What VRAM is (plain English)
VRAM — Video RAM — is the memory on your graphics card (GPU). It's separate from your system RAM. AI models have to be loaded fully into VRAM before they can run; the GPU does the computation using the model data in VRAM.
Think of it like loading a book into your head to answer a question. If the book is small, you can hold it all in memory and answer quickly. If the book is bigger than your head's capacity, you have to go read a page at a time, which is dramatically slower. Modern AI image models are "big books" — the whole model has to fit in VRAM for the GPU to use it efficiently.
If the model is bigger than your VRAM, three things can happen:
1. The tool refuses to load the model at all (most common on modern tools) 2. The tool uses "offloading" to swap parts of the model to system RAM, which works but is 5-20x slower 3. The tool crashes with a `CUDA out of memory` error
The practical implication: your VRAM capacity determines which models you can run, which workflows are feasible, and how fast everything goes.
The 2026 VRAM tier map
| Your VRAM | What runs well | What struggles | What won't run |
|---|---|---|---|
| 4–6 GB | SD 1.5 (small, older model) | SDXL requires reduced settings | FLUX, Wan video, most modern models |
| 8 GB | SDXL comfortably, SD 1.5, FLUX Dev with offloading | Video generation, FLUX at full quality | Wan video, large LLMs |
| 12 GB | FLUX Dev, SDXL + LoRA, basic video gen | Long/high-res video, 70B text models | High-end video workflows |
| 16 GB | FLUX full quality, Wan video, LoRA training | Multi-model stacks | Largest LLMs |
| 24 GB+ | Everything local — all models, all workflows, concurrent tools | — | — |
Current (2026) sweet spot for a serious hobbyist is 16GB VRAM, which handles the mainstream image gen stack (SDXL, FLUX), modern video (Wan 2.2 at reasonable quality), and LoRA training without constant memory-management workarounds. 24GB enables the long tail (video at higher resolution, larger LLMs for text generation, running multiple models concurrently).
8GB is the practical floor for modern AI image gen. Below 8GB, you're restricted to older models (SD 1.5) or heavily-constrained use of SDXL. Still usable for basic image generation; you miss out on FLUX, Wan video, and most of what's new in 2026.
The three options when your VRAM isn't enough
1. Buy new GPU. $400-2500 depending on which VRAM tier you target.
- 8GB tier (RTX 4060, used RTX 3060 12GB): $300-500
- 12-16GB tier (RTX 4070 Super, used RTX 3090): $500-900
- 24GB tier (RTX 4090, RTX 5090, workstation cards): $1500-2500+
Plus the rest of the machine (power supply upgrade sometimes, case fit check, maybe a new CPU if you're building from scratch).
2. Rent GPU time. $0.40-2.00/hr on RunPod, Lambda Labs, Vast.ai, or similar. You spin up a cloud GPU instance when you want to generate; shut it down when done. Pay only for actual usage.
Typical rates: RTX 3090 24GB at ~$0.40/hr, A100 40GB at ~$1.20/hr, H100 at ~$2.50/hr. 20 hours/month of casual use = $8-40/month. 100 hours/month of serious use = $40-200/month.
3. Use hosted companion apps. $10-30/mo flat subscription. The hosted service runs the models on their hardware; you use any device with a web browser. No VRAM constraints on your end.
The pricing model is different from GPU rental: flat monthly fee regardless of usage, generally includes a monthly credit allowance for photo/voice/video output (which has per-generation cost on the vendor side).
The honest math on which option wins
Depends on your usage pattern. Rough breakeven calculations:
20 hours/month (casual hobbyist use): - Rent GPU: $8-40/month - Hosted companion: $10-30/month - Buy GPU: $600 investment, 2.5 years to amortize vs rental, 4+ years vs hosted - Winner: hosted or rental, roughly tied; depends on specific use case fit
100 hours/month (serious hobbyist / daily use): - Rent GPU: $40-200/month - Hosted companion: $10-30/month plus any output credit top-ups - Buy GPU: $600 investment, 6 months to amortize vs rental - Winner: hosted still competitive if the output-credit model fits; buy GPU wins at this volume if you value local privacy
300+ hours/month (near-constant use): - Rent GPU: $120-600/month - Hosted companion: $10-30/month base; may hit usage caps that require higher tiers - Buy GPU: $600 investment, 1-3 months to amortize - Winner: buy GPU — local economics flip in local's favor at this volume
Edge cases where buying GPU is almost always right: - Hard privacy requirement - Already have the hardware (zero marginal cost) - Enjoy the tooling itself (hours spent isn't a cost) - 3+ year usage horizon
Edge cases where hosted is almost always right: - Just starting out, testing whether AI generation fits your workflow at all - Primarily text-based use (AI companion chat doesn't need per-user GPU) - Chromebook or Mac without discrete GPU - Want it working tonight, not in 2-3 weekends
When buying a GPU is actually right
Four situations where buying hardware is the correct call despite the upfront cost.
1. Hobbyist who enjoys the tooling. If the setup and iteration process is part of what you got into AI for, buying the hardware is an investment in the hobby rather than a cost. The time spent fiddling with ComfyUI, trying new models, building custom workflows — these are features, not bugs. Pay once, enjoy it for years.
2. Privacy requirement. If "my data must not leave my machine" is a hard constraint, local is the only answer and no hosted service can match it. Budget for the hardware; it's the price of the privacy feature.
3. High volume. 300+ hours/month of use flips the economics in local's favor fast. If you're generating constantly (either for personal creative work at scale or for a business workflow), local GPU investment amortizes in 1-3 months vs rental and vs hosted-with-top-ups.
4. Specific technical needs hosted can't meet. Custom pipelines, unusual model combinations, batch processing with specific scheduler configurations, segmentation-driven workflows, research use cases. Hosted services make product decisions; local gives you the knobs.
When hosted is actually right
Four situations where hosted companion apps are the correct call.
1. You want the output, not the hobby. If your actual goal is "generate images" rather than "learn to generate images," hosted collapses the time-to-first-usable-output from 20+ hours to 15 minutes. The subscription cost is less than the opportunity cost of the setup hours for most working adults.
2. Low-to-moderate volume. Under 100 hours/month of use, hosted monthly subscriptions are cheaper than GPU investment and competitive with rental. The predictability (flat fee, no surprise costs) is a real feature.
3. No existing hardware. If you're on a Mac without discrete GPU, a Chromebook, an older laptop, or just a device that can't fit a modern GPU, the local option isn't available without buying a whole new machine. Hosted works from any device with a browser.
4. The specific use case is "AI companion with same persona across many scenes." AI companion apps productize per-persona LoRA training and the full chat + photos + voice + video stack. For this specific use case, hosted apps aren't just easier — they're genuinely better than what most users would assemble locally.
