Tiny corp wants Etched's numbers. Jane Street led $700M at $21B without publishing any

On August 18, tiny corp, the startup behind the tinygrad framework, publicly questioned whether rival Etched has a competitive AI inference chip at all, the same day Etched shipped a chip to Jane Street ^[1][3]. Etched then confirmed a $700 million round at a valuation near $21 billion, led by Jane Street ^[2]. The gap between those two events is the story: the price was set by one buyer's private test, and critics say Etched has released no independent benchmarks to support its performance claims ^[4][15].

The repricing has been fast. Etched was valued at $5 billion in December 2025 and at $10.3 billion in a Series C that closed in July, roughly a month before the new round, which Cryptopolitan reported citing the Wall Street Journal ^[13]. That is about 4.2 times in eight months and about double in a month ^[1][2]. Even the headline number is soft: about $21 billion per TechCrunch, or $20.3 billion per Cryptopolitan's account of the Journal's reporting ^[14].

tiny corp's demand is unusually cheap to satisfy. It asked Etched for peak FLOPS, power draw, or third-party benchmarks that do not reveal proprietary technology ^[7]. Instead, writing on X, the account said it could not rule out that the company led by CEO Gavin Uberti either lacks a working chip or is propping up a poor one with "smoke and mirrors," while conceding that "it's possible they have a great chip and just very distasteful marketing" ^[5][6]. tinygrad also suggested Etched may be avoiding outright fraudulent claims by publishing no numbers at all ^[8]. Its conclusion: "Etched shipped a slogan" ^[10]. The line of attack is not new; in June 2025 the same account urged people to warn Etched investors about a pitch it called nonsense, claiming "20x gains from etching transformers into silicon" ^[9].

A second critic, robotics engineer Wesley Yue, who claimed prior roles at Kitty Hawk and Waymo, wrote that "this smells very bad. none of the claims make sense" ^[11]. His technical objection is the more useful one: a high model FLOPs utilisation number is easy to hit when peak throughput is low, and spending scarce bleeding-edge wafers on lower-performance chips does not add up ^[12]. Note that the source dates Yue's scrutiny to July 2026, which sits awkwardly beside an August 18 flashpoint; treat the timeline as reported rather than settled ^[11].

Against that, Etched has more than marketing. Jane Street said in the funding announcement, "we tested the chip and are pleased with the early results" ^[15]. Etched reported first-pass silicon success on TSMC's N4P process, employs more than 400 people, and has booked over $1 billion in orders ^[17]. It runs a 2-megawatt site in San Jose and has opened a 10-megawatt facility in Milpitas, roughly 12 megawatts combined ^[18][3]. Its systems now run any frontier model, including DeepSeek, Qwen, Mamba and Llama, shedding the early one-model-per-chip framing ^[19]. Sequoia, Andreessen Horowitz, Peter Thiel, Bain Capital Ventures, SK Hynix and Blackstone have all participated in its rounds ^[16].

So the valuation is roughly 21 times booked orders and under $53 million per employee ^[4][5]. Those are not absurd multiples for pre-revenue-scale silicon, but they are underwritten by testimony rather than data.

What to watch: whether Etched publishes a peak FLOPS figure or power envelope, which discloses nothing an experienced competitor cannot infer; whether any third party outside its investor base runs a benchmark; and whether the $1 billion order book converts into deployed capacity at Milpitas ^[7][17][18].