In 2023, two Harvard dropouts walked into investor meetings with a 30-page deck saying AI's future needed purpose-built chips, not general-purpose GPUs. Every major VC turned them down.

Three years later, Etched -- the company they founded -- is valued at $5 billion, has raised $800 million in total funding, and holds $1 billion in inference system contract orders. The investor list reads like an AI all-star roster: Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Arthur Mensch (Mistral founder), Scott Wu (Devin/Cognition founder), plus Stanley Druckenmiller and Peter Thiel.

This isn't another AI chip PowerPoint story -- TSMC has successfully taped out, and Etched is conducting chip testing with customers. Their product is called a "frontier inference cluster": chips + custom racks + software, sold as a package to customers running AI inference at scale.

Inference: AI's Biggest Cost Center

To understand Etched's opportunity, you first need to understand the structure of the AI compute market.

AI compute splits into two phases: training and inference. Training is "teaching" the model -- a massive one-time investment. Inference is "using" the model -- every user query, every Agent executing a task consumes inference compute. As Agents proliferate -- especially with affordable models like Sonnet 5 driving mass Agent deployment -- inference compute demand is exploding. Inference is currently AI companies' biggest cost center and biggest bottleneck.

Nvidia's GPUs are general-purpose compute chips that can handle both training and inference, but "can do everything" also means "not optimal at anything." ASICs (Application-Specific Integrated Circuits) are custom-designed for specific tasks and can be faster, cheaper, and more power-efficient than GPUs in inference scenarios. Etched's bet is: as inference demand far outstrips training demand, the market for purpose-built inference chips will explode independently of the GPU market.

Etched isn't alone. Cerebras completed a $5.5 billion IPO in May; Groq raised $650 million in June; Amazon, Google, and Microsoft are all building their own AI chips; OpenAI just released its first self-built chip Jalapeño (manufactured by Broadcom). The AI chip market is shifting from Nvidia's dominance to a diversified landscape of "training GPUs + inference ASICs."

China's Domestic Compute Race in Parallel

The same day Etched announced its funding news, the China Academy of Information and Communications Technology (CAICT) released the first benchmark for AI Infra operations, covering 5 mainstream domestic chips. This isn't a coincidence -- two fronts of the global AI chip race are advancing simultaneously: the US as the innovation frontier (ASIC startup explosion), China as domestic substitution (building an autonomous compute stack under Nvidia restrictions).

In June, Shenzhen Hetao Academy/ShenZhiCheng Compute Network completed full-parameter post-training of DeepSeek-V4-Pro on a domestic compute cluster -- the first time a third-party institution has achieved this. But gaps objectively remain: Nvidia's Vera Rubin ships in the second half of the year with 3.3x the performance of Blackwell Ultra, and domestic chips still have generational gaps in process technology, interconnect efficiency, and software ecosystem.

That said, inference chips may be domestic compute's breakthrough point. Training demands extreme peak performance from chips, but inference scenarios are more sensitive to cost and energy efficiency -- precisely where domestic chips can compete on differentiation. As Etched proves there's a billion-dollar market for inference ASICs, Chinese inference chip startups are also rapidly following suit.

The global AI compute landscape is undergoing a quiet restructuring. Nvidia remains strong on training, but the inference wars have only just begun -- Etched, Cerebras, Groq, in-house chip teams, and China's domestic chipmakers are all competing for this market.

For LLM companies, this is good news: falling inference costs mean expanded gross margins and better economics for Agent deployment at scale. For Nvidia, this is a challenge it must take seriously: its CUDA ecosystem moat remains deep on training, but in inference, purpose-built chips are prying open cracks with price-performance.

See you tomorrow.