5 Common Traps Beginners Fall Into When Buying an AI Computer
Don't waste money on 128GB of regular RAM or weak 8GB graphics cards. Avoid these 5 common hardware mistakes when configuring your first local AI workstation.
Buying hardware for local AI is very different from building a standard gaming PC or video editing workstation. What works well for Call of Duty or Premier Pro can be a total mismatch for running large language models like Llama 3 or DeepSeek.
Here are the 5 most frequent mistakes first-time buyers make and how to avoid them.
Trap 1: Spending Big on Regular System RAM Instead of GPU VRAM
We often see beginners configure a desktop with an AMD Ryzen 9 CPU, 128GB of DDR5 system RAM, and an 8GB NVIDIA RTX 4060 graphics card.
The problem? Regular DDR5 system RAM runs at roughly 50 to 80 GB/s bandwidth. In contrast, modern GPU VRAM (GDDR6X/GDDR7) runs at 1,000 to 1,792 GB/s—over 15 to 20 times faster!
If your model has to run on regular system RAM, output speeds drop from 30 words per second to 1 or 2 words per second. Always prioritize maximum GPU VRAM over excessive system RAM (unless buying an Apple Silicon or AMD APU system where system memory is high-bandwidth and unified).
Trap 2: Falling for High TFLOPS with Low VRAM
Retailers love to advertise raw GPU compute power in TFLOPS (Teraflops). But in local LLM inference, compute power is useless if the model doesn't fit into the card's memory in the first place.
For example, a fast 12GB graphics card cannot run a quantized 32B model, regardless of how many TFLOPS it has. Meanwhile, an older RTX 3090 with 24GB of VRAM will run that 32B model effortlessly. In local AI, memory capacity always trumps raw compute clocks.
Trap 3: Overlooking Power Supply (PSU) and Thermals
High-end AI workloads run graphics cards at 100% load continuously for minutes or hours at a time, generating immense sustained heat.
- If buying an RTX 4090 or RTX 5090 system, ensure the power supply is at least 850W to 1000W with native ATX 3.0 / PCIe Gen 5 12VHPWR connectors.
- Multi-GPU setups (like dual RTX 3090s) produce 700W+ of heat. Ensure your PC chassis has unrestricted airflow, or consider blower-style coolers or unified memory workstations (like Mac Studio or DGX Spark) which run virtually silent.
Trap 4: Ignoring the Power of Apple Silicon Unified Memory
Many lifelong Windows/Linux enthusiasts reflexively dismiss Apple Macs. However, Apple's unified memory architecture is a game-changer for local AI inference.
A Mac Studio with an M2 or M4 Max/Ultra chip allows you to assign up to 75-80% of its total 64GB, 128GB, or 256GB unified RAM directly to the GPU cores at 800+ GB/s bandwidth.
To achieve 128GB of fast memory on an NVIDIA PC, you would need four enterprise GPUs costing £10,000+. A unified memory Mac lets you run 70B parameter models on your desk in dead silence for a fraction of the cost and power.
Trap 5: Overbuying on Day One
You do not need a £10,000 server rack or four GPUs to get started with local AI. Modern 8B models (such as Llama 3.1 8B or Gemma 2 9B) are extraordinarily capable at everyday writing, brainstorming, summarization, and lightweight coding assistance.
A sensible £1,000 to £1,800 starter rig with 16GB to 24GB of VRAM will let you experiment, learn the software stack, and understand your real-world usage patterns before committing thousands of pounds to heavier hardware.
Check our £/GB VRAM Index to compare real UK prices and value across all tiers.