Etched, a San Jose-based AI chip startup, is simultaneously negotiating two venture funding rounds that would value the company at $10 billion and $20 billion, according to The Wall Street Journal. The higher figure, led by existing investor Jane Street, would quadruple the company’s prior valuation. A separate round led by Sequoia Capital targets the $10 billion mark. Neither deal had closed as of July 17, and terms remain subject to change.
The company builds inference-specific silicon designed to run deployed AI models rather than train them. According to the WSJ, Etched’s website states $1 billion in customer demand is already lined up for chips still in validation.
The Dual-Round Structure
Back-to-back financings, where a startup raises at one valuation and immediately pursues more capital at a higher price, have become a defining feature of the current AI investment cycle, per the WSJ. The structure reflects leverage that leading AI companies hold over investors competing to secure allocations before valuations climb further.
Etched was founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen. Earlier backers include Stripes, Peter Thiel, Ribbit Capital, and Primary Venture Partners. The company has raised $800 million across four previously unannounced financings, including a strategic investment from VentureTech Alliance, according to its website. It now employs over 400 engineers drawn from NVIDIA, Google TPUs, Broadcom, SK Hynix, and TSMC.
Why Inference Silicon Matters for Agent Workloads
The $1 billion demand figure is notable because Etched’s product has not yet shipped. According to MarketScale, that pipeline represents procurement commitments from operators planning inference capacity now rather than waiting for proven supply.
Etched’s A0 silicon is back from TSMC’s N4P process, and the company is validating its first rack-scale product with customers, per its website. The chip architecture targets trillion-parameter sparse mixture-of-expert models, long context windows, and agentic workloads through two core innovations: Low Voltage Inference (LVI), which runs math blocks at under half the voltage of typical AI chips to achieve over 80% sustained FLOPs utilization, and Cluster Scale Memory (CSM), a proprietary low-latency shared memory pool across the scale-up domain.
The Competitive Field
NVIDIA dominates AI chip sales, but the inference market is drawing a distinct set of challengers. The WSJ cites Cerebras Systems and Groq as companies with demonstrated commercial traction, while newer entrants including U.K.-based Fractile and SambaNova also target inference workloads. Etched is betting that purpose-built inference silicon can deliver performance and cost advantages that general-purpose GPUs cannot match at equivalent workload scales.
The timing aligns with broader infrastructure consolidation across the AI industry. Alphabet announced an $80 billion capital raise this week to fund AI infrastructure buildout, and Anthropic is negotiating a $10 billion compute lease with Meta. Enterprise buyers are locking in inference capacity months or years ahead of availability, treating it as a strategic procurement decision rather than a spot purchase.
The Pre-Revenue Valuation Question
A $20 billion valuation for a company whose first chip hasn’t shipped places Etched in rarefied territory. The bet rests on two assumptions: that inference will grow as a share of total AI compute spend (a view supported by the shift from model training to deployment), and that purpose-built silicon will outperform general-purpose GPUs at the margin where agent workloads operate. If enterprise demand for always-on, low-latency agent inference continues to accelerate, Etched’s early customer commitments could prove prescient. If inference remains GPU-adequate for most workloads, a $20 billion pre-product valuation looks like a bet on a market that doesn’t exist yet at sufficient scale.