Goldman Sachs estimates that memory costs now account for 62% of the materials bill for Nvidia's Vera Rubin superchip. That's a hefty slice, and it's not just a line item on a spreadsheet. The figure points to a quiet shift in the AI supply chain, one that could hand more leverage to memory makers and force Nvidia to rethink how it builds its next-generation hardware.
The memory bill inside Vera Rubin
The Vera Rubin superchip is Nvidia's upcoming platform, designed to power the next wave of AI training and inference workloads. But according to Goldman's analysis, the cost of the memory components—the high-bandwidth memory and related parts that feed data to the compute cores—now eats up nearly two-thirds of the total materials cost. That's a striking ratio, especially when compared to earlier chip designs where compute silicon dominated the bill.
Memory has always been a necessary cost, but it's becoming the dominant one. The estimate suggests that as AI models grow larger and demand more data throughput, the memory subsystem is no longer a supporting player. It's the main event, at least in terms of what Nvidia pays to put a superchip together.
Why memory makers stand to gain
If memory costs are that large, the companies that produce those components suddenly have more bargaining power. The facts point to a simple dynamic: when a supplier's product makes up 62% of your materials budget, that supplier matters. Memory manufacturers—the firms that make HBM and other advanced memory—could find themselves in a stronger position to set prices, secure long-term contracts, or influence design decisions.
That's a shift from the recent past, when compute logic was the scarce and expensive part. Now the balance is tipping. For Nvidia, that means production strategies may have to adapt. Locking in memory supply, negotiating better terms, or even redesigning chips to use memory more efficiently could become priorities.
What this means for the AI supply chain
The broader AI supply chain has been built around a few critical bottlenecks—advanced packaging, foundry capacity, and power delivery. Memory is now joining that list, and it might be the most consequential one. If memory costs stay this high, the entire economics of AI hardware changes. Margins get squeezed, prices for AI systems could rise, and the competitive landscape among chipmakers could shift.
Goldman's estimate is just one data point, but it's a telling one. It suggests that the next generation of AI infrastructure won't be defined solely by who designs the best compute cores. It'll also be defined by who controls the memory supply. That's a different game, and the players are already positioning for it.
For now, the question is how Nvidia responds. The company hasn't publicly detailed its memory procurement strategy for Vera Rubin, and it's unclear whether the 62% figure will hold as production ramps. But the estimate is a clear signal: memory is no longer a footnote in the superchip story. It's the headline.
