While NVIDIA's CUDA software provides a powerful lock-in for AI training, its advantage is much weaker in the rapidly growing inference market. New platforms are demonstrating that developers can and will adopt alternative software stacks for deployment, challenging the notion of an insurmountable software moat.

Google successfully trained its top model, Gemini 3 Pro, on its own TPUs, proving a viable alternative to NVIDIA's chips. However, because Google doesn't sell these TPUs, NVIDIA retains its monopoly pricing power over every other company in the market.

Google training its top model, Gemini 3 Pro, on its own TPUs demonstrates a viable alternative to NVIDIA's chips. However, because Google does not sell its TPUs, NVIDIA remains the only seller for every other company, effectively maintaining monopoly pricing power over the rest of the market.

Even if Google's TPU doesn't win significant market share, its existence as a viable alternative gives large customers like OpenAI critical leverage. The mere threat of switching to TPUs forces NVIDIA to offer more favorable terms, such as discounts or strategic equity investments, effectively capping its pricing power.

Tesla's decision to stop developing its Dojo training supercomputer is not a failure. It's a strategic shift to focus on designing hyper-efficient inference chips for its vehicles and robots. This vertical integration at the edge, where real-world decisions are made, is seen as more critical than competing with NVIDIA on training hardware.

This theory suggests Google's refusal to sell TPUs is a strategic move to maintain a high market price for AI inference. By allowing NVIDIA's expensive GPUs to set the benchmark, Google can profit from its own lower-cost TPU-based inference services on GCP.

In five years, NVIDIA may still command over 50% of AI chip revenue while shipping a minority of total chips. Its powerful brand will allow it to charge premium prices that few competitors can match, maintaining financial dominance even as the market diversifies with lower-cost alternatives.

NVIDIA's primary business risk isn't competition, but extreme customer concentration. Its top 4-5 customers represent ~80% of revenue. Each has a multi-billion dollar incentive to develop their own chips to reclaim NVIDIA's high gross margins, a threat most businesses don't face.

Beyond the simple training-inference binary, Arm's CEO sees a third category: smaller, specialized models for reinforcement learning. These chips will handle both training and inference, acting like 'student teachers' taught by giant foundational models.