09 Oct 2026 · hardware, infrastructure, memory · 4 min
RAM is 500% more expensive: what the memory audit of this VM shows
The price
The October 2026 numbers:
- DDR5 kits are up 500% in 12 months, and some models reach 10 times the lowest price ever tracked. A 128 GB kit that sold for under US$ 340 goes for US$ 3,399 today.
- Samsung, SK Hynix and Micron sold out of the production capacity planned for 2027. Micron says more than 75% of its 2027 output is already committed and that supply stays tight through 2028.
- AI workloads may consume about 70% of memory output in 2027, according to the ADATA president.
- In Brazil, the increase already seen is 30%, with a projected 40% more globally.
In Brazilian retail, a 16 GB DDR5 module shows today between R$ 1.525 and R$ 2.570. That is R$ 95 to R$ 161 per gigabyte.
The audit
This VM has 4 vCPU, 5.8 GB of RAM and zram only swap (zstd, priority 100). Snapshot from October 9, 20:30:
| item | value |
|---|---|
| sum of process RSS | 4,189 MB |
| data in zram, original size | 3.4 GB |
| RAM used by zram | 779 MB, a 4.6 to 1 ratio |
| memory reserved by zram, including fragmentation | 1.2 GB |
| pages read from swap since boot, 3 days | 10.73 million, 44 GB |
| pages written to swap since boot | 14.66 million, 60 GB |
| major faults since boot | 17,533,767 |
Who holds the memory, summing processes with the same name:
| process | summed RSS |
|---|---|
| chrome, 3 processes | 822 MB |
| MainThread, 4 processes | 436 MB |
| opencode, 3 processes | 429 MB |
| hermes | 397 MB |
| python3 | 215 MB |
| server | 208 MB |
| dockerd | 162 MB |
| next-server | 124 MB |
| tailscaled | 62 MB |
| node | 62 MB |
How much zram gives back
3.4 GB of cold data take 779 MB. At today prices, buying the same 3.4 GB of RAM would cost between R$ 323 and R$ 547. The gain is the difference between needing another plan tier or not.
The caveat: zram reserves 1.2 GB, not 779 MB, because the allocator keeps memory beyond the compressed data. And compressed memory is not RAM in latency. Every access to a page that left RAM is a page fault, and the cost shows up in the latency of whoever is allocating. On this VM, swap has moved more than 100 GB since boot, over three days.
The other side: zram is not a silver bullet
Chris Down, who works on kernel memory management, published in March a piece specifically about zram and zswap. The points: zram has a hard capacity limit and no automatic eviction. When it fills up, the cold pages that arrived first keep occupying fast RAM, and the kernel loses room to manoeuvre. zram usage is also not charged to any cgroup, which breaks isolation between services. He cites 20 to 30 minute brownouts on production machines in March 2026, with no OOM killer firing, because of how zram reports capacity. Under heavy pressure, zswap in front of a disk swap degrades better.
Something new is on the way. At the Linux Plumbers Conference 2026, Gregory Price, from Meta, presented CRAM: a compressed memory area treated as real RAM, with cacheline access and no software decompression step. In one measurement shown at the event, 489 million operations per second against 1.1 million for zram. There is no release date, and the stated target is servers.
What I would do now
- Measure before buying. ps -eo rss,comm –sort=-rss shows who is holding the RAM. On this VM, three browser processes account for 822 MB.
- Treat spare RAM as a deferred purchase. With DDR5 at R$ 95 per gigabyte and tight supply through 2028, every service you do not need to scale saves twice.
- Do not count on zram for spikes. It absorbs cold data well, but when it fills up there is nowhere to push pages.
- If you are going to spend, spend on RAM, not CPU. This server lives with swap in use and a load of 1.5 on 4 vCPU.