We investigate a network of integrate-and-fire neurons characterized by a distribution of spiking frequencies. Upon increasing the coupling strength, the model exhibits a transition from an asynchronous regime to a nontrivial collective behavior. At variance with the Kuramoto model, (i) the macroscopic dynamics is irregular even in the thermodynamic limit, and (ii) the microscopic (single-neuron) evolution is linearly stable. The investigation of networks of oscillators can provide new insights on the basic mechanisms which underlie brain functioning. In particular, the spontaneous onset of a collective dynamics is an intriguing phenomenon that can contribute to information transmission across different brain areas. Given the large number N of neurons (oscillators) present in a real brain, it is tempting to adopt a statistical-mechanics point of view and thereby investigate the behavior for N → ∞ (the so-called thermodynamic limit). Two different setups are typically invoked [1]: (i) sparse networks, characterized by a fixed number of synaptic connections; (ii) massively connected networks, where the number of connections is proportional to N . In the former case, the local field seen by the single neurons naturally fluctuates even for N → ∞ (being the sum of a fixed finite number of different input signals), consistently with the experimental evidence of an irregular background activity in the cerebral cortex [2]. The latter setup has the advantage of being, at least in principle, amenable to an exact mean-field treatment, although microscopic fluctuations survive only if inhibition and excitation balance each other [3].In this Letter, we numerically show that an irregular microscopic and macroscopic dynamics can generically arise even in an inhibitory, globally coupled network. More precisely, we consider a heterogeneous network of pulse-coupled integrate-and-fire (IF) neurons [4], each characterized by a different bare spiking frequency. This setup is similar to that of the Kuramoto model (KM) [5], where each single oscillator is identified by a phase variable φ. The analogy is so tight that it has even been shown that the pulse-coupling mechanism characterizing IF neurons reduces, in the weak coupling limit, to that of the KM, the only difference being that the coupling function is not purely sinusoidal [6]. It is therefore quite important to clarify to what extent a network of IF neurons reproduces the KM scenario for stronger coupling strengths, especially by recalling that the KM is often invoked while testing new ideas on the control of synchronization within neural contexts [7]. Finally, in order to make the model closer to a realistic setup, we include delay to account for the finite propagation time of the electric pulses.Our strategy consists in studying the macroscopic collective dynamics in the large N -limit, for different values of the coupling strength g. In the KM, it is known that for a weak enough coupling, the single oscillators rotate independently of each other. On the other hand, above a ...