They have an incentive to make their models efficient enough to serve demand and make a profit on it.

The incentive that is missing is passing on efficiency improvements as price savings to customers, when your model is still in demand because of its higher intelligence.

Agreed, efficiency is still important, but being at the "frontier of efficiency" is significantly more relevant to commodity model providers than state-of-the-art model providers. Frontier labs are incentivized to route their spend towards beating benchmarks because that's what enables them to charge a premium.