In 2011, Cory Doctorow took the stage at Chaos Communication Congress in Berlin and announced that the copyright wars were just a warm-up. The real battle, he said, lay ahead: over the right to even own a machine capable of running any code. The talk was called "The Coming War on General Purpose Computation," and for fifteen years it read like prophecy about laws, DRM, and signed bootloaders.
The prophecy came true. Only the weapon turned out different: not law, but price list.
It all started with a comment in a discussion about a vulnerable proprietary format parser: proprietary formats aren't "too complex," they simply lack public corpora for fuzzing. The thought stuck, but not the fuzzing itself—rather the phrase nearby, about how the war for the general purpose computer now runs through the price of memory. That same morning, Doctorow's 2011 speech surfaced in feeds, presented as "prophetic." Two sources coinciding in one day—a rare hint: worth checking what became of the fifteen-year-old thesis.
The check is simple: if the war really followed Doctorow's scenario—through bans, laws, and technical restrictions—then by 2026 we should find traces of regulatory victory: laws, prohibitions, legal precedents. If it followed a different scenario, the traces won't be in court decisions but in financial reports. Let's look at the reports.
Start with the price tag, because it leaves no room for interpretation. A 32GB DDR5 kit cost under $100 in mid-2025; in July 2026, the cheapest went for about $375. DRAM contract prices in Q1 2026 alone rose 90–95%—a record in TrendForce's entire observation history. By August, Tom's Hardware recorded 500% growth over twelve months: 128GB DDR5 cost $3,399, and the bottom of the price chart shot up tenfold from its historical low.
The mechanics are simple to the point of obscenity. All three memory manufacturers—Samsung, SK Hynix, and Micron—redirected over half their wafers to HBM, high-speed memory wrapped around AI accelerators. An HBM stack is multiple DRAM dies connected vertically through through-silicon vias: hungry for both wafers and yield. Commodity DDR as a result didn't just slow growth—it got starved. The resulting picture is unique: record prices with shipments that barely grew. Micron reported $41.5 billion in quarterly revenue with +346% year-over-year growth at 84.9% gross margin—and shipments in single-digit percentages. Growth almost entirely from pricing.
The safety net is signed too: Micron locked in 16 take-or-pay contracts with a revenue floor around $100 billion. SK Hynix holds roughly 57% of the HBM market and prints 72% operating margin—a company that three years ago lost money on every chip. Economists call this a "supercycle," skeptics call it "peak cycle," because in 2022–23 DDR4 fell from $5 to $1.63, and back then too it seemed like nothing would ever end.
The buyers of this magnificence are four companies, and their reports read like confessions. Microsoft attributed roughly $25 billion of its capex specifically to pricier memory; Alphabet spent $180–190 billion over the year and ultimately conducted its first major stock offering at $49.6 billion because free cash flow went negative; Amazon went full circle from $18.2 billion inflow to $7.6 billion outflow; Meta saw free cash flow drop 91% in a quarter. This isn't about "big spending in general"—reports carried phrasing like "higher component pricing," when the buyer names your price increases by name in their financial documents.
While this happens, the private buyer stands in line for the same product empty-handed. This isn't conspiracy: manufacturers find it more profitable to sell all memory to a handful of clients with take-or-pay contracts than to break it into retail batches. The shortage is real even without this—supplier inventories contracted to 2–4 weeks versus normal 13–17—but product allocation decides everything. Retail gets leftovers at spot price.
The same storm hits graphics. GDDR chips are made on the same lines and compete for the same wafers. The RTX 5090 debuted in early 2025 at $1,999 and by September 2026 averaged $5,000: 41% growth since February, 145% above MSRP, individual versions at $5,740. A simple gaming rig upgrade turns into a financial decision at the level of "second used car."
Rumors about the next generation add spice: the RTX 60-series on Rubin silicon is expected with 32GB on a 512-bit bus—meaning even the future flagship doesn't promise capacities an enthusiast considered luxury just yesterday. Every megabyte of VRAM that went to the datacenter is a megabyte a 4K game won't see, or will see at a premium and delay.
Now the most interesting part, what doesn't make headlines about prices. Parallel to the price shock ran a quiet process in software. In August 2024, PyTorch closed RFC #134015: binary builds of version 2.5 drop CUDA 11.8, and with it—support for the 2012 Kepler architecture. In January 2025, NVIDIA in CUDA 12.8 release notes stated: support for Maxwell, Pascal, and Volta "feature-complete," will be frozen in the next release. Nobody cut anything off with both hands: the driver works, the chip is sound. But the upper floors of libraries vanish, and "run anything" becomes "run what still compiles for your architecture."
There's no law here. There's the fact that machine learning rewrote the priority list: code is optimized for H100, not for the GTX 1080 in your garage. The effect for the owner is the same as a DRM chip in the motherboard: the machine stops being general purpose not because it was forbidden, but because the world moved forward without it.
This war, as befits one, had trial runs. In 2010, Sony removed the OtherOS feature from PlayStation 3 via firmware update—Linux on a purchased console. Six years of legal siege ended in 2016 with a $7 million settlement: $55 for those who actually used Linux, $9 for those who bought the console believing the promises. Formally the owner got compensation; in practice the market learned a different lesson—you can delete a purchased feature, the price a decade later measures in cents per unit.
In 2008, Apple laptops and other vendors' machines broke en masse due to NVIDIA chips with defective "bad bumps": solder made of 95% lead cracked from thermal cycling, and The Inquirer opened chips under electron microscopes to prove it. NVIDIA then wrote off $200 million, vendors announced recalls. Note the symmetry: back then hardware performance died physically, and it cost the manufacturer money. Today the hardware lives, support dies—and it costs the owner, not the manufacturer.
Doctorow's speech came true by the letter and failed in spirit. He expected the general purpose computer to be taken away by regulation: signed bootloaders, SOPA-type laws, rootkits by statute. None of this happened at mass scale—and wasn't needed. Competition for memory did what lobbyists couldn't: the PC enthusiast became an economic anachronism, a buyer claiming scarce goods without a take-or-pay contract.
There's bitter elegance in this: you can't ban the general purpose computer—Turing can't be repealed by decree. But making it unprofitable—that you can do, and it doesn't even need a law, just redirect 50% of wafers to a different product line. Ownership of computing is now regulated not by law but by queue: whoever has the contract has the machine fit for work.
What's next? The supercycle can burst—in 2028 a wave of new fabs arrives, and memory has crashed before with chemical precision. Then cheap 128GB kits will appear again at clearance, and enthusiasm will pretend nothing happened. But the infrastructural conclusion can't be undigested: a machine's universality is not a technical property but an economic regime. And regimes, as history teaches, change not by decree but by price.