We investigate the quality of security reductions for non-interactive key exchange (NIKE) schemes. Unlike for many other cryptographic building blocks (like public-key encryption, signatures, or zero-knowledge proofs), all known NIKE security reductions to date are non-tight, i.e., lose a factor of at least the number of users in the system. In that sense, NIKE forms a particularly elusive target for tight security reductions. The main technical obstacle in achieving tightly secure NIKE schemes are adaptive corruptions. Hence, in this work, we explore security notions and schemes that lie between selective security and fully adaptive security. Concretely: We exhibit a tradeoff between key size and reduction loss. We show that a tighter reduction can be bought by larger public and secret NIKE keys. Concretely, we present a simple NIKE scheme with a reduction loss of O(N2 log(ν)/ν2), and public and secret keys of O(ν) group elements, where N denotes the overall number of users in the system, and ν is a freely adjustable scheme parameter. Our scheme achieves full adaptive security even against multiple “test queries” (i.e., adversarial challenges), but requires keys of size O(N) to achieve (almost) tight security under the matrix Diffie-Hellman assumption. Still, already this simple scheme circumvents existing lower bounds. We show that this tradeoff is inherent. We contrast the security of our simple scheme with a lower bound for all NIKE schemes in which shared keys can be expressed as an “inner product in the exponent”. This result covers the original Diffie-Hellman NIKE scheme, as well as a large class of its variants, and in particular our simple scheme. Our lower bound gives a tradeoff between the “dimension” of any such scheme (which directly corresponds to key sizes in existing schemes), and the reduction quality. For ν = O(N), this shows our simple scheme and reduction optimal (up to a logarithmic factor). We exhibit a tradeoff between security and key size for tight reductions. We show that it is possible to circumvent the inherent tradeoff above by relaxing the desired security notion. Concretely, we consider the natural notion of semi-adaptive security, where the adversary has to commit to a single test query after seeing all public keys. As a feasibility result, we bring forward the first scheme that enjoys compact public keys and tight semi-adaptive security under the conjunction of the matrix Diffie-Hellman and learning with errors assumptions. We believe that our results shed a new light on the role of adaptivity in NIKE security, and also illustrate the special role of NIKE when it comes to tight security reductions.