"The machine is stronger than the human" in 1997 and the same sentence today are two different claims, and money shows the difference better than anything else does. In 1997 the sentence stood on an installation a corporation built for one task, and what it cost that corporation has not said — not then, and not in the 29 years since. In 2019 the same sentence stands on a line in a Science paper: $144 at cloud spot rates.
There are fourteen prices here, from nine matches. Four of them have a number. The other ten are empty — and that is not a hole in the research, it is the main thing this page has to show. The most famous machines on the list had their cost published nowhere, while the money around the matches — purses, fees, an estimate of what the publicity was worth — is known to the dollar and turns out to be larger than anything said about the machines themselves.
Not to be confused with the neighbouring page called The Price of a Move: that one counts game trees and processors, and its "price" is a figure of speech. Here the price is money.
Every sum is nominal, in the money of its own year. A 1997 dollar is printed on this page as a 1997 dollar and as nothing else.
Conversion to today's money is computed on the fly from the consumer price index and labelled as a snapshot taken on a date. No ready-made "in today's dollars" is stored in the data, deliberately: a figure written down once as "in such-and-such year's prices" goes on quietly ageing and nobody notices.
Every price shows the boundary it was drawn around: hardware only, the whole system, the whole project, cloud rental, or unknown. Without it the number means nothing — "$3 million for Deep Blue" and "$3 million on the Deep Blue project" are different claims.
A confidence tag sits next to each figure: measured, specification, reconstructed, order of magnitude. The tag can be clicked and will explain itself; all four are written out in full at the foot of the page.
A hollow band the full width of the axis does not mean "figure pending". It means exactly what it draws: the number could have been anywhere along this axis, and no published basis exists for preferring one point on it to another.
All fourteen rows come out of systems.js as they lie there, with their confidence tag and their basis. The full text of every record, with sources, is further down this page.
IBM never disclosed what Deep Blue cost. The "$3 million" and "$5 million" have circulated in retellings for decades and lead nowhere: the sum is not in Hsu's paper on the chess chips, not in Campbell et al., not on IBM's own page, not in Wikipedia. Neither figure has a traceable origin — which means that putting any tag beside them except an empty one would be a lie.
The other half of the same ledger, though, is known exactly. The 1997 purse: $700,000 to the winner and $400,000 to the loser, with a further $100,000 to the team from Carnegie Mellon University. And the estimate IBM itself put about: $250,000,000 worth of advertising for the 1996 match and $500,000,000 for the 1997 one.
Not one published number in 29 years.
For the 1997 match. A year earlier, $250,000,000.
The comparison is worth saying out loud. The machine that won the match has not one published figure of what it cost. The benefit of that match has a figure, and it is over four hundred times the entire purse. Snapshotted to a 2025 dollar, half a billion 1997 dollars is $1,002,831,476. The CPI caveat applies here too: the index tracks advertising rather closer than it tracks computing, but still not directly.
Deep Blue versus Garry Kasparov — Wikipedia, Wikipedia · IBM's Deep Blue Chess Grandmaster Chips — F.-h. Hsu, IEEE Micro 19(2), pp. 70–81 · Kasparov versus Deep Blue: The Rematch — J. Schaeffer, A. Plaat, ICCA Journal 20(2), pp. 95–101
In 2019 Pluribus beat professionals at six-handed poker, and its authors wrote down what the computation cost: "At current cloud computing spot instance rates, this would cost about $144 to produce" — Brown and Sandholm, Science 365(6456), p. 887. The "around $150" that circulates in retellings is a rounding of that line.
What it covers: the blueprint strategy computation and nothing else — 12,400 core hours on a single 64-core server over eight days. What it does not cover: any invoice at all (the machine was the group's own, not rented), the real-time search across the 20,000 hands played, the payments to the players ($50,000 among the participants of the first experiment), and the years of laboratory work whose code, by the authors' own statement, makes up the bulk of Pluribus. No provider and no instance type are named and spot prices move: this is the authors' arithmetic, not a bill.
"Unknown" and $144 cannot be compared directly, and that is the point. In 1997 the price could not be learned because it was a corporation's internal business, and the match was an advertising campaign worth hundreds of millions to it. In 2019 the price is stated in one line of a paper because it had become too small to be worth hiding. Twenty-two years separate those two lines, and what moved across them was not so much the price as who owned it.
Superhuman AI for multiplayer poker — N. Brown, T. Sandholm, Science 365(6456), pp. 885–890
The CPI measures something other than what is being compared here. It is a basket of housing, food, fuel and medical care, in which computing equipment is a fraction of a per cent. Running Deep Blue's or Watson's price forward through the CPI answers the question 'how much bread did this machine cost', not 'how much cheaper has computing become'. The difference is enormous. Nordhaus (Journal of Economic History, 2007, Table 10) set the U.S. government's official hedonic indexes beside the price of the same work actually performed: the official BEA index gives a real price decline of 18.7% a year over 1969–2004, while the performance-based measure gives 50.7% a year over 1969–2005. Understated by roughly a factor of three — and the gap is not about precision but about the question: the official index prices a computer, not a task.
Separately: the line 'computing keeps getting tens of per cent cheaper every year' stops being true after 2011. The official price indexes for computing hardware — BEA's index for computers and peripherals (B935RG3A086NBEA) and BLS's producer price index for electronic computers (PCU334111334111) — stopped falling around 2011 and have been rising since 2020. The claim holds for 1959–2010 and fails for 2011–2026. On a site that compares machines from 1979, 1997, 2011 and 2016, the caveat cannot be dropped: the slope changes exactly in the middle of the period on display.
The factor is the 2025 annual average of CPI-U (U.S. city average, all items, not seasonally adjusted, series CPIAUCNS) divided by the annual average for the year of the sum. Annual averages, not December readings: monthly points give a different answer and a wider spread. The series was snapshotted on 12 August 2026; it does not refresh itself, and BLS revises recent points, so a year from now these numbers will differ.
The contract asks for base: 2026. No annual average CPI for 2026 exists, and none will until January 2027, because the year is not over. The base is the last complete annual average, 2025. Where the site says '2026 dollars', read '2025 dollars'.
| What | Year | Nominal | Factor | Snapshot |
|---|---|---|---|---|
| DECsystem-1090, base package through practical configurations | 1978 | $620,000 – $2,000,000 | no factor | — |
| Pentium Pro 200 MHz, 256 KB L2 — chip price at introduction | 1995 | $1,325 | no factor | — |
| 1997 match: winner's share of the purse | 1997 | $700,000 | ×2.0057 | $1,403,964 |
| 1997 match: loser's share of the purse | 1997 | $400,000 | ×2.0057 | $802,265 |
| 1997 match: Carnegie Mellon's addition for the team | 1997 | $100,000 | ×2.0057 | $200,566 |
| Advertising return on the 1996 match, IBM's estimate | 1996 | $250,000,000 | no factor | — |
| Advertising return on the 1997 match, IBM's estimate | 1997 | $500,000,000 | ×2.0057 | $1,002,831,476 |
| Watson's hardware, as estimated by Tony Pearson of IBM | 2011 | $3,000,000 | ×1.4312 | $4,293,737 |
| 2016 match: the purse | 2016 | $1,000,000 | ×1.3414 | $1,341,390 |
| 2016 match: Lee Sedol's $150,000 fee and $20,000 for his win | 2016 | $170,000 | ×1.3414 | $228,036 |
| 2017 match: the purse, split four ways | 2017 | $200,000 | ×1.3134 | $262,682 |
| Pluribus: the blueprint computation at cloud spot rates | 2019 | $144 | ×1.2593 | $181 |
| Pluribus: payments to the first experiment's players | 2019 | $50,000 | ×1.2593 | $62,964 |
The factor is recomputed on this page at every load: the annual average CPI-U for the base year divided by the annual average for the year of the sum, straight out of the series snapshot in YVC_CONFIG. The snapshot holds no annual average for 1978, 1995 or 1996 — it carries only the years the records mention — and this page will not invent them. Consumer Price Index for All Urban Consumers: All Items in U.S. City Average, annual, not seasonally adjusted — U.S. Bureau of Labor Statistics, FRED, Federal Reserve Bank of St. Louis · Two Centuries of Productivity Growth in Computing — W. D. Nordhaus, The Journal of Economic History 67(1), pp. 128–159 · Price Index for Private Fixed Investment: Information Processing Equipment: Computers and Peripheral Equipment — U.S. Bureau of Economic Analysis, FRED, Federal Reserve Bank of St. Louis · Producer Price Index by Industry: Electronic Computer Manufacturing — U.S. Bureau of Labor Statistics, FRED, Federal Reserve Bank of St. Louis
Carnegie-Mellon University · Monte Carlo, Monaco
There is nothing to convert.
The price of the machine BKG 9.8 played on is unknown: no source names its model or configuration. It belonged to the university and was not bought for the match.
Nominal 1978 dollars. The snapshot holds no annual average CPI for 1978, so no conversion is offered here.
Datapro's September 1978 report lists the DECsystem-1090 base package on a KL10-D processor at $620,000 and the KL10-D processor alone at $395,000, and says practical configurations start around $700,000 and run past $2,000,000. This prices a machine of the same class, not the one that played the match.
Digital Equipment DECsystem-10 — Datapro Research Corporation, Datapro Reports, September 1978; scanned on bitsavers
University of Alberta · The Computer Museum, Boston, USA
There is nothing to convert.
Neither SGI's press release nor the Chinook project pages give a price, and no 1994 Silicon Graphics price list for the Challenge XL could be found. The machine came from the match sponsor, Silicon Graphics itself, so it was probably never purchased for the event at all.
Silicon Graphics World Checkers Championship — Silicon Graphics, Inc., a Silicon Surf press release, 1994
IBM · New York, Equitable Center
There is nothing to convert.
IBM never disclosed what Deep Blue cost. The '$3 million' and '$5 million' figures appear in popular retellings but lead nowhere: neither Hsu's paper, nor Campbell et al., nor IBM's own page, nor Wikipedia names a sum. The only firm money around the match is the purse — $700,000 to the winner and $400,000 to the loser, with a further $100,000 to the team from Carnegie Mellon University.
Deep Blue versus Garry Kasparov — Wikipedia, Wikipedia · IBM's Deep Blue Chess Grandmaster Chips — F.-h. Hsu, IEEE Micro 19(2), pp. 70–81
NEC Research Institute · Princeton, New Jersey, USA
There is nothing to convert.
Buro's report does not give a price for the machine, and none of his other texts do either. It belonged to the NEC Research Institute.
Nominal 1995 dollars. The snapshot holds no annual average CPI for 1995, so no conversion is offered here.
Intel's announced price for the 200 MHz Pentium Pro with 256 KB of L2 cache at the processor's introduction on 1 November 1995. This is the chip alone, not the machine, and by August 1997 it had fallen considerably — but no price list dated to the match could be found.
200 MHz Intel Pentium Pro Benchmarks at 366 SPECint92 — HPCwire, HPCwire, 3 November 1995
IBM Research · Yorktown Heights, New York — a set built inside IBM's research lab
Nominal 2011 dollars. Snapshot of 2026-08-12: the factor to a 2025 dollar is 1.4312, which makes it $4,293,737.
An estimate, not a price. Tony Pearson, an IBM senior consultant and master inventor, put the cost of Watson's hardware at around three million dollars; Computerworld carried the estimate on 21 February 2011. It covers the servers alone — not the DeepQA development, not the corpus, not the team.
Can anyone afford an IBM Watson supercomputer? — Lucas Mearian (the estimate is Tony Pearson's, of IBM), Computerworld, 21 February 2011 · IBM Watson — Wikipedia, Wikipedia
There is nothing to convert.
IBM never disclosed what the DeepQA programme cost in total. The scale of the effort is known — some two dozen researchers over several years — but not the figure.
Building Watson: An Overview of the DeepQA Project — D. Ferrucci, E. Brown, J. Chu-Carroll, J. Fan, D. Gondek et al., AI Magazine 31(3), pp. 59–79
Google DeepMind · Seoul, the Four Seasons hotel — though the computing was elsewhere
There is nothing to convert.
There is no invoice. The hardware was Google's and ran in Google's own cloud; no rental price and no internal rate for 48 first-generation TPUs was ever published — the first-generation TPU was never sold or rented to anyone outside.
In-Datacenter Performance Analysis of a Tensor Processing Unit — N. P. Jouppi et al., ISCA '17, Proceedings of the 44th Annual International Symposium on Computer Architecture; arXiv:1704.04760 · Mastering the game of Go without human knowledge — D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou et al., Nature 550(7676), pp. 354–359
There is nothing to convert.
What AlphaGo cost to build was never disclosed. Google's 2014 purchase price for DeepMind does not belong here: that is the price of a company, two years before the match, and it says nothing about this match's hardware or labour. The firm money is the $1 million purse, which DeepMind gave to charity after winning, and Lee Sedol's $170,000 — $150,000 for playing and $20,000 for the game he won.
AlphaGo versus Lee Sedol — Wikipedia, Wikipedia
Carnegie Mellon University; computing at the Pittsburgh Supercomputing Center · Pittsburgh, Rivers Casino — though the thinking happened half an hour away, on the Bridges supercomputer
There is nothing to convert.
There is no invoice, and that is the point. Bridges was built with United States National Science Foundation money; time on it is allocated by XSEDE grants rather than sold. Neither the Pittsburgh Supercomputing Center nor the authors published an internal rate per machine hour, nor an estimate of what 25 million core hours would have cost in a commercial cloud. The only firm money around the match is the purse: $200,000 split among the four professionals. The machine finished $1,766,250 up in chips, but that is a scoreline, not money — nobody was playing his own.
Libratus: The Superhuman AI for No-Limit Poker — N. Brown, T. Sandholm, Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), pp. 5226–5228 · Carnegie Mellon Artificial Intelligence Beats Top Poker Pros — Byron Spice, Garrett Allen, Carnegie Mellon University News, 31 January 2017
Carnegie Mellon University and Facebook AI Research · the paper names no venue: play went through an interface built for the experiment
Nominal 2019 dollars. Snapshot of 2026-08-12: the factor to a 2025 dollar is 1.2593, which makes it $181.
The figure comes from the primary source verbatim: 'At current cloud computing spot instance rates, this would cost about $144 to produce' — Brown and Sandholm, Science 365(6456), p. 887. The 'around $150' that circulates in retellings is a rounding of this one. What it covers: the blueprint computation only, 12,400 core hours on a 64-core server over eight days. What it does not cover: any actual bill — the machine was the group's own, not rented; the real-time search across the 20,000 hands played; the payments to the players ($50,000 among the participants of the first experiment, $2,000 each plus a possible $2,000 bonus in the second, funded by Facebook); and the years of laboratory work whose code, by the authors' own statement, makes up the bulk of Pluribus. No provider and no instance type are named, and spot prices move — this is the authors' arithmetic, not an invoice.
Superhuman AI for multiplayer poker — N. Brown, T. Sandholm, Science 365(6456), pp. 885–890
There is nothing to convert.
Nobody published what the playing machine cost. The configuration is known — two Xeon E5-2695 v3 processors and under 128 GB of memory — but not the price paid for it. By 2019 those processors were off Intel's price list: the chip was five years old, bought second-hand or already sitting in the lab.
Superhuman AI for multiplayer poker — N. Brown, T. Sandholm, Science 365(6456), pp. 885–890 · Intel Xeon Processor E5-2695 v3 — Intel, Intel Product Specifications (ARK), SKU 81057
FAR AI, MIT and UC Berkeley for the attacking team; KataGo itself is David Wu's engine · the KGS Go Server, ordinary online conditions
There is nothing to convert.
There was no money around these games at all: no purse, no sponsor, no stake — ordinary online play on a public server. FAR AI never published what training the attacking policy cost, and converting 2,223 V100 GPU-days into cloud dollars is not something this record will do: no provider and no rates are named, and any such figure would be our invention rather than a found number.
Adversarial Policies Beat Superhuman Go AIs — T. T. Wang, A. Gleave, T. Tseng, K. Pelrine, N. Belrose, J. Miller, M. D. Dennis, Y. Duan, V. Pogrebniak, S. Levine, S. Russell, ICML 2023; arXiv:2211.00241 (v1 — 1 November 2022, v3 — 18 February 2023, v4 — 13 July 2023)
What the price is drawn around
hardware only The price of the iron and nothing else: a machine or a part of one. No development, no salaries, no electricity.
the whole system The installation as it stood: nodes, memory, disks, switch — everything needed to make it run.
the whole project Everything the work cost: people, years, data, hardware. The widest boundary and the one least often published.
cloud rental Somebody else's capacity at list price. It may be an invoice, or it may be a calculation of what it would have cost.
unknown The boundary cannot be named: the source will not say what it counted — or there is no number at all.
Confidence
M · measured A direct measurement of this quantity, with a published method. If the source will not say where the meter was, it is not an M.
P · specification The maker's own published figure: a TDP, a price list, a datasheet. Stated, but never measured in the installation described.
R · reconstructed Derived from the configuration: a TDP times a count, the length of a game, a comparable machine. The derivation is in the method field.
O · order of magnitude Good for scale only -- ten times, a hundred times -- and not for setting two numbers against each other.
The question mark means the tag itself is in doubt: read the letter and then discount it.
The first thing that does not follow: "computing got N times cheaper". Four points across forty-four years will not support any curve, and the four that exist are drawn around different boundaries: the price of a chip, the price of a machine of the same class, an outside estimate of some servers, and an arithmetic exercise at cloud spot rates. Lining them up pretends they are measurements of the same thing.
The second is about the slope. The claim that computing gets tens of per cent cheaper every year holds for 1959–2010 and fails afterwards: the official price indexes for computing hardware stopped falling around 2011 and have been rising since 2020. Five of the nine matches on this page fall on that plateau. "Hardware got cheaper, so machines started winning" simply does not work over the stretch after 2011.
What the money does show reliably is whose machine it was. 1979: the university's, and not bought for the match. 1994: supplied by the match sponsor, who also built it. 1997: on one side a corporate installation whose price is the corporation's own business, on the other an ordinary PC in an NEC laboratory whose price nobody wrote down either. 2011: a corporation's servers, priced from outside. 2016: Google's own cloud, and the first-generation TPU was never sold to anyone at all. 2017: a supercomputer built with National Science Foundation money, whose time is allocated by grant rather than sold. 2019: eight days on one server, converted into dollars by the authors themselves. 2023: a public Go server, with no money around it whatsoever. The price stopped being a corporate secret not because anyone disclosed it, but because it got too small to be worth keeping.
Data and tags: systems.js. The watts of the same nine machines are on the Watt League page; the size of the games and the count of processors are on The Price of a Move.