Etched has raised US$700 million at a US$21 billion valuation as the artificial intelligence chip start-up begins shipping its inference hardware to customers, stepping up its bid to challenge Nvidia Corp (NASDAQ:NVDA, XETRA:NVD) in one of the fastest-growing areas of AI computing.
The latest financing was led by quantitative trading firm Jane Street, with participation from Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, Neo, Primary, Stripes, Positive Sum and Blackstone.
Etched said the funding would help accelerate production as it moves from chip development into commercial deployment.
The San Jose-based company has also named Jane Street as its first customer, having shipped its first rack-scale system to the trading firm in July.
Jane Street is now deploying the hardware in its data centre after testing the technology, marking an important transition for Etched from developing silicon to running customer workloads.
Etched said it has secured more than US$1 billion in customer contracts covering AI companies and cloud providers and has now raised US$1.9 billion in total.
Focus on AI inference
Unlike AI processors designed to handle a broad range of computing workloads, Etched is concentrating on inference - the process by which a trained AI model generates responses and performs tasks once deployed.
Demand for inference computing has become increasingly important as AI models move from development into widespread commercial use and applications process greater numbers of user requests.
Etched is developing complete inference systems covering chips, server racks, software and manufacturing rather than selling an individual processor alone.
Its architecture incorporates what it calls Low Voltage Inference, designed to increase the amount of computation possible within a given power envelope, and Cluster Scale Memory, which creates a shared memory pool across a cluster of processors.
The company says the approach can improve throughput and latency for large AI models, although its performance claims against competing hardware have not yet been independently established.
Chief executive and co-founder Gavin Uberti said getting hardware into customers' hands and running real workloads had been a priority since the company's formation.
“Jane Street putting this cluster into production is proof of what we've built,” Uberti said in the company's announcement.
Building an Nvidia alternative
Etched's progress adds another competitor to a market overwhelmingly shaped by Nvidia's data-centre GPUs.
The start-up has recruited heavily from established semiconductor companies, with around 15% of its roughly 400 employees previously working at Nvidia, according to the Wall Street Journal.
Etched received its initial production silicon from Taiwan Semiconductor Manufacturing Company earlier this year and is already developing 3 generations of hardware in parallel.
The US$21 billion valuation represents a substantial vote of confidence in the idea that specialised inference processors can capture part of the expanding AI infrastructure market.
The next test will be whether Etched can scale production and translate its early contracts into widespread deployments while competing with Nvidia and a growing group of companies developing alternative AI accelerators.