One of the example applications proposes to stimulate patterns of neurons in visual cortex to produce vision. To even get to the point where you can train a monkey to detect phosphenes, Neuralink had to (a) develop a way of producing flexible electrodes that would measure activity from a single neuron, were robust enough to be implanted, would survive the punishing in-vivo environment, and retain their ability to measure and stimulate for long periods of time (none of those steps can be adequately done in in vitro models). (b) They then had to make sure that their robot could implant them to the proper depth (despite brain movement), and that during implantation they could maintain connection to the interface system. (c) They had to build a ASIC that would capture a 1000 of channels of neural data and wirelessly transmit them, while ensuring that the SNR was sufficient for use (again, no in vitro model). (d) They had to test the encapsulation of the ASIC to make sure that it, too, could survive in vivo.

They could certainly have done all these things serially, but it would have taken a lot longer. In the end, where they are is still far from a useful product - they have monkeys that can see phosphenes, they have electrodes that last 1 year, and they have a reasonable ASIC - but compared to the academic timescale, where we've been working on these individual pieces 1 PhD (6 years) at a time, they've made really remarkable progress.