Etched
Etched
One-line summary: Private (unlisted) transformer-inference ASIC startup — founded 2023 by gavin-uberti and rob-wachen — building full inference racks on TSMC 4nm; tracked here not as a tradeable itself but as load-bearing evidence for the inference-ASIC wave's effect on public tickers (TSM demand broadening, NVDA inference-share threat).
What it is
A chip-and-systems company building complete racks for AI inference (chip, boards, power delivery, custom interconnect, contract manufacturing in Taiwan). Claims two technical bets: low-voltage inference (running at "under half the voltage of any other AI chip" to escape thermal throttling and pack more usable FLOPs per watt) and cluster-scale memory (a fully custom interconnect stack cutting chip-to-chip latency >5x vs. Nvidia's ~4,000ns, letting a whole scale-up cluster's SRAM+HBM act as one pool). First product ("Sohu" rack) taped out working on first attempt; claims >$1B customer demand and ~$800M raised.
Why it matters to stock-market
Etched is private — no direct trade. Its wiki value is as a datapoint in public-ticker chains:
- TSM: Etched is a TSMC N4 customer and deliberately built on a different supply chain than Nvidia's Rubin (4nm vs 3nm, different HBM generation) — first-party evidence that the ASIC wave is additive to foundry/HBM demand, not substitutive. Canonical chain: inference-asic-wave-to-tsm-demand-broadening.
- NVDA: a live entrant in the transformer-ASIC threat to Nvidia's inference share, and another data point for cuda-moat-erosion-at-inference (explicitly refuses to support CUDA/PyTorch/arbitrary graphs).
- Speed commoditization: rob-wachen claims "an entirely new wave of AI chips... are all going to be able to hit these speeds," shifting differentiation to concurrency-per-megawatt — a tension input for inference-speed-as-a-pricing-premium and inference-demand-to-wafer-scale-advantage (Cerebras).
Sourcing caveat (read first)
All load-bearing performance and demand claims come from the founders on a podcast hosted by a disclosed Etched investor — 2026-06-30-podcast-invest-like-the-best-etched-building-ai-hardware-to-make-inference: Patrick O'Shaughnessy: "I'm a big Etch Investor. I've been involved for a long time... So I'm incredibly biased in this conversation." Treat every claim (voltage, latency, $1B demand, 40-day bring-up) as unverified promotional until independently benchmarked. No independent corroboration in the wiki yet.
Key facts (all self-reported, same source)
- Founded 2023 by Uberti (ex-kernels engineer, Harvard dropout) and Wachen; CTO Mark Ross (ex-Cypress Semi CTO); rack lead Brian Leyler (built Nvidia HGX/DGX systems).
- Fabbed at TSMC N4; own factory in Taiwan; CM (not JDM) model; "different HBM than Rubin."
- Claims: >$1B customer demand, ~$800M raised, silicon-to-rack-inference in 40 days (vs a named-anonymous competitor's 10 months), chips running at "under half the voltage of any other AI chip."
- Strategy: no compiler, no CUDA/PyTorch/ONNX support — kernels-first for "under 100 models that actually matter"; HFT firms cited as culturally aligned early believers/hires.
Open questions
- Independent benchmarks of the concurrency-per-megawatt claim (order of magnitude vs GPUs at a given interactivity) — nothing public yet.
- Does gen-2 stay off the leading node, or does scaling force it into contention with Rubin for the same wafers/HBM (which would break the "additive" claim)?