There is a big question whether to do cluster analysis at all on the HB utilities. Many might prefer the direct approach of Latent Class rather than the 2-stage approach of doing HB first (which necessarily does some smoothing toward global or segment means) followed by cluster.
If you do plan to do HB utilities followed by cluster analysis on those utilities, most analysts would prefer to use the normalized zero-centered diffs. Those have equal sums of differences between best and worst levels across attributes per each respondent. In other words, they try to put each respondent on the same scale.
With raw HB utilities, respondents who answer very consistently and are easy to predict have large magnitude utilities. Respondents who answer less consistently or using more in-depth processing heuristics that are harder to predict get lower magnitude utilities.
If you use the raw utilities in cluster analysis, the degree to which respondents are consistent and easy to predict will be a significant influencer of segment membership. Most folks view that as undesirable.