FHE · A little slower
The gloves were too thick.
Computing on locked data works. In software it is too slow for a real rack. Each locked step adds noise. Past a budget, decryption fails. Managing that budget is most of the runtime cost: not the useful work.
Encryption injects randomness so locked data does not leak the open files. Homomorphic operations grow that randomness. Eventually the scheme must bootstrap: refresh the lock so computation can continue. Bootstrapping is correct. It is also expensive. A production rack does not pause after every useful step.
Versus conventional lattice FHE, the operation-level gap is on the order of ~1000×. That axis is why software missed the rack.
Versus doing the same work in the clear, end-to-end overhead is ~50×. That axis is what a CFO feels. They never share a sentence here because they do not measure the same thing.
Umbra runtime is deterministic locked execution with no bootstrapping and no noise refresh by design. Umbra does not claim to eliminate the cost of encrypted computation. It claims to have made that cost payable. Card figures are measured. Node figures are estimated.