Everyone Wants to Build AI Apps. One Investor Is Betting on the Toll Roads Instead.

The race to build AI applications is visible, crowded, and producing clear winners and losers every quarter. The race to control the infrastructure those applications run on is quieter, more capital-intensive, and, according to one investor with five positions in the space, where the durable value is actually accumulating.

How It Works

The thesis distinguishes between two layers of the AI stack that get conflated: training and inference.

Training — building large foundation models — has a few dominant players and limited startup opportunity. The capital requirements are enormous and the moats have largely been established. The Microsofts and Anthropics of the world have trained their models; the marginal entrant is at a structural disadvantage.

Inference — running those trained models efficiently for paying customers at scale — is a different problem. As the Entrepreneur profile frames it: “an efficiency problem, not a brute-force one, and efficiency problems are where specialized architecture beats general-purpose hardware.” That’s where the opportunity lives.

Five companies have raised significant capital on this thesis:

• Cerebras Systems went public in May at $185/share, raising $5.55B at a jump to ~$95B market cap. Cerebras builds wafer-scale chips specifically optimized for neural network inference.
• Groq rejected a reported $20B Nvidia licensing offer to pursue its inference cloud business, then raised $650M. The bet: that proprietary inference hardware has more upside than licensing to the incumbent.
• Positron raised $230M Series B in February, then $875M more in September — one of the largest private inference infrastructure fundraises.
• d-Matrix closed a $275M Series C at a $2B valuation.
• Tenstorrent secured $1.4B Series E (August 2026), pushing its valuation to $5.76B.

The combined capital deployed into these five companies exceeds $8B in 2026 alone. The investment metaphor driving this thesis: “I’d rather own the toll roads than bet on which car wins the race.”

What This Means for You

The inference infrastructure buildout has two practical implications for founders and operators thinking about AI economics.

First: the inference pricing environment is becoming genuinely competitive. Cerebras, Groq, and the others are building hardware that competes on tokens-per-second and cost-per-million-tokens against Nvidia-based cloud infrastructure. That competition benefits anyone running high-volume inference workloads — prices will fall and alternatives to the standard cloud AI APIs will multiply.

Second: the application layer is more volatile than it looks. Models are commoditizing faster than most app-layer founders anticipated. An application built on a specific model’s capability advantage can find that advantage eroded within a product cycle. Infrastructure positions are stickier: once a major cloud provider or AI platform commits to a specific inference chip architecture, the switching cost is high.

Source: Entrepreneur

Content created with AI assistance and reviewed for accuracy.