The ledger remembers what the mind forgets. In 2018, Bitmain, then the king of crypto mining ASICs, was valued at $12 billion. Its chips were purpose-built for SHA-256, a single algorithm. Within two years, the market shifted, power efficiency gains plateaued, and the company’s valuation collapsed by over 80%. Today, Etched, a startup building an ASIC exclusively for the Transformer neural network architecture, has seen its valuation double to $21 billion in a single funding round led by Jane Street. The company has yet to ship a mass-produced chip. The parallels are not lost on anyone who has watched hardware cycles before.
Context: The AI inference market is the new frontier. Training costs are falling, but inference—the actual deployment of models—now accounts for 80% of total AI compute spending. For every dollar spent on training, four are spent on running models. The standard solution, NVIDIA’s GPUs, are general-purpose accelerators that excel at both training and inference but are not optimized for the latter. Enter Etched, with its Sohu chip, a 5nm-class ASIC that claims to deliver 10x the inference throughput per watt of an H100 for Transformer models. The promise is simple: if you only run Transformers, why pay for the overhead of a GPU? The valuation of $21 billion says the market believes this logic is worth betting on—but the logic itself carries hidden assumptions that demand scrutiny.
Core: First, let’s deconstruct the valuation. At $21 billion, assuming a conservative 10x price-to-sales ratio in a mature market, Etched needs to generate over $2 billion in annual revenue within the next 3-4 years. That implies selling hundreds of thousands of chips, each priced at a premium over GPUs, or securing massive token-sharing contracts. The only way to reach that scale is to win large customers: cloud hyperscalers (AWS, Google, Microsoft), captive AI labs (OpenAI, Anthropic), or at least a consortium of financial firms. Jane Street’s lead investment is strategic—it is a high-frequency trading firm with real-time inference needs. But one quant powerhouse does not a $2 billion revenue stream make.
Now, examine the technology. The Sohu chip is an ASIC—application-specific integrated circuit. For inference, it can indeed be more efficient than a GPU because it hardwires the matrix multiplication and attention mechanisms central to Transformers. But efficiency comes at the cost of flexibility. The moment the dominant AI model architecture strays from the pure Transformer—toward state-space models like Mamba, mixture-of-experts (MoE) variants, or hybrid architectures—the Sohu chip’s value proposition evaporates. The crypto ASIC analogy is apt: Bitmain’s Antminer S9 was incredible for Bitcoin mining, but useless for Ethereum. Similarly, if the next generation of AI models (GPT-5, Gemini 2) incorporates non-Transformer components, Etched’s chips become paperweights. The assumption that Transformers will remain dominant for 3-5 years is not unfounded, but it is a bet on technological stasis in a field that changes quarterly.
Beyond architecture, the engineering challenges are immense. Etched’s chip likely uses TSMC’s 5nm or 4nm process, requires HBM memory, and depends on CoWoS advanced packaging. All of these are the very same resources NVIDIA needs for its Blackwell and Rubin chips. Supply chain allocation is the hidden tax on every AI chip startup. If TSMC prioritizes NVIDIA’s orders, Etched’s delivery timeline slips. The company has not publicly disclosed a foundry commitment or a wafer capacity agreement. Without that, the $21 billion valuation is a bet on a factory that hasn’t guaranteed a seat at the table.
Software is another chasm. NVIDIA’s CUDA ecosystem is a 15-year fortress of libraries, operators, and deployment tools. Etched must build a compiler stack, an inference framework, and a hardware abstraction layer that allows developers to drop in their models without rewriting code. The cost of porting a single production model to a new hardware platform is often over $1 million and months of engineering time. Cloud providers like AWS and Google have their own chips (Trainium, TPU) precisely because they can amortize that cost across millions of workloads. Etched has no such captive user base. Its software team is likely small compared to the incumbents. The risk of a buggy or incomplete software stack is high, and it would kill adoption regardless of raw hardware performance.
Contrarian: The bull market in AI hardware is creating a euphoric willingness to pay for potential. Etched’s $21 billion valuation is not just a bet on the company; it is a bet on the entire thesis that inference will be dominated by specialized ASICs rather than general-purpose GPUs. But the contrarian view is that NVIDIA’s next generation, Blackwell, already narrows the gap. Early benchmarks suggest that for inference, Blackwell achieves up to 5x the efficiency of the H100, and Rubin may push that further. If the gap between a GPU and an ASIC shrinks to 2x, the ASIC loses its economic advantage because the GPU offers flexibility and a mature ecosystem. The ASIC’s advantage is a window, not a permanent moat.
Moreover, customer concentration is a silent risk. Jane Street’s involvement suggests that Etched’s initial use case is financial inference: low-latency, high-frequency trading models. That market is large but capped. For Etched to justify $21 billion, it must also capture cloud inference for general AI workloads. But cloud hyperscalers are already building their own chips. Why would they buy from Etched when they can vertically integrate? Google’s TPU, Amazon’s Trainium, and Microsoft’s Maia are all custom ASICs that are tightly integrated with their respective cloud services. Etched could become a niche supplier, not a general infrastructure player. The valuation assumes the latter, but the evidence suggests the former.
Takeaway: The next 12 months will be decisive. Watch for three signals: a confirmed foundry deal with TSMC or Samsung, an independent benchmark from MLPerf or a similar body, and a non-financial customer announcement (preferably a cloud provider or an AI lab). Until then, $21 billion is a price tag on hope, not results. Code doesn’t lie, but valuations do. Macro tides turn. Be ready for the shift.