On AI Chip Infrastructure as Foundational Capital

What We Observe

• The AI compute bottleneck is structural, not cyclical. Global demand for training and inference hardware far exceeds available supply. Lead times for high-end AI chips extend beyond twelve months, and hyperscale cloud providers are spending tens of billions annually to secure capacity. This is not a temporary shortage but a reflection of exponential growth in AI workloads against the physical constraints of semiconductor manufacturing.
• The market has concentrated almost entirely on a single vendor for AI compute. This concentration creates systemic risk for the industry and economic opportunity for companies offering architecturally differentiated alternatives. Enterprises and sovereign AI programs are actively seeking supply diversification.
• Cerebras Systems has taken a fundamentally different approach to chip design by building the largest chip ever made: a wafer-scale processor purpose-built for AI training and inference. Rather than incremental improvements to existing architectures, Cerebras designed from first principles to eliminate the memory bandwidth and interconnect bottlenecks that limit conventional processors.
• Groq has developed a novel inference-first chip architecture called the Language Processing Unit. Founded by Jonathan Ross, the architect of Google's Tensor Processing Unit, Groq's LPU delivers inference speeds that are orders of magnitude faster than existing solutions with deterministic latency and predictable cost per token.

Why It Matters

The AI industry is undergoing a buildout comparable in scale to the construction of the internet's physical infrastructure in the 1990s. Every major technology company, government, and enterprise is racing to secure AI compute capacity. The companies that provide the foundational hardware layer, the silicon that makes AI systems possible, will occupy structural positions in the technology economy for decades. Just as the internet required routers, fiber optic cables, and data centers before applications could scale, the AI economy requires purpose-built training and inference chips before the promise of artificial intelligence can be fully realized.

How We Think About It

Bacalia evaluates AI hardware investments through the same infrastructure framework we apply to marinas, real assets, and autonomous systems. We ask consistent questions: Is the service essential? AI compute is becoming as fundamental as electricity. Is supply constrained? The hardware shortage is severe and multi-year. Are barriers to entry high? Designing and manufacturing custom silicon requires billions in capital, years of engineering, and deep semiconductor expertise. Is there a path to durable cash flows? AI chip companies with differentiated architectures can generate recurring revenue from cloud deployments, enterprise contracts, and sovereign AI programs. Cerebras and Groq each score strongly across this framework. Cerebras addresses the training bottleneck with wafer-scale compute that delivers performance conventional chips cannot match. Groq addresses the inference bottleneck with an architecture optimized for speed and cost efficiency at scale. Together, they represent complementary positions in the AI hardware stack.