Concentration of capability is the systemic tendency for advanced computational infrastructure — particularly AI accelerators and large-scale training clusters — to accumulate in the hands of a small number of corporations, nation-states, and research institutions. Unlike earlier computing eras, where Moore's Law democratized performance gains across all hardware purchasers, the shift to domain-specific accelerators creates discontinuities: a TPU pod or a ten-thousand-GPU cluster is not merely faster than consumer hardware; it enables qualitatively different kinds of model training and scientific discovery. This concentration is not an accident of market structure; it is baked into the physics of specialized hardware. The fixed costs of designing a new chip generation, the supply-chain control of advanced semiconductor manufacturing, and the operational complexity of large-scale distributed training create barriers that cannot be overcome by open-source software or algorithmic ingenuity alone. The result is a two-tier system: those who can train at the frontier, and those who can only infer from models they did not create.
The open-source movement has won every battle about software, but it is losing the war about hardware. And hardware, in the age of AI accelerators, is the only battlefield that matters.