VMware targets costly DRAM with new memory tiers

VMware is expanding its memory-tiering technology as high DRAM and server prices lead customers to postpone hardware purchases. The company says the feature can reduce CPU consumption by 30 percent and total ownership costs by 40 percent, but it requires enterprise-grade NVMe drives and trades some DRAM capacity for storage latency.

Memory tiering moves infrequently used “cold pages” out of DRAM and onto NVMe. When a virtual machine needs more memory than the host has available, the hypervisor can place those less-active pages on the NVMe tier instead of forcing the CPU to wait for memory to be freed. VMware’s implementation is built into its ESX hypervisor and is designed to keep applications running while reducing the amount of physical DRAM required per server.

The technology first appeared in Cloud Foundation 9, VMware’s private-cloud software bundle. The VCF 9.1 update adds database performance improvements, dashboards showing how much tiering is taking place, and additional management controls. VMware says the feature currently supports about 75 percent of workloads.

Recommended reading

Exchange Online outage disrupts email and authentication

Marcus Vance • Aug 31, 2026 • 4 min read

The storage requirements are substantial

VMware allows more than one memory-to-NVMe allocation, but its recommendation is considerably more conservative than the maximum supported ratio:

The NVMe devices must support at least 100,000 writes per second and a lifetime write rating of 7,300 TB. VMware also recommends a minimum endurance rating of three Drive Writes Per Day (DWPD). That requirement rules out using ordinary consumer SSDs as a drop-in extension of server memory. The storage tier has to tolerate sustained write activity without becoming the new bottleneck or wearing out prematurely.

VMware’s claimed savings also depend on how effectively workloads can use the tier. The company says memory tiering can let customers consolidate workloads onto fewer servers, reducing CPU cycles by 30 percent and total ownership costs by 40 percent. Those figures are VMware’s claims; the supplied reporting does not provide independent benchmarks, test conditions, workload sizes, or a breakdown of whether the calculation includes the cost of high-end NVMe drives.

If DRAM costs several times more than the servers it is installed in, adding fast storage may be cheaper than adding memory or buying another host. That does not make the storage tier equivalent to DRAM. The trade-off is attractive only when a workload has enough cold memory to move without creating unacceptable latency.

“We have a two-to-three year roadmap,”

More VM types are planned

VMware intends to extend support beyond the current workload set. Morera said the roadmap includes Microsoft’s Radius cloud-native application platform and “monster VMs,” which VMware defines as virtual machines able to run with up to 960 vCPUs and 16TB of memory across multiple physical hosts.

The company expects some of those changes to arrive in VCF 9.2, which it will likely introduce around May 2027. That is a planned release window, not a shipping date confirmed in the supplied material.

Memory tiering already works with vMotion, VMware’s mechanism for moving virtual machines between clusters with almost no downtime. VMware plans to improve tiered-VM migration performance, although Morera described the remaining obstacles only as unidentified “physics problems.” The phrase leaves unresolved whether the limitation involves storage bandwidth, page movement, network transfer, or another part of the migration path.

Support for Radius and large multi-host VMs could broaden the feature’s usefulness, but those are also the workloads where memory placement and migration costs may matter most. A tier that works acceptably for lightly accessed pages in a conventional VM may be harder to manage when a large VM has high working-set churn or when vMotion has to move tier metadata and cold pages between clusters.

A new reason to stay with VCF

The timing gives VMware a reason to promote Cloud Foundation. Customers facing expensive DRAM and long server lead times can try to increase host density instead of immediately expanding their fleets. The approach may also extend the useful life of existing servers, according to VMware’s position, by allowing them to run more workloads without a matching increase in DRAM.

That argument arrives as some organizations are already reconsidering VMware’s role in their infrastructure. Foxconn is replacing VMware with Arcfra across factories and offices, making memory tiering potentially useful to customers who remain on VCF but less relevant to organizations actively moving away from the platform. Separately, the rapid exploitation of a VMware vCenter flaw across 47 countries has made the operational cost of staying current with the stack harder to ignore; our earlier coverage tracked that exploitation campaign.

VMware says customer interest is strong: two conference sessions on memory tiering filled 500-seat rooms. Attendance shows demand for an answer to the memory-cost problem, not that the claimed savings have been independently validated. The question is whether the NVMe endurance, capacity, and performance requirements make the tier cheaper in a customer’s actual configuration once the storage hardware and operational overhead are included.