The hypergraph Zarankiewicz's problem, introduced by Erdős in 1964, asks for the maximum number of hyperedges in an $r$-partite hypergraph with $n$ vertices in each part that does not contain a copy of $K_{t,t,\ldots,t}$. Erdős obtained a near optimal bound of $O(n^{r-1/t^{r-1}})$ for general hypergraphs. In recent years, several works obtained improved bounds under various algebraic assumptions -- e.g., if the hypergraph is semialgebraic. In this paper we study the problem in a geometric setting -- for $r$-partite intersection hypergraphs of families of geometric objects. Our main results are essentially sharp bounds for families of axis-parallel boxes in $\mathbb{R}^d$ and families of pseudo-discs. For axis-parallel boxes, we obtain the sharp bound $O_{d,r}(tn^{r-1}(\frac{\log n}{\log \log n})^{d-1})$. The best previous bound was larger by a factor of about $(\log n)^{d(2^{r-1}-2)}$. For pseudo-discs, we obtain the bound $O_r(tn^{r-1}(\log n)^{r-2})$, which is sharp up to logarithmic factors. As this hypergraph has no algebraic structure, no improvement of Erdős' 60-year-old $O(n^{r-1/t^{r-1}})$ bound was known for this setting. Futhermore, even in the special case of discs for which the semialgebraic structure can be used, our result improves the best known result by a factor of $\tildeΩ(n^{\frac{2r-2}{3r-2}})$. To obtain our results, we use the recently improved results for the graph Zarankiewicz's problem in the corresponding settings, along with a variety of combinatorial and geometric techniques, including shallow cuttings, biclique covers, transversals, and planarity.