Considering that the LTC6268 buffer op-amp has an input referred noise of 4.3 nV/\$\sqrt{Hz}\$ and your signal has a bandwidth of up to 100 kHz, the effective RMS noise seen at the input (that your signal has to overcome) is nearly 1.4 μV RMS. This is massively larger than your input amplitude (stated to be 1 nV) so, how can you expect this design to ever work?

In addition, the reference inputs on your instrumentation amplifiers need to be tied to 0 volts. I expect that if you looked at the DC output from the middle gain stage it will be swamping the final stage.

a signal that comes from a hydrophone with around 1nV amplitude and with a bandwidth of 10 to 100 kHz

Considering that the LTC6268 buffer op-amp has an input referred noise of 4.3 nV/\$\sqrt{Hz}\$ and your signal has a bandwidth of up to 100 kHz, the effective RMS noise seen at the input (that your signal has to overcome) is nearly 1.4 μV RMS. This is massively larger than your input amplitude (stated to be 1 nV) so, how can you expect this design to ever work?

Apart from that, the middle stage instrumentation amplifier has a lower value equivalent input noise compared to the front-end op-amp so, the op-amp buffer is pointless (even if the noise is still far too high to make a reasonable amplifier for your application).

when I run the simulation I get the 40 times gain on the first op-amps that amplifies but the second one de-amplifies the signal

In addition, the reference inputs on your instrumentation amplifiers need to be tied to 0 volts. I expect that if you looked at the DC output from the middle gain stage it will be swamping the final stage thus preventing any reasonable amplification.