Intel Reveals GPU D-Series AI Inference Chip With 480GB Memory and 350W Power Design - Konsulteer

Skyler Monroe (0:10)

Hey everyone, welcome back to The AI Hardware Show. I'm your host Skyler Monroe. And if you're into chips, silicon, and the hardware that's actually making the AI revolution happen, you're in exactly the right place. Big thanks to our sponsors keeping this show running. Ada, helping businesses integrate AI into their workflows. Ago Consulting, that's Ago, your go-to for silicon development from AI to SOAT and Zen Semiconductor. Building the processors and AI fabric behind modern data centers, with ventures like their Sierra RISC-V CPUs and Loom AI fabric, plus Ada for deploying AI at scale in any business. Today we've got a stacked episode, Intel's back with a serious inference chip, a startup just raised $350 million to take on Nvidia. A scrappy Chinese chip company is making some bold claims, TSMC is absolutely crushing it financially, and we're looking at what's driving the semiconductor equipment market. Let's get into it. All right, first up, Intel. And look, I know Intel has had a rough few years in the AI chip space. They've been playing catch up to Nvidia for a while now, and the gaudy story hasn't quite landed the way they hoped. But they're not sitting still, and this new announcement is worth paying attention to.

Intel has revealed what they're calling the GPUD series, a new AI inference chip, and the headline numbers are genuinely impressive. We're talking 480 gigabytes of memory and a 350 watt power design. Let me put that memory figure in perspective for you, because 480 gigabytes is a lot, like a lot a lot. When you're running inference on large language models, so we're talking about the process of actually using an AI model to generate responses, not training it, but running it, memory capacity is one of the most critical constraints. The model has to fit in memory. If it doesn't fit, you're doing all kinds of expensive workarounds like model sharding across multiple chips, which adds latency and complexity.

So when Intel says 480 gigabytes on a single chip solution, they're basically saying you can run some seriously large models without breaking them apart across multiple devices. To put that in context, Nvidia's H100 comes in 80 gigabyte configurations, and even the HBM packed versions of their latest hardware Macs out well below what Intel is claiming here.

Now, the type and speed of that memory matters a lot, too.

We don't have all the bandwidth specs yet, and bandwidth is really what determines how fast you can actually move data to and from that memory. Think of it like this. Memory capacity is the size of your desk, but memory bandwidth is how fast you can grab things off that desk and actually use them. You can have a huge desk, but if you're moving in slow motion, you've got a problem. The 350 watt TDP is interesting, too.

That puts it in a similar power envelope to high-end data center, GPOs, which is good. It means it can slot into existing data center power and cooling infrastructure without requiring a complete overhaul. Some of the more exotic AI accelerators out there are pushing 500, 600, even 700 watts, and that creates real deployment headaches for operators. 350 watts is aggressive but workable. Intel is clearly positioning this as a practical deployable inference solution, not just a benchmark monster. The inference market specifically is huge and getting bigger.

As more companies move from training models to actually running them in production, serving users, powering applications, the demand for efficient inference hardware is exploding. Training gets a lot of the press because it's dramatic and expensive. But inference is where the volume is. Every single query to an AI assistant, every image generation, every recommendation, that's inference. And if Intel can carve out a real slice of that market, it would be a meaningful comeback story for them. I'm watching this one closely. More details will come, but the D-Series has my attention. Okay, story two. And this one is big in a different way. It's about money and market dynamics. TensorWave, a data center infrastructure startup, just raised $350 million.

Their stated mission, help break Nvidia's monopoly on AI chips. Bold words. Let's talk about what's actually going on here. So TensorWave is building out data center infrastructure centered around AMD's Instinct GPUs, specifically the MI300X and newer variants. They're not designing their own chips. They're essentially betting that AMD's hardware, combined with the right software stack and data center operations, can offer a compelling alternative to Nvidia-based cloud AI services. And $350 million is a serious vote of confidence from investors that this is a real opportunity. Now, why does this matter? Because Nvidia's grip on the A, iCompute market is genuinely extraordinary. We're talking about market share in the 70 to 80% range for AI training and inference accelerators.