Etched is weighing fresh funding offers at valuations of up to $50 billion as it prepares to push its inference chips into wider deployment.
According to TechCrunch, the four-year-old startup is considering investment offers valuing it at between $40 billion and $50 billion, only months after raising $700 million at a $21 billion valuation.
Rather than selling another general-purpose accelerator, Etched is developing complete AI systems around custom inference silicon built specifically to run trained models quickly and cheaply.
Inference has become one of commercial AI’s highest costs because every user prompt has to be processed somewhere. Etched reckons a narrower architecture can beat Nvidia by concentrating silicon on that job rather than trying to do everything.
The company claims its processors can generate more tokens faster and at lower cost than Nvidia hardware. Those claims still need proving at scale. Etched manufactured test silicon at TSMC this summer and says it has secured $1 billion in customer orders.
Quant trading firm Jane Street led Etched’s most recent $700 million funding round and has already taken delivery of an early system.
For a trading operation where shaving tiny amounts of time from a computation can translate into serious money, specialised inference hardware has rather more appeal than another colourful AI demo.
Etched co-founder and chief operating officer Robert Wachen said the company designed two new components from scratch specifically to accelerate inference.
The startup is not planning to stop at selling chips. It is building full hardware systems around them, which means racks, interconnects, power delivery, cooling and software all have to work together.
Etched already operates a new 10MW data centre in Silicon Valley and has opened a facility in Taiwan to coordinate production close to TSMC.
That places it much closer to being a systems supplier than another chip startup with a clever block diagram and an optimistic PowerPoint deck. The company employs about 400 people, with roughly 15 per cent having previously worked at Nvidia.
That staffing gives Etched useful accelerator experience, although hiring former Nvidia engineers is rather easier than dislodging Nvidia from a market where its hardware, networking and software stack are already deeply embedded.
Etched’s strategy depends on specialisation paying off. A chip designed narrowly for inference can devote more silicon to the workload, potentially improving throughput and power efficiency. If model architectures or inference techniques shift in directions Etched’s hardware handles poorly, a specialised design can age much faster than a broader accelerator.
The cocaine nose jobs of Wall Street are apparently pricing Etched as though its hardware advantage is already proven while the company is still moving from test silicon and early systems towards large-scale deployment.
