NVIDIA and AI-chip startup d‑Matrix announced a joint AI chip system that integrates inference-focused ASICs into NVIDIA’s stack. d‑Matrix targets low-latency, low-power inference using memory-centric 3D IMC to cut data movement, aiming to lower per-call cost for high-volume KV‑cache, long‑context and agent workloads. Model companies are increasingly pursuing custom inference silicon — OpenAI with Broadcom’s Jalapeño, Anthropic reported in talks with Samsung, and startups including Etched, Cereb

NVIDIA and AI-chip startup d‑Matrix announced a joint AI chip system that integrates inference-focused ASICs into NVIDIA’s stack. d‑Matrix targets low-latency, low-power inference using memory-centric 3D IMC to cut data movement, aiming to lower per-call cost for high-volume KV‑cache, long‑context and agent workloads. Model companies are increasingly pursuing custom inference silicon — OpenAI with Broadcom’s Jalapeño, Anthropic reported in talks with Samsung, and startups including Etched, Cerebras, Groq, DeepSeek and Zhipu evaluating self‑developed inference chips. Market takeaway: successful inference ASICs would compress cost, latency and energy for high-frequency AI apps and could shift demand away from general‑purpose GPUs; failure to deliver those advantages would keep demand concentrated on NVIDIA GPUs. Key execution risks: partnership terms remain undisclosed and d‑Matrix faces production, customer validation and software‑integration hurdles. Watch NVIDIA, Broadcom and Marvell as the most directly relevant listed plays; d‑Matrix is private.