NVIDIA Rubin Ramping Up: Cloud Giants Lock In All Memory Supply Through 2027 via Long-Term Agreements
Claire Weston
Cloud providers have locked up all available 2027 long-term memory capacity and begun negotiating 2028 supply; Nvidia's Vera Rubin ramp in H2 2026 will tighten the memory squeeze through 2027, with shortages expected to exceed 2026 levels.
Why is 2027 memory capacity already spoken for?
Nvidia's next-gen AI accelerator Vera Rubin is set to ramp in H2 2026. To secure supply, major cloud providers have locked up all available 2027 long-term agreement (LTA) capacity.
This means → 2027 memory output is fully claimed before production begins. Non-AI customers face an extremely narrow margin of available supply.
Multiple memory-module makers have received OEM notices: no additional supply beyond previously committed volumes.
Why are 2028 negotiations starting this early?
Suppliers had refused to discuss 2028 capacity before May 2026, but have recently begun accepting 2028 Q1 orders. Some HBM — high-bandwidth memory designed specifically for AI chips — and server capacity has already been provisionally allocated.
In plain terms = cloud providers suffered through the 2025–H1 2026 shortage and would rather queue two years early than wait passively again.
Most LTAs follow a structure: customer commits expected volumes → supplier adjusts expansion plans → final pricing is set only before shipment. Few require upfront deposits.
Will cloud providers fund dedicated memory fabs?
Rumors had surfaced that cloud providers might subsidize memory makers to build dedicated lines and cover equipment costs.
Industry insiders say upstream suppliers are already highly profitable; new fab and equipment budgets are self-funded, with no need to depend on a single customer.
This reflects a fundamental shift: memory makers hold stronger pricing power in this AI cycle than at any previous point.
How much supply space is being squeezed from non-AI buyers?
Server memory and standard memory — the kind used in ordinary PCs and phones — share highly similar architectures. Prioritizing AI means standard and PC memory bear the brunt of capacity reallocation.
Together, these segments account for roughly 60%–70% of total DRAM. PC memory's share has fallen from 11%–12% to about 9%.
Apple continues to lock Q3 capacity for product launches; other brands are forced to pull production schedules forward, compressing available supply further.
AI servers are downgrading configs — what does that signal?
128 GB had been the mainstream spec for AI servers, but some buyers have shifted to 64 GB or 96 GB to control costs.
This means → memory pricing is high enough that even AI buyers are cutting configurations. The shortage is not just a volume problem — cost is slowing the upgrade cycle itself.
When will this shortage cycle actually ease?
Jensen Huang recently confirmed that SK Hynix, Samsung, and Micron have all completed HBM4 qualification, begun production, and fully support the Vera Rubin platform.
He added that shortages need to be managed rationally across systems while supply scales up.
Put simply = the industry consensus is converging on one view: 2027 will be tighter than 2026. Whether conditions ease after 2028 depends on whether the three major memory makers can expand fast enough to match AI compute demand growth.
Content is for reference only, not financial advice.