Etched Hits $10.3 Billion Valuation With $300 Million Series C for AI Inference Chips

Money is still pouring into alternative AI silicon. Etched, a startup building chips purpose-built for AI inference, closed a $300 million Series C on Thursday led by Sequoia Capital, with Andreessen Horowitz, SK Hynix, and existing investors participating. The round values the company at $10.3 billion, up from roughly $5 billion in December. The jump reflects how aggressively investors are betting that the inference market—the compute-intensive phase where trained models actually generate outputs—will fragment beyond Nvidia’s GPU empire.

Etched was founded by three Harvard dropouts who built custom chips and memory components designed to accelerate inference on any model without recompilation. The company says its architecture sidesteps the CUDA lock-in that has kept most AI workloads tethered to Nvidia’s ecosystem. In a market where training GPUs grab headlines, inference is the faster-growing bottleneck: as models get deployed across billions of queries, the cost and latency of serving them is becoming the dominant infrastructure concern.

Why Inference Silicon Is Attracting Billion-Dollar Bets

The inference market is fundamentally different from training. Training is a batch process run in massive data centers; inference is real-time, latency-sensitive, and distributed across edge devices, cloud endpoints, and enterprise servers. Nvidia dominates training with its H100 and Blackwell architectures, but inference workloads have different memory-access patterns and throughput requirements that leave room for specialized designs. Etched argues its architecture handles arbitrary model weights without retraining, a claim that matters to teams running open-weight models from Llama to Mistral to custom fine-tunes.

SK Hynix’s participation signals that memory makers see inference as a groywhere high-bandwidth memory interfaces will define winners. The South Korean chip giant already supplies HBM stacks for Nvidia’s training GPUs; investing in Etched suggests it wants a piece of the inference stack as that market matures. a16z’s involvement adds crypto-world credibility—Etched counts several AI and crypto-native funds among its backers, and the firm has made inference-chip investments a bet on the post-GPU era of AI.

The Valuation Leap Signals a Shift in AI Hardware Narrative

Etched’s $10.3 billion valuation puts it in the same tier as semiconductor startups that took a decade to build. The speed of the climb—from $5 billion to $10.3 billion in seven months—reflects a market that has stopped waiting for proof of scale. Investors are pricing in the scenario where inference becomes the dominant AI compute category by 2027, driven by agentic applications that run models continuously rather than in occasional batch jobs.

That narrative carries risk. Etched does not yet have a shipping product at scale, and competing against Nvidia’s software moat is a multi-year slog. But the capital being deployed into the inference layer—from custom silicon to sparse-attention algorithms to speculative-decoding frameworks—suggests the industry believes the current GPU-centric model is temporary. For now, Etched’s valuation is a bet on that transition arriving faster than the incumbents can adapt.