With clean cathodes in UHV it is shown experimentally that (i) the arc craters are formed successively and displaced without spatial interruption, (ii) there are only a few active craters at one instant for currents below the spot-splitting limit (generally 1-2), and (iii) the crater formation is much faster than is compatible with the heat conduction time scale, the formation time amounting to only a few nanoseconds. From these observations and with the measured values of crater radii and formation times the heating mechanism of the craters is estimated. The calculations show that Joule heating is insufficient to explain the short time scale, therefore ion impact heating is concluded to be the dominant process. Also this energy source is effective only by heating thin layers of a well defined thickness (about 0.1 mu m) which are removed immediately after melting. Therefore, the melting front proceeds faster into the interior than by heat conduction. This model explains the high heat conduction losses at the cathode as measured by Daalder (1977), and also the reason for the arc spot movement.