Next-Generation AI Chip Company Challenges NVIDIA's Dominance Through Six Pathways, Shifting Focus from Peak Compute to Data Movement Efficiency

July 26th, Deedy, an investor in Anthropic, pointed out that all startups developing next-generation AI chips are trying to challenge NVIDIA's dominance in different ways, aiming to tackle the fundamental issue of "data movement." Six differentiated challenge approaches include: eliminating DRAM (Groq, already acquired by NVIDIA for $20 billion), eliminating interconnect (Cerebras, IPO'd at around $48 billion market cap), eliminating compute/memory separation (d-Matrix), eliminating server-centric architecture (Majestic), eliminating generality (Etched, Taalas, MatX), and eliminating the $4 billion lithography machine (Substrate).

The IPO and acquisition prices of Cerebras and Groq have set a pricing benchmark for the entire AI chip track, while Etched's valuation doubled this week to $10.3 billion, erupting from stealth mode in just three weeks, further confirming that capital is rapidly reassessing the pricing in this track. Among the 18 major startups, private companies have a combined paper value of around $58 billion, with a public market cap of about $48 billion, covering inference, training/new architecture, systems, fabs, and lithography, among other subdirections.

Each approach is challenging the core assumption of NVIDIA's GPU architecture — that the movement of data between the compute unit and memory consumes both time and energy. Groq replaces DRAM with SRAM to completely eliminate memory level latency, Cerebras eliminates chip-to-chip interconnect bottlenecks at the wafer level, d-Matrix performs calculations directly in memory, and Etched sacrifices generality to design hardware specifically for the Transformer architecture. The AI chip startup ecosystem has evolved from a singular GPU replacement narrative to a multi-point attack at the architecture level. NVIDIA's acquisition of Groq actively engages in this process, indicating its deep understanding that the core of next-generation competition is no longer peak computing power but the complete reconstruction of data movement efficiency.