The Complete RTX 5090 Guide for Local AI (2026)

The NVIDIA GeForce RTX 5090 is the fastest consumer graphics card you can buy for running AI at home, and by a wide margin. With 32 GB of VRAM and a colossal 1,792 GB/s of memory bandwidth, it runs models that used to require a multi-GPU server or a cloud rental — Qwen 3 32B, DeepSeek R1 32B, even a quantized Llama 70B — fully on a single card, faster than anything else in its class.

But it is also a $1,999, 575-watt beast that demands a serious power supply and real cooling, and it is not the right buy for everyone. This complete guide explains exactly what the 5090 delivers for local LLMs, why its raw specs translate into real-world speed, which models it runs and how fast, how it compares to the RTX 4090 and cheaper alternatives, what to build around it, and — honestly — whether it is worth the money for your use case.

Table of contents

• The RTX 5090 at a glance
• What is Blackwell? The architecture
• Why these specs matter for AI
• The single most important number: memory bandwidth
• What models can the RTX 5090 run?
• Real-world performance
• Setting up your 5090 for AI
• Beyond LLMs: image, video and everything else
• RTX 5090 vs RTX 4090 vs the field
• The 5090 vs datacenter GPUs (A100/H100)
• Power, cooling and PSU requirements
• Building a rig around the 5090
• Running two 5090s: 64 GB of VRAM
• Fine-tuning on the 5090
• Common myths about the 5090 for AI
• Is the RTX 5090 worth it?
• Who the 5090 is really for
• Buying advice
• Capstone: the ultimate single-card AI workstation
• Frequently asked questions

The RTX 5090 at a glance {#at-a-glance}

The 5090 is the flagship of NVIDIA's Blackwell generation (the RTX 50-series), built on the GB202 die. Here are the numbers that matter for AI, and what each one means.

The full, always-current spec page with compatible models lives at the RTX 5090 GPU page. Two numbers dominate everything that follows: 32 GB of VRAM (how big a model fits) and 1,792 GB/s of bandwidth (how fast it runs). The rest — CUDA cores, clock speeds, ray-tracing units — barely matters for LLM inference, which is a point worth internalizing before you spend anything.

To put those two numbers in context: the previous flagship, the RTX 4090, had 24 GB and 1,008 GB/s. So the 5090 adds a third more memory and nearly doubles the bandwidth in a single generation — an unusually large jump, and one aimed squarely at AI workloads rather than gaming, where the gap is smaller. That is why the 5090 has become the default aspiration for serious local-AI builders, and why it is worth understanding in detail before you buy. The sections below unpack each spec, translate it into real-world model support and speed, and — crucially — help you decide whether you actually need this much card.

What is Blackwell? The architecture {#blackwell}

The 5090 is built on NVIDIA's Blackwell architecture (the GB202 die), the generation after the RTX 40-series' Ada Lovelace. For gaming, Blackwell brings new ray-tracing and DLSS features; for AI, three things matter.

First, GDDR7 memory. This is the new memory standard that gives the 5090 its enormous 1,792 GB/s bandwidth — a big jump over the 4090's GDDR6X at 1,008 GB/s. Since bandwidth sets generation speed, this is the single most important AI upgrade in the card.

Second, fifth-generation Tensor Cores with native FP4 support. Blackwell can do math in 4-bit floating point (NVFP4) natively, roughly doubling the throughput of low-precision workloads compared to the previous generation. As the ecosystem adopts FP4 inference and training, the 5090 is positioned to accelerate it in hardware rather than emulating it — a forward-looking advantage that will matter more over the card's lifetime.

Third, 32 GB of capacity. Blackwell's flagship finally breaks past the 24 GB that capped consumer cards for years, which — as the what-models section shows — unlocks the 32B model tier and single-card 70B at low quant.

The rest of Blackwell's feature list (ray tracing, DLSS 4 frame generation, display outputs) is largely irrelevant to AI. That is the recurring theme of buying a GPU for local models: the AI-relevant specs are memory and bandwidth, and Blackwell advances exactly those. If you want to see how the generations stack up numerically, every card's specs are on the GPU benchmarks page.

Why these specs matter for AI {#why-specs-matter}

Gaming benchmarks and AI benchmarks reward completely different things, and buying a GPU for AI using gaming logic is how people waste money. For local LLMs, two things matter, in this order:

1. VRAM decides what you can run at all. A model must fit in VRAM to run fast. If it does not fit, part of it spills to system RAM and runs 10× slower — the single worst outcome in local AI. The 5090's 32 GB is its headline feature: it clears the 24 GB ceiling of the 4090, which is exactly enough to comfortably hold a 32B model at 4-bit with a large context, or to squeeze a 70B model at aggressive quantization. That extra 8 GB over the 4090 is not incremental; it moves you into a new model tier. 2. Memory bandwidth decides how fast it runs. Once a model fits, generation speed is governed almost entirely by how fast the GPU can stream the model's weights through its cores — memory bandwidth. Not compute. This is the counterintuitive heart of local AI performance, and it deserves its own section.

Notice what is absent from that list: raw compute (TFLOPS), CUDA core count, and clock speed. These matter for training and for the "prefill" phase (reading your prompt), but for the token-by-token generation that dominates a chat, bandwidth is king. That is why the benchmarks page predicts speed primarily from bandwidth, and why a card's AI performance often does not track its gaming performance. If you take one idea from this guide, take that one — it is developed further in VRAM Requirements Explained.

To make this concrete: imagine two cards, one with more compute cores but 24 GB and 1,000 GB/s, and the 5090 with 32 GB and 1,792 GB/s. For gaming, the core-heavy card might win. For running a 32B language model, the 5090 wins twice over — it fits the model that the 24 GB card cannot hold fully, and it streams tokens ~78% faster thanks to bandwidth. This is why seasoned local-AI users learn to read a spec sheet from the bottom up: VRAM first (can it hold my model?), bandwidth second (how fast?), and everything else a distant third. Internalize that ranking and you will never overpay for the wrong card again — the entire philosophy behind our Best GPU Buyer's Guide.

The single most important number: memory bandwidth {#bandwidth}

Here is why bandwidth rules. To generate each token, the model must read all of its weights from VRAM. A 4-bit 8B model is about 4.7 GB; to produce one token, the GPU reads all 4.7 GB. To produce 100 tokens, it reads that 4.7 GB one hundred times. So the theoretical ceiling on tokens per second is roughly bandwidth ÷ model size in memory.

Real-world efficiency is around half the theoretical ceiling (overheads eat the rest), which is why the 5090's ~1,792 GB/s yields ~213 tok/s on an 8B model rather than the ~380 the naive math suggests. But the ratio holds: the 5090's 78% bandwidth advantage over the 4090 translates into real speedups, and — crucially — that advantage grows with model size. On a small 8B model both cards are fast enough that the gap is ~30%; on a big 70B model, where memory bandwidth is the binding constraint, the 5090 pulls ~70% ahead. If you mostly run large models, the 5090's bandwidth is the reason to buy it. This roofline model is exactly what the site's benchmarks methodology uses to estimate speed for any GPU-and-model pair.

What models can the RTX 5090 run? {#what-models}

Thirty-two gigabytes is a genuinely powerful amount of VRAM. Here is what fits, at the default Q4_K_M quantization, with headroom for context.

The practical picture:

• Small and mid models (8B–14B) fly. At 200+ tok/s the text appears faster than you can read it, with gigabytes to spare for enormous context windows.
• 32B models are the sweet spot. Qwen 3 32B and DeepSeek R1 32B fit fully at ~20 GB with room for a big context — near-frontier reasoning at high speed, on one card. This is the tier the 5090 was made for.
• MoE models run beautifully. A Qwen 3.5 35B-A3B uses ~21 GB but only computes ~3B per token, so it is both smart and fast.
• 70B is possible with compromise. A 70B at full Q4 needs ~40 GB and will not fully fit, but at Q3 (~30 GB) it runs on the single card — the first consumer GPU that makes single-card 70B realistic. Quality drops a little at Q3, but the model runs.

For any specific model, the VRAM Calculator gives the exact number, Can I Run It? gives a yes/no, and the Hardware Analyzer ranks everything the 5090 can run by speed. To actually run them, see the Complete Ollama Bible or Complete LM Studio Bible.

Use the quantization headroom. Most cards force you to the smallest quant that fits; the 5090's 32 GB often lets you go the other way — running a model at a higher-quality quantization than a smaller card could. Where a 16 GB card runs a 14B model only at Q4, the 5090 can run it at Q6 or Q8 with room to spare, squeezing out a little more accuracy. For a 32B model, you might run Q4_K_M with a massive context, or step down in size and up in quant. This flexibility — choosing quality over just "does it fit" — is a quiet luxury of a big-VRAM card. The trade-offs are in Quantization Explained; the rule of thumb is that a bigger model at Q4 usually beats a smaller one at Q8, but with 32 GB you rarely have to make that compromise in the first place.

Real-world performance {#performance}

Benchmarks in tokens per second, at Q4_K_M, give the clearest picture of what daily use feels like. Remember that anything above ~10 tok/s reads faster than a person, so the 5090 is comfortably fast across the entire range it can hold.

The standout is the combination of speed and size: the 5090 runs a 32B reasoning model at ~50 tok/s, which is faster than most people can read a careful, thinking answer — so even DeepSeek-style step-by-step reasoning feels responsive. For live, community-measured numbers across many GPUs and models, and to submit your own, use the benchmarks platform. Our companion article RTX 5090 vs 4090 for local LLMs drills into the head-to-head.

There is a second speed dimension worth noting: prompt processing (prefill). When you paste a long document or a big codebase, the model must read all of it before answering — and unlike token generation, this phase is compute-heavy, so the 5090's ample compute (not just bandwidth) makes it fast here too. On a card with weak compute, a huge prompt causes a long pause before the first token; on the 5090, even a 100K-token context is processed quickly. This matters enormously for RAG and long-document work, where you routinely feed the model a lot of text. So the 5090 is fast in both phases: quick to read your prompt (compute) and quick to write the answer (bandwidth) — which is why it feels so responsive even on demanding tasks that would make a lesser card stutter. The context windows guide explains how prompt length affects this.

Setting up your 5090 for AI {#setup}

Getting the 5090 running for local AI is refreshingly simple once the hardware is installed. The steps:

1. Install the latest NVIDIA driver. Blackwell needs a recent driver; get the newest Game Ready or Studio driver from NVIDIA. On Linux, install the current proprietary driver for your distribution. Verify with nvidia-smi — it should report the RTX 5090 and 32 GB. 2. Install your runner. Ollama or LM Studio bundle the CUDA they need — you do not have to install CUDA separately for basic inference. For fine-tuning or custom PyTorch work, install a CUDA toolkit version that supports Blackwell. 3. Pull a model that shows off the card. A 32B is the 5090's sweet spot:

ollama run qwen3:32b        # 32B reasoning, fully on-GPU, ~50 tok/s
ollama run deepseek-r1:32b  # step-by-step reasoning at readable speed

4. Confirm it is fully on the GPU. Run ollama ps — it should read 100% GPU. If it shows any CPU, the model spilled (too big, or context too large); the Ollama tuning section explains the fixes, though with 32 GB you will rarely hit this on a 32B model. 5. Enable the speed wins. Turn on flash attention and KV-cache quantization (near-free on Blackwell) to fit even longer contexts. On the 5090 these let you run a 32B model with a very large context comfortably.

That is it — you now have one of the fastest private AI setups money can buy. Add Open WebUI for a polished interface and you have a personal ChatGPT that outruns many cloud endpoints.

Beyond LLMs: image, video and everything else {#beyond-llms}

The 5090 is not just an LLM card — its 32 GB and compute make it a do-everything local AI workstation. The same hardware that runs Qwen 32B also runs:

• Image generation. Stable Diffusion and newer image models love VRAM; 32 GB lets you run the largest models at high resolution, batch many images, and use heavy workflows in ComfyUI without memory juggling. This pairs naturally with Open WebUI's image generation.
• Video generation. Local text-to-video and image-to-video models are extremely memory-hungry; the 5090 is one of the few consumer cards that can run them at meaningful resolutions and lengths.
• Speech and audio. Whisper transcription, voice cloning, and text-to-speech all run fast, enabling a fully local voice assistant.
• Simultaneous workloads. With 32 GB you can hold a small LLM and an image model at once, or serve a model while you work — the headroom that a 12–16 GB card simply does not have.

This versatility is part of the value case: one 5090 replaces what used to be several specialized setups, and it does all of it privately and offline. If your interest is a single machine that handles the full spectrum of local AI, the 5090 is purpose-built for it.

RTX 5090 vs RTX 4090 vs the field {#comparison}

The 5090 does not exist in a vacuum. Here is where it sits against the cards people actually cross-shop.

Reading this table honestly:

• vs RTX 4090: the 5090 is ~30% faster on small models, ~70% faster on large ones, and adds 8 GB of VRAM (a new model tier) — for ~$400 more MSRP and 125 W more power. If you run big models or fine-tune, it is worth it. If the 4090's 24 GB already covers you, the 4090 remains a superb, slightly cheaper option. The full comparison is in RTX 5090 vs 4090.
• vs RTX 5080: the 5080's 16 GB caps you at ~14B models; the 5090 doubles that. Different leagues.
• vs used RTX 3090: a used 3090 gives you 24 GB for a fraction of the price and is the budget enthusiast pick — slower, but astonishing value.
• vs Apple Silicon: a maxed-out Mac's unified memory can hold bigger models than a 5090 (up to 128 GB), quietly and efficiently, but at lower bandwidth so slower per token. For the largest models on one machine without a power-hungry rig, Apple is compelling — see Best Mac for Local LLMs.

To weigh any two cards yourself, the GPU benchmarks page and Best GPU Buyer's Guide are the tools; if budget is the constraint, best GPU under $500 covers the other end.

The three cards people most often weigh against the 5090:

The 5090 vs datacenter GPUs (A100/H100) {#datacenter}

It is worth understanding where the 5090 sits relative to the datacenter cards you hear about — the A100, H100, and H200 that power cloud AI. The short version: for inference of models that fit, the 5090 is remarkably competitive; for training and the very largest models, datacenter cards are in a different league.

Datacenter GPUs win on three things: much more VRAM (an H100 has 80 GB; an H200 has 141 GB), so they hold far bigger models; NVLink, a high-speed interconnect that makes multi-GPU scaling efficient in a way consumer PCIe cannot match; and sustained reliability for 24/7 server duty. That is why frontier models are trained and served on them.

But they are eye-wateringly expensive (tens of thousands of dollars each) and not sold for desktops. For a single person running a model that fits in 32 GB, a 5090 delivers a large share of an H100's inference experience at a fraction of the price, because inference of a fitting model is bandwidth-bound and the 5090's bandwidth is genuinely high. The practical takeaway: you do not need datacenter hardware for local AI. If you occasionally need to run something bigger than 32 GB allows, renting an H100 by the hour for that job is far cheaper than owning one — and the Cost Calculator shows when renting beats buying.

Power, cooling and PSU requirements {#power}

The 5090's performance comes with real power demands that you must plan for — this is not a drop-in upgrade for a modest system.

For AI specifically, there is good news on running costs: inference does not peg the card at 575 W continuously. Generating text is memory-bound, so the GPU often draws well below its peak during decode; the full 575 W is more a training/gaming figure. Still, if you run models many hours a day, electricity adds up — the electricity cost of running LLMs 24/7 guide does the math, and you can even undervolt the card to cut power draw with minimal speed loss. Plan the PSU for the peak, but expect real-world AI power draw to be gentler.

Tuning for efficiency: undervolting

Here is a trick most 5090 owners running AI should use: undervolting. The card ships with voltage headroom for peak gaming stability, but AI inference does not need it. Using a tool like MSI Afterburner (Windows) or the equivalent on Linux, you can lower the voltage curve and/or set a power limit — say, cap the card at 400–450 W instead of 575 W — and lose only a few percent of inference speed while dramatically cutting heat, fan noise, and electricity use.

For a machine that runs models many hours a day, this is a meaningful win: quieter operation, lower temperatures (which helps longevity), and a smaller power bill, all for a negligible performance hit because inference is memory-bound rather than compute-bound. It also eases the cooling and PSU burden. If you plan to leave your 5090 serving models on a home server, undervolting is close to mandatory. The electricity cost guide quantifies the savings, which over a year of heavy use are far from trivial.

Building a rig around the 5090 {#building}

If you are building a machine with the 5090 as its heart, the rest of the parts should not bottleneck it — but for AI, they matter less than you might think.

• CPU: a mid-range modern CPU is plenty; LLM inference barely touches it once the model is on the GPU. Do not overspend here.
• System RAM: 32 GB is a comfortable minimum, 64 GB is nice if you want to run huge models with CPU offload or juggle several. It does not need to be fast.
• Storage: a fast NVMe SSD matters for loading models quickly (a 20 GB model off a slow disk is painful), but not for inference speed. Get 2 TB+ — models pile up.
• Motherboard: any board with a full-speed PCIe x16 slot. If you might add a second GPU later, plan the slot layout and PSU now.
• PSU: as above, 1000–1200 W quality unit.

The overall guidance is to put your money in the GPU and keep the rest sensible. A balanced PC build guide walks through complete parts lists at different budgets, and the reverse build tool can spec a machine around a specific model you want to run. If you would rather not build, prebuilt "AI" desktops with a 5090 exist, though you pay a premium.

One planning note if you might go dual-GPU later: check the motherboard's PCIe lane layout and slot spacing before you buy. Two triple-slot 5090s need physical room and adequate lanes; a board that drops the second slot to a slow x4 link, or spaces the slots so the cards suffocate each other, will hamper you. It is far cheaper to buy the right board now than to rebuild later. That said, most people never add a second card — one 5090 is enough for the vast majority of local-AI work.

A note on laptops. There are mobile "RTX 5090" laptop GPUs, but do not confuse them with the desktop card: the laptop version has far less VRAM (typically 24 GB or less) and much lower power and bandwidth. It is still a strong mobile AI option, but the desktop 5090's 32 GB and 1,792 GB/s are in a different class. If portability matters, a high-VRAM Mac or a mobile workstation is worth comparing — but for maximum local-AI capability, the desktop card wins decisively.

Running two 5090s: 64 GB of VRAM {#dual-gpu}

Because Ollama, LM Studio, and vLLM can split a model across multiple GPUs, two 5090s give you 64 GB of combined VRAM — enough to run a 70B model at full Q4 fully in VRAM, or a 122B MoE, on a single machine. This is genuine workstation territory at consumer prices.

The caveats are real, though: two 5090s means ~1,150 W of GPU power (a 1600 W PSU and serious cooling), a motherboard and case that fit two triple-slot cards, and the understanding that a second card lets you run bigger models, not necessarily faster ones — inference is split, not doubled in speed. For most people, one 5090 running a 32B model is the better experience than two running a 70B. But if you specifically need 70B-class models locally, dual 5090s are the most accessible path. The trade-offs, and when a second GPU actually pays off, are covered in Multi-GPU Inference: Worth It?.

On cost, a dual-5090 build lands around $4,000+ for the two cards alone, plus a beefier PSU, board, and cooling — real money, but still a fraction of a single 80 GB datacenter GPU, and it delivers 64 GB of fast VRAM you fully own. For a professional who needs 70B-class models locally every day, that math can work out favorably against years of cloud rental. For everyone else, it is worth pausing to ask whether a single 5090 running a 32B model — which is genuinely excellent — meets the need before doubling down. The Cost Calculator helps you compare a dual-card rig to cloud and single-card options honestly.

Fine-tuning on the 5090 {#fine-tuning}

Inference is memory-bandwidth-bound, but fine-tuning is where the 5090's raw compute and 32 GB VRAM also shine. The extra memory lets you fine-tune larger models, use bigger batch sizes, and fit longer training sequences than a 24 GB card — and Blackwell's fast low-precision support speeds up the math.

With QLoRA (quantized LoRA), the 5090 can fine-tune surprisingly large models on a single card — comfortably a 14B, and with care a 32B — teaching a model your domain, style, or task. This is the difference between renting cloud GPUs for every experiment and iterating freely on your own hardware. The beginner path is Fine-Tune Your First LLM in 1 Hour; the deeper methods are in Fine-Tuning with Datasets (SFT, LoRA, DPO) and how to fine-tune a local LLM. Pair it with a dataset and you can produce a specialized model that outperforms a much larger general one on your task — then convert it to GGUF and serve it in Ollama.

Common myths about the 5090 for AI {#myths}

A few persistent misconceptions lead people to buy the wrong card — or the right card for the wrong reasons.

"More CUDA cores mean faster AI." Not for generation. Token generation is memory-bandwidth-bound, so the 5090's speed advantage comes from its GDDR7 bandwidth, not its core count. A card with more cores but less bandwidth would be slower at inference. Cores matter for training and prompt processing, not the streaming of tokens you watch on screen. "The 5090 will make any model fast." Only models that fit in its 32 GB. A 120B model still spills to system RAM and crawls. The card raises the ceiling to ~32B (or 70B at Q3), but it cannot defy the fit-in-VRAM rule — nothing can. Match the model to the memory. "I need a 5090 to run local AI." You do not. Excellent local AI runs on 8–16 GB cards; a Qwen 3.5 9B on an 8 GB card is genuinely capable. The 5090 is for people who specifically want the biggest models or the fastest speed. Many are better served by a cheaper card — see best GPU under $500. "It'll cost a fortune to run." Less than you think for AI. Inference is memory-bound and often draws well under the 575 W peak, and undervolting cuts it further. It is not a space heater running flat-out during a chat. "Gaming reviews tell me how good it is for AI." They do not. Gaming performance tracks compute and rasterization; AI inference tracks VRAM and bandwidth. A card can be a modest gaming upgrade but a large AI one, or vice versa. Judge it on the AI benchmarks, not frame rates.

Clearing these up saves money and disappointment. The recurring lesson — VRAM and bandwidth are what matter — is the thread running through the whole Best GPU Buyer's Guide.

Is the RTX 5090 worth it? {#worth-it}

The honest answer depends entirely on what you do. Here is a straight framework.

There is also the cloud comparison to weigh. A 5090 rig is a real upfront cost, but if you use AI heavily, it pays for itself against per-token or per-hour cloud pricing — and gives you privacy and no usage limits. If your usage is light or bursty, renting a GPU by the hour or using a cloud API may be cheaper. The Cost Calculator and cloud vs local comparison let you plug in your usage and see the break-even. For most serious enthusiasts and professionals who run local AI daily, the 5090 is the "buy once, cry once" card that removes hardware from the equation for years.

Who the 5090 is really for {#profiles}

Beyond the buy/don't-buy checklist, it helps to see yourself in a profile. The 5090 is the right call for some people and an expensive mismatch for others.

• The serious enthusiast. You run local AI daily, want the best experience, and value never thinking about whether a model fits. The 5090 is your card — it makes 32B models effortless and removes hardware friction for years.
• The AI developer. You build apps, test models, and need fast iteration on your own machine. The 32 GB lets you run big models and fine-tune without renting cloud GPUs for every experiment. A clear yes.
• The content creator. You use image and video generation alongside LLMs. The 5090's VRAM is one of the few consumer options that handles the largest generative-media models — a strong fit.
• The researcher / student in ML. You need to fine-tune, run experiments, and understand models hands-on. The 5090 is a capable personal research rig, cheaper than cloud over a long project.
• The privacy-focused professional. You handle confidential data — legal, medical, financial — and cannot use cloud AI. A 5090 running a 32B model on-premise gives you frontier-adjacent capability with total data control, the case in private AI for business.
• The small business. One 5090 box with Open WebUI can serve a small team a private assistant, replacing per-seat subscriptions. Size it against the Cost Calculator.

And who should not buy it: casual users, people who mostly run models up to 14B, anyone on a tight budget (a used 3090/4090 or a 5080 is better value), and those who prioritize silence and low power (a Mac may suit better). There is no shame in a cheaper card — the best GPU is the one that runs your models well, which the Hardware Analyzer will tell you.

Buying advice {#buying}

A few practical notes on actually acquiring one:

• Price reality. The $1,999 MSRP is often optimistic; real-world prices vary with demand, and premium models (better coolers, factory overclocks) cost more. For pure AI, the cheapest 5090 that fits your case and has a solid cooler is fine — you are buying the VRAM and bandwidth, which are the same across models.
• Availability. Flagship cards can be scarce at launch and around demand spikes. If you are watching for a price or a restock, set up GPU alerts.
• New vs used. The 5090 is new enough that used deals are rare; the used market is where the 3090 and 4090 shine (buying a used GPU).
• Founders vs partner cards. For AI it does not matter much; prioritize cooling and physical fit over brand.
• Don't overpay for gaming features. Ignore ray-tracing and gaming-overclock marketing; for AI you care about the 32 GB and 1,792 GB/s, which every 5090 has.

If you are not sure the 5090 is the right tier, run your target models through the Hardware Analyzer against a few cards first, or read the full Best GPU Buyer's Guide — sometimes the honest answer is that a 4090, a used 3090, or a Mac serves you better for less.

On timing. There is always a "next card" rumored, and it is easy to wait forever. The pragmatic view: the 5090's 32 GB and 1,792 GB/s are enough to run the models people actually run locally today extremely well, and model efficiency keeps improving, so a 5090 bought now will stay capable for years. If you have a concrete need — you are hitting VRAM walls, renting cloud GPUs repeatedly, or starting a project — buying now and getting years of use beats waiting for a hypothetical. If your current card already runs your models fine, there is no urgency; upgrade when you hit an actual limit, not on spec-sheet FOMO. On resale. High-end NVIDIA cards hold their value comparatively well, and demand from both gamers and AI users supports the used market. That softens the effective cost of ownership: a 5090 used for two years and sold on has a real residual value, unlike a cloud subscription, which leaves you nothing. Factor that into the buy-vs-rent math in the Cost Calculator.

Capstone: the ultimate single-card AI workstation {#capstone}

Let us put it all together into a concrete blueprint — the machine a serious local-AI enthusiast builds around a 5090.

The hardware. A 5090, a mid-range modern CPU, 64 GB of system RAM, a 2 TB NVMe SSD, a quality 1200 W ATX 3.1 PSU, and a case with strong airflow. Money in the GPU, sensible everywhere else — the PC Build Guide has full parts lists. The software stack. Ollama as the engine (or add vLLM if you will serve others), Open WebUI as the interface, and a local embedding model for RAG. Optionally ComfyUI for image generation and Whisper for voice — all of which the 32 GB card runs alongside a language model. The models. Qwen 3 32B or Qwen 3.5 27B as the daily driver (fast, smart, fits with huge context), DeepSeek R1 32B for the hardest reasoning, qwen3-coder for programming, and a small fast model for instant tasks. A 70B at Q3 for the occasional heavyweight job. What you get. A private, offline, no-subscription AI platform that runs 32B reasoning models at ~50 tok/s, chats with your documents, generates images, transcribes voice, powers your coding assistant, and lets you fine-tune your own models — reachable from every device in your home over Tailscale. It is, genuinely, a personal AI lab. The upfront cost is real, but so is the payoff: no per-token fees, no data leaving your walls, and no capability locked behind someone else's paywall. Run your intended models through the Hardware Analyzer to confirm speeds, and compare the total cost to years of cloud subscriptions with the Cost Calculator.

Frequently asked questions {#faq}

Is the RTX 5090 good for AI?

It is the best consumer GPU for local AI, full stop — the most VRAM (32 GB) and the most bandwidth (1,792 GB/s) of any GeForce card, which together decide how big a model you can run and how fast. See its full profile.

How much faster is the 5090 than the 4090 for LLMs?

Roughly 30% faster on small models and up to ~70% faster on large ones, because its bandwidth advantage matters most for big, memory-bound models — plus it adds 8 GB of VRAM. Details in RTX 5090 vs 4090.

What size models can the RTX 5090 run?

Comfortably up to 32B at 4-bit with a big context, MoE models up to ~36B total, and a 70B at 3-bit quantization on the single card. Smaller models run extremely fast. Check any model with the VRAM Calculator.

Can it run a 70B model?

Yes, with a caveat: a 70B at full Q4 needs ~40 GB and does not fully fit 32 GB (it will offload and slow down), but at Q3 (~30 GB) it runs entirely on the card. Two 5090s (64 GB) run 70B at full Q4 comfortably.

What power supply do I need?

1000 W minimum, 1200 W recommended, ideally ATX 3.1 with a native 12V-2x6 connector. The card draws up to 575 W, though real-world AI inference often draws less. See electricity costs.

Is the 5090 worth it over a 4090 or used 3090?

If you run big models daily or fine-tune, yes. If a 24 GB card already runs your models, a 4090 or used 3090 is better value. Weigh it with the Best GPU Buyer's Guide.

Should I buy a 5090 or rent cloud GPUs?

Heavy daily use favors buying (privacy, no limits, pays off over time); light or bursty use may favor renting by the hour. The Cost Calculator shows your break-even point.

Does the 5090 help with fine-tuning?

Yes — its 32 GB and fast low-precision compute let you QLoRA-fine-tune larger models (up to ~32B with care) on a single card. See Fine-Tune Your First LLM.

Can the 5090 do image and video generation too?

Yes, and it excels at it — 32 GB of VRAM runs the largest Stable Diffusion models, heavy ComfyUI workflows, and memory-hungry video models that smaller cards cannot. It is a full local AI workstation, not just an LLM card. See the beyond-LLMs section.

Do I need Blackwell/FP4, or is it just marketing?

For today's inference, the practical benefits are the GDDR7 bandwidth and 32 GB capacity. Native FP4 is forward-looking: as FP4 inference and training tooling matures, the 5090 accelerates it in hardware. It is a genuine advantage over the long life of the card, not just a spec-sheet line.

Is the RTX 5090 overkill for a beginner?

Usually yes. A newcomer is better served starting on an existing GPU or a 16 GB card, learning what models they actually use, and upgrading if they hit a wall. The Complete Beginner's Guide and Hardware Analyzer are the right starting points.

How long will a 5090 stay relevant for AI?

Its 32 GB and high bandwidth should keep it a strong local-AI card for years — model efficiency keeps improving (smaller models get smarter, quantization improves), so what fits and runs well on 32 GB only gets better over time. It is a durable investment for someone committed to local AI.

Can I use the 5090 on Linux for AI?

Yes — Linux with a discrete NVIDIA GPU is the highest-performance local-AI platform. Install the current proprietary driver, confirm with nvidia-smi, and run Ollama or vLLM. Many serious local-AI setups run headless Linux.

Verdict and where to go next

The RTX 5090 is the no-compromise choice for local AI: 32 GB of VRAM opens up the 32B-and-up model tier, and 1,792 GB/s of bandwidth makes those models genuinely fast. It is expensive, power-hungry, and more than most people need — but for anyone who runs large models daily, fine-tunes, or simply wants the best, it removes hardware as a limiting factor for years. If your needs are lighter, a 4090, a used 3090, or a Mac will serve you well for less; the point is to match the card to the models you actually run.

Once you have your GPU, put it to work:

• Run models: the Complete Ollama Bible and Complete LM Studio Bible.
• Choose models: the Complete Qwen Guide and Complete DeepSeek Guide — both sing on a 5090.
• Build a private assistant: the Complete Open WebUI Guide.
• Plan the whole rig: the PC Build Guide and Best GPU Buyer's Guide.

---

See the RTX 5090's full profile and compatible models → GPU page Check which models it can run, at what speed → Hardware Analyzer Compare a 5090 rig's cost to the cloud → Cost Calculator