Why does running my prompt inside an encrypted VM not actually keep it private?

Confidential computing via Intel TDX and AMD SEV-SNP is the leading approach for private AI inference: user memory is encrypted and the cloud operator is architecturally blocked from reading it. TDXRay, published at IEEE S&P 2026 by UC San Diego, CISPA, and Google, showed that a host can reconstruct user prompts word-for-word at over 90 percent similarity by monitoring tokenizer cache access patterns without ever decrypting guest memory. A separate October 2025 disclosure showed that a $1,000 DDR5 bus interposer can extract attestation keys from both TDX and SEV-SNP hardware outright, making attestation forgery practical with commodity equipment. The only known software mitigation, oblivious RAM, imposes latency overhead that makes production inference uneconomical. An IDC survey from December 2025 found 84 percent of organizations adopting confidential computing cite attestation validat