The rumored modules are 64-bit. 

https://www.micron.com/products/memory/dram-components/lpddr5/part-catalog/part-detail/mt62f768m64d4ek-026-wt-b

It's just a question of nomenclature.

Soldered LPDDR5s modules with 64-bits have half the speed of 32-bit ones (which themselves are double 16-bit width channels, actually). The RAM remains at the frequency indicated as 32 bits, but the communication bus is twice as large but half as fast.

It's quite unlikely the Switch 2 CPU could even handle the parallel data lines from 64-bit DRAM SO-DIMMs at these data rates, anyway.

Edit - but as I mentioned before, this is based on what the prototype looks like and its two modules, the final product might be different or have a custom memory controller, who knows.

Is Micron using different nomenclature from Samsung here? I ask, because -- for example when talking about the Steam Deck OLED's memory Samsung uses the x32 terminology rather than x64. 

Micron also has single component LPDDR5 modules with only 16bit bus width.

Example: https://www.micron.com/products/memory/dram-components/lpddr5x/part-catalog/part-detail/mt62f1g16d1ds-023-it-b

But then Samsung seems to use the x64 organization too. For example, 

https://semiconductor.samsung.com/dram/lpddr/lpddr5x/k3kl3l30em-bgcv/

If the Steam Deck OLED has 4 modules that give a total of 102 GBps, that would be 25.5 GBps per module. But that is on what they're calling x32 "organization." So the implication is something that Samsung labels as x64 has the potential to give double that (at the same clock-rate.) 

Also how does the Rog Ally achieve 100 GBps with dual-channel LPDDR5-6400? 

Edit: Never-mind, I totally misread your post. Also Asus is probably listing the Rog Ally as dual channel even though it is technically quad channel just to preserve the traditional meaning. 

Edit 2: 

Digging in further, I don't think the module count matters, it is what each module supports. 

Here is a Lenovo Legion Go motherboard with two modules. 

Source: https://www.ifixit.com/Guide/Lenovo+Legion+Go+Chip+ID/167631?srsltid=AfmBOorF5zqjkVxbspdRpmCiZaNDOAqvOSfSvD6WIOeUIh3C-1hZgssA

The modules are the orange squares. 

It still supports quad-channel and has a total memory bandwidth of about 120 GBps. The memory it uses is  K3KL3L30CM-BGCT 

Compare that to the Rog Ally board that has four modules. 

But the modules are of 32x "organization" K3LKBKB0BM-MGCP . 

Switch 2 could have something similar to the Legion Go. Two 64x "organization" modules or in Micron's language "bus-width" that support quad-channel. 

Also, as a test, I ran AIDA64 on my Galaxy Book 2 Pro with ostensibly "Dual-Channel" LPDDR5 - 5200 and I am getting ~67 GB/s mem-reads (can't see writes since it is the trial version.) 

It will be on a 128-bit memory bus.
The number of channels is meaningless as it's not using sticks in a traditional PC sense... Rather it can initiate transactions to a number of chips concurrently.

The memory is likely on both sides of the PCB as well if each LPDDR5X chip is the normal 32bit chips.

It does matter because in practice each module can't sustain the maximum bandwidth simultaneously for every channel using the most common bank modes. It would need to operate at half the command clocks unless you double the burst length but that still falls short of the theoretical maximum. See this for reference.

I don't know how the Legion GO Samsung's RAM operates but the above is true for Micron DRAM with 4 dies per module as per the rumored spec sheets.

That being said, we're discussing hardware comparable in price to the PlayStation 5 Pro vs. a company known always to release underpowered hardware. We've been in this discussion multiple times over the past generations, so yeah. The games end up looking good but the hardware is almost always the cheapest possible.