Supercomputers/Analysis/Cost per petaflop
Analysis 15
What a petaflop costs
57 systems in this dataset have both a stated cost and a measured Rmax. Restated in 2025 dollars, a GFlop/s of measured performance cost a median of $3.1 million on the 3 machines that first ran before 2000, and $3.8 on those that first ran from 2020: about 800 thousand times less. The fall was fastest early. Before 2010 the price fell tenfold every 3.7 years; since 2010 it has taken 10.3. The cheapest measured performance on record is El Capitan, at $0.42 per GFlop/s.
A stated cost is not an invoice. The figures here are what operators, funders, vendors and the trade press said a machine cost, and they cover different things: 41 are contract or purchase prices, 14 are grant awards that may also pay for operations or storage, and 2 are third-party estimates. A difference of two or three times between two machines can be a difference in what was counted. The trend survives that, because the points span 7 orders of magnitude.
This is also the measured world only. Of the 129 systems with a stated cost, 69 have no measured Rmax or carry the Rmax of a later configuration, and 3 costs paid for an upgrade rather than a machine. 55 of the 69 are AI or mixed AI and HPC systems, whose costs often include buildings and power and whose performance is quoted at low precision, so they cannot be placed on this axis.
Researched with AI assistance, checked by scripts. These rows were researched by AI agents reading public operator, laboratory and vendor sources, appended by scripts and checked by automated validation; no subject-matter expert has reviewed them (how this site is made). A mechanical second-reader pass has since fetched every cited source: 1,350 of 1,581 systems have a readable citation that names them, and 64 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.
- Contract or purchase price
- Grant or funding award
- Third-party estimate
Table view: Stated cost per GFlop/s of measured performance
| First light | Systems | Median 2025 US$ per GFlop/s | Cheapest | Dearest |
|---|---|---|---|---|
| 1990 to 1999 | 3 | $3.1 million | ASCI Blue-Pacific $89,034 | KISTI Cray Y-MP C916 $5.3 million |
| 2000 to 2004 | 6 | $9,088 | System X $885 | CPTEC NEC SX-6 $169,401 |
| 2005 to 2009 | 12 | $299 | System G (Virginia Tech) $106 | ASC Purple $5,983 |
| 2010 to 2014 | 11 | $61 | Stampede $7.6 | Gordon $131 |
| 2015 to 2019 | 11 | $13 | Sunway TaihuLight $3.9 | Quinde I $35 |
| 2020 to date | 14 | $3.8 | El Capitan $0.42 | Nautilus $11 |
How this was calculated
Cost per GFlop/s is the stated cost in US dollars, restated in 2025 dollars, divided by the measured HPL Rmax in GFlop/s. Costs come from the stated-cost field of each system's profile, each with its own source; Rmax comes from the system's row. Costs first stated in another currency were converted by the source that reported them, not by us.
Inflation uses the US consumer price index for all urban consumers, annual average (FRED series CPIAUCSL). A cost is deflated from the earlier of the year it was stated and the year the machine first ran, since a figure reported years later was still spent when the machine was built, and a contract announced before first light was spent around the award. Systems are plotted at their first-light year.
Left out: costs that paid only for an upgrade, expansion or refresh of an existing machine (3), and systems recorded as upgraded, whose Rmax describes a later configuration. A machine expanded in place and not marked upgraded keeps its latest Rmax against its original cost, which makes it look cheaper than it was. The scope tag is read from each cost's note: a third-party estimate where the source is an estimate, a grant or funding award where the note says so, and a contract or purchase price otherwise.
The trend is the least-squares slope of the base-ten logarithm of cost per GFlop/s on first-light year: over all 57 systems it falls tenfold every 5.8 years (R² 0.87); on the 21 systems that first ran before 2010, every 3.7 years (R² 0.90); on the 36 since, every 10.3 years (R² 0.68). The later fit is looser, partly because its points span grant awards, contracts and street prices in a narrower range of years.
Record lows
Each system here was cheaper per GFlop/s than every system before it in this dataset.
| First light | System | Country | Stated cost (US$ m, nominal) | 2025 US$ per GFlop/s | Cost figure |
|---|---|---|---|---|---|
| 1991 | Touchstone Delta (Caltech) | USA | 18.0 | $3.1 million | Contract or purchase price |
| 1998 | ASCI Blue-Pacific | USA | 93.0 | $89,034 | Contract or purchase price |
| 2000 | ASCI White | USA | 110 | $41,652 | Grant or funding award |
| 2001 | Lemieux | USA | 36.0 | $15,082 | Grant or funding award |
| 2003 | System X | USA | 5.20 | $885 | Contract or purchase price |
| 2007 | Encanto | USA | 20.0 | $245 | Grant or funding award |
| 2008 | System G (Virginia Tech) | USA | 1.20 | $106 | Contract or purchase price |
| 2010 | Trestles (SDSC) | USA | 2.80 | $61 | Grant or funding award |
| 2011 | HokieSpeed | USA | 1.40 | $17 | Grant or funding award |
| 2013 | Stampede | USA | 27.5 | $7.6 | Grant or funding award |
| 2016 | Sunway TaihuLight | China | 270 | $3.9 | Contract or purchase price |
| 2021 | HiPerGator AI | USA | 33.0 | $2.2 | Contract or purchase price |
| 2023 | Gaia | Brazil | 14.6 | $2.2 | Contract or purchase price |
| 2024 | El Capitan | USA | 600 | $0.42 | Grant or funding award |
The read
The price of measured performance has not fallen at a steady rate. The steepest stretch is the move from vector machines and custom parallel systems to commodity clusters in the late 1990s and 2000s: the median cost per GFlop/s fell by a factor of 10,236 between the pre-2000 group and the 2005 to 2009 group. Since 2010 each tenfold fall has taken about 2.7 times as long as before. The data here cannot say why, and the post-2010 points are noisier, so read this as a slowdown in the price of FP64 performance on the machines that report it, not as a law.
In the latest group the three lowest prices belong to El Capitan ($0.42), Santos Dumont ($1.4), and ABCI 3.0 ($1.6). Grant-funded machines sit higher, partly because an award often pays for more than the hardware: since 2010, the 7 award figures have a median of $17 per GFlop/s against $9.3 for the 27 contract or purchase prices.
What this cannot measure is the AI build-out. Those machines are bought for low-precision arithmetic, so a cost per FP64 GFlop/s undervalues them, and their stated costs usually include land, buildings and power. The build pipeline covers that spending in megawatts instead.