From $10B to $21B in a Month: Jane Street Leads $700M Bet on Etched's Transformer-Only Silicon

In a cycle where AI chip valuations have been described as “irrational” even by the people writing the checks, Etched just set a new standard for speed. On Tuesday, August 18, 2026, the San Jose–based startup announced it had raised $700 million at a $21 billion valuation — more than double the $10.3 billion it was worth in mid-July. The round was led by Jane Street, the famously quantitative trading firm, and the reason for the breakneck re-rating is unusual: Jane Street didn’t just write a check. It installed Etched’s first shipped AI inference cluster, evaluated the hardware in production, and then led the round.

What happened

Etched’s valuation trajectory this year reads like a compressed history of the AI silicon boom. In December 2025, the company raised $500 million at a $5 billion valuation, bringing its total funding at that point to roughly $800 million across several unannounced rounds. On June 30, 2026, it emerged from what it called a second stealth phase with a working transformer-specific chip called Sohu, over $1 billion in customer demand, and its first inference rack delivered to Jane Street. In July, Sequoia Capital led a $300 million Series C at a $10.3 billion valuation. Less than a month later, Jane Street’s $700 million round landed at $21 billion — bringing Etched’s total raised to approximately $1.9 billion.

The Wall Street Journal, which profiled the company under the headline “A $21 Billion ‘Kids in Chips’ Startup,” reported that around 15% of Etched’s roughly 400 employees are former Nvidia engineers and executives. The company was founded by Harvard dropouts — CEO Gavin Uberti among them — who skipped the usual career ladder and went straight to building application-specific silicon.

The bet: transformers are the destination, not a layover

Etched’s core thesis is deceptively simple. Nearly all frontier AI today — ChatGPT, Claude, Gemini, Llama, DeepSeek — runs on the transformer architecture. Yet the GPUs these models run on, overwhelmingly Nvidia’s, are general-purpose machines: programmable hardware designed to handle whatever workload a data center might see. That flexibility costs silicon area, power, and money.

Sohu takes the opposite approach. Etched hardcodes the transformer architecture directly into the chip, stripping out the programmability that general-purpose GPUs carry as overhead. Because every transistor is dedicated to transformer math, the company claims dramatic efficiency gains for inference — the phase where trained models actually serve users, and which increasingly dominates AI compute budgets. Etched has claimed that a single eight-chip Sohu server can serve a Llama-70B-class model at over 500,000 tokens per second, throughput figures it has marketed as roughly ten times the speed and cost efficiency of GPU-based systems. The chip is manufactured on TSMC’s N4P process, and the company co-designs the full stack — chips, racks, software, and manufacturing methods — to squeeze out latency and power costs.

The catch is symmetry with the upside: a chip that only runs transformers only has a market as long as transformers remain the dominant architecture. If the industry’s center of gravity shifts — toward diffusion-based language models, state-space architectures like Mamba, or some yet-uninvented design — Etched’s silicon becomes an expensive paperweight. Investors at $21 billion are explicitly betting that transformer inference demand is durable for at least the length of a data center depreciation cycle.

Why a trading firm is the lead investor

Jane Street’s role is the most telling detail in the story. The firm is not a traditional venture investor; it is one of the world’s largest quantitative trading operations, with an insatiable appetite for low-latency computation and a long history of treating hardware as a competitive weapon. According to reporting around the round, Jane Street installed Etched’s first shipped cluster system, ran it, and was impressed enough to lead the financing.

That matters for two reasons. First, it converts Etched’s marketing claims into something closer to field validation — a sophisticated, performance-obsessed customer deployed the hardware before funding the company. Second, it signals who the early buyers of inference-specialized silicon may be: not just AI labs and hyperscalers, but financial firms, and anyone whose economics improve directly with token throughput per dollar and per watt.

Context: the inference ASIC wave

Etched is the most richly valued representative of a broader shift. Groq, Cerebras, and SambaNova have all pushed inference-first architectures; Cerebras has itself raised at multi-billion valuations on wafer-scale chips. What distinguishes Etched is the purity of its specialization — competitors like Groq built low-latency tensor processors that still run varied model families, while Sohu is welded to the transformer blueprint itself — and the velocity of its re-rating, which has no obvious precedent in chip industry history. Even by the standards of 2026’s AI infrastructure frenzy, doubling a $10.3 billion valuation in under a month is an outlier.

Nvidia, for its part, is not standing still: its latest accelerator generations keep improving inference throughput, and its CUDA software moat remains the industry default. Etched’s answer is that ~15% of its staff, per the Journal, are Nvidia veterans who know exactly what they are competing against. Meanwhile, with over $1 billion in contracted demand and a total of roughly $1.9 billion raised, the company now has the capital to attempt volume delivery — historically the graveyard stage for chip startups, where fabrication commitments, yield problems, and software ecosystems kill companies with working prototypes.

What to watch

Three open questions will decide whether $21 billion looks cheap or absurd in hindsight. Can Etched convert its first Jane Street deployment into a fleet of repeat customers before Nvidia’s next generation closes the efficiency gap? Will transformer dominance hold through the useful life of Sohu’s taped-out silicon? And can a 400-person company build the compiler, kernel, and serving software stack that took Nvidia two decades and thousands of engineers to mature?

For now, the market has rendered its interim verdict: specialized inference silicon is no longer a fringe bet. The money behind Etched — Sequoia, Jane Street, and a payroll seeded with Nvidia alumni — believes the GPU’s general-purpose era is entering its long twilight, and it is willing to pay a doubling premium, month over month, to find out.

---

Figures in this article are drawn from Reuters, TechCrunch, The Wall Street Journal, Semiconductor Reports, and Etched’s own disclosures as of August 18–22, 2026.