What this is

Intel Arc Pro B60 Dual 48G surfaced this week on Swiss e-commerce site Digitec/Galaxus, with an all-in price around $3,000 (CHF 2,500). It's a dual-GPU workstation card, and the key spec is 48GB of VRAM—that number is just enough to run full-blooded 70B-parameter open-source models locally (think DeepSeek-V3, Qwen3).

Until now, getting this level of local VRAM basically meant choosing the NVIDIA RTX 6000 Ada or A6000, both priced above $8,000 per card. Intel's positioning here is explicit: lower the hardware barrier to "local LLM inference" and sidestep NVIDIA's CUDA ecosystem moat.

Industry view

The local-AI community's reaction is "watching, but not ready to pull the trigger." Original post author reto-wyss rated it directly as "not particularly competitive"—the deciding factor is price-to-performance; for it to truly shine, the price can't exceed 2.5x a single B60.

The bull case: Intel is finally pushing professional cards through retail channels. Earlier messaging was "OEM-only, not for individual buyers"—going retail publicly shows Intel is eager to build a user base. Keeping a non-CUDA option alive is structurally good news for companies handling sensitive data locally (finance, healthcare, government).

The bear case carries more weight: a $3,000 price tag + software stack far less mature than NVIDIA's (mainstream inference frameworks like vLLM and SGLang have only preliminary Intel GPU optimization; oneAPI/OpenVINO ecosystems aren't mature yet) + no CUDA compatibility layer means most enterprise IT won't add it to procurement lists in the short term. Local-AI hobbyists can buy it to "play around," but it's not production-ready.

Impact on regular people

  • For enterprise IT: Having a backup beyond NVIDIA is welcome, but if you're seriously considering procurement, we'd suggest waiting six months—see if pricing loosens and whether oneAPI/OpenVINO catches up.
  • For working professionals: AI engineers, take note—if your company won't approve an $8,000+ A6000, the dual B60 brings local experimentation costs down to roughly a third of that. But we recommend checking benchmarks first to confirm inference speed can actually beat an RTX 4090 before pulling the trigger.
  • For the consumer market: NVIDIA's solo dominance may get a corner pried loose. If Intel keeps shipping and improves the software stack, complete local-AI workstation bundles next year could drop 20–30% in price—a trend worth watching.