NVIDIA to Build GPUs on Intel Foundry from 2028: Report

On the specs front, “Feynman” isn’t here to play small. NVIDIA’s eyeing some ridiculous new memory standards—think HBM4e or maybe even HBM5 if they can pull it off—and an insane amount of memory capacity per chip. Rumor has it, this beast will be able to train those trillion-parameter AI models everyone in Silicon Valley drools over.
Now, here’s the kicker. Nvidia’s got money in Intel (they bought a 5% stake recently), but most of the action is still happening at TSMC. So it’s not a full marriage, more like an open relationship where Intel gets to handle some bonding (literally, chiplet interconnects) and packaging, possibly in the US. But what’s the downside? Well, splitting manufacturing like this can lead to headaches (integration, delays, or even finger-pointing when things go sideways) and you just know it’s going to make these GPUs even harder to get—or more expensive—when launch time arrives.
Don’t kid yourself—this Feynman thing isn’t for gamers who just want better frame rates. It’s aimed squarely at AI overlords, research labs, and anyone else with a data center budget that’d make your head spin. Average folks aren’t going to see one of these unless their boss’s boss’s boss signs the check. Even then, good luck finding one in stock.
As for the UAE? If you’re reading this in Dubai or Abu Dhabi thinking, “am I ever going to see this at my local store?”—don’t hold your breath. But, with how fast the region’s appetite for AI and cloud is growing, you might see this hitting shelves at retailers like Datcart sooner than you’d expect (assuming supply chain chaos doesn’t mess it up… again). Price? Don’t even ask. NVIDIA’s keeping that close for now—and we all know what that means.