Note: We updated our Google TPU compute model to apply a one-quarter lag between sales and data center deployments, in line with other chip types. We also added revenue shares to the model for each TPU version in Q2 2026 (rather than extrapolating). Finally, we reweighted the TPU revenue shares by a price-to-power ratio for each chip, rather than allocating IT power to chips directly by their revenue share (though this has a small effect). These changes decreased the aggregate current H100-equivalents across the Google data centers that we cover by 15%, and the aggregate projected H100-equivalents by 14%. Most individual data centers decreased, by up to 31%, though Google Papillion increased. The decrease is mainly because of the one-quarter lag.