We investigate constraints on early dark energy (EDE) using ACT DR4, SPT-3G 2018, Planck polarization, and restricted Planck temperature data (at $\ell<650$), finding a $3.3\sigma$ preference ($\Delta\chi^2=-16.2$ for 3 additional degrees of freedom) for EDE over $\Lambda$CDM. The EDE contributes a maximum fractional energy density of $f_{\rm EDE}(z_c)=0.163^{+0.047}_{-0.04}$ at a redshift $z_c=3357\pm200$ and leads to a CMB inferred value of the Hubble constant $H_0=74.2^{+1.9}_{-2.1}$ km/s/Mpc. We find that Planck and ACT DR4 data provide the majority of the improvement in $\chi^2$, and that the inclusion of SPT-3G pulls the posterior of $f_{\rm EDE}(z_c)$ away from $\Lambda$CDM. This is the first time that a moderate preference for EDE has been reported for these three combined CMB data sets. We find that including measurements of supernovae luminosity distances and the baryon acoustic oscillation standard ruler only minimally affects the preference ($3.0\sigma$), while measurements that probe the clustering of matter at late times - the lensing potential power spectrum from Planck and $f \sigma_8$ from BOSS - decrease the significance of the preference to 2.6$\sigma$. Conversely, adding a prior on the $H_0$ value as reported by the SH0ES collaboration increases the preference to the $4-5\sigma$ level. In the absence of this prior, the inclusion of Planck TT data at $\ell>1300$ reduces the preference from $3.0\sigma$ to $2.3\sigma$ and the constraint on $f_{\rm EDE}(z_c)$ becomes compatible with $\Lambda$CDM at $1\sigma$. We explore whether systematic errors in the Planck polarization data may affect our conclusions and find that changing the TE polarization efficiencies significantly reduces the Planck preference for EDE. More work will be necessary to establish whether these hints for EDE within CMB data alone are the sole results of systematic errors or an opening to new physics.