Nvidia is weighing a reduced-memory version of its Rubin Ultra chip, The Information reports. The idea marks a sharp reversal for a company that has pushed memory capacity upward on every data center flagship since the A100.

The report gives no final figure for the cut. It frames the change as an option under discussion, not a shipping decision.

Why would Nvidia ship less memory?

Every recent Nvidia flagship has raised the memory ceiling. The H100 carried 80 GB of HBM3. The Blackwell generation roughly doubled that with a two-die package. Rubin Ultra, expected in 2027, was widely expected to continue the climb.

More memory per GPU drives up cost in two places at once. HBM supply remains tight, and the number of stacks per package sets a floor on the bill of materials. A leaner configuration would cut both.

That matters for Nvidia's margin arithmetic. The company now sells most of its compute through a small number of hyperscale buyers. Those buyers negotiate hard, and they count total cost per token, not peak specs.

What does this change for buyers?

For cloud operators, the question is simple: does less on-package memory force more data movement across the interconnect? If yes, real workload throughput drops even as the per-chip price falls.

Three groups face different outcomes:

• Hyperscalers could gain a cheaper SKU for inference, where capacity per GPU matters less than aggregate bandwidth.
• AI labs training frontier models would likely stay on the full-memory part regardless of price.
• Server OEMs would need new board designs if the package layout changes.

The Information report does not specify which customer segment the reduced part would target.

A supply-side signal, not just a product tweak

The memory question also reflects the HBM market. SK hynix, Samsung and Micron control supply, and their allocation decisions now shape GPU roadmaps as much as transistor budgets do.

A company with Nvidia's bargaining power does not trim a flagship spec lightly. When it weighs a cut, the reason is usually that the supply chain cannot deliver the volume the demand curve wants.

Nvidia has not commented publicly. The Information attributes the plan to people familiar with the discussions.

Any decision would land far ahead of the Rubin Ultra launch window. Chip configurations freeze years in advance, so a change under discussion now signals planning for 2027 volume, not a near-term product.

For the broader market, the signal is the one worth watching: the binding constraint on AI compute has shifted from logic to memory, and even the market leader is now designing around it.