Zama, a cryptography startup, says it has cracked a major speed barrier for fully homomorphic encryption. CEO Rand Hindi announced the company achieved 1,000 confidential transfers per second on GPUs — a performance level that, if verified, could bring privacy-preserving computation closer to real-world use.
What fully homomorphic encryption does
Fully homomorphic encryption, or FHE, lets you compute on encrypted data without ever decrypting it. That means a cloud server could process sensitive information — financial transactions, medical records, private messages — while the data stays encrypted the whole time. The catch has always been speed. Early FHE schemes were millions of times slower than plaintext operations, making them impractical for anything beyond research.
The performance milestone
Zama, the company behind the FHE library Concrete, has been working on GPU acceleration. Hindi’s claim of 1,000 confidential transfers per second represents a roughly 100x improvement over previous GPU-based FHE speeds, according to the company. The tests used Nvidia GPUs, though Zama did not specify which model. The transfers are a benchmark for confidential payments — a use case where FHE could let a bank or payment processor verify transactions without seeing the amounts or parties involved.
The claim has not been independently verified. Zama has not released the full benchmark methodology or third-party audit results. Cryptography researchers often treat such performance announcements with caution until they can reproduce the results.
Why GPUs matter for FHE
FHE operations are highly parallelizable, which makes GPUs a natural fit. But the mathematical overhead is enormous — each operation on encrypted data requires many more computations than the equivalent plaintext operation. Zama’s approach uses a technique called programmable bootstrapping, which refreshes the noise that accumulates during computation. The company says its GPU implementation reduces the latency of that bootstrapping step significantly.
Other groups, including researchers at Intel and Microsoft, have also worked on hardware acceleration for FHE. Zama’s claim puts it ahead of publicly known results, but the field moves fast.
What’s next for Zama
Hindi did not announce a product launch date or a specific customer deployment. The company has not said when the technology will be available in a production environment or whether it plans to open-source the GPU code. Zama’s existing library, Concrete, is open source, but the GPU optimizations may be proprietary.
For now, the claim stands as a promise. Independent verification and a clear path to production will determine whether 1,000 transfers per second becomes a real-world tool or just another research milestone.
