Intel has finally launched the GPU it has been teasing for months. The Arc Pro B70 — codenamed "Big Battlemage" — is a desktop card with 32GB of VRAM, 32 Xe2 cores, and a starting price of $949 for the Intel reference design.
The positioning is explicit: this is an AI card, not a gaming card.
What Makes the Arc Pro B70 Different
The spec that matters is the memory. 32GB of VRAM is the threshold that allows running larger language models locally without offloading to system RAM — a bottleneck that kills inference performance on consumer GPUs with 8-12GB.
At $949, the B70 sits below Nvidia's professional AI cards (which start significantly higher) but above consumer gaming cards. Intel is carving out a middle tier: serious AI inference capability without the enterprise price tag.
A B65 Pro variant with 20 Xe2 cores is also available, but only through partner manufacturers at varying prices.
The Timing
Intel's AI GPU push comes as the market for local AI inference hardware heats up. The rise of quantized models — smaller, more efficient versions of large language models that run on consumer hardware — has created real demand for GPUs with large VRAM pools.
Nvidia dominates this market, but supply constraints on high-VRAM consumer cards (the RTX 4090 with 24GB has been a favourite for local AI work) have created an opening.
The Software Question
Hardware specs are only half the story for AI workloads. Nvidia's CUDA ecosystem — its programming model, libraries, and tooling — is deeply embedded in AI development workflows. Most frameworks (PyTorch, TensorFlow, JAX) are optimised for CUDA first.
Intel's oneAPI software stack is functional and improving, but the ecosystem gap remains the biggest barrier to adoption. Developers who need to run existing AI workloads with minimal friction will still reach for Nvidia first.
The Arc Pro B70 is Intel saying it's serious about AI hardware. Whether developers follow depends on software support catching up.
