GPT-5.5 was presented as co-designed with Nvidia's newest accelerator generation, information that entered the launch material as a product characteristic.
Co-designing model and hardware means shaping the model architecture around what that equipment executes best, and shaping the equipment around that model's format. The gain shows up in efficiency per watt, which is direct operating cost.
The practice existed in research for a while; the novelty is it becoming a commercial launch argument, which indicates the gain grew large enough to sell to end customers.
The move explains much of what would follow in the sector: companies buying equity in manufacturers, assembling teams to design their own chips, and acquiring companies that etch models directly into silicon.
For operators, there's a trade-off worth watching. A model optimised for specific hardware tends to perform worse on other hardware, which raises the cost of migrating between cloud suppliers and reduces, in practice, the portability that existed on paper.