The Last Personal Computer

I have built or rebuilt a desktop every few years for as long as I have had money. This year I priced one out, closed the tab, and started writing this instead. Not because the parts were expensive, though they were. Because for the first time the prices did not look like a bad year. They looked like a decision somebody else had made about what my computer is for.

Here is the whole argument. For forty years the personal computer got cheaper because the silicon that ran the big machines trickled down to your desk a few years later at a fraction of the price. That trickle has reversed. Every gigabyte of memory, every advanced wafer, every spare megawatt now has a bidder with a nine-figure purchase order, and that bidder is not you. The machine you would buy, the models you would run on it, and the electricity you would run them with are all being re-priced for someone else. The personal computer is not dying. It is being repossessed.

Most writing on this stops at the facts, because the facts are safe and the next step is not. I am going to take the next step. The facts are in the notes at the bottom, numbered, so you can check them. The rest is me extending the lines, and I will say so each time. If you want a balanced take, there are a hundred of those. This is a bet.

1. Memory, or: you are not a strategic customer

Three facts, then the extrapolation.

One. On 3 December 2025 Micron announced it was leaving the consumer memory business; Crucial, the brand on the RAM in half the PCs you have ever built, shipped its last parts in February. The company said the exit would improve supply for larger, strategic customers in faster-growing segments.¹ Two. The three companies that make nearly all the world's DRAM have, according to DigiTimes, already allocated their entire 2027 output, and expect to meet only 60 to 70 percent of demand. Samsung told investors that supply stays tight through 2028, with 2027 the worst of it, and that it has multi-year agreements with its top five data-center customers.² Three. SK hynix's chief executive, on the day his company listed on Nasdaq, told Reuters:

We forecast that next year will be the worst year in the industry's history from the supply perspective.

He expects demand to exceed what the company can build beyond 2030.

Everything after this sentence is mine.

The retail memory market is no longer a market. It is an overflow pipe. When a fab's next two years are spoken for by five customers on multi-year terms, the channel you and I buy from gets whatever is left after the contracts, at whatever price clears. Contract prices rose about 60 percent in one quarter this year, and the slowdown since is not relief; it is PC makers hitting the ceiling of what they can pass on.⁴ Extend that line. The 32 GB kit that cost around $150 in 2025 and around $450 this autumn⁵ is a $600 part in 2027, and it does not come down in any way that matters before 2029 if the dissenters are right that a glut arrives in 2028,⁶ and not at all if they are wrong. The RAM you have is the RAM you will have. Treat it like that.

Here is the twist I think almost nobody is pricing in. The one thing that could rescue your wallet is cheap Chinese memory. China's own DRAM and NAND makers are ramping exactly the commodity parts the big three are walking away from, and in any other decade that flood would have cratered prices by 2028. I do not think it reaches you. The same instinct that writes data-center exemptions into a chip tariff will keep Chinese memory out of Western consumer channels on security grounds, because a stick of RAM with a Chinese die is politically the same object as a Chinese phone. So the builders get first call on Korean and American memory, you get what is left, and the cheap stuff that would have saved you is contraband. One sentence for the next three years of the memory market: scarcity by contract on one side, scarcity by statute on the other, and you in the middle.

Second-order effects I expect, in order of confidence. Base configurations shrink: the 8 GB laptop comes back, sold as an AI PC with an NPU that runs nothing you care about. Soldered memory becomes universal, because soldered memory is cheaper to buy in bulk and impossible to upgrade, and the makers will call that thin. The used market gets strip-mined: a working 2022 desktop is worth more as 64 GB of DDR5 than as a computer, and someone will build the business that harvests them. And memory becomes a line in the terms of service. "Performance may vary with available system resources" is going to mean "we moved this feature to the cloud, because your RAM is ours now".

2. Graphics cards, or: the Micron playbook comes for GeForce

The fact that matters is a ratio. In the quarter that ended 25 January 2026, Nvidia's gaming business did $3.7 billion and its data-center business did $62.3 billion. Gaming is 5.5 percent of the company, and it fell 13 percent on the quarter while the company as a whole grew 20.⁷ The chief financial officer wrote the sentence you should frame:

We expect supply constraints to be a headwind to Gaming in the first quarter of fiscal 2027 and beyond.

The street already knows. The RTX 5090 listed at $1,999. This autumn one survey found the cheapest US card at about $4,900; the big retailers had no stock at all and marketplace listings ran to $9,659; a tracker that follows 81 cards saw prices rise on 59 of them in a single month.⁸ The constraint is GDDR7, the memory soldered next to the chip, made by the same three companies that sold 2027 to the data centers, who would rather make HBM for customers who never haggle.⁹

Now the part I cannot cite. Micron just showed every component maker the move: when a consumer line is a few percent of revenue and the strategic customers are begging, you do not cut the consumer line, you exit it and call it focus. I think Nvidia does the same thing, slowly, and never announces it. Not an exit. A hollowing. GeForce becomes one halo card a generation, priced above $3,000 and never sold at list, plus a long tail of parts built from whatever the data-center lines reject. The real GeForce becomes GeForce Now: the brand survives as a streaming service, and owning an Nvidia card becomes a thing enthusiasts say the way audiophiles say they own a reel-to-reel. The mid-range falls to AMD and Intel, who need the same memory and have less leverage to get it, so the mid-range becomes the old high-end at the old high-end's price.

People keep asking when the waitlist clears. There is no waitlist for you, because you were never in the queue. The queue is the 2027 order book, and it is full of people who buy by the rack and sign for three years. A years-long wait is not a market failure. It is the published schedule, and the memory makers have said on their own earnings calls that it runs to 2028 at the earliest.

What eventually comes down the pipe is used data-center cards: the H100s and Blackwells rotating out of hyperscalers in 2028 and 2029. I expect that flood, and I expect it to arrive with a catch. The Chip Security Act, which cleared the House Foreign Affairs Committee 42 to 0 in March, would build location verification into export-controlled chips before they ship.¹⁰ It is not law. But the demos exist, the appetite is bipartisan, and I think the ex-data-center card you buy in 2029 phones home whether or not Congress ever passes anything, because the hyperscalers will want it for their own asset tracking. A second-hand GPU that has to be re-registered is not a second-hand GPU. It is a lease with extra steps.

3. Everything else, or: the $500 computer is over

The repricing is general, and I will keep the facts short because the pattern is the same every time. Memory is now about a third of a PC's bill of materials, up from under a fifth a quarter earlier, according to HP, and every maker has done the arithmetic: Dell and HP raised prices by double digits, Lenovo voided every open quote on 1 January, and IDC expects the average PC up 17 percent this year with shipments down 11.¹¹ Sony put the PS5 Pro at $899.99 and the disc PS5 at $649.99; Microsoft put the Series X at $799.99 and said its memory and storage costs are more than 2.5 times what they were.¹² Raspberry Pi raised prices three times this year; the 16 GB board is $220.¹³ A 25 percent Section 232 tariff on advanced AI chips took effect in January, with exemptions for data centers, research, startups and consumer uses; the broad second phase, on all semiconductors, is undecided.¹⁴ And Windows 10 support ended in October 2025 with somewhere between 240 and 400 million PCs unable to move to Windows 11.¹⁵

Extend the lines. The $500 PC does not come back; the entry price for a machine that does anything interesting drifts to four figures and stays there. Consoles go first: I think the next generation is a decoder box with a controller, because a console that streams needs no GDDR, and the memory it would have used is worth more sold to someone else. The pro tier is the only one with local silicon, and it is priced like a PC. Phones become the only computer most people own, which was already true in usage and becomes true in ownership: the laptop turns into a thing a household owns one of, like a printer.

My strangest bet, and the one I am most confident in: Apple ends up as the last company shipping a personal computer that is actually personal. Not out of virtue. Apple signs memory contracts like a hyperscaler, designs its own silicon, and can hold prices by cutting base configurations while everyone else cuts features. The Mac becomes the Leica: the last fully owned, fully local machine, at a price that makes the choice for you. The Windows side goes the other way. I expect the consumer Windows license to become a subscription inside two years with a cloud PC in the bundle, so that the cheapest Windows computer is a thin client to a Windows that runs in Azure. The 240 to 400 million orphaned Windows 10 machines are the largest Linux desktop cohort in history waiting to happen, and I think a real share of them take it, which does nothing for the market and a great deal for the people who keep those machines alive.

The Raspberry Pi line is the one that makes me sad. The $35 computer was a bet that tinkering should be cheap. At $77.50 for 2 GB it is still a bargain against everything else, but the trend says the on-ramp for a kid with a soldering iron now costs what a textbook costs, and every year the ramp gets a little steeper while the locked tablet gets a little cheaper. You can guess which one the school buys.

4. Open models, or: nobody banned them, they did something smarter

The facts here are a list of small moves, none of which is a ban. Meta, the patron that made open weights fashionable, built its next frontier model closed and shipped Muse Spark in April with no weights to download; it has not said whether a new Llama ever comes.¹⁶ The Department of Homeland Security ordered federal civilian agencies to remove DeepSeek technology this month.¹⁷ The FRONTIER Act would put audits, incident reporting and Commerce emergency powers on the largest developers; an executive order in June set up a voluntary 30-day pre-release review; and the EU began enforcing its general-purpose model obligations in August, with fines up to 3 percent of global turnover, a partial exemption for open source, and none for models judged to carry systemic risk.¹⁸ Anthropic's $1.5 billion settlement with authors was approved in July, roughly $3,000 a title.¹⁹ And in May 2025 the card networks cut off the largest community hub for open image models over mature content it hosted; it survives on crypto.²⁰

Put the pieces on one table and the shape is a ratchet. Each click makes releasing a serious open model slightly more expensive, slightly more exposed, slightly less rational, and ratchets do not turn backwards. Open weights do not get outlawed. They get orphaned.

Now the clicks I expect next, which is the part worth your attention. Rentals get a face: to rent a serious GPU in the cloud you will verify your identity the way you do to open a bank account, first as fraud prevention and then as policy. Hosting gets liability: the hubs that host weights get treated like the hubs that host files, and I expect the big one to geo-fence Europe inside eighteen months and to require sign-in for anything over a size threshold. App stores get a rule: an app that runs a local model above some size needs a review, then a license, and the old truth that you cannot ship what Apple will not list becomes the truth for models. Training gets a price: the $3,000-a-title number becomes the floor for a clean dataset, so the only labs that can train on anything are the dozen that can write a ten-figure cheque, which is to say the dozen with no reason to give the result away. And on the other side of the Pacific, China's proposed export rules, which as drafted leave published open-source work alone,²⁵ stop doing that the first time a Chinese open model embarrasses an American one in a way Washington cares about. By then Beijing will have worked out that open weights are its best soft-power export, and a government that has worked that out does not give the export away for free.

So by 2028 I expect the open ecosystem to have no patron on either side. What survives is what survived the end of every other commons: small models that run on what you already own, fine-tunes of the last generous release, and a folk culture of people who keep 2026-era weights the way people keep vinyl pressed from masters nobody can re-cut. The best open image and video models you will ever be able to download may already be out. If one matters to you, you should have a copy. Not a bookmark. A copy.

5. Power, or: the data center is on your bill

Two numbers. In PJM, the grid for thirteen states and Washington, the market monitor estimated that data centers caused 63 percent of the 2025/26 capacity price increase, about $9.3 billion that flows to customers, and the next auction cleared at the cap.²¹ Nationally, residential rates rose 7.3 percent between April 2025 and April 2026, above what the government had forecast.²² The counter-argument is that PJM's market design is the real culprit, and I half buy it: market design is the mechanism, and the thing pulling the lever is a 500-megawatt campus that did not exist three years ago.²³

Where I think it goes. Households get time-of-use pricing as the default and data centers get fixed long-term contracts, which means you pay the peak and they pay the average, and the regulator calls that fairness because the campus brought load. Demand response comes to your thermostat and your car charger: the utility turns you down so the campus stays up, and you get a credit for it. Rooftop solar and a battery stop being an environmental choice and become the only hedge a household has against a bill that is now partly a subsidy to the building down the road. The four largest buyers of compute guided to roughly $700 to $725 billion of capital spending this year, about three quarters more than last year, and that money comes out of markets that used to clear at consumer prices: memory, wafers, transformers, electricians, megawatts.²⁴ You are a price-taker in all of them, and your electricity bill is where that shows up with a utility logo on it.

A timeline I would bet on

This is the section the analysts will not write, so here it is. The dates are guesses. The directions are not.

2027

Memory's worst year, by the makers' own forecast. PC shipments fall by double digits for a second year. The next flagship consumer GPU launches above $3,000 and does not sell at list, ever. The first major console revision ships without a disc drive or a local install option for most titles. A federal frontier-model law passes in some form, with state preemption, and the first EU fines land on a general-purpose model provider. Cloud GPU rentals start asking for identity verification above a spend threshold.

2028

The AI PC pitch quietly ends; the NPU stays on the spec sheet and nobody mentions it. Consumer Windows becomes a subscription with a cloud PC in the bundle. Apple holds Mac prices by cutting base memory and calling it a feature. The weight-hosting hubs geo-fence Europe and require sign-in. The open image and video models everyone still uses are 2026 releases. Either the memory glut arrives, as Bloomberg Intelligence thinks it may,⁶ or it does not; either way list prices for consumer hardware do not come back, because the makers have learned what the market will bear.

2029

The ex-data-center GPU flood arrives, used Blackwells by the pallet, with location checks inside them and a registration step to change hands. Consoles are streaming-first; the local tier is a luxury line. Indie PC gaming is half its 2024 size. Linux desktop share has doubled on the back of the orphaned Windows 10 fleet. For most households the personal computer is a phone plus a thin client.

2030

The thin decade is a fact. The people with 2025-era hardware and 2026-era open weights are the ham-radio community of computing: real, skilled, respected, and irrelevant to the market. Owning a general-purpose computer that does serious work locally is a hobby with a price tag, like film photography. And the phrase personal computer starts to sound like home stereo: a thing your parents had.

I give the whole picture maybe 60 percent. Each line is less certain than that; the direction is more. If you want to prove me wrong, the way to do it is a memory glut that lasts, a GeForce line that gets cheaper, and one big open model released in 2028 by a company that could have kept it. I would love all three. I am not betting on any of them.

What I am doing about it

1. Not buying the flagship. A $6,000 card to render things I can rent by the second is a bad trade, and I say that as someone who wanted one. The card is not the point. What it renders is the point.
2. Keeping the old machine alive. The Windows 10 box runs a current Linux now and will for years. If you need RAM, buy it before the next contract round, not after. Storage too, while the numbers still make sense.
3. Archiving the commons. Local copies of the open models and weights I care about, on a drive that is not a subscription. The ones already released cannot be un-released. The ones not yet released may never be.
4. Renting compute by the creation, not by the month. This is where I stop pretending to be neutral. This post lives on Wurt.app, and the reason is the thesis: if owning the GPU no longer makes sense, the question becomes who rents you one on fair terms. Wurt runs in the browser, free to start, with credits you buy as needed and no subscription. Picwurt makes pictures up to 4K from a prompt or your own photo, with up to nine reference images. Vidwurt makes short videos with sound on supported models, clips up to 30 seconds on Sqwurt and Seedance 2.5, Quick Animate for a one-tap clip from a picture, and Edit Any Video for changing a clip you upload by describing the change. VIP makes free HD clips up to 10 seconds for qualifying accounts, and eligible new accounts can get free starting credits. The card does its work a long way from your desk. You keep the file. That is the deal the next decade is going to offer everyone, and this is one place where the terms are already fair.
5. Making things now. The models are good, the lights are on, and nothing above says either of those is guaranteed. Use the window. Start free, and if you still want the hardware, I compared the cards in best GPU for AI art.