Etched Exits Stealth With Working Sohu Transformer ASIC, $800 Million Raised at a $5 Billion Valuation and Over $1 Billion in Chip Orders
Etched unveiled working A0 silicon for Sohu, its transformer-only inference ASIC, disclosing $800 million raised at a $5 billion valuation and more than $1 billion in signed customer contracts.
Overview
AI chip startup Etched emerged from stealth on June 30, 2026, unveiling a rack-scale inference system built around Sohu, its transformer-only application-specific integrated circuit, and disclosing that it has raised $800 million in total funding, according to Crypto Briefing. On its own site, the company says its first Sohu silicon has already returned from the foundry: “Earlier this year our A0 silicon came back from TSMC N4P,” Etched writes, adding that it is “busy validating our first rack-scale product with customers to fulfill $1B in demand.”
What We Know
Etched described the June 30 disclosure as the point at which it moved from a design-stage startup to one with working hardware and signed orders. The company has raised $800 million across four previously unannounced financing rounds, including a strategic investment from VentureTech Alliance, per Etched. The most recent round brought in $500 million at a $5 billion post-money valuation and was led by Stripes, according to Data Center Dynamics.
The chip is fabricated on TSMC’s N4P process node, and Etched reported first-pass silicon success on its initial design, according to Crypto Briefing. The company says it has “kicked off production to fulfill over $1B in customer contracts,” per Etched, a figure of more than $1 billion in signed commitments that Data Center Dynamics also reported.
What sets Sohu apart from a general-purpose GPU is that it is a transformer-only design. It implements transformer attention as fixed-function silicon rather than as programmable matrix-multiply instructions, and carries 144GB of HBM3E per chip, according to an analysis by Spheron. Etched has validated the chip on Llama, DeepSeek, Qwen, and Mamba models, Crypto Briefing reported.
The company’s headline performance claim is that a single eight-chip Sohu server delivers roughly 500,000 tokens per second on Llama 70B, a figure reported both by Spheron and by Tech Company News, which framed it as a large multiple of the throughput of comparable Nvidia H100 or B200 systems. These are vendor-supplied benchmarks that have not been independently verified.
Etched is led by chief executive Gavin Uberti, co-founder Chris Zhu, and president Robert Wachen, according to Tech Company News. “Production is the product,” co-founder Rob Wachen told Data Center Dynamics, arguing that the companies that matter will be those that can translate technology into systems manufactured and operated at massive scale. To that end, the startup has stood up a factory in Taiwan and built a 2MW data center, a test house, and a prototyping lab at its California headquarters, per Data Center Dynamics.
The roster of backers and advisers spans quantitative-trading firms and prominent AI researchers, including Jane Street, Hudson River Trading, Peter Thiel, Andrej Karpathy, and Geoffrey Hinton, according to Data Center Dynamics. The team numbers more than 400 engineers drawn from NVIDIA, Google TPUs, Broadcom, SK Hynix, and TSMC, per Etched.
What We Don’t Know
Etched says first rack shipments are scheduled for this summer, and it has laid out a goal of reaching gigawatt-scale capacity by 2027, according to Crypto Briefing, but as of the announcement the systems had not yet shipped. The performance comparisons against Nvidia hardware are Etched’s own claims and have not been confirmed by third-party benchmarking; published secondary analyses differ on how the throughput figures should be normalized against GPU baselines. The identities of the customers behind the more than $1 billion in contracts have not been disclosed.
Analysis
Etched’s central bet is architectural: that the transformer will remain the dominant neural-network design long enough for a chip that hard-codes it into silicon to pay off. That is the same wager that makes Sohu faster than a GPU on transformer workloads and also riskier, because a specialized ASIC cannot pivot to a new architecture the way a programmable GPU can. The move from taped-out design to A0 silicon, signed contracts, and an in-house Taiwan factory marks a shift from the crowded field of pre-production inference-chip startups toward the far smaller group that has working hardware. Whether it can convert more than $1 billion in orders into shipped, operating racks this summer is the test that will separate Etched’s claims from its results.