Major tech companies are engaged in a historic capital expenditure cycle, projected to reach nearly a trillion dollars across the supply chain. This spending is not just for immediate needs but is a long-term strategic play, involving prepayments for data centers, power infrastructure, and semiconductor manufacturing capacity for as far out as 2029.
Leading AI labs like Anthropic and OpenAI are in a desperate race to secure gigawatts of compute capacity to support both model training and explosive inference demand. Anthropic's conservative strategy has left it compute-constrained and scrambling for capacity, while OpenAI's more aggressive, diversified approach has secured it a more robust supply pipeline.
The AI boom is creating critical chokepoints throughout the semiconductor supply chain. TSMC is struggling to allocate wafer capacity among competing AI and mobile customers, memory vendors are raising prices due to HBM demand, and the ultimate long-term constraint is projected to be ASML's limited production capacity for essential EUV lithography tools.
The market for AI compute is defined by high costs, with a gigawatt of capacity renting for approximately $10 billion per year. While the total cost of ownership for GPUs is lower than current rental prices, scarcity drives spot prices to extreme highs. This pressure is forcing AI labs to increase their own prices, which is expected to dramatically improve their currently thin gross margins.
The analysis highlights a widening compute gap between US-based AI labs and their Chinese counterparts, driven by the massive infrastructure investment in the West. While China is aggressively pursuing an indigenous semiconductor supply chain, with predictions of domestic EUV capability by 2030, it currently lacks the scale to compete with the AI buildout in the US.
Keep pulling the thread on Dylan Patel.