D-matrix joins the AI chip completion w/ its inference chip on PCI-e card named Corsair released. The core innovation here is tiered memory w/ SRAM “performance” memory on die, and DRAM “capacity” memory. This is a slight tweak to the in-memory compute paradigm that many AI chip startups are adopting (e.g. Groq and Cerebras). Others with 100% SRAM can deliver great performance, but it comes at great costs and less flexibility. You need to have many chips linked together at minimum to get the inference server to work. Similar to others startups, D-matrix is targeting on inferencing bc it is way easier to break in vs. training. And we expect the space of inferencing can be more commoditized and competitive vs. $NVDA's CUDA moat. A nice advantage and tweak is that it is based on PCI-e card form factor. One D-matrix Corsair card has 2GB SRAM and 256GB DRAM. Customers can mix and shift numbers of Corsair cards, up to 8 cards in a server. Competitors like Groq requires at least 64 Groq chips linked together to work because one Groq chip only has 230MB of SRAM. By adding DRAM in, one single D matrix card can also serve larger models too. And customers can start with one or 8 cards that are lower in CAPEX and flexible in design, much more similar to Nvidia GPGPU. For low latency fast inference on small models, it can generate tokens in 2ms/token because its SRAM runs at 150TB/S which is 15x+ faster than H100's 10TB/S HBM bandwidth. For slow inference that doesn't require snappy generations, it can leverage slow but large & cheap DRAM to host a larger model. An 8 card server can host 1trillion+ parameter model. For the software stack, it has its own called d-Matrix's Aviator. Supermicro $SMCI is the partner to make the D-matrix chip into a server and a finished rack system. Not a surprise, given SMCI is the leader in accelerator servers due to its engineering focus and flexible architecture. One Corsair card consumes 600w in a dual slot air cooled form factor. This is similar to H100's power consumption range. Overall, it looks very similar to Qualcomm's PCI-e AI inference card. But it is more polished w/ tiered memory and software stack. Potentially, $QCOM should acquire it although QCOM right now is partnering with Cerebras where Cerebras WSE handles the training and QCOM handles the inferencing.