I assume they use garden variety MOSFETs like 2N7002, which cost a few pennies when bought in bulk.
The gate capacitance is 20 pF. One logic gate output is typically driving several inputs, so the typical load is several-fold of this capacitance. One could check the schematics for the maximum fan-out in the real circuit. The longest, one foot long traces on the PCB add another few tens of pF. So the capacitance seen by an output of a logic gate will probably be in the range of 20-100 pF most of the time. But it is the slowest gates which determine the speed of the whole circuit, so the worst case could be much worse.
These transistors themselves are extremely fast and can switch on and off in about 3 ns. Their channel resistance is in single Ohm range even at the lowish gate voltages, so the RC time constant is also very small 100pF * 3 Ohms = 0.3 ns. Thus this is also not an issue. The slowness of the circuit comes from elsewhere.
First, from the resistors which pull the logic gate outputs to the supply voltage. There is one for each logic gate, so a total of one thousand of these, and in the first approximation half of them conduct at any given time. To keep power consumption low, these resistors must have relatively high resistance values. If we limit the current for the logic to 1 A total, that is 2 mA for each of the 500 resistors. Therefore the resistors must be 2.5 kOhm for 5 V supply voltage. Thus charging of the nodes of the circuit through these 2.5 kOhm resistors is almost a thousand times slower than discharging the same nodes through the 3 Ohm channel resistance of a fully open transistor. Still, this is not too bad -- on the order of 250 ns of delay per gate.
But the frequency for the whole circuit is determined by delay not through just one, but through many gates in series, plus there are probably parts of the circuit topology which are not optimal for implementing them with discrete transistors, where the delay does not follow from such a simple reasoning as above.
The speed can probably be easily increased by an order of magnitude if one were willing to spend 50W instead of 5W for the circuit. Beyond that, one would want a CPU circuit designed specifically for implementation in discrete transistors, not a discrete element copy of a monolithic chip.