Fault-tolerant quantum computation requires qubit measurements to be both high fidelity and fast to ensure that idling qubits do not generate more errors during the measurement of ancilla qubits than can be corrected. Towards this goal, we demonstrate single-shot readout of semiconductor spin qubits with 97% fidelity in 1.5μs. In particular, we show that we can engineer donor-based single-electron transistors (SETs) in silicon with atomic precision to measure single spins much faster than the spin decoherence times in isotopically purified silicon (270μs). By designing the SET to have a large capacitive coupling between the SET and target charge, we can optimally operate in the “strong-response” regime to ensure maximal signal contrast. We demonstrate single-charge detection with a signal-to-noise ratio (SNR) of 12.7 at 10 MHz bandwidth, corresponding to a SET charge sensitivity (integration time for SNR=2) of 2.5 ns. We present a theory of the shot-noise sensitivity limit for the strong-response regime which predicts that the present sensitivity is about one order of magnitude above the shot-noise limit. By reducing cold amplification noise to reach the shot-noise limit, it should be theoretically possible to achieve high-fidelity, single-shot readout of an electron spin in silicon with a total readout time of approximately 36 ns.