In February 1996, IBM spent millions building Deep Blue — a 30-processor, 480-custom-chip supercomputer that was the 259th most powerful machine on the planet — to beat Garry Kasparov at chess. It lost the match that year, won the rematch in 1997, and entered history as proof that brute computational force could overcome human genius. The question this project asks is blunt: was any of that hardware necessary? The answer is no, and it isn't close. The author ported Stockfish — today's strongest open-source engine, rated around 3700 Elo — to a 200 MHz Pentium Pro with 64 MB of RAM, a machine that would have cost roughly $4,000-$5,000 at launch in late 1995. After rewriting the code for 32-bit arithmetic and optimizing for the Pentium's architecture, the resulting engine plays at an estimated 3300 Elo at classical time controls. Kasparov peaked just under 2800. A 500-Elo gap translates to roughly a 9-out-of-10 expected score. Deep Blue, for reference, was rated about 2700. The experiment runs two tracks. First, take Crafty 11.13, a competitive 1996 engine, and progressively bolt on thirty years of algorithmic knowledge: fail-soft search, late move reductions, correction history, NNUE neural-net evaluation. On the Pentium Pro at 10 seconds per move, Crafty climbs from 2383 Elo to 2974 — a 591-point gain from pure software improvements on identical hardware. Second, port Stockfish itself, tune it for the Pentium's constraints, and watch it hit 3180 at 30 seconds per move. The head-to-head clincher: modern engine on old hardware versus old engine on modern hardware, 100 games. Score: 78 wins, 17 draws, 5 losses for the time-traveling software. The core finding decomposes cleanly. Of the roughly 1335 Elo points gained between 1996 engines and modern Stockfish, 720 come from better algorithms — smarter move selection, better position evaluation, NNUE — and 615 from faster hardware. Software wins by a comfortable margin. The critical insight is that chess engines improve the same way humans do: not by looking at more positions (though they do that too), but by looking at better positions and evaluating them more accurately. The engineering is charming and real. The Pentium Pro boots from a 1.44 MB floppy disk running a custom bare-metal OS whose only job is to initialize the network card and await chess queries via HTTP. The author connected it to the internet with a Realtek RTL8139C Ethernet card so readers can actually play against it in the browser. The code was largely written by LLMs (Fable 5.1, Opus 5.5, GPT-6), which the author discloses plainly — the article's prose is human. What makes this more than a stunt is the controlled methodology. Each algorithmic improvement is isolated and measured independently, with Elo estimates at multiple time controls and error bars throughout. The extrapolation to classical time controls is transparent about its assumptions (diminishing returns on the +70 Elo per tripling curve). This is empirical computer science done on vintage iron, and the rigor is better than most published benchmarks. The deeper lesson is about the mythology of hardware supremacy. IBM's Deep Blue narrative — that you needed a room-sized supercomputer to beat the world champion — was always partly marketing. Thirty years of algorithmic progress have made the hardware almost beside the point. A $1,000 depreciated workstation from 1997 running software that understands chess better would have demolished Kasparov without breaking a sweat. The millions IBM spent bought spectacle, not necessity.
