Zama builds FHE libraries. H33 runs FHE in production at 2.21 million operations per second. Four FHE engines, ZK-STARK proofs, and Dilithium signatures — delivered as a single REST API call. No parameter tuning. No cryptography PhD required.
Fully homomorphic encryption (FHE) is encryption that lets a server run computation over encrypted data (ciphertext) without decrypting it. The plaintext never appears on the machine doing the work, and the answer comes back encrypted. Both Zama and H33 implement this primitive; H33's FHE engines are a supporting expression of the FHE Platform, which computes on encrypted data. This page is a comparison of how each delivers it.
FHE exists to protect data in use — the one state where encryption traditionally had to be dropped so a program could read the values. FHE keeps data encrypted through the computation itself.
FHE is not the same as a zero-knowledge proof. FHE runs computation over encrypted data (ciphertext) without decrypting; the ZK Platform proves a statement is true without revealing the underlying data. TFHE (Zama) and BFV/CKKS (H33) are open FHE schemes from the broader research community; both companies implement them independently and neither invented them.
How an open-source FHE library compares to a production FHE platform with integrated ZK and post-quantum cryptography.
| Feature | Zama | H33 |
|---|---|---|
| FHE Engines | 1 (TFHE) — single scheme for all workloads | 4 (BFV-128, BFV-256, CKKS, FHE-IQ auto-router) — automatic engine selection in <500ns |
| Per-Operation Latency | 124ms (64-bit add) | 35.25µs per auth (3,200x faster) |
| Throughput | 189K bootstraps/sec (8x H100 GPU) | 2.21M auth/sec (CPU only, Graviton4) |
| ZKP Verify Latency | 123–467ms | 0.059µs (2,000,000x faster) |
| GPU Required | Yes — H100 coprocessors for performance | No — CPU-only (ARM Graviton4) |
| Hardware Cost | ~$15K/mo per GPU coprocessor | ~$2/hr spot instance |
| Post-Quantum Signatures | None | Dilithium + Kyber + FALCON + SPHINCS+ |
| ZKP Security | Not PQ-secure | SHA3-256 STARK — post-quantum secure |
| Side-Channel Protection | “Not yet implemented” | AI agent (1.14µs real-time detection) |
| Total Products | 5 | 38 |
| Pricing Model | Patent license for commercial use, contact sales | Self-service API, $0.001/auth at scale, BotShield free tier |
| Compliance | None | SOC 2 + HIPAA + ISO 27001 pending + HATS |
| Deployment Model | Self-hosted library (Rust / Python) | Managed REST API — one call, full stack |
| Python FHE Compiler | Concrete (TFHE only) | H33-Compile (4 engines + FHE-IQ auto-routing) |
Zama's TFHE takes 124ms for a 64-bit add. H33 completes a full biometric authentication — FHE match, ZK proof, and Dilithium signature — in 35.25µs. That is 3,200x faster per operation, on CPU only, with no GPU infrastructure required.
Zama gives you building blocks. H33 gives you a production service. No parameter tuning, no noise budget management, no key rotation infrastructure to build. Send a request, get cryptographically verified results. Deployed and monitored on AWS Graviton4 with sub-millisecond FHE batch latency.
Zama provides FHE computation only — no signatures, no ZKP, no side-channel protection. H33 combines lattice-based FHE, SHA3-256 STARKs, Dilithium + Kyber + FALCON + SPHINCS+ signatures, and a real-time AI side-channel agent into a single API call. Every layer is post-quantum secure.
Zama requires 8x H100 GPUs (~$15K/month) to reach 189K bootstraps/sec. H33 achieves 2.21M auth/sec on a single Graviton4 spot instance at ~$2/hr. That is 11x higher throughput at a fraction of the cost — with no GPU procurement, no CUDA dependencies, no driver updates.
FHE with a library vs. FHE with an API — the developer experience difference.
# Zama: manage parameters, keys, circuits from concrete import fhe @fhe.compiler({"x": "encrypted"}) def match_biometric(x): return (x - template) ** 2 circuit = match_biometric.compile(inputset) circuit.keys.generate() encrypted = circuit.encrypt(biometric_data) result = circuit.run(encrypted) decrypted = circuit.decrypt(result) # You manage: parameters, noise, keys, # deployment, scaling, monitoring...
// H33: one call, full cryptographic stack const result = await h33.authenticate({ biometric: capturedTemplate, securityLevel: 'h33-128', mode: 'standard' }); // result.verified → true / false // result.attestation → Dilithium-signed proof // result.zkProof → ZK-STARK verification // result.fheEngine → 'BFV' (auto-selected) // // FHE batch: ~937µs (32 users) // ZK proof + PQ attestation included // No parameters, no keys, no circuits
Head-to-head benchmarks on production workloads. All H33 numbers from Graviton4 c8g.metal-48xl (96 vCPUs, CPU only).
| Metric | Zama | H33 |
|---|---|---|
| Per-Operation Latency | 124ms (64-bit add) | 35.25µs per auth (3,200x faster) |
| Throughput | 189K bootstraps/sec (8x H100 GPU) | 2.21M auth/sec (CPU only) |
| ZKP Verify | 123–467ms | 0.059µs (2M x faster) |
| FHE Batch (32 users) | N/A (single-user) | 937µs |
| Dilithium Attest | N/A (no PQ sigs) | 189µs (1 per batch) |
| GPU Required | Yes (H100 coprocessors) | No (ARM Graviton4 CPU) |
| Hardware Cost | ~$15K/mo per coprocessor | ~$2/hr spot instance |
| Benchmark Variance | Not published | ±0.71% (120s sustained) |
Quantum computers will break RSA, ECDSA, and classical ZKPs. Only H33 secures every cryptographic layer against that threat.
| Layer | Zama | H33 |
|---|---|---|
| FHE | ✓ TFHE (lattice-based) | ✓ BFV + CKKS (lattice-based) |
| Digital Signatures | ✗ None | ✓ Dilithium + Kyber + FALCON + SPHINCS+ |
| Zero-Knowledge Proofs | ✗ Not PQ-secure | ✓ SHA3-256 STARK (PQ-secure) |
| Side-Channel Protection | ✗ “Not yet implemented” | ✓ AI agent (1.14µs real-time) |
| Key Exchange | ✗ Not included | ✓ ML-KEM (Kyber) hybrid |
Both platforms offer encrypted biometric processing — but the production readiness gap is significant.
Both platforms address encrypted computation on blockchain — with very different approaches and throughput.
Running machine learning models on encrypted inputs without decryption.
@h33.compile, get an FHE circuit with 4-engine auto-routing via FHE-IQZama focuses on FHE tooling. H33 is a full-stack production platform.
| Dimension | Zama | H33 |
|---|---|---|
| Total Products | 5 (TFHE-rs, Concrete, Concrete ML, fhEVM, fhEVM coprocessor) | 38 products (FHE engines, ZK verifiers, PQ signatures, biometrics, blockchain, detection, storage, video, search, and more) |
| Pricing Model | Patent license required for commercial use; contact sales for pricing | Self-service API; $0.001/auth at scale; BotShield free tier |
| Blockchain Pricing | $0.005–$1.00 per on-chain FHE operation | Credit-based pricing with volume discounts |
| Free Tier | Open-source libraries (BSD), commercial license required for production | 1,000 free operations/month, no credit card |
| Compliance | None | SOC 2 + HIPAA + ISO 27001 pending + HATS certified |
What FHE is, what it is not, and where it sits relative to ZK and H33-74.
| What does FHE compute? | Computation run directly over encrypted data (ciphertext) — the server returns an encrypted result and never sees plaintext. This is the primitive owned by the FHE Platform; H33's engines are a supporting expression of it. |
| What does it not do? | It does not prove a statement without revealing data (that is zero-knowledge), and it does not decrypt during computation. |
| How is it different from ZK? | FHE runs computation over encrypted data without decrypting; the ZK Platform proves a statement is true without revealing the underlying data. |
| How does H33-74 fit in? | FHE produces the encrypted computation; H33-74 anchors the evidence of it as a portable post-quantum attestation. Neither owns the other. |
| When should you choose FHE? | When data must stay encrypted while it is processed — encrypted search, private ML inference, biometric matching, cross-party analytics. Not when you only need to prove a fact (use ZK) or when data can safely be decrypted in a trusted enclave. |