The Large Hadron Collider (LHC) is now under construction at the European Organisation for Nuclear Research (CERN). This 27 km long accelerator requires 1248 superconducting dipole magnets operating at 1.9 K. The cold mass of the dipole magnets is closed by a shrinking cylinder and two end covers at both extremities of the cylinder. The end covers are welded, cast or powder metallurgy (PM) dished-heads equipped with a number of protruding nozzles for the passage of the different cryogenic lines. Structural materials and welds must retain high strength and toughness at cryogenic temperature. AISI 316L-type austenitic stainless steel grades have been selected because of their mechanical properties, ductility, weldability and stability of the austenitic phase against low-temperature spontaneous martensitic transformation. 316LN is chosen for the fabrication of the end covers, while the interconnection components to be welded on the protruding nozzles will be fabricated from forged 316L or 316LN, and welded 316L tubes. Autogenous welds between the nozzles and the interconnection components will be performed by the automatic orbital TIG technique. Several thousands of welds are foreseen. When welding together grades of slightly different composition, or grades issued from different fabrication methods (cast, PM, cold or hot rolled, forged, etc.), phenomena such as variable weld penetration, “off-centre welding” and “arc wander” may possibly appear, resulting in uncontrolled formation of non-axisymmetric welds. Such deflections of the weld pool are difficult to correct for an automatic process and may affect the soundness of the weld. A large and systematic campaign of welding tests associated with video recording of the melt pool has been carried out. Hot metal deflections have been precisely quantified. The results are interpreted in terms of the different content of soluble surface-active elements of the various steel batches and the directions of the thermocapillary flow arising from these different contents. This interpretation gives a quantitative prevision of the hot metal deflections. Possible corrections applicable to automatic welding processes are discussed.