How Birmingham AV adapts to the AI computing shift

Something quiet has happened to the definition of a capable PC over the past eighteen months. Local language models, in browser image generation, agentic assistants running under Ollama or LM Studio, and Copilot Plus features baked into Windows 11 have pulled the baseline up together. The Birmingham AV product line has been reworked around that shift, and this piece explains what changed.

What changed on the demand side

Two years ago the typical enquiry was for a general purpose office tower or a mid range gaming rig. Today a growing share of buyers name a specific model or tool: Llama 3.3, DeepSeek R1, Mistral Small, Stable Diffusion XL, Flux.1, ComfyUI, Ollama, or an agentic framework like AutoGen. Others do not name any of that, but they mention wanting a machine that will "still be useful in three years."

Both groups want the same thing underneath. Enough memory to hold a serious model, enough VRAM to run it at a useful quantisation, and enough fast storage to load weights without a coffee break.

The new floor: 32GB, 12GB, 2TB

The Birmingham AV AI ready product line commits to three numbers as standard, not as an upsell.

• 32GB of system RAM as standard, upgradeable to 64GB on the same board.
• 12GB of GPU VRAM as the absolute floor, with most builds shipping 16GB.
• 2TB of NVMe storage on PCIe Gen4, because model weights are large and slow storage bottlenecks everything.

These are not aspirational figures. They are the point below which the machine cannot honestly be called AI ready. A tower with 16GB of DDR4 and integrated graphics is still fine for spreadsheets, video calls, and web browsing. It will not run a 13 billion parameter model at a usable speed, and it will crawl the moment a Stable Diffusion checkpoint is loaded.

Why 12GB of VRAM is the honest cut off

Video memory is the hardest constraint in local AI work. A 7 billion parameter model at 4 bit quantisation lands around 4 to 5GB of VRAM, which sounds like plenty on an 8GB card until context length grows or a second workload joins. 12GB opens up 13B class models comfortably. 16GB is where SDXL, Flux.1, and 20B class quantised models sit without swapping to system RAM.

Buyers who cross the 12GB threshold report a very different experience. The machine stops feeling like it is compromising and starts feeling like it is capable.

The recommended build: Ryzen 7 plus RTX 5070 Ti

For the majority of buyers the answer is a Ryzen 7 paired with an RTX 5070 Ti. Complete builds in this configuration are listed on the Birmingham AV eBay store, with variations from 32GB up to 64GB of RAM and from a single 2TB NVMe to dual drive setups with additional 4TB SATA storage.

The core specification looks like this.

• Ryzen 7 5700X: 8 cores, 16 threads, 65W TDP, PCIe Gen4. Feeds a Blackwell class GPU without bottlenecking, on the mature AM4 platform where DDR4 memory is affordable.
• RTX 5070 Ti 16GB: best value tier for VRAM per pound in the current generation, with strong support in PyTorch, llama.cpp, ComfyUI, and Ollama.
• 32GB DDR4 3600MHz, headroom to 64GB on the same board.
• 2TB NVMe Gen4 SSD, typically a Samsung 990 Pro or WD SN850X class drive.
• 750W 80 Plus Gold PSU from Corsair, Seasonic, or be quiet!
• Mid tower case with mesh front, three intake fans, and a tower CPU cooler.

This is the machine that pulls double duty. It runs local inference and image generation without complaint, and it plays modern titles at 1440p high settings on the same evening.

What was dropped from the range

The shift has also meant retiring configurations that no longer make sense. Towers with 16GB of RAM as the top tier have been dropped from the AI ready line. GPUs below 12GB of VRAM are still stocked for buyers who explicitly want a general purpose or gaming only machine, but they are no longer marketed as AI capable, because they are not. SATA only primary storage has been phased out in favour of NVMe.

Cooling and sustained load

AI workloads pin the GPU at full load for long periods. This is different from gaming, where thermal peaks are short. A card that runs cool during a two hour raid will run hot during an eight hour training session, and heat kills components over time.

Every AI ready build ships with tower CPU cooling, at least three case fans in a balanced configuration, and a mesh front case. Fan curves are tuned for sustained load. This is the kind of detail that does not appear on the spec sheet but decides whether the machine is healthy in three years.

FAQ

Is 32GB of RAM really necessary if I am not running AI?

For pure office and light gaming use, 16GB is still workable. The 32GB standard reflects the reality that most buyers will, at some point in the next three years, run something that benefits from the extra memory. It is future proofing at a modest cost delta.

Can I add a second GPU later for larger models?

The recommended build has PSU headroom and PCIe lanes for a modest second card, and some buyers pair the 5070 Ti with a used RTX 3090 for 24GB plus workloads. Flag the intention at time of purchase so a suitable case and PSU are specified.

What about Copilot Plus and NPUs on the CPU?

The Copilot Plus certification currently sits on ARM based laptops and a small number of newer x86 chips with dedicated NPUs. For local model work at any useful size the GPU still does the heavy lifting, so NPU presence is a nice extra rather than a requirement.

Do I need Windows 11 Pro or will Home do?

Home handles everyday AI tools such as Ollama, LM Studio, and ComfyUI without difficulty. Pro is preferred where BitLocker, Hyper V, remote desktop, or joining a domain matter. Ubuntu 24.04 LTS is available on request.

About Birmingham AV

Birmingham AV has sold over 87,000 items on eBay since 2017, with 24,756 buyer feedbacks at 98.9% positive. Every system ships with a twelve month warranty. Birmingham AV Ltd is registered at Companies House under number 12383651, VAT registration GB 348755066, based in Bromsgrove, Worcestershire. One of the highest volume refurbished PC operations on eBay UK, we build every machine to a specification we would use ourselves.