vSphere Memory Tiering: How to Cut the Cost of Your Next Server Refresh
Server memory has become the most expensive line on a refresh quote, and in 2026 it is the line that keeps moving. vSphere memory tiering, built into VMware Cloud Foundation 9 and vSphere Foundation 9, lets a host use enterprise NVMe as a second tier of memory alongside DRAM. Done properly, it halves the DRAM you need to buy, or doubles the virtual machines a host can carry, with the workloads none the wiser. Here is how it works, what it needs, where it fits, and how to size it so the saving is real.
Memory tiering lets an ESX host treat local NVMe as memory. DRAM stays tier 0 for the pages a virtual machine is actively using, and NVMe becomes tier 1 for the pages it is not. Most workloads keep only 10 to 30 percent of their memory active, so at the default one to one ratio a host with 512 GB of DRAM and 512 GB of NVMe presents 1 TB of memory, and VMware's own worked example cut the cost of a 1 TB node by almost 40 percent. It is included in VCF 9 and VVF 9 at no extra charge, needs mixed use enterprise NVMe rated at 3 drive writes per day or better, and is not in vSphere Standard or the reported vSphere Express. If you are about to buy servers, size the memory with tiering in the design, not after.
Why this matters in 2026
Memory makers have moved wafer capacity to the high bandwidth memory that AI accelerators need, and standard server DRAM has paid for it. Counterpoint Research projected that a 64 GB DDR5 RDIMM, the workhorse module of the enterprise data centre, would cost twice as much by the end of 2026 as it did at the start of 2025, and Samsung was reported to have raised DDR5 pricing by around 30 percent for the second quarter of 2026 alone. Network World; Guru3D. Virtualization Review put it bluntly in July 2026: DRAM can now account for 30 to 70 percent of total host hardware cost, against 15 to 20 percent historically, and distributor quotes for enterprise RDIMMs are valid for days rather than weeks. Virtualization Review, July 2026.
That changes the refresh maths. A dual socket host that used to be priced by its CPUs is now priced by its DIMMs, and the cheapest DIMM is the one you do not buy. Memory tiering is the lever that lets you not buy it without giving up capacity.
How vSphere memory tiering works
ESX has always known which memory pages a virtual machine is touching and which it is not. Memory tiering uses that knowledge. The hypervisor keeps hot pages in DRAM, moves cold pages to a dedicated NVMe device, and brings them back the moment the guest asks for them. The guest operating system sees one contiguous memory space and needs no agent, driver or change. Tiering starts as host memory consumption approaches about 80 percent of DRAM, only virtual machine memory is tiered, never kernel memory, and the CPU cost is around an eighth of a core on a small host. Broadcom, memory tiering considerations and best practices.
The key number is active memory, the share of a workload's memory it is actually using at any moment. Broadcom's guidance is that most workloads sit between 10 and 30 percent, and that a host is a good candidate when its active memory has never exceeded 50 percent of DRAM. Consumed memory, the figure most teams watch, is always higher than active memory and is the wrong number to size tiering on. Broadcom, capacity planning considerations.
What you need
| Requirement | Detail |
|---|---|
| Licence | VMware Cloud Foundation 9 or vSphere Foundation 9. Memory tiering is generally available in both, at no extra charge, licensed per core like everything else. Broadcom's vSphere 9 comparison also lists it for Enterprise Plus. It is not in vSphere Standard, and the reported vSphere Express leaves it out. VMware, vSphere Foundation feature comparison. |
| Version | ESX 9.0 or later. Tiering was a technology preview in vSphere 8 Update 3 and became production ready with VCF 9.0 in June 2025. VCF 9.1 added cluster level configuration through vSphere Configuration Profiles, software mirroring and encryption of the NVMe tier. |
| NVMe devices | Local, dedicated to tiering, never shared with vSAN. Mixed use enterprise NVMe rated at 3 drive writes per day or better, Broadcom Compatibility Guide endurance class D and performance class F or G. Capacity at least equal to the host's DRAM. |
| Redundancy | Two NVMe devices of equal capacity per host for software mirroring in VCF 9.1, or a hardware RAID pair. Without it, an NVMe failure takes the host's virtual machines down. |
| Ratio | Default one to one, which doubles presented memory. Adjustable per cluster after reviewing historical active memory; VMware's launch material mentions ratios up to one to four for suitable workloads such as VDI. |
| Operating rules | No memory overcommit on a tiered host. Active memory kept under 50 percent of DRAM. Workloads that need all their memory in DRAM go on separate hosts or get a full reservation. |
Where it fits, and where it does not
Strong fits. General virtual server estates, where consumed memory is high and active memory low. VDI, which VMware's own testing scaled from 160 to 320 desktops per host at a 90 percent CPU target. Databases with large buffer pools and modest active sets, which is most of them: VMware's HammerDB test ran six more SQL Server instances on a tiered host and reached about 9 million transactions a minute against 5.8 million without tiering. Any estate where hosts hit their memory ceiling long before their CPU ceiling, which is the common case. VMware, June 2025.
Poor fits. Latency sensitive workloads where every memory access matters, such as trading systems and real time analytics. Virtual machines using fault tolerance, or physical GPU and vGPU passthrough. Very large virtual machines, which Broadcom suggests should be excluded from tiering with a full memory reservation. Workloads whose active memory genuinely approaches their allocation, where there is nothing cold to tier.
In practice the answer is rarely all or nothing. Tiering is enabled per cluster, so the estate splits into tiered clusters for the general estate and a smaller untiered cluster for the exceptions. That split is the heart of the design work.
The saving on a refresh is largest when tiering is in the design from the start. On refreshes we have been involved in, memory has been the line that decided the budget, and designing the new hosts around half the DRAM plus mirrored NVMe has taken a large share of it out, or let the same budget buy headroom the estate did not have before. Retrofitting tiering to hosts that were already bought full of DRAM still improves density, but the money has already left.
The second saving: cores
Broadcom licenses VMware per core, and so does Microsoft for SQL Server and Windows Server Datacenter. When a host is memory bound, cores sit idle and are still paid for. Tiering lets those cores carry more virtual machines, so an estate that needed twelve hosts may need nine, and every host not bought is sixteen cores or more of VCF or VVF, SQL Server and Windows licensing that is never renewed. VMware's own example on a 64 core host put the equivalent saving at 12 cores and roughly $28,200 a year on SQL Server subscription alone. Your figure depends on your licensing, which is exactly the modelling to do before the refresh order goes in.
How to size it
- Measure active memory, not consumed memory. Pull per host and per virtual machine active memory from vCenter over at least a month that includes month end and any batch peaks. The hosts to tier are those that never exceeded 50 percent of DRAM.
- Sort the workloads. Latency sensitive, fault tolerant, GPU passthrough and very large virtual machines go on the untiered list. Everything else is a candidate.
- Choose the ratio by cluster. Start at one to one. VDI and development estates can go further; databases that matter should stay conservative or keep more DRAM.
- Specify the NVMe properly. Mixed use, 3 drive writes per day or better, on the compatibility guide, at least the size of the host's DRAM, two per host for mirroring, in a slot and form factor the server actually has. This is where quotes go wrong: a read intensive drive that happens to be cheaper is not supported.
- Design the new hosts around it. Half the DRAM of the DRAM only design, or the same DRAM and twice the capacity, with the NVMe pair added. Price both against the quote you were given.
- Run it in one cluster first. Confirm active memory stays inside DRAM under real load, then roll out. VCF 9.1 applies the configuration host by host with live migration and no reboot.
How C4C helps
We design server refreshes for VMware estates and we do not resell VMware licences, so the advice is not steered by what is on the quote. We measure your active memory, separate the workloads that can tier from the ones that cannot, design the hosts with the right DRAM and NVMe, and put a tiered and an untiered design side by side with the licensing saving modelled on your real core counts. Through our partner network we then point you to the server vendor and VAR that will give you the sharpest price and the right expertise, chosen on the evidence of the many refreshes and RFPs we have run. If you are also facing a VMware renewal, the two decisions belong in one plan, because the bundle you license decides whether you can tier at all. VMware strategy and renewal support.
About to buy servers for a VMware estate?
Tell us roughly how many hosts, how much memory per host and when the order needs to go in. Within one working day we will tell you whether memory tiering fits your estate, what it would take off the quote, and which bundle you need to license to use it. We do not resell VMware or servers. We get you the best design and the best price from those who do.
Prefer email? Reach us directly at hello@c4cgroup.co.uk.
Frequently asked questions
What is vSphere memory tiering?
vSphere memory tiering lets an ESX host use local enterprise NVMe as a second tier of memory alongside DRAM. Active pages stay in DRAM and cold pages move to NVMe, and virtual machines see one memory space with no changes. At the default one to one ratio a host presents twice its DRAM. It became generally available in VMware Cloud Foundation 9.0 and vSphere Foundation 9.0 in June 2025.
Which VMware licence includes memory tiering?
VMware Cloud Foundation 9 and vSphere Foundation 9 both include memory tiering at no extra charge, licensed per core like the rest of the bundle. Broadcom's vSphere 9 comparison also lists it for vSphere Enterprise Plus. It is not in vSphere Standard, and the reported vSphere Express package leaves it out.
How much can memory tiering save on a server?
VMware's worked example priced a 1 TB host at $55,878 with DRAM only and $33,792 with 512 GB of DRAM plus 512 GB of NVMe, almost 40 percent less, in June 2025. DRAM prices have risen sharply since, so the gap is wider now. VMware also cites a financial services customer that cut host memory spend by 50 to 65 percent. Your saving depends on your active memory and the NVMe you specify.
What NVMe drives does memory tiering need?
Local NVMe devices dedicated to tiering and never shared with vSAN, mixed use enterprise drives rated at 3 drive writes per day or better, listed in the Broadcom Compatibility Guide at endurance class D and performance class F or G, with capacity at least equal to the host's DRAM. Two per host are needed for mirroring.
Does memory tiering slow down virtual machines?
Broadcom states the performance difference is negligible for workloads whose active memory stays within DRAM, which it puts at under 50 percent of DRAM. Access to the NVMe tier takes tens of microseconds against nanoseconds for DRAM, so latency sensitive workloads, fault tolerant virtual machines, GPU passthrough and very large virtual machines are better kept on untiered hosts or given a full memory reservation.
Can you use memory tiering with vSAN?
Yes, on the same host, but not on the same drive. The NVMe devices used for memory tiering must be dedicated to tiering and separate from the vSAN devices, which usually means a different drive pair in a vSAN ReadyNode configuration.
Does memory tiering reduce licensing costs?
Indirectly, and often significantly. A memory bound host leaves paid for cores idle. Tiering lets those cores carry more virtual machines, so an estate needs fewer hosts, and every host avoided is sixteen cores or more of VMware, SQL Server and Windows Server licensing that is never bought or renewed. VMware's example on a 64 core host put the saving at the equivalent of 12 cores.