§03//BLOG
§ log.016 — the memory eaters
AI data centers now consume 70% of global memory output. Console prices have doubled, studios are closing, and games that will never be announced are being killed at the pitch stage. All because of RAM. Meanwhile China, locked out of the memory boom by US export controls, is building cheaper models and a domestic chip supply chain. The United States is cannibalizing its own consumer economy to feed an arms race it might not win.
A PlayStation 5 launched at $399 in 2020. Six years later it costs $599. Xbox has raised its price three times since 2025. Nintendo hiked the Switch. Valve's Steam Machine, a box barely as powerful as the six-year-old PS5, launched at $1,049. In May 2026, Americans bought fewer PlayStations than in any May since the year 2000, before the PlayStation 2 had even reached American shores. Xbox recorded its worst May in the brand's history. A director of a Windows handheld company publicly described his own product category as dead.
The cause is not inflation. It is not a pandemic supply chain residue. It is AI eating the world's memory supply.
A single Nvidia Blackwell Ultra chip requires 288 GB of memory. That is eighteen PlayStation 5s worth of RAM, on a fancier die but manufactured on the same production lines by the same companies. One rack of 72 AI chips equals 1,300 PS5s. Microsoft's first cluster for OpenAI was 64 racks: 83,000 PS5s of memory. OpenAI's Stargate campus will hold 100 times that, 8.3 million PlayStations worth, enough for every household in Australia. AI data centers are now estimated to be swallowing 70% of the world's memory output. Seventy percent of all DRAM on Earth goes into services we barely knew existed five years ago.
the chain reaction
The mechanism is straightforward. Three companies, Samsung, SK Hynix, and Micron, control almost all memory production. They are earning 70 to 80% margins on AI-grade HBM versus a historical norm of 30 to 40% on consumer DRAM. HBM revenue per wafer is 3 to 5 times higher than conventional DDR5. So they pivoted. SK Hynix, Samsung, and Micron have collectively shifted 93% of production toward HBM. A single HBM3E stack sells for $60 to $100. The same amount of DDR5 sells for $5 to $10. When capacity is constrained, rational manufacturers prioritize the product that prints money.
Every wafer allocated to an HBM stack is a wafer denied to the LPDDR5X module in a mid-range phone, the SSD in a laptop, or the RAM in a game console. It is a zero-sum game. New fabrication plants take three to four years to build, and the memory companies have little incentive to risk it. They are currently printing cash. An AI crash during those three to four years would turn a $20 billion fab into a monument to bad timing.
the memory pipeline, 2026
[3-5x margin] [70% of global output]
┌──────────┐ ┌──────────────┐
│ HBM fabs │──────────────────────────▶│ AI datacenters│
└──────────┘ └──────────────┘
│ ▲
│ same production lines │
▼ │
┌──────────┐ shrinking pool ┌──────────────┐
│ DDR5/NAND│─────────────────────────▶│ everything │
│ (consumer)│ 30% of what's left │ else on earth │
└──────────┘ └──────────────┘
The price effect has been severe and fast. Within three months at the end of last year, the price of a single 16 GB stick of RAM rose almost 300%. Xbox CEO Asha Sharma stated that console storage component prices are now over five times what the company paid two years prior, and are expected to double again by the 2027 holiday season. On some high-end graphics cards, the VRAM now accounts for up to 80% of the total card cost. The GPU chip itself, the component that Nvidia spends billions designing, costs less than the memory soldered around it.
the four levers
When hardware costs explode, console makers have four mechanisms to recover margin. They are pulling all of them.
Lever one: subscriptions. Xbox attempted an enormous Game Pass price hike, pulled it back slightly after backlash, but the new price is still higher than before. Nintendo has begun raising online service costs in Japan. The monthly subscription is becoming essential rather than optional.
Lever two: game prices. The $60 price point held for fifteen years. It moved to $70 in 2020. It is now moving to $80, only five years later. The $100 game is visible on the horizon. Sony's CFO has publicly signaled that current pricing is not sustainable, and the direction is upward, not downward.
Lever three: control. Sony announced that from 2028 onward they will seize distribution of physical game discs entirely. GTA 6, perhaps the biggest game launch in history, ships a "physical edition" that contains nothing but a download code in a box. Nintendo's Switch 2 cartridges increasingly do not contain the game. Elden Ring ships as a "key card" that downloads the entire game to your internal storage. When games are digital-only and tied to your account, the secondhand market ceases to exist. You cannot lend a game to a friend. You cannot buy it used for $5 on eBay. Sony is currently selling Immortals of Aveum, a commercial flop, for £69.99 digitally while the physical copy trades for £5 on the resale market. The digital-only future eliminates that price gap, and the consumer absorbs the difference.
Lever four: cost-cutting. This is the most consequential one. Xbox shut down five game studios including the teams behind Hellblade and Psychonauts. 3,200 jobs were eliminated in a single round of layoffs. Reportedly half of id Software, the studio that invented the first-person shooter with Doom and Quake, was cut. EA is building a real-time advertising platform that allows advertisers to purchase billboard space inside games while players are actively playing. The monetization tail is wagging the design dog.
Console entry price trajectory (USD, launch MSRPs)
Nominal launch prices, not adjusted for inflation. PS5 disc drive now $599 (2026 hike). PS6 rumored $999.
the spiral
This is not a simple story of rising prices and falling demand. Console gaming depends on momentum. A new platform either sells tens of millions of units out the gate, proving to developers that it is worth building for, or it fails to reach escape velocity. The Wii hit 101 million sales and became a magnet for every developer on earth. The Wii U bottled its launch, reached 13.5 million lifetime, and nothing saved it.
If the PS6 launches at $1,000, which is what current rumors suggest, and only 5 to 8 million people buy one in the first year while 80 million sit on PS5s, which platform does a developer choose for their next game? The old one. The new console never reaches escape velocity. The install base becomes a melting ice cube. Consoles break, players drift to PC or mobile, and nothing refills the pool. Publishers are already killing games at the pitch stage because projected install bases have shrunk. Games that you and I will never hear about are being cancelled right now because of RAM prices.
The games that do get greenlit will be safe ones. Obsidian ended work on a new unannounced RPG to focus on the next Fallout, a known entity. Assassin's Creed Black Flag is being remade cheaply and filled with in-game monetization. Innovation requires risk, and risk requires a large enough install base to justify the investment. When the install base shrinks, the only economically rational strategy for a publisher is to recycle what already worked. Remaster after remaster, made by fewer and fewer people.
the worry
What makes this genuinely frightening is not any single price hike. It is the meta-pattern. Microsoft, Amazon, Google, and Meta are spending more money than they are making from AI because the fear of losing the AI race has overtaken the fear of debt. This is the first time since the dot-com bubble that some of the richest companies on the planet are not behaving like businesses. They are not optimizing for profit. They are optimizing for survival in a race that may not have a finish line.
The memory companies, caught in the middle, have quietly decided to forget about the consumer entirely. Micron shut down its consumer memory brand Crucial to focus on data center customers. The margins are two to three times higher and the customers pay whatever the asking price is. There is no rational reason to serve the consumer market when the data center market is structured this way.
The chain runs as follows: AI panic leads to infinite memory demand, which leads memory companies to abandon consumers, which doubles hardware prices, which reduces consumer purchases, which shrinks install bases, which causes game cancellations, which closes studios, which produces fewer games, which removes reasons to buy a console, which further shrinks the install base, which triggers more cancellations. The spiral feeds itself. The only thing that breaks it, new memory fabs, will not produce meaningful volume before 2028. Even then, most of that new capacity is already pre-sold to AI customers.
Meanwhile, Nvidia has started capping how many hours per month you are allowed to play on GeForce Now, their cloud gaming service. The same company that manufactured the chips that consumed your console's memory supply is now offering to rent you the graphics card you can no longer afford to own, for a monthly fee, with a usage cap. You do not own the hardware. You do not own the games. You rent access to both, metered and monitored, until the subscription lapses.
the china comparison
Here is where the picture becomes more complex, and arguably more concerning.
The United States has spent the last two years attempting to choke China's AI ambitions through export controls. In December 2024, the US banned HBM exports to China above the HBM2E generation. Chinese companies had been stockpiling. They procured roughly 13 million HBM stacks before controls took effect. SemiAnalysis estimates China exhausted that stockpile by late 2025. China's domestic HBM producer, CXMT, is 3 to 4 years behind Samsung and SK Hynix on process node. The Ascend 910C, Huawei's flagship AI chip, runs on domestic HBM that delivers 25% less bandwidth than the imported version it replaced.
By every metric the US prioritizes, compute performance, memory technology, and fabrication capability, China is behind. Significantly behind. The export controls were designed to keep it that way.
But the model side tells a different story. Cut off from frontier chips and frontier memory, Chinese labs did what constrained systems always do. They optimized for efficiency. DeepSeek's R1 model triggered a US tech stock rout in January 2025, not because it was the most powerful model but because it was trained at a fraction of the cost anyone thought possible. Chinese open-source models now account for 41% of Hugging Face downloads. DeepSeek, Zhipu's GLM 5.2, and Moonshot's Kimi K3 operate at 60 to 90% lower cost than comparable Anthropic and OpenAI models while sitting within 6 to 9 months of the US frontier on benchmarks. GLM 5.2 came within a percentage point of Anthropic's Opus 4.8 on agentic tasks at roughly one-fifth of the cost. Lindy, Cursor, Airbnb, and Siemens, all US and European companies, are migrating traffic to Chinese models because the price-performance ratio has become impossible to ignore.
DeepSeek is now developing its own AI chip to reduce dependence on both Nvidia and Huawei. CXMT plans to allocate 60,000 wafers per month to HBM3 production by late 2026. Huawei is targeting 600,000 Ascend 910C chips. China's 15th Five-Year Plan, covering 2026 through 2030, explicitly targets domestic HBM3E production and self-reliance for domestic AI needs.
two approaches to the AI buildout
US CHINA
── ────
infinite capital constrained capital
spend more than you earn optimize cost per token
cannibalize consumer markets consumer markets unaffected
(70% of memory → AI) (can't buy HBM anyway)
frontier models, frontier cost near-frontier, 10-90% cheaper
Nvidia monopoly domestic chip dev (Ascend, DeepSeek)
race-driven panic necessity-driven efficiency
The irony is sharp. US export controls were designed to slow China down. Instead they forced China to build the one thing the US does not have: a domestic, self-contained AI supply chain that does not depend on TSMC, Samsung, or SK Hynix. China cannot buy the memory that is destroying the US consumer electronics market, so they are immune to the very crisis the US created. Chinese gamers are not paying five times more for console storage because Chinese consoles do not use imported HBM. Chinese phones are not jumping in price because Chinese OEMs source from domestic NAND and DRAM makers. The US walled China out of the feast, and in doing so, insulated China from the famine.
Meanwhile, the US is consuming its own consumer economy to feed an AI arms race in which the primary competitors, OpenAI, Anthropic, and Google, are American companies competing against each other. Not against China. The US is not racing China. It is racing itself, and the casualty is everyone who is not an AI company.
what actually worries me
I am not worried about AI. I build with it every day. I am not worried about China. Their models being cheap and open is a net positive for the world.
I am worried about the misallocation. The United States is directing the majority of its advanced manufacturing capacity, memory, fabs, packaging, and engineering talent, toward a single objective: making AI slightly more capable, slightly faster, for companies that are not profitable and may not be for years. The consumer hardware market, the gaming industry, the smartphone market, and the PC market are all being squeezed to feed a fire that shows no sign of self-sustaining. Microsoft, Amazon, and Google are spending more than they make from AI. If the AI revenue materializes, the bet pays off and this period looks like a brilliant land grab in hindsight. If it does not, if the bubble pops, we will have burned five to seven years of consumer hardware progress, thousands of game development jobs, and entire console platform lifecycles for nothing.
The memory companies are not going to save us. They have no incentive to. The new fabs being built in Texas by Intel and others will not produce meaningful volume before 2028, and most of that capacity is pre-sold to AI customers. Micron shut down its consumer brand. The message is unambiguous: if you are not an AI data center, you are not the customer. You are the externality.
China, ironically, may end up in the more resilient position. Not because they are ahead, but because they were forced to build a supply chain that serves their entire economy rather than one hyped sector. When the AI bubble cools, whenever that is, China will still have domestic memory production, domestic chips, and domestic models that run at a fraction of the cost. The US will have Stargate, a cratered console market, and a generation of games that were never made.
my prediction
1. The PS6 launches above $899, possibly at $999, in 2028 at the earliest. It will sell less than half of the PS5's first-year volume. At least one major third-party publisher will publicly commit to PS5 as the lead platform for their next flagship title, citing install base economics.
2. At least one major console brand exits hardware by 2030. The economics of a $1,000 console with a shrinking install base and a melting user base do not work. The survivor pivots to a cloud and subscription model where the consumer rents access rather than owning hardware.
3. Chinese AI models reach price-performance parity with US frontier models within 12 months. Not benchmark parity. Cost-per-useful-token parity. When that happens, US enterprise AI spending will shift toward Chinese infrastructure, and US labs will be forced to cut prices in a way that makes the current burn rate unsustainable.
4. The memory shortage does not ease before mid-2028. When it does ease, it will not be because demand dropped. It will be because new HBM4 fabs finally come online, and those fabs will be pre-sold to AI customers. The consumer shortage will outlast the headline shortage by another 12 to 18 months.
I keep thinking about that handheld director who called his own industry dead. He was not being dramatic. He was being precise. The question is not whether console gaming survives. It is what we sacrificed it for, and whether the thing we sacrificed it for was real.
$ echo "the us is eating its own future to feed a machine that isn't hungry yet"
the us is eating its own future to feed a machine that isn't hungry yet