Hold on to Your Hardware

Marius argues that AI data center buildout is cannibalizing the consumer hardware supply chain, and that the era of cheap, available components is ending. The piece is part warning, part call to action: buy and maintain hardware now, because the market is structurally shifting away from serving individual consumers.

The demand side of this dynamic is filled in by Zitron's May 2026 follow-up (hyperscalers committing ~$700B in 2026 capex on top of $900B already sunk) and the Axios polling (Q1 2026 data-center cancellations at record levels despite the capex commitments — community resistance plus financial pressure both biting).

The supply crisis

The numbers are stark. OpenAI's Stargate project reportedly needs ~900,000 DRAM wafers per month — roughly 40% of global DRAM output going to a single project. Micron has exited consumer markets entirely, leaving Samsung and SK Hynix as the only major DRAM suppliers. That's a duopoly.

Storage is in the same shape. Western Digital's entire 2026 HDD production is sold out. Kioxia's 2026 NAND production is sold out. The Silicon Motion CEO put it bluntly: "We're facing what has never happened before: HDD, DRAM, HBM, NAND… all in severe shortage in 2026."

This isn't a temporary spike. Western Digital's consumer revenue is 5% of sales. Cloud and enterprise account for 89%. When there's a shortage, consumer products are what gets cut. The incentive structure doesn't recover — serving data centers is simply more profitable than selling RAM sticks to individuals.

Visible impacts

The shortage is already hitting consumer products:

• Raspberry Pi 5 (16GB) jumped 70%, from $120 to $205
• Steam Deck OLED goes in and out of stock
• PlayStation 6 launch may slip to 2028–2029
• HP launched a laptop subscription service — monthly payments, no ownership

These aren't isolated incidents. They're symptoms of the same structural shift: consumer hardware is becoming a low-priority byproduct of enterprise demand.

The rental future

Marius draws the line from hardware scarcity to a world where consumers don't own compute at all. You'd rent processing power through cloud platforms the way you already rent communication (Meta), music (Spotify), films (Netflix), and storage (Google Drive). The difference is that renting compute means losing the ability to run software on your own terms — what programs you execute, what data stays local, whether you can operate offline.

This connects directly to titit-local-ai's argument for running LLMs locally. If you can't buy hardware, you can't run local models. The case for local AI depends on hardware being available and affordable. Marius is saying that assumption is eroding.

Geopolitics of chips

Hardware access is already a political tool. US export controls restrict Nvidia GPU sales to China. Enforcement has busted smuggling operations moving $1B+ worth of Nvidia chips through gray markets. If governments can restrict GPU access between nations, they can — in principle — restrict it between citizens and companies.

Chinese manufacturers CXMT (4th largest DRAM maker, preparing a $4.2B IPO) and YMTC are expanding aggressively and could provide a counterweight. Major PC manufacturers are already qualifying them as suppliers. But geopolitical restrictions may limit their reach in Western markets, especially for US buyers.

The economics underneath

The ai-subsidy-economics picture is relevant here. AI companies are spending at a pace that consumes global hardware supply, funded by investor subsidies that may not last. If the AI bubble deflates (ai-bubble-pale-horses), the hardware shortage could ease as data center demand drops. But the structural damage — manufacturers exiting consumer markets, supply chains reoriented toward enterprise — may not reverse quickly. Samsung and SK Hynix won't rebuild consumer product lines overnight just because data center orders soften.

The subprime-ai-crisis analysis predicts that the AI spending bubble will eventually pop. If it does, the aftermath includes excess data center capacity and a consumer market that's been starved of investment for years. The recovery path for consumer hardware availability is unclear.

What to do

Marius's advice is practical: buy hardware now if you need it, maintain what you have, extend device lifespans through cleaning and repair, and stop assuming prices will normalize. The second-hand market becomes more important as new production prioritizes enterprise. Resist subscription models that trade ownership for convenience.