Dr. Ian Cutress, chief analyst at More Than Moore, describes semiconductor manufacturing as assembling a complete “matched set” of components. Every piece needs to be ready before a product can ship, which means Nvidia could have plenty of AI silicon waiting on HBM without freeing up a lick of useful manufacturing capacity for making more GeForce cards.
“Typically vendors have to make wafer orders from people like TSMC and wait six months from start of manufacturing to end,” Cutress told me. That’s six months just to get the silicon made, never mind lining up the right memory, packaging, and everything else needed to build a finished graphics card. Even if Nvidia decided to prioritize GeForce production tomorrow, those extra GPUs wouldn’t materialize overnight.
Nvidia spent decades pushing graphics memory technology forward, but that’s hardly what memory manufacturers are obsessing over these days.
“When I speak to memory vendors, there are only two words on their mind. HBM and DDR (or LPDDR),” Cutress says. “We never hear them talk about GDDR.”
A disassembled RTX 3090 Founders Edition showing off its GDDR6 memory. The more recent RTX 50-series is onto an advanced version of GDDR7.
Thiago Trevisan/IDG
GDDR isn’t necessarily disappearing or in short supply (though Cutress isn’t sure that the next generation of faster GDDR8 memory will arrive this decade after developing at breakneck speeds over prior generations), but it’s startling how little the memory technology that helped build Nvidia’s gaming empire comes up in Cutress’s conversations with the companies making it.
Don’t just take mine or Ian’s words for it — look at Nvidia’s spending commitments. In its August financial filing, the company disclosed $279 billion in supply and manufacturing-capacity commitments, mostly supporting its data center business. In October, Nvidia announced another AI manufacturing expansion, including a $1.1 billion advance supporting semiconductor testing facilities.