Upstart chipmakers keep challenging Nvidia. This time it's Microsoft-backed D-Matrix - CNBC

In the increasingly competitive AI chip market, D-Matrix, a startup located near NVIDIA's Silicon Valley headquarters, is challenging the industry leader with its novel chip architecture. The company'... Read original →

D-Matrix’s SRAM-centric Corsair chip disrupts AI hardware orthodoxy by sidestepping DRAM—a move that trades large-model support for immunity to memory supply chain volatility, especially under tightening U.S. export controls on advanced DRAM from the U.S., Korea, and Taiwan, China. This architecture pressures TSMC to allocate more 6nm/4nm capacity to inference ASICs while forcing Micron and SK Hynix to innovate in low-power HBM variants. NVIDIA will likely counter by accelerating Grace-based heterogeneous integration and deepening CUDA lock-in. Competitors like Cerebras may retreat to training-focused niches. Over the next 18 months, inference workloads will fragment: edge-optimized, small-language-model deployments will fuel demand for custom ASICs, yet SRAM’s density limits cap their scalability in data-center-scale LLMs. The real battleground shifts to chiplet interconnect standards and near-memory compute architectures.