Etched, the AI chip startup founded by three Harvard dropouts, reached a $10.3 billion valuation in its latest funding round, with backing from major investors including Lightspeed Venture Partners and others backing the company's bold bet against GPU dominance.

The company has built custom silicon and memory components designed to accelerate AI inference without relying on traditional GPUs. Etched's architecture targets a specific pain point in AI deployment: GPUs excel at training but consume massive power and cost for inference workloads. The startup claims its chips deliver faster inference speeds across any AI model while reducing computational overhead.

The Harvard founders built Etched around a counterintuitive thesis. While Nvidia's GPUs have become the default infrastructure for AI, they argue inference represents a distinct problem requiring specialized hardware. Etched's approach involves redesigning memory hierarchies and compute patterns specifically for inference tasks, potentially offering substantial advantages in latency and power efficiency.

The $10.3 billion valuation puts Etched in rarified air for a hardware startup, reflecting investor appetite for alternatives to the GPU monopoly. Companies like Tesla, Meta, and Google have all explored custom silicon to reduce reliance on Nvidia, but most remain heavily dependent on the GPU maker. Etched enters a crowded field that includes Cerebras, Graphcore, and others attempting to challenge Nvidia's stranglehold on AI infrastructure.

Lightspeed's backing carries weight in venture circles. The firm has backed successful infrastructure companies before, signaling confidence in Etched's technical approach and market timing. The valuation also reflects broader recognition that AI compute infrastructure remains a bottleneck limiting deployment at scale.

Etched faces the classic hardware startup challenge: building chips requires massive capital expenditure, long development cycles, and manufacturing relationships. Proving performance advantages in real-world deployments against entrenched competitors presents another