Abstract Seismic reflection and refraction data from Hikurangi Plateau (southwestern Pacific Ocean) require a crustal thickness of 10 ± 1 km, seismic velocity of 7.25 ± 0.35 km/s at the base of the crust, and mantle velocity of 8.30 ± 0.25 km/s just beneath the Moho. Published models of gravity data that assume normal crust and mantle density predict 5–10-km-thicker crust than we observe, suggesting that the mantle beneath Hikurangi Plateau has anomalously low density, which is inconsistent with previous suggestions of eclogite to explain observations of high seismic velocity. The combination of high seismic velocity and low density requires the mantle to be highly depleted and not serpentinized. We propose that Hikurangi Plateau formed by decompression melting of buoyant mantle that was removed from a craton root by subduction, held beneath 660 km by viscous coupling to slabs, and then rose as a plume from the lower mantle. Ancient Re-Os ages from mantle xenoliths in nearby South Island, New Zealand, support this hypothesis. Erosion of buoyant depleted mantle from craton roots by subduction and then recycling in plumes to make new lithosphere may be an important global geochemical process.