To understand the concentration forming around large language models, it is not enough to look only at model quality. What produces monopoly power is who has access to the computational infrastructure needed to train these models.

At the most basic level, the ability to train large models through access to GPU chips determines who the main players will be. Companies that can secure enough GPUs, combine them at scale, and build the necessary software stack gain an advantage in the model race.

Hardware alternatives to GPUs are also expanding quickly. Options such as Google TPU, AWS Trainium and Inferentia, AMD MI300, Intel Gaudi, Cerebras, Groq, and mobile NPUs have made important progress, especially on the inference side.

But the main issue driving concentration is model training. In large-scale LLM training and heavy workloads, no real substitute has yet emerged outside GPUs and, to some extent, TPUs. CPUs, mobile NPUs, and other chip solutions cannot play the same role in large-scale model training.

Companies with access to GPUs train larger models; larger models attract more users; more users generate more data, revenue, and investment capacity. That revenue then goes back into accumulating more GPUs. The result is a closed loop in which technical advantage feeds financial advantage, and financial advantage feeds technical advantage.

The software ecosystem around GPUs, memory architecture, high-speed interconnects, and enterprise toolchains all reinforce this same structure. NVIDIA’s advantage is not just that it produces powerful GPUs; it also controls the software platform and programming model built around that hardware. Unless the firm’s software advantage, embodied in the CUDA platform, is broken, market balance will not shift even if alternative chips compete at the hardware level.

In short, at the center of today’s oligopolistic AI market is not ownership of an LLM alone, but compute power, that is, access to advanced computational infrastructure. Competition may appear to be opening up at the level of customer-facing applications, but real power is still accumulating in the model-training layer. And at that layer, access to advanced GPUs remains the single most important driver of concentration. Who can access which computational infrastructure largely determines who succeeds.