HN Debrief

Historical memory prices 1960-2026

  • Hardware
  • Infrastructure
  • AI
  • Economics

The page is a historical price series for computer memory, normalized to dollars per gigabyte and shown on a log scale from 1960 through 2026. It mixes very old memory technologies with modern DRAM, NAND flash, and storage, and the main signal is obvious even before the comments: memory got exponentially cheaper for decades, then that trend bent hard in the 2010s and snapped upward recently. People kept coming back to two basics that many readers missed. First, inflation changes the early decades far less than the dramatic shape suggests because the chart is logarithmic. Second, using dollars per GB is not a category error. It is the standard way to compare a commodity resource across time, even if it says nothing about how much memory a normal machine needed in a given year.

Treat this as a supply-cycle chart, not a chart of end-user pain. If you buy hardware, watch for product mix tricks and capacity expansion timing, because the next move in RAM prices depends more on what fabs actually ship than on the headline graph shape.

Discussion mood

Mostly frustrated and skeptical. People liked the long historical view, but many thought the chart was easy to misread, softened the current pain with log scaling and old DDR3 data, and failed to reflect how much more RAM modern software and AI infrastructure now consume.

Key insights

  1. 01

    Recent DRAM points may understate current pain

    The current end of the DRAM series appears to include DDR3 listings, including a 2 GB stick in 2025, which drags down the apparent present-day price per GB. That makes the chart more optimistic than what buyers of mainstream DDR4 or DDR5 systems are actually paying, and explains why some people's lived experience looks worse than the graph.

    Do not use aggregate memory-price charts to budget current hardware without checking the product mix underneath. Separate obsolete generations from the memory you can actually deploy today.

      Attribution:
    • fleebee #1
    • kaelwd #1
    • SirMaster #1
  2. 02

    This looks like a cycle that has lasted too long

    Memory was described as a classic boom-bust industry where synchronized fab expansion causes overproduction, price crashes, and retrenchment. What feels different now is not the existence of a cycle but its persistence. AI-era demand has kept memory tight for longer than the industry expected, while producers hesitate to overinvest because they remember how brutally past expansions ended.

    If you run infrastructure plans on the assumption that RAM will mean-revert quickly, build in more slack. The next price break depends on real new capacity showing up, not just on demand cooling headlines.

      Attribution:
    • vkazanov #1
    • gruntled-worker #1
  3. 03

    DRAM scaling is hitting a physical wall

    The slowdown is not just market structure. DRAM stores charge in a capacitor, and once features got into the 10 nanometer range, shrinking further became much harder because the capacitor geometry itself limits how much charge can be stored reliably. That means newer DRAM nodes bring progress, but not the old kind of effortless density gains that made prices fall decade after decade.

    Expect memory cost improvement to be lumpier and slower than the old trendline suggests. Software and hardware roadmaps that assume RAM will keep getting dramatically cheaper need to be revisited.

      Attribution:
    • sehansen #1
  4. 04

    Cheap RAM let software stop caring

    Several comments put the blame for modern bloat on runtime stacks, browser engines, large buffers, background services, and frameworks that insulated developers from memory limits. As long as RAM kept falling in price, there was little pressure to ask what an application cost in memory. Higher prices can force that question back into product decisions, not just low-level engineering.

    If you ship software to cost-sensitive users, start profiling memory now instead of waiting for customers to complain. Framework choice and default buffer sizes can move system requirements more than hand-tuned code.

      Attribution:
    • microgpt #1
    • skywhopper #1
    • cryptonym #1
  5. 05

    Cloud pricing already trained teams to conserve RAM

    One commenter argued that parts of the 'cloud native' movement were shaped by expensive RAM long before this consumer-facing spike. Smaller companies often designed production systems, including PostgreSQL deployments, around 8 to 16 GB footprints because cloud instance pricing made additional memory disproportionately costly. That means memory scarcity has already been influencing architecture in quieter ways.

    Review whether your current service design reflects actual technical needs or old cloud pricing pressure. Some systems were optimized around rented-RAM economics, not around the best architecture.

      Attribution:
    • aftbit #1
  6. 06

    Price per GB is the right base metric

    The strongest defense of the chart was that dollars per GB is a neutral resource metric, not a claim about usefulness. Once you switch to 'price per common task' or 'price per average workstation,' you bake in subjective assumptions about workloads, operating systems, and what counts as acceptable performance. Those are useful follow-on analyses, but they belong on top of the base data, not in place of it.

    Keep commodity and workload metrics separate in your own reporting. Start with a clean unit price, then layer on workload-specific views for procurement or product planning.

      Attribution:
    • appreciatorBus #1
    • tbrownaw #1
    • arter45 #1

Against the grain

  1. 01

    The rollback is not as deep as it feels

    Several people pushed back on the popular line that RAM prices have reverted all the way to 2010. On the chart itself, today's nominal high-water mark lines up closer to 2018, and once inflation is included the comparison moves forward again. The stronger claim is not a 16-year regression. It is that prices have reversed enough to break the assumption of steady improvement.

    Avoid using dramatic historical comparisons in budget discussions unless you specify nominal versus inflation-adjusted numbers. The strategic issue is the broken trend, not the exact nostalgia year.

      Attribution:
    • Aurornis #1
    • pron #1
    • AbsurdCensor #1
  2. 02

    Most users still do fine on 8 to 16 GB

    The loudest complaints came from developers and power users, but multiple comments noted that ordinary office and web workloads remain perfectly workable on 8 GB or 16 GB machines. Build jobs, containers, large codebases, and local development stacks are what push memory needs up fast. That makes the RAM crunch real, but concentrated in certain workflows rather than universal.

    Segment hardware standards by role instead of raising specs across the board. Many employees do not need a developer-class memory budget.

      Attribution:
    • Schlagbohrer #1
    • ack_complete #1
    • pmontra #1
    • ryukoposting #1
    • zahlman #1

In plain english

AI
Artificial intelligence, software systems that generate or analyze content in ways that mimic tasks usually associated with human intelligence.
DDR3
Double Data Rate 3, an older generation of DRAM used in older PCs, servers, and embedded systems.
DDR4
Double Data Rate 4, the DRAM generation that preceded DDR5 and is still used in many existing systems.
DDR5
Double Data Rate 5, the current mainstream newer generation of DRAM for PCs and servers.
DRAM
Dynamic Random-Access Memory, the main short-term memory used in computers and many electronic devices.
GB
Gigabyte, a unit of digital storage or memory capacity roughly equal to one billion bytes.
NAND
A type of flash memory used in solid-state drives and other storage devices.
PostgreSQL
A popular open source relational database system.

Reference links

Memory price datasets and archives

Industry background on memory markets

Related technical references

  • Moore's law
    Referenced in debate over whether the slowdown in falling memory prices should be tied to the end of Moore's law or related scaling limits.
  • Ethernet frame
    Mentioned during a side discussion about binary versus decimal units and typical network packet sizes.

Examples and cultural references