The Hardware Coup: Why the AI Hardware Paradigm Is Changing Forever

The AI hardware market is currently undergoing a massive shift from general-purpose GPUs (Generalist) to ASICs (Specialist) optimized for specific tasks. Over the course of June alone, major companies simultaneously entered the custom chip market, accelerating this trend.

The moves by major companies are as follows:

• OpenAI: Partnered with Broadcom and Celestica to develop its first custom chip, 'Jalapeño', in just 9 months. Plans to deploy at a scale of 10GW by 2029.
• Etched: Unveiled 'Sohu', a 4nm chip dedicated to transformer models. Claims it can run the Llama 70B model at a speed of 500,000 tokens per second, and secured $800 million in investment.
• Amazon: Began discussions to sell its Trainium chip, previously used only for its own cloud, to external data centers.
• SambaNova: Launched the SN50, a 5th-generation chip aimed at reducing enterprise AI costs.

This evolution in hardware is threatening the moat of software companies. Previously, software optimization was the key competitive advantage, but now optimization is being performed directly at the hardware design stage.

Ultimately, the outcome of future AI competition will depend not just on algorithms, but on who dominates the physical layer—custom silicon design, securing power grids, and building efficient hardware.