The chip that makes model maths fast enough to sell.
Accelerators run the huge parallel matrix multiplications that models are made of. A modern data-centre GPU pairs thousands of compute units with very fast on-package memory.
Memory bandwidth, not raw arithmetic, is usually the binding constraint during inference: the weights must be streamed to the compute units for every token generated.
Supply of these parts, and of the packaging and memory they need, sets the pace of the entire industry.
The accelerator market has one obvious barometer: NVIDIA at 210.96, down 8.42% on the day. When supply or demand for these chips shifts, this is the line that moves first.