How it's built.
The Deflation Index measures cost per unit of capability: what a dollar buys in compute, bandwidth, kilowatt-hours, and stored energy. It is not a cost-of-living measure. The distance between the two is the subject. Mechanically: a weighted geometric mean of four technology-cost series, indexed to 1990 = 100, with CPI as the reference line.
The formula
The index is a weighted geometric mean: DIt = ∏ Si,twi, where each Si,t is a sector cost series indexed to 100 at its 1990 base and each wi is its weight. Geometric, not arithmetic, because the series span orders of magnitude — an arithmetic average quietly becomes a chart of its slowest component. Late-start sectors hold at 100 until their first measured year. The arithmetic variant is published alongside as a sensitivity.
The four series
| Sector | Metric | Weight | Starts | Sources |
|---|---|---|---|---|
| Computing | compute · storage · memory | 29.41% | 1990 | Cost-of-computing record, Komorowski / Backblaze, McCallum |
| Communications | $/Mbps·mo transit | 23.53% | 1998 | DrPeering (Norton), TeleGeography |
| Energy | $/kWh solar LCOE | 29.41% | 2010 | IRENA |
| Transportation | $/kWh Li-ion pack | 17.65% | 2010 | BloombergNEF, IEA |
The weights balance GDP contribution, enabling effect, and deflationary force — and that third factor makes the index partly weighted by the thing it measures. We disclose and size that circularity in the weighting doc: removing the factor entirely leaves the composite at −99.98% (published: −99.97%).
The line, and its reference
The index is drawn against CPI-U (BLS, CPIAUCSL), indexed to 1990 = 100, so a nominal-dollar series has its context. Through 2025: technology -99.97%, CPI +150%. M2 (FRED, M2SL) is published in the data for anyone who wants the monetary comparison — the argument that comparison suggests is an essay, kept separate from the measurement.
The v4.0 correction
In 2026 we audited our own index and found the headline was too small: arithmetic averaging muted the fastest series, some early anchors were back-extrapolations rather than measurements, and some labels didn't match their series. v4.0 rebuilt the index on three rules — every series is the single metric its label claims, every datapoint traces to a published source, every series starts when defensible measurement starts. The full account is here; the audit memo is in the repository, next to the data it criticizes.
What it isn't
Not a counter-CPI, not a cost-of-living measure, not investment advice. It measures the technology input layer — raw units of capability — and deliberately excludes housing, healthcare, education, and everything else you actually spend most of your money on. What it actually measures covers the distinction.