Jeff Geerling published a piece this week arguing that DRAM pricing is killing the hobbyist SBC market, and the reaction on Hacker News was immediate, with over 500 comments from people who build things on small boards for a living or a hobby. The point landed because it names something the community has been circling around for a few years: the affordable single-board computer, the device that made Raspberry Pi a cultural force in hobbyist computing, is becoming structurally impossible to manufacture at the prices we got used to.
This is worth unpacking carefully, because the surface-level framing, “DRAM is expensive right now,” undersells what’s actually happening.
How the $35 Pi Was Possible in the First Place
The original Raspberry Pi launched in 2012 at $25 with 256MB of RAM soldered directly to the board. The Pi 2 in 2015 hit the canonical $35 price point with 1GB. The Pi 4 in 2019 started at $35 for the 1GB variant and scaled to $75 for 8GB. These prices were achievable partly because of the Raspberry Pi Foundation’s nonprofit structure and clever negotiation, but mostly because the DRAM market from 2012 through most of 2019 was either oversupplied or recovering from oversupply.
DRAM is manufactured by three companies that control roughly 95 percent of global supply: Samsung, SK Hynix, and Micron. Capital expenditure for a single leading-edge fab runs into the tens of billions of dollars, which creates a brutal cycle. When prices are high, all three race to add capacity. When that capacity comes online simultaneously, prices crash. When prices are too low to justify further investment, supply tightens and prices recover. The Pi 4 shipped into a favorable part of that cycle. The Pi 5, which launched in late 2023 at $60 for 4GB and $80 for 8GB, did not.
What AI Infrastructure Changed
The DRAM cycle argument would predict prices eventually come back down. The complication in 2025 and into 2026 is that AI infrastructure demand has inserted a new, structurally different buyer into the market.
High Bandwidth Memory, the stacked DRAM used in AI accelerators like NVIDIA’s H100, H200, and B200 series, is manufactured by the same companies that make LPDDR4 and LPDDR5. It uses the same underlying DRAM die technology, then adds Through-Silicon Via stacking to achieve the bandwidth those chips require. HBM commands dramatically higher margins than commodity mobile DRAM. A set of HBM3e stacks for a single H200 GPU costs multiples of what an entire Pi 5 costs in memory alone.
When Samsung and SK Hynix have to choose between allocating fab capacity to high-margin HBM for hyperscaler AI orders and low-margin LPDDR5 for consumer electronics and SBCs, the calculus is not complicated. SK Hynix has publicly stated that it has been prioritizing HBM capacity expansion aggressively since 2024. That capacity comes from somewhere, and a meaningful fraction of it comes from LPDDR production.
The result is that the LPDDR5 packages that Raspberry Pi 5 needs, package-on-package modules bonded directly to the SoC, are both in tighter supply and priced higher than they would be in a world without sustained AI infrastructure buildout. This is not a short-term disruption in the same way the 2021-2022 chip shortage was. The demand signal driving HBM investment is durable.
The PoP Problem Compounds Everything
SBCs like the Pi 5 use Package on Package memory, where the DRAM is stacked directly on top of the application processor during manufacturing. This is necessary for the compact board size and the low-latency memory access that makes a small board usable. It also means you cannot substitute. You cannot decide to use a cheaper DRAM vendor’s chips if your board layout assumes a specific PoP package from a specific supplier at a specific height.
This is why Raspberry Pi discontinued the 1GB and 2GB Pi 5 variants. At current LPDDR5 pricing, a 4GB PoP package eats enough of the BOM cost that building a cheaper board means either absorbing a loss or selling at a price that defeats the purpose. The $35 price point requires DRAM costs that are not currently available for the memory type the Pi 5 needs.
Competitor boards face the same problem. The Orange Pi 5 and Radxa Rock 5 series, both built around the Rockchip RK3588 family, also use LPDDR5 or LPDDR4X in PoP configurations. They have not escaped the pricing pressure. Several have quietly increased prices or dropped lower-RAM SKUs over the past eighteen months.
Boards That Avoided This, and Why
Boards that still use LPDDR4 or DDR4 in non-PoP configurations have some flexibility. The BeagleBone AI-64, built around the TDA4VM processor, uses soldered LPDDR4 and occupies a different price and performance tier. Some industrial-oriented SBCs use standard SO-DIMM slots, which lets them absorb price swings more easily by swapping memory suppliers.
The trade-off is physical size, power consumption, and latency. A board with a DDR4 SO-DIMM slot is not going to be Raspberry Pi sized. For many project contexts, that is fine. For the canonical use case of a Pi, something small enough to hide behind a monitor or inside a custom enclosure, the PoP form factor is essentially mandatory.
RISC-V boards like the Milk-V Pioneer or the StarFive VisionFive 2 are sometimes floated as alternatives, but they have the same DRAM dependency. The ISA does not change the memory cost structure. A RISC-V board that needs 4GB of LPDDR5 is paying the same prices.
What the Foundation Can Actually Do
The Raspberry Pi Foundation’s structure as a charity with an educational mission creates a genuine tension here. Their stated goal is providing affordable computing for education. Their commercial arm, Raspberry Pi Ltd, has to fund that mission while also funding the next generation of hardware development.
During the 2021-2022 shortage, the Foundation made the decision to prioritize commercial and industrial customers over hobbyists to keep the company solvent. That decision was criticized at the time but was probably necessary. The current situation is different in that it is not a shortage, it is a price floor. Supply is available at a price that simply does not permit the board to be sold at $35 or even $45.
Eben Upton has been candid in interviews over the past year that the pricing environment for the memory Pi 5 needs is not what anyone hoped. The CM5 and Pi 500 product lines suggest a deliberate move toward segments where higher per-unit prices are more acceptable. That is a rational response to the cost structure, but it is a concession that the original market positioning is harder to maintain.
What This Means if You Build Things
For practical project work, a few things follow from this situation. Existing Pi 4 boards are still capable machines for most tasks, including running Discord bots, home automation, lightweight web services, and media servers. If you have them, they are not going anywhere. The Pi 4’s LPDDR4X memory is cheaper than LPDDR5, and the boards have a long support tail.
For new purchases, the math favors buying sooner rather than waiting for prices to drop significantly. The structural factors driving LPDDR5 pricing are not going to resolve quickly. AI infrastructure investment is not decelerating.
For use cases that require more headroom, a used x86 mini PC in the NUC or Beelink tier often delivers more compute per dollar now than it did two years ago precisely because it runs standard SO-DIMM memory and has benefited from its own price normalization. That is not a Pi replacement in the size and power category, but for anything running as a home server it deserves honest comparison.
The harder question is what comes after Pi 5 in the education-focused, sub-$50 tier. If LPDDR5 pricing stays elevated, the next board either needs a different memory strategy, a different SoC partner willing to absorb more of the cost, or an explicit acceptance that entry-level starts at $60 now. The last option is not the end of the world, but it is worth naming clearly rather than pretending the $35 board is coming back.
The original promise was never really about $35. It was about a device cheap enough that a student, a maker, or a tinkerer could buy one without it being a significant financial decision. Whether that threshold is $35 or $45 or $55 depends on context. What Geerling is pointing at is that current cost structures are pushing well past whatever that threshold is, and the force driving it is not temporary.