Did we just get the first **real** AI PC?
The Big News
The race for the best AI hardware just got real. Nvidia just announced a GPU with 96GB of VRAM, which is about 4x more than what Intel and AMD currently offer on the consumer level.
That might not sound exciting unless you’re into machine learning, but trust me, it completely changes what you can do with AI, locally, without the cloud.
For reference, with my M1 MacBook (24GB unified memory), I can run a pretty low level LLM with no problem, but if I wanted to chat with a private AI as good as ChatGPT, the laptop might explode before it could even generate an em dash.
Since this is such a big leap, we haven’t heard anything in response from Intel and AMD yet. Does this mean Nvidia is the only way to go to run high end LLMs privately? Let’s talk about it.
What Specs Do You Actually Need to Run Local AI?
For a baseline, we’ll be looking at the power needed to run a DeepSeek AI model in the DeepSeek Unchained app.
Absolute bare minimum
The least amount of VRAM you can get away with is unfortunately not likely to get you where you want to go. With only 4GB VRAM (DeepSeek Copper), your LLM can get caught in a spiral with its answers and then hallucinate you asking a question you never asked, which is just a frustrating waste of time.
So, technically the bare minimum is 4GB VRAM, but, in reality, you’re gonna need at least 8GB VRAM to start getting anywhere.
8GB+ VRAM can get you:
- Basic conversation abilities
- Simple task completion
- Very basic coding help
- Limited context retention
- Struggle with complexity
- Frequent inconsistencies
Optimal specs
A GPU with 16GB VRAM is workable, but we’re really starting to cook with 36+ GB VRAM. In DeepSeek Unchained, that gets you the Gold model, which is pretty great as far as local models go.
With DeepSeek Gold (36+ GB VRAM) you get:
- Strong overall performance
- Good reasoning and analysis
- Capable coding assistance
- Good context retention
- Handles complexity well
- Occasional minor inconsistencies
Cloud-quality AI, locally
This is where we get to the dream of running ChatGPT level AI completely off the cloud. With the RTX 6000, you can run DeepSeek’s best model locally, no cloud required, with performance on par with ChatGPT.
With this tank of a GPU, you can get the highest quality DeepSeek model available locally:
- Near ChatGPT level performance
- Excellent reasoning and analysis
- Strong coding abilities
- Consistent context retention
- Nuanced understanding of complex topics
- Reliable task completion
Which AI PC Setup Is Right for You?
| User Type | Recommended GPU | What It Can Run |
| Entry-Level | MSI GeForce RTX 4060 (8GB VRAM) | DeepSeek Bronze or Silver with light coding and chat |
| Power-User | NVIDIA RTX 4090 (24GB VRAM) | DeepSeek Gold for strong AI workflows |
| Pro Workstation | NVIDIA RTX Pro 6000 (96GB VRAM) | DeepSeek Platinum for offline ChatGPT performance |
Nvidia’s RAM is a game changer. AMD and Intel will follow suit if they have any sense at all. With 96GB of RAM anyone can run DeepSeek Unchained on their own server, and reap all the benefits of private AI. Even if you can’t afford the top of the line, you will be able to run some level of DeepSeek Unchained on your machine with an AI PC.
This new development from Nvidia is an exciting sneak peek into the future of private AI. We’re looking at the transition from concepts to reality in terms of what we can do with local LLMs.
