Etched just doubled its valuation to $21 billion, led by Jane Street. The news broke like a flash grenade in the AI chip world. A 32-year-old startup with a single product roadmap – Sohu, a Transformer-only ASIC – now commands a price tag that rivals some of the most established players in the industry.
Chasing the alpha until the trail goes cold—that's the game. But here's the rub: the chip hasn't shipped yet. No production silicon. No public benchmark. No verified customer deployments. Yet investors are throwing capital at a binary bet: that the future of AI inference will be dominated by a single, inflexible architecture.
I've been around long enough to remember the ETHDenver hype cycle in 2017. I was there, chasing Vitalik's off-record comments about scalability, publishing a rapid interpretation within 45 minutes. The energy was electric, but the technical reality was messy. The same pattern repeats here: a narrative of 'specialization over general purpose' is selling like hotcakes, but the engineering details are still buried under hype.
Let me break down what this $21 billion valuation actually means – and why you should be skeptical.
Context: The ASIC Bet
Etched's Sohu chip is not a general-purpose GPU. It's a custom ASIC designed from the ground up to run Transformer models. Think of it as a dedicated highway for one specific type of traffic. The theoretical efficiency gain is enormous: a single ASIC can be 10x more power-efficient than a GPU for the same inference workload. But the trade-off is zero flexibility. If the AI industry shifts away from Transformers – say, to State Space Models like Mamba, or to hybrid architectures – the chip becomes a paperweight.
Jane Street, the quant trading giant, led this round. That's a strategic signal. Jane Street runs massive low-latency inference workloads for trading. They need speed, and they need it now. An ASIC that cuts latency by an order of magnitude is a direct competitive advantage. But Jane Street is one firm. The $21 billion valuation implies the market expects a dozen more Jane Streets to line up.
Core: The Valuation Math
Let's do the math. At $21 billion, investors are essentially pricing Etched as if it already has a lock on 10-15% of the AI inference market in the next 3-5 years. That's a bold assumption. The inference market is currently dominated by NVIDIA's GPUs, with a software ecosystem that's a decade in the making. CUDA is not just a library; it's a moat. Every ML engineer knows how to deploy on NVIDIA. Switching to an ASIC means rewriting stacks, rebuilding pipelines, and trusting a new supply chain. The switching cost is high.
Based on my audit experience during the DeFi Summer of 2020, I saw how liquidity mining APYs masked real user retention. Projects pumped TVL with incentives, but when the rewards stopped, users vanished. The same dynamic is at play here: the 'demand for dedicated AI hardware' is real, but it's not clear if it's sustainable or just a temporary spike driven by NVIDIA's supply constraints. If NVIDIA can close the gap with its next-generation Blackwell or Rubin architectures, Etched's advantage could shrink to a point where the ASIC premium isn't worth the lock-in.
And let's talk about the supply chain. Sohu likely uses TSMC's 5nm or 4nm process, with HBM memory and CoWoS packaging. These are the same resources that NVIDIA and the hyperscalers are fighting over. TSMC's capacity is finite. If Etched doesn't have a long-term capacity agreement (LTA), they could face delays that kill the momentum. I've seen this movie before: a hot chip startup announces a massive valuation, then misses its tape-out deadline, and the valuation collapses.
Contrarian: The Unreported Angle
Here's what almost every article is missing: the risk of architecture obsolescence. Etched's entire thesis is that Transformers will remain the dominant architecture for the next 5-10 years. But the AI research community is actively exploring alternatives. Mamba, a State Space Model, achieves linear-time inference and matches Transformer quality on many benchmarks. Google's Pathways, Meta's research on mixture-of-experts, and the rise of multimodal models are all pushing the boundaries. If any of these become the new standard, Etched's ASIC loses its raison d'être.
Moreover, the customer concentration risk is real. Jane Street is a quant firm, not a cloud provider. The big money in AI inference comes from OpenAI, Anthropic, Meta, Google, and Microsoft. If Etched can't land a single hyperscaler as a customer, its revenue ceiling is limited to niche financial and low-latency applications. That's a billion-dollar market, maybe, but not a $21 billion one.
Another blind spot: software. ASICs are only as good as their compiler, operator library, and integration with popular inference frameworks like vLLM, TensorRT, or ONNX Runtime. Building a software stack that rivals CUDA is a herculean task. Even Google's TPU, which is far more mature, still struggles with developer adoption outside of internal Google teams. Etched is starting from scratch.
Takeaway: What to Watch
The next 12 months will determine whether Etched is a moonshot or a mirage. I'm watching three signals:
- Mass production schedule: If they don't announce a firm tape-out date and a TSMC capacity agreement within 6 months, the valuation is built on sand.
- Independent benchmarks: Any benchmark from MLPerf or a third-party lab will reveal whether the Sohu chip actually beats H100/B200 by a meaningful margin. If the gap is less than 3x, the ASIC story weakens.
- Customer announcements: Beyond Jane Street, who else is buying? If the next big customer is another hedge fund, not a cloud provider, the market will cap the upside.
Chasing the alpha until the trail goes cold. Etched's trail is hot right now, but it's also thin. The $21 billion valuation is a bet on the future of AI inference, not a reflection of present reality. Keep your eyes on the silicon, not the news.
I've seen this script before – in DeFi, in NFTs, in the Terra collapse. The hype cycle always peaks before the product ships. The ones who make money are the ones who exit before the inevitable correction. But if Etched delivers, the rewards will be enormous. Either way, the game is on.
Chasing the alpha until the trail goes cold – that's my mantra. And right now, the trail is blazing.