In this paper, we consider the over-the-air computation of sums. Specifically, we wish to compute $M\geq 2$ sums $s_m=\sum_{k\in\mathcal{D}m}x_k$ over a shared complex-valued MAC at once with minimal mean-squared error ($\mathsf{MSE}$). Finding appropriate Tx-Rx scaling factors balance between a low error in the computation of $s_n$ and the interference induced by it in the computation of other sums $s_m$, $m\neq n$. In this paper, we are interested in designing an optimal Tx-Rx scaling policy that minimizes the mean-squared error $\max_{m\in[1:M]}\mathsf{MSE}_m$ subject to a Tx power constraint with maximum power $P$. We show that an optimal design of the Tx-Rx scaling policy $\left(\bar{\mathbf{a}},\bar{\mathbf{b}}\right)$ involves optimizing (a) their phases and (b) their absolute values in order to (i) decompose the computation of $M$ sums into, respectively, $M_R$ and $M_I$ ($M=M_R+M_I$) calculations over real and imaginary part of the Rx signal and (ii) to minimize the computation over each part -- real and imaginary -- individually. The primary focus of this paper is on (b). We derive conditions (i) on the feasibility of the optimization problem and (ii) on the Tx-Rx scaling policy of a local minimum for $M_w=2$ computations over the real ($w=R$) or the imaginary ($w=I$) part. Extensive simulations over a single Rx chain for $M_w=2$ show that the level of interference in terms of $ΔD=|\mathcal{D}_2|-|\mathcal{D}_1|$ plays an important role on the ergodic worst-case $\mathsf{MSE}$. At very high $\mathsf{SNR}$, typically only the sensor with the weakest channel transmits with full power while all remaining sensors transmit with less to limit the interference. Interestingly, we observe that due to residual interference, the ergodic worst-case $\mathsf{MSE}$ is not vanishing; rather, it converges to $\frac{|\mathcal{D}_1||\mathcal{D}_2|}{K}$ as $\mathsf{SNR}\rightarrow\infty$.