Intel has quietly become the third serious option for self-hosted AI. The Arc Pro B60 puts 24GB of VRAM on dedicated GPU hosting at £129/month – undercutting the 24GB RTX 3090 (£159) on price while bringing a modern architecture and a maturing software stack. For VRAM-bound inference on a budget, it is the new value entry point.
Table of Contents
Why a Third Vendor Matters
For years self-hosting meant NVIDIA, full stop, with AMD a distant second. A credible third option from Intel does two things: it pushes prices down at the 24GB tier, and it reduces the platform risk of betting everything on one vendor’s roadmap and allocation. The B60 is Intel signalling it intends to compete for the open-source LLM hosting buyer, not just gaming.
24GB at £129: What It Runs
The 24GB buffer is the headline. It is enough to comfortably run the models that make up the bulk of real self-hosting workloads:
- Llama 3.1 8B / Qwen 2.5 14B – full precision, with concurrency
- Mistral 7B, Gemma 2 9B – fast, with long context headroom
- 32B models at 4-bit – fits with room for a usable context window
- SDXL and Flux.1 – image generation without 16GB-card constraints
See the tokens per second benchmark for throughput across models and the benchmarks section for deeper comparisons.
Intel’s AI Software in 2026
The practical question with Intel has always been software. In 2026 the picture is much healthier: the oneAPI and IPEX-LLM stacks support the common inference paths, llama.cpp has a SYCL backend, and OpenVINO covers a growing set of models for optimised inference. It is not yet as turnkey as the CUDA ecosystem – you will hit fewer “just works” moments – but for mainstream LLM and image workloads the deployment path is real and documented.
24GB Inference, Lower Cost
Run mid-size models on a dedicated Arc Pro B60. Flat monthly pricing, full root access, UK hosting.
Browse GPU ServersB60 vs RTX 3090
| Arc Pro B60 | RTX 3090 | |
|---|---|---|
| VRAM | 24GB | 24GB |
| Price/month | £129 | £159 |
| Architecture | Modern (2025-class) | Ampere (2020) |
| Software maturity | Good, improving fast | Excellent (CUDA) |
| Best for | Budget VRAM-bound inference | Widest model/tooling compatibility |
The RTX 3090 still wins on ecosystem breadth – if you need a specific niche library or the absolute least friction, CUDA is unbeaten. The B60 wins on price and is the smarter pick when your stack is mainstream (vLLM, Ollama, llama.cpp, ComfyUI).
Verdict
The Arc Pro B60 is the most interesting budget development in the lineup. At £129 for 24GB it is the cheapest modern 24GB option, and Intel’s software has crossed the line from “experimental” to “viable for mainstream workloads.” If your stack is standard and your bottleneck is VRAM-per-pound, the B60 deserves a serious look.
Compare the full range in our GPU comparisons hub and follow hardware launches in the news section.
See also: the 2026 self-hosting GPU map, the 32GB Radeon AI Pro R9700, and open-source LLM hosting.
