The Memory Crisis Is Killing the Consoles That Barely Use Any Memory

Excerpt: PLAION just pushed the NeoGeo AES+ back a full year to September 2027, blaming "unprecedented worldwide demand for memory chips, driven by investment in AI infrastructure." There's one problem with that explanation as a memory story: the AES+ is a hardware-accurate replica of a 1990 machine with 64KB of RAM. It buys less working memory than a smart lightbulb. And it's dying of the memory shortage anyway — because the scarce input isn't gigabytes, it's the obsolete little chips nobody wants to fab anymore. The same mechanism now reaches the cartridges themselves.

An original LostInConsoles investigation into why the AI memory crunch is now delaying and killing the smallest, least memory-hungry console hardware first — and why physical cartridges may be the next casualty.

The short version: For two years the story of the AI memory crunch in gaming has been a price story. DRAM triples, so systems get more expensive. That story is now too small. In September 2026 the crunch crossed into something nastier: it started cancelling products. And the counter-intuitive part — the part almost nobody has assembled — is which products are going first. Not the memory-hungry ones. The ones that use almost no memory at all. The NeoGeo AES+ carries 64KB of RAM and is delayed a year. A physical Evercade cartridge carries up to 1GB of flash and its maker says its entire business model is now about stockpiling "cartridge chips." The reason is that the memory market's victims are not defined by how much memory they buy, but by which node that memory is fabbed on.

The event: a 1990 console delayed a year, over memory

On 9 September 2026, PLAION REPLAI announced that the NEOGEO AES+ — a hardware-accurate recreation of SNK's 1990 AES, developed under license from SNK — would move its worldwide release to Thursday 16 September 2027. (Observed fact: the company's own press release.)

That is a full year of delay. And the reason PLAION gave is unambiguous. From the press release:

"Since announcing the NEOGEO AES+, PLAION REPLAI has been working with manufacturing partners to increase production capacity. However, unprecedented worldwide demand for memory chips, driven by investment in AI infrastructure, has created well documented industry-wide supply [constraints]… Taking additional time allows us to secure more memory chips while expanding our manufacturing."

The press release pairs memory scarcity with a manufacturing scale-up ("The decision to delay follows an extraordinary response to the announcement"). Pre-orders remain secure and the price is unchanged — this is a delay, not a price rise. The press framing that followed (Tom's Hardware, Shacknews, and others) collapsed this into "RAM shortage." Fine. But read that against the spec sheet and the whole thing gets strange.

The inversion: it barely has any memory

Here is the NeoGeo AES+ hardware, per the published specs:

• 64KB RAM
• 84KB VRAM
• 2KB audio RAM
• games read straight from ROM (the cartridge is the storage)
• the optional CD upgrade adds 7MB

Add it up: the entire working memory of the console is under 200 kilobytes. A 2026 midrange phone ships with 12 gigabytes of RAM — roughly 60,000 times more. The AES+ is not a memory-hungry device. It is nearly a memory-free device. Its design philosophy is the exact opposite of the thing that AI data centers are outbidding the world for.

So on the face of it, the vendor's explanation shouldn't hold. A machine that buys less memory than a thermostat cannot be the victim of a DRAM shortage — DRAM is sold by the terabyte and the AES+ buys by the kilobyte. This is the objection a sharp reader raises immediately, and it looks like it should kill the story.

It doesn't. It is the story.

The mechanism: it's not about gigabytes, it's about the node

The AI boom doesn't just make memory expensive. It makes specific kinds of memory extinct.

The memory that AI accelerators eat — HBM, high-layer 3D NAND, DDR5 — is fabbed on the same mature manufacturing lines that produce the cheap, low-density, legacy parts that small hardware runs on: SDRAM, SRAM, NOR flash, SLC NAND, DDR3/LPDDR3, eMMC. When a fab can sell a wafer of HBM to an AI buyer at several times the revenue of a wafer of legacy SDRAM, the economics decide the allocation. The legacy lines get converted or wound down. The obsolete little parts simply stop being made — not because demand for them collapsed, but because supply was reallocated to something more profitable.

This is documented as a structural shift, not a commodity blip. TrendForce's own research headline from June 2026 says it outright:

"Contract Prices Surged More Than 100% in 1H26; Structural Shortages to Keep NOR Flash and SLC NAND Prices Rising in 2H26."

That is the sentence that explains the NeoGeo AES+. NOR flash and SLC NAND are exactly the low-density, mature-node memory that a cartridge-based retro console lives on. Prices for those specific parts more than doubled in six months — and TrendForce calls the shortage structural, meaning it doesn't self-correct next quarter.

So the AES+ isn't being starved of "memory" in the abstract. It's being starved of cheap legacy memory on a node that the AI build-out is draining. It buys by the kilobyte — and kilobytes are precisely what's going out of production.

The second exposure: the cartridges are memory devices too

If the AES+ were just one delayed console, this would be a footnote. What makes it a pattern is the part the industry keeps missing: in a cartridge ecosystem, every single cartridge is itself a piece of memory hardware.

Consider Evercade. Blaze Entertainment's systems play games off physical flash-based cartridges that run 128MB to 1GB of flash each, sold for around $30, on hardware that costs $70–$200. In a July 2026 interview with FRVR, Blaze CEO Andrew Byatt described what the memory crunch has done to that model. He said the company had spent essentially all its cash hoarding components years in advance:

"We bought as much as we could in terms of cash flow… We bought as many cartridge chips, as much RAM, and as much processors and all the rest of it."

And the priority was explicit. Cartridges, he said, are "the lifeblood of the business" — so Blaze front-loaded cartridge chips for the next year and locked in 2027 supply, while treating console hardware as a "wait and see" item to buy as it goes.

That is a completely different shape of exposure from a normal console. An ordinary console maker has one bill of materials per box. A cartridge platform has two: the console BOM, and a media BOM for every cartridge in a growing catalog. When NOR/NAND flash prices more than double, the second bill gets hit not once but once per game, forever, across the whole back catalogue. And because Evercade deliberately keeps cartridge pricing consistent across 128MB and 1GB carts, the platform can't simply let its biggest-size carts float upward — the whole "buy a physical thing on a shelf" promise depends on that consistency.

Blaze has, for now, absorbed this by buying years ahead. It is a stockpile, not a solution. The stockpile is the story: a company chose to convert its cash into a warehouse of obsolete memory chips because the chips were the one input it could not count on getting later.

The pattern is bigger than one vendor

The AES+ and Evercade are the vivid cases, but they aren't isolated. In August 2026, the open-hardware maker PINE64 announced it was halting production of its Linux hardware line — boards, tablets, e-readers — blaming the DRAM/eMMC shortage directly on the AI build-out, expecting its existing PineNote and PineTab2 stock to run out within about three months and saying nothing new is coming before mid-2027. That's not a price rise. That's a product line going dark.

And the retro/niche segment has a whole casualty list now forming around the same cause: Retro Games Ltd's "The Spectrum" (delayed, with costs cited on chips and plastics), the Analogue Pocket (sitting at $239.99), Analogue's N64-based Analogue 3D, the Panic Playdate, Atari's 2600+/7800+ line, AYANEO's retro handhelds, and more. Some of these are primary-sourced, some are reported; treat the individual ones as provisional and the cluster as the signal. (Observed facts: the announcements. Inference, labeled: that they share one root cause — the vendors each name component/memory supply, not a single demand story.)

The pattern is consistent across every one of them, and it is the opposite of what a naive reading of "memory shortage" predicts: the smallest buyers of memory are the ones being cut off first. Not because they need a lot — because they need the wrong kind.

Falsification — the honest counter-cases

Objection 1 (the strong one): "A 64KB machine can't be a memory-shortage victim — memory is a cover story." We tested this hard, and it's the reason this article exists. The AES+ genuinely uses negligible working memory, and a raw commodity-DRAM shortage genuinely cannot explain its delay. But the objection only defeats a bad version of the story. The correct version isn't "AES+ ran out of RAM"; it's "AES+ is built on mature-node specialty memory, and that segment is being discontinued and repriced by the AI build-out — TrendForce's >100% NOR/SLC-NAND contract-price move is the receipt." The vendor named memory chips; the mechanism named by the data is the legacy node. So the objection doesn't falsify the thesis — it clarifies it. The inversion survives.

Objection 2: "It's really just demand — PLAION wants to make more consoles, and 'memory' is a euphemism." Partly fair: PLAION does say it is expanding manufacturing in the same breath. But you cannot build more consoles without the parts, and the vendor explicitly ties the delay to securing memory chips ("Taking additional time allows us to secure more memory chips while expanding our manufacturing"). Demand and component scarcity are complementary here, not competing. Treated as a co-factor; does not overturn the component cause.

Objection 3: "One delayed replica console isn't a market trend." Weakened by independent corroboration at a different product class (PINE64's boards/tablets on DRAM/eMMC) and a different exposure (Evercade's cartridge chips). Three unrelated vendors, three different memory types, one direction. The trend claim rests on that spread, not on the AES+ alone. Acknowledged limitation: not every retro casualty is primary-verified, and some price moves carry tariff/FX noise.

What the data says it means (analysis, not fact)

• Observed facts (high confidence): the AES+ delay to 16 Sep 2027 and PLAION's stated cause; the AES+ spec (64KB/84KB/2KB, ROM-fed); TrendForce's >100% 1H26 NOR/NAND contract-price rise and "structural" framing; Evercade carts at 128MB–1GB/~$30 and Byatt's "cartridge chips… lifeblood" quote; PINE64's halt on DRAM/eMMC through mid-2027.
• Inference (moderate, labeled): the binding constraint for retro/niche hardware is mature-node specialty memory, not commodity DRAM; cartridge platforms face double exposure (console BOM + per-cartridge flash BOM); the smallest memory buyers are structurally first in line to be cut.
• Speculative (explicit): that this becomes an extinction event for the smallest retro makers; that the cartridge-revival era is ending. TrendForce sees the shortage as structural into 2027 and beyond, which makes some attrition likely — but how far it goes is not knowable from here, and we flag it as such.

Bottom line

The NeoGeo AES+ is a 1990 console rebuilt in 2026 that carries less memory than a disposable gadget — and it just got delayed a year because the world doesn't make enough obsolete memory anymore. That sounds like a paradox. It's actually the cleanest evidence yet of what the AI memory crunch really is.

It isn't a shortage of memory. It's a shortage of the specific, unfashionable, low-margin memory that small hardware is built on — the SRAM, the NOR flash, the SLC NAND, the legacy SDRAM that AI has no use for and fabs no longer want to make. And because cartridges are memory, the crunch reaches the retro market twice: once when you can't build the console, and again when you can't afford to build the games.

The consoles that use the least memory are going first. That is the inversion, and it is where the story goes next.

Sources

• PLAION REPLAI — "The NEOGEO AES+ worldwide release date moves to 16th September 2027" (September 9, 2026): presse.plaion.com
• TrendForce — "Contract Prices Surged More Than 100% in 1H26; Structural Shortages to Keep NOR Flash and SLC NAND Prices Rising in 2H26" (June 16, 2026): trendforce.com
• Kotaku — "RAMageddon Is Hammering Retro Handheld Makers Like Evercade": kotaku.com
• Tom's Hardware — "Hardware-accurate NeoGeo AES+ delayed to late 2027": tomshardware.com
• GamesRadar — "The Neo Geo AES+ now isn't arriving until September 2027": gamesradar.com
• Shacknews — "NEOGEO AES+ delayed to 2027": shacknews.com
• IT'S FOSS — "Pine64 halts Linux hardware manufacturing over DRAM/eMMC prices" (August 2026): itsfoss.com
• Gaming.net — "NeoGeo AES+ delayed to September 2027 as PLAION cites memory chip supply": gaming.net
• Evercade / Blaze Entertainment — cartridge format and memory sizes, official product information: evercade.co.uk
• Wikipedia — Neo Geo (hardware specification: 64KB RAM / 84KB VRAM / 2KB audio RAM): en.wikipedia.org

Evercade's cartridge-flash sizes (128MB–1GB) and CEO Andrew Byatt's "cartridge chips… the lifeblood of the business" quote are as reported by Kotaku, drawing on the Blaze/FRVR interview. TrendForce's figures describe a reported industry-wide contract-price move, not a confirmed contract.

LostInConsoles is an independent investigations outlet focused on the global console hardware business. Quotes and dates are drawn from vendors' own releases and named interviews; causal readings that go beyond what sources state are labeled as analysis. Prices in USD unless noted.