How they use Nebius AI Cloud
Zama measured TFHE programmable bootstrapping performance on Nebius H100 GPUs. In September 2025, it reported single-GPU latency of 796 microseconds for Booleans and 945 microseconds for 4-bit integers. On eight H100 GPUs, throughput reached 189,000 bootstraps per second for 4-bit integers.
Boolean throughput reached 223,440 bootstraps per second on eight Nebius H100 GPUs, a reported 1,655-fold improvement over its 2021 Concrete-core CPU baseline measured on AWS. It also reported 64-bit encrypted addition in 8.7 milliseconds and multiplication in 32 milliseconds on eight H100 GPUs.
The TFHE-rs v1.4 announcement reported further latency and throughput benchmarks for encrypted 64-bit addition, multiplication and division on Nebius with eight H100 GPUs. Zama’s repository also defines Nebius single-H100, eight-H100 and L40 runner configurations for Slab. Its H100 Terraform configuration specifies a 200-GiB network SSD boot disk and a pinned Ubuntu 24.04 image with CUDA 12.8 and driver 570.
The generic four-GPU test backend points to Nebius’s L40 Terraform script and lists a two-GPU Nebius L40 configuration first among its fallbacks. Its generic one-or-two-H100 backend configures a single Nebius H100, with ordered fallbacks to Scaleway two-H100, Scaleway single-H100 and Hyperstack single-H100 configurations.
The repository’s weekly AFL++ GPU fuzzing workflow for harness-cuda-compute configures Nebius L40 as the scheduled and manual default. The workflow is restricted to an internal mirror and skips all jobs in the public repository.
Sources
7 sourcestfhe-rs/.github/workflows/benchmark_gpu.yml at main · zama-ai/tfhe-rs · GitHub
Listed 11 Nov 2026 · github.com · Web page
Zama’s manual GPU benchmark workflow offers selectable Nebius single-H100 and eight-H100 benchmark environments.
Open original sourceTFHE-rs v1.4 GPU performance measurements
Published 31 Oct 2025 · Zama · Web page
Zama reports TFHE-rs v1.4 latency and throughput measurements for encrypted 64-bit addition, multiplication and division on Nebius with eight H100 GPUs.
Open original sourceSlab Nebius runner backends
Listed 28 Oct 2026 · GitHub · Web page
Zama’s Slab configuration sets its generic one-or-two-H100 backend to a single Nebius H100, with ordered fallbacks to Scaleway two-H100, Scaleway single-H100 and Hyperstack single-H100 configurations.
Open original sourcetfhe-rs/ci/slab.toml at main · zama-ai/tfhe-rs · GitHub
Listed 28 Oct 2026 · github.com · Web page
Zama’s Slab configuration defines Nebius single-H100, eight-H100 and two-L40 runner backends. Its generic four-GPU test backend uses the Nebius L40 Terraform script and lists the two-GPU Nebius L40 configuration first among its fallbacks.
Open original sourceCompany context
zama.orgZama develops open-source fully homomorphic encryption software, which allows applications to compute on encrypted data. Its Zama Protocol helps developers and financial institutions issue, transfer and manage confidential assets on public blockchains. Products include confidential tokens, encrypted USDC and USDT, yield vaults and confidential swaps, supporting uses such as payments, payroll, trading and asset tokenization. Its developer tools include TFHE-rs, the Concrete Python compiler and Concrete ML for machine learning on encrypted data.
Recent updates
11 updatesZama spin-off Raycash launches confidential stablecoin payments on Ethereum
Raycash, a stablecoin payments startup spun out of Zama, launched its Ethereum mainnet app in private beta. The app supports confidential USDC payments and private yield through Morpho.
The BlockZerion Wallet adds Zama confidential tokens
Zerion integrated Zama confidential tokens on Ethereum, allowing users to manage encrypted positions alongside public holdings. Users can reveal their balances through signed, time-limited permits.
ZamaZama opens confidential Merkl reward claiming
Participants in Merkl campaigns on Zama’s confidential vaults can now claim cZAMA rewards in the Zama app. Payouts use confidential transfers to preserve the privacy of positions and reward amounts.
Zama