Inside the IBM FlashSystem 900 All-Flash Array

The SM950 can be had at $300 for 512GB with 2.5/1.5 GB/s sequential read/write and 300K IOPS. RAID together 6 of them in parallel and you'll have something with better specs and storage than the 2.9TB module ($46K) for just $1.8K (albeit with less endurance, though not sure how much less).

I don't. I think they are using an FPGA because it is convenient and they are willing to throw away some of the margins on their product for ease of implementation. That particular FPGA has a PCIe hard IP, and they probably implement the NAND read/write/balancing using its LUTs and interface using its digital I/O. A consumer SSD (or really any product) might start off life like this, and then tape out to an ASIC before release if it is going to ship in any significant volume. The ASIC will have drastically lower power and, once sufficient volume hits, much higher margins.

I don't buy the programmable argument either, you can implement programmable cores (this is what basically any high feature NIC is like Chelsio, SolarFlare, etc) while still baking in the PCIe, NAND interfaces, offloads like CRC/erasure coding.

Those FPGAs each cost several hundred dollars. Imagine if each PCI device in your computer (sound, network, video) had a full Xeon on board. Not entirely dissimilar.