Standard Bank caps spending as Google opens a path for models to keep data encrypted
Standard Bank separates AI tools by licence, developer account and use case, while Google's HEIR toolchain aims to run trained models on inputs that remain encrypted.
Artificial Intelligence··Morning
Standard Bank manages spending through three separate paths
Standard Bank's chief information officer described AI spending managed in three groups. General tools arrive with existing software licences; coding tools such as Cursor, GitHub Copilot and Claude Code run through per-developer accounts with soft ceilings intended to stop bills growing without control. Applications built inside the bank are partitioned by use case on Amazon Bedrock. The executive also rejected measuring capability only by the number of tokens consumed.[1]
HEIR runs the model without decrypting the data
Google's open-source HEIR compiler and toolchain compiles a pre-trained model to run on homomorphically encrypted inputs. A developer marks in Python which data types must stay encrypted; the first demonstrations convert PyTorch models through torch_mlir into MLIR and then compile them. The demonstrations cover private recommendations, card-fraud detection, network-intrusion detection and hotword recognition. The conversion path is not yet automatic.[2]
The spending and processing limits remain visible
Standard Bank's soft ceilings on developer accounts constrain coding-tool spending at the individual level, while partitioning internal Bedrock applications by use case separates which work consumes which resources. For HEIR, timings reported by InfoQ readers reach 80 ms for a 64-bit equality test, 100 ms for addition or subtraction and up to 8 seconds for division. The overhead reaches roughly a thousandfold on some tasks. Google published benchmarking code but no official figures for large language models.[1], [2]
