Why garbled-circuit MPC is the right tool for institutional privacy

There are four practical families of technology for keeping data confidential while still computing on it: fully homomorphic encryption, zero-knowledge proofs, trusted execution environments, and multi-party computation. Each solves a real problem. They are not interchangeable, and for institutional finance the differences are decisive.

What institutions actually need

Strip away the cryptography and the requirements list of a bank, fund or payment company is short and rigid:

1. Latency that feels like the Internet. A transfer that takes seconds to confirm is a product; one that takes minutes is a demo.
2. No trusted third party. A privacy system that depends on trusting one operator, one hardware vendor or one cloud is a single point of failure with a compliance department.
3. Standards-based encryption. Security teams approve AES. Novel ciphers with young cryptanalysis histories face years of review.
4. A disclosure path. Regulators and auditors must be able to see what they are entitled to see, or the system is unusable by regulated entities.
5. Sane economics. Privacy that requires GPU farms gets priced out of high-volume payments before it starts.

How the four families measure up

FHE computes directly on ciphertext with elegant generality, but encrypted operations run orders of magnitude slower than native ones, practical deployments lean on GPU hardware to compensate, and the underlying lattice schemes are young by AES standards. It is a fine research direction and a hard sell for a payments desk.

ZK proofs let one party prove a statement about their own data, and they do it superbly; rollup scaling is built on them. But a proof covers a single source of private data. An order book, an auction or any computation across many parties' secrets is outside the model, and client-side proving adds its own latency.

TEEs are fast, because computation runs in plaintext inside the enclave. The trust model is the price: you are trusting a hardware vendor, and the published attack record against production enclaves keeps growing. For an institution, "the data was decrypted, but inside Intel's box" is a sentence the risk committee has to sign.

Garbled-circuit MPC computes on encrypted data across many parties' inputs at once, with near-native circuit evaluation speed, on ordinary CPUs, using standards-based AES-256 for the data itself. No party ever holds the plaintext, so there is no vendor to trust and no enclave to breach. Disclosure is a first-class feature: the on-chain access list (ACL) grants auditors scoped decryption of exactly the records they are entitled to.

That combination, low latency, no trusted third party, battle-tested encryption, a compliance path and CPU economics, is the institutional requirements list, item by item.

The numbers behind the claim

This is not a theoretical ranking. Measured end-to-end on Arbitrum over 15,000 confidential transfers (July 2026): 500 cTPS sustained, 750 cTPS peak, at $0.14 per million transfers. The same family of technology has processed more than 125,000,000 confidential transactions in production as the privacy layer of COTI mainnet.

That is why we built Bubble on garbled-circuit MPC. Not because the other families are wrong, but because this is the one whose trade-offs line up with what regulated finance actually requires.

For the side-by-side comparison with sources, see the technology taxonomy. For the construction itself, read the gcVM paper.