Both companies build on fully homomorphic encryption. That's where the similarity ends. Zama is an FHE library company building toward blockchain confidentiality. H33 is a post-quantum security infrastructure company delivering FHE, ZK proofs, and post-quantum signatures as a production API. This comparison covers every dimension—12 categories, with honest assessments of where each company leads.
Table of Contents
FHE Engines
Zama
1 scheme: TFHE. Boolean gates and integer arithmetic on encrypted data. TFHE-rs is the core Rust library. Concrete compiles Python programs to TFHE circuits. Single approach for all workloads—general-purpose, gate-by-gate bootstrapping.
H33
4 engines: BFV-128, BFV-256, CKKS, FHE-IQ auto-router. BFV handles exact integer arithmetic and biometric matching. CKKS handles approximate float arithmetic and ML inference. FHE-IQ auto-selects the optimal engine in <500 nanoseconds based on workload, security tier, and hardware. Developer makes one API call—routing is automatic.
| Dimension | Zama | H33 |
|---|---|---|
| FHE schemes | TFHE (1) | BFV-128, BFV-256, CKKS, FHE-IQ (4) |
| Integer arithmetic | Yes (bootstrapped) | Yes (BFV, no bootstrapping needed) |
| Float arithmetic | Limited | CKKS (native approximate) |
| Auto-routing | None | FHE-IQ (<500ns decision) |
| Developer experience | Learn TFHE paradigm | One API call, engine selected automatically |
Performance
These are different workloads. Zama does general-purpose FHE computation. H33 does production authentication pipelines. But the performance gap is structural—and the numbers speak for themselves.
| Metric | Zama | H33 |
|---|---|---|
| Per-operation latency | 124ms (64-bit add, CPU) | 38.5µs per auth |
| Throughput | 189K bootstraps/sec (8× H100 GPUs) | 2.17M auth/sec (single Graviton4) |
| ZKP verification | 123–467ms | 0.059µs cached STARK |
| GPU required | Yes (H100 for competitive perf) | No |
| Hardware cost | ~$120K/yr (8× H100 cluster) | ~$2/hr spot (~$17K/yr) |
| Speed ratio | H33 is 3,200× faster per operation on CPU alone | |
Why the gap is structural
H33's Montgomery NTT, Harvey lazy reduction, NTT-domain fused inner products, and SIMD batching (32 users per ciphertext) are purpose-built optimizations that TFHE's gate-by-gate bootstrapping architecture cannot match. TFHE bootstraps after every operation. H33's BFV pipeline uses shallow circuits with no bootstrapping needed—the entire auth completes in a single forward pass.
Zero-Knowledge Proofs
Zama
ZKPoK only—Zero-Knowledge Proof of Knowledge. Proves that a ciphertext was correctly formed. Does not prove what was computed or whether policies were followed. Generation time: 6.9 seconds for a single FheUint64 encryption proof.
H33
Two STARK engines. Lookup STARKs for precomputable verification (0.059µs cached). AIR STARKs (Algebraic Intermediate Representation) for open-ended computation proofs. Both hash-based (SHA3-256), both post-quantum secure, no trusted setup.
| Capability | Zama | H33 |
|---|---|---|
| Proof type | ZKPoK (encryption correctness) | Lookup STARK + AIR STARK + STARK-IQ auto-router |
| Auto-routing | N/A (single proof type) | STARK-IQ selects Lookup vs AIR automatically per workload |
| Proof generation | 6.9 seconds | 0.059µs (cached lookup) |
| Post-quantum secure | No (Zama acknowledges this) | Yes (SHA3-256 hash-based) |
| Trusted setup | No | No |
| Proves computation correctness | No | Yes (AIR STARK) |
| Proves policy compliance | No | Yes |
Post-Quantum Signatures
Zama
None. Zero signature implementation. Relies on host blockchain's native signatures—Ed25519 for Solana, ECDSA for Ethereum. Both are quantum-vulnerable. Zama's FHE is lattice-based (inherently post-quantum), but their ZKP is not PQ-secure (they acknowledge this).
H33
Complete PQ signature stack. Dilithium (ML-DSA-65/87, FIPS 204). Kyber (ML-KEM-768/1024, FIPS 203). FALCON-512. SPHINCS+. H33-3-Key nests three independent signature families (Ed25519 + Dilithium + FALCON) with temporal binding. 291µs per Dilithium attestation.
| Signature | Zama | H33 |
|---|---|---|
| Dilithium (ML-DSA) | None | ML-DSA-65/87 (FIPS 204) |
| Kyber (ML-KEM) | None | ML-KEM-768/1024 (FIPS 203) |
| FALCON | None | FALCON-512 |
| SPHINCS+ | None | Yes |
| Nested hybrid | None | H33-3-Key (3 independent families) |
| Attestation latency | — | 291µs (Dilithium sign+verify) |
Blockchain & DeFi
Zama
fhEVM for Ethereum confidential smart contracts. Zama Protocol mainnet live. Confidential ERC20 transfers. ~20 TPS on CPU, targeting 500–1,000 TPS with GPU acceleration. $150M funded, with partnerships across Ledger, Fireblocks, and OpenZeppelin.
H33
10 Solana smart contracts spanning identity, authentication, DeFi, and document management. Soulbound biometric NFTs—FHE-encrypted biometrics bound to non-transferable tokens. Dilithium-verified on-chain transactions. 7-stage token economics with automated burn mechanism. Biometric SCIF DeFi wallet where the private key is protected by FHE—the wallet itself is post-quantum. Solana Shield: first post-quantum privacy layer for Solana.
| Dimension | Zama | H33 |
|---|---|---|
| Chain | Ethereum (fhEVM) | Solana |
| Primary use | Confidential transactions | Post-quantum identity + DeFi |
| Smart contracts | fhEVM library | 10 deployed contracts |
| TPS | ~20 CPU / ~500–1K GPU target | Solana-native throughput |
| User signatures | ECDSA (quantum-vulnerable) | Dilithium (quantum-proof) |
| Biometric wallet | None | SCIF DeFi wallet (FHE-protected key) |
ML on Encrypted Data
Zama
Concrete ML. Compile sklearn, PyTorch, and XGBoost models to FHE circuits. Encrypted LLM fine-tuning via LoRA. CIFAR10 classification: ~4 minutes per image at 88.7% accuracy. Strong tooling for data scientists who want to train on encrypted data.
H33
CKKS engine for encrypted float arithmetic—dot products, similarity search, ML inference. FHE-IQ auto-routes between BFV (exact) and CKKS (approximate) based on workload. AI-Blind: one API call, model processes encrypted data, never sees plaintext. Plus 3 native AI security agents—harvest detection (0.69µs), side-channel detection (1.14µs), crypto health monitoring (0.52µs).
| Capability | Zama | H33 |
|---|---|---|
| ML compiler | Concrete ML (Python → FHE) | No (API-first) |
| Framework support | sklearn, PyTorch, XGBoost | CKKS inference API |
| Encrypted inference | ~4 min/image (CIFAR10) | Sub-millisecond (biometric match) |
| Auto engine selection | No | FHE-IQ routes BFV/CKKS automatically |
| Security agents | None | 3 native agents (<1.2µs each) |
Biometrics
Zama
Research demo only. Single-user iris matching via FastAPI. Not production, not batched, not optimized.
H33
Production FHE biometric pipeline. 32 users per ciphertext via SIMD batching (4,096 slots ÷ 128 dimensions). Constant-time verification: ~967µs whether matching 1 or 32 users. Templates stored encrypted—256KB/user vs 32MB/user unoptimized (128× reduction). The server never sees the biometric. Ever.
| Metric | Zama | H33 |
|---|---|---|
| Status | Research demo | Production |
| Users per ciphertext | 1 | 32 (SIMD batched) |
| Verification latency | Seconds (unbatched) | ~967µs (constant-time) |
| Template size | Unoptimized | 256KB/user (128× reduction) |
| Server sees plaintext biometric | Yes (FastAPI demo) | Never |
Security Posture
| Certification / Capability | Zama | H33 |
|---|---|---|
| Side-channel protection | "Not yet implemented" (their README) | AI agent (1.14µs) |
| SOC 2 | None | In progress (100% via Drata) |
| HIPAA | None | Compliant |
| ISO 27001 | None | Pending |
| Standards body | None | HATS v1.0 (AI trustworthiness certification) |
| Security audit | Trail of Bits (blockchain only) | Full production audit |
Products
Zama 5 products
H33 38 products across 13 categories
Pricing
Zama
BSD-3-Clause-Clear license with mandatory patent licensing. Free for research only. Any commercial deployment requires negotiating a separate patent license with Zama's legal team—all Zama technologies are patented and the open-source license explicitly excludes commercial patent rights. On top of the license fee: blockchain protocol fees of $0.005–$1.00 per operation, plus self-hosted infrastructure costs of ~$15K/month per coprocessor operator. GPU acceleration (H100s) adds another ~$30K+/year per node.
H33
Self-service API. Credit-based pricing from $0.05/credit (free tier, 1,000 credits) down to $0.001/credit (Enterprise+, 3M credits/month). $0.001 per authentication at scale. BotShield free tier: 2,500 challenges/month, no card required.
| Dimension | Zama | H33 |
|---|---|---|
| Free tier | Research use (BSD-3-Clause-Clear) | 1,000 credits + BotShield 2,500/mo |
| Commercial license | Patent negotiation required | Self-service API key |
| Per-operation cost | $0.005–$1.00 (blockchain) | $0.001 per auth (at scale) |
| Infrastructure | ~$15K/mo per coprocessor (self-hosted) | Managed API (no infrastructure) |
| GPU costs | ~$120K/yr (8× H100) | $0 (CPU only) |
What Zama Does Better
Honesty matters more than marketing. Here's where Zama leads.
1. Python Compiler UX
Concrete lets data scientists write Python and get FHE. You write a function, decorate it with @fhe.compiler, and Concrete compiles it to an FHE circuit. That's a genuinely powerful developer experience for researchers who think in Python. H33-Compile (coming Q2 2026) does the same — but routes to 4 FHE backends instead of one, with automatic engine selection via FHE-IQ.
2. Ethereum Ecosystem Depth
fhEVM has a production mainnet, partnerships with Ledger, Fireblocks, and OpenZeppelin, and a well-funded protocol team. If you're building confidential smart contracts on Ethereum, Zama has the deepest integration. H33's blockchain play is Solana-native—different chain, different ecosystem.
3. Team and Funding
$150M raised. ~200 people. ~50% PhDs. CTO Pascal Paillier invented Paillier encryption. That's real academic and financial firepower. The depth of cryptographic research talent at Zama is world-class.
What H33 Does Better
- 3,200× faster on CPU, no GPU required. 38.5µs per auth vs 124ms per FHE operation. Single Graviton4 vs 8× H100 cluster.
- Complete PQ stack (FHE + ZK + Signatures + KEM) vs FHE only. Dilithium, Kyber, FALCON, SPHINCS+, and H33-3-Key nested hybrid signatures. Zama has zero post-quantum signatures.
- 38 products vs 5. 13 categories covering FHE, ZK, signatures, biometrics, key management, healthcare, fraud, device attestation, video, AI compliance, procurement, bot protection, and standards.
- Production biometrics vs research demo. 32 users per ciphertext, constant-time verification, 128× template size reduction. Server never sees the biometric.
- $0.001/auth vs $0.008–$0.80/operation. Self-service API key vs patent license negotiation. No GPU costs. No coprocessor infrastructure.
- 10 smart contracts + biometric SCIF wallet vs smart contract library. Soulbound biometric NFTs, Dilithium-verified transactions, 7-stage token economics, Solana Shield privacy layer.
- SOC 2 / HIPAA / HATS vs no compliance certifications. Production security audit, HIPAA compliant, SOC 2 in progress, HATS AI trustworthiness standard.
- Side-channel protection vs "not yet implemented." AI-powered side-channel detection in 1.14µs. Zama's own README acknowledges the gap.
- 2 STARK engines (Lookup + AIR) with auto-routing vs encryption-only ZKPoK. 0.059µs cached verification. Proves computation correctness, policy compliance, identity, and audit trails—not just encryption correctness.
- 4 FHE engines with auto-routing vs 1 scheme. BFV-128, BFV-256, CKKS, and FHE-IQ. Developer makes one call; the engine is selected in <500ns.
Conclusion
Zama is building the FHE library for blockchain confidentiality. H33 is building the post-quantum security infrastructure for everything else.
If you need to make Ethereum transactions private, Zama is purpose-built for that. They have the team, the funding, the Ethereum partnerships, and a Python compiler that researchers love.
If you need production-grade encrypted authentication, biometrics, fraud detection, compliance, key management, device attestation, or any security operation that needs to survive quantum—H33 has 38 products, 4 FHE engines, 2 STARK engines, and a complete post-quantum signature stack, all in one API call.
The single-sentence summary
Zama encrypts computation on Ethereum.
H33 secures everything else against quantum.
3,200× faster. CPU only. $0.001 per auth.
See the Numbers Yourself
Get your free API key and benchmark H33's full pipeline. 1,000 free credits, no credit card required.
Get Free API Key