Cheap exotic silicon with no open LLM stack is a paperweight (Cambricon/Ascend/Iluvatar/Coral/Grayskull/FPGA)

Cambricon MLU270/370, Huawei Ascend, Iluvatar MR-V100, Google Coral, Tenstorrent Grayskull and FPGA cards appear cheap on eBay but have no llama.cpp/vLLM/GGUF path - they cannot run an open 7B model without writing your own kernels.

Neuware (Cambricon), CANN/MindIE (Huawei), CoreX SDK (Iluvatar) are closed vendor stacks; Coral is INT8 vision-only; Grayskull is legacy (no LLM demos in tt-metal); FPGAs have no turnkey LLM flow.

Verify an open serving path exists for the exact model size BEFORE buying. For Coral/Grayskull/FPGA/Cambricon the answer today is 'no' - treat as teaching/vision cards only.