A 1996 Pentium Pro Running Stockfish Would Have Crushed Kasparov

Thirty-one years ago this week, on November 1, 1995, Intel shipped the Pentium Pro. A year and change later, IBM’s Deep Blue took a game off Garry Kasparov — the first computer to do so — before losing the match, then winning the 1997 rematch with hardware worth several million dollars: 30 processors at 120 MHz plus 480 custom chess chips, crunching 200 million positions a second as the world’s 259th fastest supercomputer.

Nicholas Carlini, the Google researcher behind a long line of delightful overachiever experiments, asked an uncomfortable question on the anniversary: was Deep Blue even necessary? His answer, demonstrated rather than argued: he has a 200 MHz Pentium Pro with 64MB of RAM sitting next to him, running a custom port of Stockfish, and by his assessment it would beat 1996 Kasparov easily — “almost certainly” not even close.

The point is not that IBM wasted money. The point is where the winning came from. Deep Blue won on specialized silicon and brute force; three decades of software — the search refinements and evaluation craft inside Stockfish — mean the same feat now fits inside a 31-year-old desktop that cost a few thousand dollars. The expensive part of that historic match, it turns out, was the part that aged out first. Algorithmic improvement compounded; the VLSI co-processors became museum pieces.

There is a reason this belongs on a gaming site and not just a hardware blog. Chess was the original “can a machine beat a human” spectacle — the esports boss fight of 1997, watched by people who had never played a competitive match in their lives. And the lesson Carlini packages is the one every player eventually learns the hard way: skill compounds where hardware plateaus. The engine you can run on a museum piece today would humiliate the era’s greatest champion, not because someone built a bigger machine, but because somebody rebuilt the search. Every “impossible” skill ceiling in games has the same shelf life. Ask the bosses of the fighting games that were solved, or the speedrun categories that were “finished” and then cut in half. The machine was never the story. The method was.

If you want the full walkthrough — including how a Stockfish port gets squeezed into 64MB of RAM — Carlini’s write-up is the best hour you will spend this week, and a fitting companion piece to the vibe-coded launcher saga. It is also a quiet answer to the question of what today’s AI milestones will look like from 2057: enormous, expensive, and eventually runnable on whatever the Pentium Pro of that era turns out to be.