Behind Etched's $10B Valuation: A Bet on Transformer's 'Ultimate Form'

$300 million Series C funding, $10.3 billion valuation, founded in 2022, three Harvard dropouts. This set of data is striking enough for any AI chip company. Etched's Sohu chip, an ASIC designed specifically for the Transformer architecture, is challenging a repeatedly validated "common sense": AI chips must follow a general-purpose path, otherwise they will die under ecosystem lock-in.

Let's look at the facts first. Etched's valuation has surpassed several established AI chip companies—Cerebras (~$4 billion), Groq (~$2.8 billion), nearly double SambaNova's. Yet the Sohu chip has not yet entered mass production. Behind this "expectation pricing" is investors' judgment on structural changes in the AI inference market: As model scales leap from hundreds of billions to trillions of parameters, the inference efficiency ceiling of general-purpose GPUs is approaching.

Core Conflict: Generality vs. Extreme Efficiency

Etched's logic can be broken down into three layers:

• Technical Layer: The core computation of Transformer inference is matrix multiplication in the Attention mechanism (Q·K^T · V). The Sohu chip hardwires this computation path, eliminating unnecessary instruction scheduling and cache overhead on GPUs. Measured data shows Sohu's energy efficiency ratio (Token/Joule) in Llama 3 70B inference is 3-5 times that of H100.

• Market Layer: AI applications are shifting from training to inference. OpenAI's GPT-4o inference cost has dropped to 1/10th of training costs, but inference usage is still growing exponentially. The cost advantage of specialized chips in inference scenarios directly corresponds to cloud providers' profit margins.

• Ecosystem Layer: Etched is compatible with PyTorch and TensorRT, using CUDA's compilation layer for adaptation rather than reinventing the wheel. This lowers migration costs but heavily relies on NVIDIA's software stack—if NVIDIA modifies underlying interfaces, Etched faces compatibility risks.

# Simplified Attention calculation in Transformer inference
def attention(Q, K, V):
    # Q, K, V shape: [batch, heads, seq_len, dim]
    scores = torch.matmul(Q, K.transpose(-2, -1))  # Core matrix multiplication
    weights = torch.softmax(scores / math.sqrt(dim), dim=-1)
    output = torch.matmul(weights, V)
    return output

In this code, torch.matmul requires massive general-purpose computing resources on a GPU, while Sohu maps this operation directly through hardware, theoretically achieving utilization close to the theoretical limit.

Benchmarking Overseas Cases: Google TPU's Success and Limitations

Google TPU is the most successful case of specialized chips. TPU v1 (2016) was designed specifically for inference, performing excellently in Google Search and RankBrain. But TPU's success was built on Google's internal closed-loop ecosystem: self-developed framework (TensorFlow), self-developed models (BERT, PaLM), self-developed data centers. As an independent company, Etched must convince external customers to pay for a "single architecture."

[!tip] Key Difference

TPU was never sold extensively to external customers but provided as a cloud service. Etched sells chips directly, facing NVIDIA's channel barriers and customer trust issues—will customers invest tens of millions in procurement for a chip that hasn't mass-produced and depends on the Transformer architecture?

Looking at Porter's Five Forces, the risks Etched faces:

• Supplier Bargaining Power: TSMC's advanced process capacity is tight. As a startup, Etched has low priority in production scheduling, potentially leading to delivery delays.
• Threat of Substitutes: The Transformer architecture is not the end state. State Space Models (SSMs) like Mamba and RWKV are challenging Transformer's dominance. If SSMs become mainstream, Sohu's specialized advantage drops to zero.
• Existing Competitors: NVIDIA's B200 GPU continues to improve inference performance and possesses a complete CUDA ecosystem. AWS's Trainium2 and Microsoft's Maia 100 are also optimizing for inference scenarios.

An Extreme Hypothesis

If within the next two years, the Transformer architecture proves to be the "final form" (like x86's dominance in PCs), then Etched's valuation logic holds. But the speed of technical disruption in AI far exceeds the semiconductor industry. Specialized chips take 18-24 months from design to tape-out, while architectural innovations emerge every 6 months.

Etched has chosen a "narrow path": It must bet that Transformer's dominance lasts long enough, while simultaneously tearing open a gap in NVIDIA's "iron wall."

Financials

Original Link: https://techcrunch.com/2026/07/23/ai-chip-startup-etched-defies-skeptics-hits-10-3b-valuation-from-big-name-investors/