Build a PC for local AI in 2026: the VRAM-first guide
VRAM decides everything, and 2026's memory crunch repriced it. What each tier runs, and the builds that still pencil out.
In April 2025, Nvidia launched the RTX 5060 Ti 16 GB at $429. Fifteen months later the same card sells for $589, and it is still the cheapest sane way into local AI. That is the 2026 GPU market in one sentence: nothing got faster, everything got pricier, and the parts that hold memory got pricier fastest.
You can still build a desktop that runs serious AI models at home. But the advice that circulated for two years, grab a used RTX 3090 for $700 and laugh at the cloud, has an expired price tag on it. This guide is the July 2026 version: why VRAM is still the only spec that decides anything, what the memory shortage did to each tier, and the builds that still make sense, with honest numbers on the ones that no longer do.
VRAM decides what you can run. Buy it first
A language model runs at GPU speed only while all of it sits in the GPU’s own memory. The moment a model spills into system RAM, generation drops from reading speed to a crawl. So the video card’s memory capacity, its VRAM, is not one spec among many. It is a hard line between the models your PC can run and the ones it cannot, and no amount of GPU speed on the other side of that line buys it back.
The math is friendly: at the 4-bit quantization nearly everyone runs, a model needs about 0.6 GB per billion parameters, plus a few GB of headroom for context. An 8B model is a 5 GB download. A 32B model needs about 20 GB. That is why a five-year-old 24 GB card runs models a brand-new 12 GB card cannot touch, and why this guide sorts everything by VRAM instead of by benchmark scores. Two footnotes before the shopping starts. First, Nvidia remains the smoothest road: virtually every AI tool ships CUDA support first, so the price premium buys you the ecosystem, not just the silicon. Second, plan for 32 GB of system RAM. Model files stage through system memory on the way to the GPU, and 16 GB machines swap painfully. That second footnote used to be a $90 afterthought. It is now a plot point.
The memory shortage rewrote every price tag
Through late 2025 and 2026, AI datacenter buildouts absorbed most of the world’s DRAM supply. Samsung, SK Hynix, and Micron pointed their fabs at the high-bandwidth memory that datacenter accelerators use, because those buyers pay more per wafer than consumers ever will. By one industry accounting, memory now makes up more than 80% of a graphics card’s bill of materials. Every part of a local-AI build that stores bits repriced accordingly.
The numbers are blunt. The RTX 5090 launched at a $1,999 MSRP and now lists at $4,329, with partner cards above $5,000. The RTX 5070 Ti, MSRP $749, has spent 2026 between $900 and $1,250. Plain DDR5 went along for the ride: 32 GB kits that cost around $90 in mid-2025 now start above $300, and Newegg’s own explainer tells builders to expect thin stock into 2027. Even the cavalry got canceled: Nvidia’s 24 GB Super refresh, the cards that would have made this whole guide cheaper, is delayed indefinitely because it depends on the same 3 GB GDDR7 modules everyone is fighting over. Two consequences for you: buy for what you need now, not for a refresh that may not come, and treat anything with lots of memory at a fair price as the deal, even if the silicon around it is old.
What each VRAM tier actually runs
Before spending, know what each tier buys. The table below uses 4-bit model sizes, the format that quantization has made the default for local use, and assumes you leave a few GB free for context. One entry worth flagging: OpenAI’s gpt-oss-20b was designed to run within 16 GB, which is a big part of why 16 GB is the new floor worth aiming for.
Do not buy an exact fit. A model that technically loads with 200 MB to spare will choke the moment you paste a long document, because context costs memory on top of the weights. If you want to check a specific model against a specific card before buying, our hardware calculator does that math interactively.
Three builds for July 2026 (one is a warning)
These are complete-tower numbers at street prices, not wishful MSRPs. The GPU line is exact and cited; the rest is a fair estimate for boring, reliable parts, which is exactly what you want. The CPU genuinely does not matter much for AI inference, so a mid-range chip from either vendor is fine, and the money saved belongs in memory.
The entry build is the one to copy if you are new to this. Sixteen gigabytes covers the models most people actually run day to day, the 14B-to-30B class that handles drafting, coding help, and summarizing well, and it generates images without drama.
The sweet-spot build hinges on the used market, so be honest about what changed there. The RTX 3090’s 24 GB made it the value king of local AI for years, and functionally it still is: nothing else puts 24 GB of CUDA-compatible memory in a slot for close to a thousand dollars. But the crowd noticed. Used units average around $1,050, up from roughly $800 a year ago, for a card that is now five years old with no warranty. It remains the right buy; it is no longer a steal. Buy from a seller with returns, and budget an 850 W power supply for its 350 W appetite. The same logic rules out the used RTX 4090, which offers the same 24 GB for over $2,200: you would pay double for speed while gaining zero model capacity. Tinkerers with good airflow can even run two used 3090s for 48 GB total; llama.cpp splits a model across them without any special bridge, though the pair pulls 700 W under load.
And the 5090 build is the warning. It is a marvel of engineering stuck at the worst price-per-gigabyte on this page. Its 32 GB runs the same 32B-class models the $2,050 build runs, just faster, and the next milestone up, a 70B at 43 GB, does not fit anyway. Unless you also game at 4K or fine-tune models for money, the extra $3,800 buys speed you will notice and capability you will not.
AMD went from asterisk to value play
For years the honest local-AI advice was “just buy Nvidia,” and for tinkerers it mostly still is. But 2026 moved the line. The RX 9070 XT, AMD’s mainstream 16 GB card, launched at $599 and, remarkably for this market, actually sells near it, making it the rare 2026 card priced like its own spec sheet. On the software side, AMD’s ROCm stack now supports these cards officially, and independent benchmarks on the 9070 XT show llama.cpp running well over both ROCm and the vendor-neutral Vulkan path. Ollama and LM Studio, the apps most people start with, support AMD out of the box.
The asterisk that remains: inference is solved, the frontier is not. Fine-tuning stacks, cutting-edge serving engines, and day-one support for newly released models still land on CUDA first and reach AMD weeks or months later, sometimes never. If your plan is running chat models and image generators through mainstream apps, the 9070 XT is the best price-per-gigabyte of new silicon on this page. If your plan involves the word “experiment,” pay the Nvidia tax. Intel’s 12 GB Arc B580 extends the same trade further down the price ladder, with the same caveat applied twice.
Past 24 GB, buy unified memory, not a bigger GPU
Here is where the discrete-GPU ladder simply ends. Above 24 GB your options are a $4,300 card that still cannot hold a 70B model, dual used cards with dual heat, or workstation GPUs priced like small cars. Meanwhile a different architecture quietly took over the big-model story: unified memory, where CPU and GPU share one large pool and a model can use nearly all of it.
The PC version is AMD’s Ryzen AI Max+ 395, the chip PC vendors call Strix Halo, with up to 128 GB of unified LPDDR5X. Mini PCs built on it, like the GMKtec EVO-X2, have listed between roughly $2,000 and $3,300 as the memory market repriced, and AMD’s own developer version of the box runs $3,999. What that buys: a 70B model at 4-bit, or the 65 GB gpt-oss-120b, loaded and answering, on a machine that idles quietly on a shelf. The trade is bandwidth. Unified memory moves bits at a fraction of a discrete card’s speed, so big models run at patient-reading pace rather than chat-app pace. It is the same capacity-over-speed bargain that makes high-memory Macs the other big-model machines. If your ambition is the largest models rather than the fastest tokens, stop climbing the GPU ladder and buy memory instead.
Buy the tier you need, not the tier you fear
The 2026 market punishes both waiting and overbuying, so decide by workload and stop. If you want a first local-AI box, build the $1,500 entry machine; 16 GB runs the open models worth running without apology. If you know you want 32B-class assistants or light fine-tuning, hunt a used 3090 and accept that the bargain era ended. If you dream in 70B and up, skip the $4,000 GPU entirely and put the money into a unified-memory box. And if you already own a 12 GB card, you are not stuck: the 8-to-14B class got startlingly good this year, and an upgrade will cost less on the far side of the shortage.
Two don’ts to close. Don’t buy an 8 GB card for AI in 2026; it was the budget pick two years ago and it is a dead end now. And don’t wait for the refresh cycle to save you; the analysts watching DRAM supply expect tightness deep into 2027. The machines got expensive, but the models got small and good. Build for where they meet.
