Just a guess:
Theoretically Cocoon could implement software-based confidential compute via tensor masking. This would cut performance by about 5x but would allow models to run almost in "home" on RTX cards. You'd need a server-grade CPU with TEE support for this, but such a setup costs around $2-3k, unlike NVIDIA Hopper series GPUs (H100, H200) that go for $20-30K.