Edit/Append: I've had this idea [1] forever (since the 1990s, possibly earlier... don't have notes going that far back). Imagine the simplest possible compute element, the look up table... arranged in a grid. Architectural optimizations I've pondered over time lead me to a 4 bits in, 4 bits out look up table, with latches on all outputs and a clock signal. This prevents race conditions by slowing things down. The gain is that you can now just clock a vast 2d array of these cells with a 2 phase clock (like the colors on a chessboard) and it's a universal computer, Turing complete, but you can actually think about it without your brain melting down.
The problem (for me) has always been programming it and getting a chip made. Thanks to the latest "vibe coding" stuff, I've gotten out of analysis paralysis, and have some things cooking on the software front. The other part is addressed by TinyTapeout, so I'll be able to get a very small chip made for a few hundred dollars.
Because the cells are only connected to neighbors, the runs are all short, low capacitance, and thus you can really, REALLY crank up the clock rates, or save a lot of power. Because the grid is uniform, you wont have the hours or days long "routing" problems that you have with FPGAs.
If my estimates are right, it will cut the power requirements for LLM computing by 95%.
[1] Every mention of BitGrid here on HN - https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...