I almost wonder, then, if MS planned to use Van Gogh in Project Keystone? I mean why else order AMD to make such a GPU? Unless they have plans for a future Surface device featuring AMD in it.
Okay, so maybe the power is there, and the means to have that performance deliverable in a portable format that can stay persistent in the level of performance is also there, or getting there, and we can maybe see something of a Series S-type spec soon in the handheld portable space, be it from Valve with a SteamDeck 2 or some other company...at least in terms of TF.
But there are other components to overall performance to consider. You mentioned the Phoenix APU; clearly it has more TF than Series S. But does it have a comparable pixel fillrate? Texture fillrate? Geometry culling and rasterization rate? Those are things that would be influenced by amount of ROPs, TMUs, and the clock speed of the frontend, respectively (though RDNA 3 increased prims/clock to 12, and shaders are now dual-issue instead of single-issue like RDNA 2, so clocks can be more modest or even lower and still provide comparable or higher TF and geometry culling/raster performance to an RDNA 2 design with notably higher clocks, to be fair).
Then there's other things like what memory configuration would they be able to fit in such a device. Chances are, they won't use GDDR6 or GDDR6x, and HBM-based memories are probably too expensive not to mention, RDNA 3 GPUs may not be designed with HBM interfaces in mind to begin with. So they are going to be stuck with LPDDR5x or LPDDR6, and I don't think a SO-DIMM design gets them the memory bandwidth they'd need to provide comparable performance to a Series S in a portable device targeting 2025 or even 2026.
Dell have their new CAMM memory designs that are meant to supersede SO-DIMM; it's an interesting design that would address the bandwidth limitations that could impact future laptops and other portable devices, while being simpler to integrate vs. soldering the RAM onto the motherboard. But it's also proprietary, limited in product usage (currently) and not JEDEC-standard. There's a chance a company like Valve could license a CAMM design out from Dell for a Steam Deck 2, but it's a limited chance.
Any such portable would also need sufficient decompression I/O; IIRC the Steam Deck doesn't actually have a dedicated decompression I/C chip like the Series systems, although it probably supports faster SSDs than Series S or X's expansion cards do. That isn't necessarily enough to resolve the typical bottlenecks of storage I/O, and while things like DirectStorage are meant to address those, they do it in a hardware-agnostic fashion with mostly software-driven solutions, so some amount of hardware overhead on the device utilizing it will determine the overall effectiveness of the DirectStorage routines and functions. A portable handheld targeting a mass consumer market isn't going to have the type of built-in hardware to both provide sufficient gaming performance on the go AND overhead for I/O data decompression, and while I don't think the decompression chip in Series systems (or even the decompression I/O system of the PS5) cost that much to produce, the R&D associated with them is probably not cheap. Do companies like Valve spring for that R&D in a Steam Deck 2 targeting something like Series S performance as a baseline? Who knows.
That reminds me of the EverCade retro handheld. The one where they've gotten actual publishers to release flash cart compilations of retro games to run on the system, it's a pretty neat idea and helps clean up the muddiness WRT the legality of those types of devices, by working directly with the publishers owning the games those devices are meant to play, and locking down the medium to a format that only provides authenticated access.
Sony putting out a retro-themed portable for PS1/2/3/PSP/Vita games they can sell for a profit on its own (since it's targeting a specific niche that would be willing to pay more), they can even design with shells mimicking the classic consoles (adds collectability incentive), a locked-down media format for physical games released in packs (and maybe a storefront fork from PS Store for digital versions of individual games), functional controller ports for legacy PS peripheral devices...that could be a pretty decent market to appeal to.