Okay, so the noise here is just unbelievable, but we gotta talk about this Intel Arc Pro B70 thing. It’s actually, like, a big deal for AI workloads, especially when you look at the price. I mean, Intel’s Arc Pro B70, the 32GB version, is priced at around $949 to $1000, right? That’s a reference price of $950. And it’s got 32GB of GDDR6 memory, which is a lot of memory for that price point.

Now, the big news, the actual headline, is its performance in the DeepSeek R1 AI LLM. The Arc Pro B70, in a quad-GPU setup, it was tested, and it actually beat NVIDIA’s RTX 5090D and the RTX 4090D. At 256 concurrency levels, the Arc Pro B70 GPUs were up to 7.5% faster than the 5090D. And it was 48.7% faster than the RTX 4090D. That’s a huge difference, especially for the 4090D. The Arc Pro B70 delivered up to 2320.76 tokens per second. That’s a very specific number, very strong.

The NVIDIA RTX 5090D, that card launched at $2299. And some of the full RTX 5090 cards, not the D version, they are selling for like $4,000, even up to $6,569.99. So the Intel card is literally a quarter of the price, or even less, compared to some of those NVIDIA options. The RTX 5090D has 32GB of GDDR7 memory, yes, that’s faster memory, but the Intel card is still pulling ahead in these specific AI benchmarks. The RTX 4090D, that one came out at $1599. It has 24GB of GDDR6X memory. So the B70 has more VRAM than the 4090D, which is a key factor for LLMs.

Intel’s Arc Pro B70 has 256 Xe Matrix Extensions Engines, those are the XMX engines, designed for AI acceleration. And it has 32 Xe-cores. The card’s TBP, the Thermal Board Power, is 230W for the reference design. This is a dual-slot card, using a single 8-pin power connector. It launched on March 26, 2026, with Newegg listing it for pre-order on April 24, 2026. It even became the best-selling workstation GPU on Newegg back in April. That’s real market traction.

Puget Systems, a workstation builder, they also did some real-world benchmarks, showing a single Arc Pro B70 can comfortably run 8B-class models. For DeepSeek R1 Distill 8B, a single B70 hit 66.9 tokens per second with one user. And with a four-card setup, you get 128GB of VRAM, which is enough for 27B to 35B tier models in FP16 mode. That’s pretty impressive for the total cost.

This is a big deal because NVIDIA has dominated the AI hardware space, right? But Intel is making a real play here with a value proposition. The memory capacity on the Intel Arc Pro B70 is a huge advantage for these large language models. More memory means larger models can run locally, or more complex tasks can be handled.

And speaking of NVIDIA, the stock, NVDA, it moved up by 3.47% on July 8, 2026. This was because of news that China might let them sell high-performance H200 processors there. That’s a big market, a really big market, so any positive news out of China for NVIDIA is going to move the needle. But it also highlights the regulatory pressures on NVIDIA’s high-end AI chips. Intel’s play with the Arc Pro B70 is interesting because it’s not necessarily competing at the absolute top-tier performance of something like an H200, but it’s offering a very compelling price-to-performance ratio for accessible AI development. Why would you pay so much more for less performance in specific LLM tasks?

I mean, I bought AMD stock, ticker AMD, at $160.00 a share on January 15, 2024, because I thought they were undervalued in the AI space. I’m holding that until it hits $250.00, or if they totally screw up their next earnings call, then I’m out. This Intel news, it just shows how competitive the GPU market is getting, not just for gaming, but for AI, and that’s where the real money is right now. The landscape is shifting, and Intel is actually in the conversation. It’s not just NVIDIA and AMD anymore. This is a real threat to NVIDIA’s dominance in certain segments, especially for smaller businesses or researchers who need a lot of VRAM without breaking the bank. The power consumption of the RTX 5090D is 575W maximum, compared to the Arc Pro B70’s 230W. That’s also a significant difference for operational costs.