVIDIA GeForce GTX 1070, 3, 3, 151.00 W

That 3 under pcie.link.gen.max means a PCIe 3.0 board — and for single-GPU inference, that’s not a problem. PCIe bandwidth matters while the model loads from disk and barely registers during generation; single-GPU token generation runs within measurement noise across PCIe 3.0/4.0/5.0, because decode traffic over the bus is tiny once weights are resident in VRAM. (Multi-GPU tensor parallelism is the exception — that’s a different article.)

2. Power supply. Read the sticker on the PSU’s side panel. A 3090 is a 350W card, and NVIDIA’s official system recommendation is 750W. The wattage number alone isn’t the whole story, though — see the transient-spike section below.

3. Physical clearance. Most used 3090s are ~313mm long, triple-slot, and heavy. Measure from the PCIe slot to the drive cage. Mid-tower cases from roughly 2018 onward usually fit; compact and OEM cases usually don’t.

4. System RAM. 32GB is the comfortable floor for a 24GB GPU — model loading, OS overhead, and any CPU-offloaded MoE layers want headroom (our system RAM guide has the full breakdown). If your machine has 32GB+ already, the DRAM crisis just handed you a $500+ head start over any new build. If it has 16GB, you can still run GPU-resident models fine and add RAM later when prices normalize — don’t pay crisis prices for headroom you may not need.

Pass all four and the upgrade case is nearly unbeatable: $1,050–$1,343 for the card, done.

The problem you’ll actually hit: transient spikes on an older PSU

Here’s the failure mode that catches upgraders. You install the used 3090 in a tower with a 650W PSU — which looks fine on paper against a 350W card and a 100W platform — and the machine hard-resets the moment a long generation kicks off. No blue screen, no log entry, just an instant power cut. That’s the PSU’s over-current protection tripping on the 3090’s millisecond-scale transient spikes, which shoot far above the card’s 350W sustained rating. Ampere cards are notorious for this on older, budget-tier units.

Two fixes, in order of cost:

$ sudo nvidia-smi -pl 280
Power limit for GPU 00000000:01:00.0 was set to 280.00 W from 350.00 W.

Power-limiting the card to 280W tames the spikes and, as we measured in our power-limiting guide, costs only ~3% of tokens per second — inference is bandwidth-bound, not power-bound. Try this first; for many 650W systems it’s the whole fix, and it cuts your power bill besides. At the EIA’s 2026 average residential rate of 18.02¢/kWh, a 350W load costs about $0.063/hour, so a capped card saves real money on a machine that runs 24/7.

If the resets persist — or your PSU is a decade old, non-modular, or from a brand you’ve never heard of — spend $110 on a Corsair RM850e (ATX 3.1, rated for exactly this class of transient). That takes the “GPU + PSU” upgrade path to roughly $1,160–$1,450 total, still $800–$1,200 under a comparable new build.

When buying new is actually right

Upgrading loses in five specific situations, and pretending otherwise wastes your money in the other direction.

Your PC is an OEM box. Dell, HP, and Lenovo towers frequently use proprietary PSU form factors and connectors — a standard RM850e physically won’t mount, and adapters for the motherboard’s nonstandard power header are a fire-risk lottery. If the PSU can’t be replaced and the existing one is under 600W, the upgrade path is closed.

You’re on a laptop or mini PC. No slot, no upgrade. (External GPU enclosures exist but pay a latency and bandwidth tax that undercuts the whole point of a fast card.)

You want more than 24GB of VRAM. Two cards means a board with proper slot spacing, lane bifurcation, and a 1,200W-class PSU — at that point you’re rebuilding anyway. Same if you’re eyeing 128GB-class unified memory: a ~$2,200 mini PC like the GMKtec EVO-X2 from our prebuilt guide runs model sizes no single consumer GPU holds, and no upgrade gets your old tower there.

Your platform predates 2015 or so. Very old boards can lack above-4G decoding and choke on model load times over PCIe 2.0 — plus DDR3-era machines usually top out at 16GB. Workable, but you’re bolting a $1,300 engine onto a bicycle.

You can sell the old machine well. If your current PC is recent enough to fetch real money, selling it whole and buying purpose-built sometimes nets out — run the numbers honestly, including the RAM you’re giving away at 2026 replacement prices.

If you land in the buy-new column, our GPU buyer’s guide by budget and prebuilt workstation guide are the two follow-on decisions, and Mac vs PC is worth five minutes before you commit to either.

What to put in the slot: three upgrade tiers

Budget, new-with-warranty: RTX 5060 Ti 16GB. Lowest tracked US price was $762 on September 5 per GPUPrix — Amazon listings dipped to $624 — against a median around $805, 88% over its $429 MSRP per Wccftech. Painful inflation, but it’s the cheapest modern 16GB card, sips 180W (no PSU drama), and 16GB covers the excellent 14B–26B class. The used 4060 Ti 16GB is the same idea cheaper and slower. Pick this tier only if $1,300 is genuinely out of reach — the 24GB tier is where local AI gets good.

The default: used RTX 3090, $1,050–$1,343. 24GB, 936 GB/s, 161 tok/s on GPT-OSS 20B. Every dollar-per-token comparison we’ve run this year — including the three-way against the 4090 and 5090 — ends with the 3090 on top.

The speed splurge: used RTX 4090, ~$2,268. Same 24GB capacity, trading between $2,100–$2,400 on eBay per BestValueGPU. It buys 225 tok/s on the same benchmark and roughly 2× the prompt-processing speed — worth it mainly if you hammer long-context coding agents through something like Continue.dev with a local backend. For chat and general use, the 3090 and an $1,100 pocket difference wins.

Don’t want to buy anything yet? Rent the exact card first. An hour of RTX 3090 or 4090 time on RunPod costs well under a dollar, and running your actual workload beats any spec-sheet reasoning — our rent vs buy breakdown covers the break-even math.

The decision, compressed

Your old PC’s job in a local AI rig is to feed power and PCIe lanes to a GPU, and almost any tower from the last eight years does that job indistinguishably from a new one. In a normal year, buying new costs a convenience premium. In 2026, with 64GB of DDR5 averaging $1,118 and GPU street prices 88–135% over MSRP, it costs a crisis premium — you’d be rebuying, at the market’s worst prices in a decade, parts whose contribution to tokens per second is zero.

Audit the machine you own. If it passes, put $1,050–$1,450 into a used 3090 (plus PSU if needed) and skip the rest of the build sheet entirely.

FAQ

Will my old CPU bottleneck a used RTX 3090 for LLM inference? Not meaningfully, as long as the model fits in VRAM. Token generation is bounded by the GPU’s 936 GB/s memory bandwidth; the CPU handles tokenization and orchestration, which any quad-core from the last decade manages. The CPU matters only when you offload layers to system RAM — which is a sign you need more VRAM, not more CPU.

Is PCIe 3.0 x16 enough, or do I need a PCIe 4.0/5.0 board? PCIe 3.0 x16 is enough for single-GPU inference. The generation affects model load time (a few extra seconds) and does not measurably change tokens per second once weights are in VRAM. Don’t buy a new motherboard for this.

My PC has a 650W PSU — can I run a 3090 without replacing it? Often yes, with nvidia-smi -pl 280 capping the card at 280W, which costs about 3% of inference speed and tames the transient spikes that trip over-current protection. If the machine still hard-resets under load, a $110 Corsair RM850e ends the problem.

Should I add RAM as part of the upgrade? Only if you’re below 16GB or you plan to run CPU-offloaded MoE models. At August 2026’s $1,118 average for a 64GB DDR5 kit, buying headroom “just in case” is the single worst-value purchase in a build. Run what your current RAM allows and revisit when the DRAM market normalizes.

Recommended Gear

Products linked in this article:

• Used RTX 3090 24GB — the default upgrade: 24GB, 936 GB/s, $1,050–$1,343 used
• Corsair RM850e — the ~$110 ATX 3.1 PSU fix for transient-spike resets
• RTX 5060 Ti 16GB — the budget 16GB path at $624–$805
• Used RTX 4090 24GB — the ~$2,268 speed splurge for agentic workloads

Sources

• RTX 3090 Price Tracker US — Best Value GPU
• RTX 3090 Used GPU Price & Fair Asking Range — ResalePrices
• RTX 4090 Price Tracker US — Best Value GPU
• RAM price tracking 2026: 64GB DDR5 kit averages $1,118, up 485% — Tom’s Hardware
• RTX 5060 Ti 16GB median price surges to $805, 88% above MSRP — Wccftech
• GeForce RTX 5060 Ti 16GB price history, lowest US price $762 — GPU PRIX
• GPT-OSS 20B/120B performance on consumer hardware (tg128 results) — llama.cpp discussion #15396
• LLM performance and PCIe lanes: key considerations — Rost Glukhov
• Corsair RM850e Fully Modular Low-Noise ATX Power Supply — Corsair
• NVIDIA GeForce RTX 3090 specs (24GB, 936 GB/s, 350W) — TechPowerUp
• How much does electricity cost in 2026? (EIA: 18.02¢/kWh residential average) — EnergySage

Last updated September 6, 2026. GPU and RAM prices in this article were verified on that date and move weekly — check current listings before purchasing.