This article covers a hardware launch. Specs and pricing are sourced from manufacturer announcements and independent hardware press.
What’s New This Week
Intel launched the Arc Pro B70 on 25 March 2026 and it landed with a number that’s hard to ignore: $949 for 32GB GDDR6 at 608 GB/s. Nvidia’s closest competitor, the RTX Pro 4000 Blackwell, costs $1,800 and ships with 24GB. The Arc Pro B70 undercuts by ~47% and gives you more memory. That gap is large enough to change procurement decisions for inference workloads.
Changelog
| Date | Summary |
|---|---|
| 28 Mar 2026 | Initial publication. |
On 25 March 2026, Intel launched the Arc Pro B70 and Arc Pro B65. The gaming press noted both cards and moved on – they’re not for gaming. The inference story is worth examining more carefully.
The Numbers
The Arc Pro B70 has 32 Xe Cores running at 2800 MHz, delivering 22.9 TFLOPS FP32 and 367 TOPS INT8. The memory subsystem is the headline: a 256-bit bus carrying 32GB of GDDR6 at 19 Gbps gives you 608 GB/s of bandwidth. Power draw runs between 160W and 290W, with the reference design at 230W.
Price: $949.
For context:
- Nvidia RTX Pro 4000 Blackwell: $1,800, 24GB VRAM
- AMD Radeon AI Pro R9700: $1,299, 32GB VRAM
Intel undercuts both. Against Nvidia specifically, you get 33% more VRAM at 47% lower cost. Against AMD, you get the same 32GB at $350 less.
A second card is coming. The Arc Pro B65 matches the B70 on memory – 32GB, 608 GB/s – but reduces to 20 Xe Cores, 197 TOPS INT8, and a 200W typical power draw. Price is TBA, with an expected mid-April launch. The B65 looks like the more interesting card for inference-only workloads where raw FP32 throughput matters less than memory capacity and bandwidth.
Why Memory Matters for Inference
Running large language models locally is almost always a memory problem before it’s a compute problem. Model weights have to fit in VRAM. Quantised 7B models are manageable on 8GB or 16GB cards. Getting to 13B, 30B, or larger – especially in quantisations that preserve quality – requires 24GB or more. Context windows compound this: the KV cache grows linearly with sequence length, and that growth all lands in VRAM.
608 GB/s of bandwidth matters for time-to-first-token. Memory bandwidth determines how fast you can move weights and activations, which directly affects latency on long context requests. Intel claims advantages in cost-per-token and time-to-first-token, particularly for large context windows – though independent benchmarks on inference-specific workloads are still thin.
Multi-GPU scaling is supported, which matters for anyone building inference infrastructure rather than single-workstation deployments. Two B70s would give you 64GB of pooled memory for under $2,000.
The Catch
Intel’s GPU software ecosystem is not Nvidia’s. CUDA dominates inference frameworks. Most production inference tooling – vLLM, TensorRT-LLM, llama.cpp with CUDA backends – is written and optimised for Nvidia first. Intel has OpenVINO and Level Zero, and llama.cpp supports SYCL for Intel Arc, but the maturity gap is real.
This isn’t a new problem for Intel GPU compute, and it’s a legitimate reason to wait before committing budget. Driver stability and software support on the Arc Pro line need scrutiny before production deployment.
That said, the price point changes the calculus for experimentation. At $949, a B70 is a low-stakes test. If OpenVINO support for your inference stack is adequate – and for many workloads, it will be – you have 32GB of bandwidth-rich VRAM for significantly less than the Nvidia alternative.
What It Means
The premium Nvidia charges for VRAM has been the single largest barrier to local inference on workstations and small servers. That premium is now explicitly under competitive pressure. AMD has been pushing on this front with the Radeon AI Pro line, and now Intel is undercutting both of them.
Whether the Arc Pro B70 is the right choice depends on your inference framework and tolerance for software ecosystem risk. But the pricing forces a conversation that wasn’t happening six months ago: you can now spec a local inference box with 32GB of dedicated VRAM and meaningful bandwidth for under $1,000.
If you’re building out local inference infrastructure, the local AI hardware comparison and build guide are worth reviewing alongside this launch. The B70’s spec sheet fits comfortably into the upper tier of what those pieces recommend – the software ecosystem question is the variable that needs resolving.
Watch the B65 pricing in April. If Intel prices it below $800, the inference value case gets stronger still.