1 SEP 2026 — VMware is pitching memory tiering as the answer to DRAM prices, claiming 40 per cent lower cost of ownership by moving cold pages out of memory and onto NVMe. The NVMe it requires is rated at three drive writes per day and 7,300 TB of lifetime writes. The saving is real, but it converts a capital purchase into a consumable that wears out.
What the feature does
Memory tiering shipped in VMware Cloud Foundation 9.0 and identifies cold pages — regions of DRAM allocated to a virtual machine but not actively read or written — and relocates them to NVMe storage. The hypervisor pages them back when touched.
The stated requirements for the NVMe device are a minimum of 100,000 writes per second, at least 7,300 TB of lifetime writes and an endurance rating of at least three drive writes per day. VMware recommends a 1:1 ratio of DRAM to NVMe and says 1:4 is possible.
The claims are a 30 per cent reduction in CPU cycles and 40 per cent lower total cost of ownership through densification. VCF 9.1 added database performance work and management dashboards, and 9.2 is expected around May 2027.
Read the endurance numbers as a running cost
The specification anticipates this workload writing 12 TB daily, which is three drive writes per day on a 4 TB device. That is a consumable rate, and the 7,300 TB lifetime figure divided by that daily rate gives roughly 600 days of specified life.
DRAM does not wear out. An NVMe device operating at its rated endurance does, on a schedule you can calculate in advance, and the replacement lands in operating expenditure rather than in the capital budget the saving was measured against.
None of that makes the 40 per cent claim wrong. The comparison is between a one-time purchase and a subscription to a consumable, and an honest accounting includes a drive-replacement line item that the 40 per cent saving omits.
The 75 per cent is the number to plan around
Memory tiering currently supports about three quarters of workloads. The quarter it does not cover is not a random sample.
Tiering works when a workload has cold pages. That describes an over-provisioned application server, and not an in-memory database. A database that holds a working set in memory precisely so it never touches storage is the workload with the least cold memory and the most sensitivity to a page fault.
In short, the feature claws back wasted memory and offers little where memory is the core resource. A fleet-wide 40 per cent figure will not reproduce on a fleet whose expensive hosts are the ones running the workloads tiering cannot help.
The CPU claim is the more interesting one
A 30 per cent CPU reduction seems backwards for a feature that adds paging, but the saving comes from what the processor is no longer doing.
A host that has run out of memory stalls. Threads wait on allocation, the scheduler churns, and cycles are burned on work that produces nothing. Giving the hypervisor somewhere to put cold pages removes that stall, so the cycles saved are the ones that were being wasted rather than cycles the paging itself avoids.
This applies only to a host that has run out of memory. On a host with headroom, tiering adds work and saves nothing, because there was no stall to remove. The gains are concentrated on hosts already running at the edge of their memory, which are also the hosts where a mistimed page fault is most visible to a user.
Why the rooms were full
Two sessions on this feature filled 500-seat rooms at VMware Explore, which is not the usual reception for a memory-management talk. DRAM contract prices have been rising through the AI build-out, and every operator planning a hardware refresh has discovered that the memory line has moved more than the rest of the bill.
That makes tiering a demand-side response to a supply shortage. It does not add memory to the market; it lets a buyer purchase less of it and make up the difference with a component that is currently cheaper per gigabyte.
We reported that the memory crunch pushed a consumer handset price up. The same input cost is arriving in the data centre, and the responses look different only because the buyers there have an architectural option a phone maker does not.
What to check before committing
Three questions decide whether this works for a given estate. First, what proportion of allocated memory across the fleet is actually cold? Measure it, because the answer is usually higher than administrators expect and lower than a vendor slide implies.
Second, what the drive-replacement schedule costs over the hardware's life at the rated endurance, added back to the tiering saving.
Third, what the roadmap dependency is worth. Expanded support for very large virtual machines, up to 960 vCPUs and 16 TB of memory across multiple hosts, sits on a two-to-three-year horizon, and vMotion compatibility exists today with performance improvements still planned. If your most expensive machines are also your largest, you are buying a feature promised in a future release rather than one that ships today.
View author profile → · Editorial policy
