Hook
Thirty months of development. Three million dollars. Zero published benchmarks.
That ratio is the only hard datum the Soda Labs funding announcement actually contains. The project โ a privacy-computing middleware layer โ closed a $3 million seed round led by NextBlock. The accompanying text states the team has spent roughly two and a half years engineering a hybrid stack that fuses garbled circuits (GC) with secure multi-party computation (MPC), runs on standard cloud CPUs, and preserves connectivity to existing public blockchains. Each of those four claims is technically falsifiable. None of them has been falsified, because none of them has been tested in public view.
I have spent enough time inside cryptographic review cycles to know what a two-and-a-half-year development timeline with no benchmark disclosure usually means. Sometimes it means a team that is slow, rigorous, and uninterested in marketing. Sometimes it means a team with a working prototype and a broken understanding of how to communicate it. And sometimes โ more often than the industry likes to admit โ it means a system that performs acceptably on a benchmark the team selected for itself and degrades the moment anyone asks it to compute something adversarial. You cannot distinguish these three cases from a press release. You can only distinguish them from code, gas metrics, and an audit report. That gap is what this article is about, and it matters more in privacy computing than almost anywhere else.
Context
To evaluate what Soda Labs claims, you must first map the terrain it is entering. Privacy computation in a blockchain context is not one technology. It is at least four, and they are not interchangeable.
The first is fully homomorphic encryption (FHE) โ the ability to compute directly on ciphertext. It is the theoretical maximum: the data never decrypts, so the trust model reduces to the hardness of the underlying lattice problem. The cost is performance. FHE operations run orders of magnitude slower than plaintext, and bootstrapping โ the procedure that refreshes noise in a ciphertext โ is computationally brutal. Projects like Zama and Fhenix have spent years chipping at this wall, and the honest framing is that they have made FHE usable for narrow workloads, not fast in any general sense.
The second is the trusted execution environment (TEE) โ hardware-isolated computation, typically Intel SGX or an equivalent. It is fast because it is not really cryptography; it is a hardware trust assumption. Secret, Oasis, and Phala built significant ecosystems on this premise. The problem is that the premise leaks. SGX has a documented history of side-channel vulnerabilities, and every one of them converts a "trusted" enclave into a partially observable one. The system's unintended consequences arrive not through the cryptography but through the silicon underneath it.
The third is zero-knowledge proofs (ZK) โ proof systems that let a prover demonstrate a statement is true without revealing the witness. ZK is magnificent for verifiable computation, but it is not a general privacy engine. It proves things about data; it does not typically let multiple parties jointly compute over hidden inputs in a straightforward way.
The fourth is secure multi-party computation (MPC), and within it, the specific primitive Soda Labs leans on: garbled circuits, also called Yao's protocol. The idea is elegant. Two parties want to compute a function without revealing their inputs. You encode the function as a Boolean circuit, "garble" each wire with encryption keys, and let one party evaluate the garbled circuit against keys for their own input while the other party supplies keys for theirs. Neither learns the other's input. It is one of the oldest and most studied constructions in the field.
Soda Labs' claimed innovation is the combination: use GC for Boolean-circuit portions, MPC for arithmetic and multi-party portions, and run the whole thing on commodity cloud CPUs. This is a deliberate positioning choice, and it is the single most interesting fact in the entire announcement.
Core
Let me be precise about what GC-MPC buys you and what it costs, because the trade-offs are where the project either succeeds or quietly dies.
Garbled circuits are communication-bound, not compute-bound. This is the crucial property. When you evaluate a garbled circuit, the dominant cost is the transfer of garbled wire labels between parties, and that cost scales with the size of the circuit โ specifically with the number of AND gates, which require the expensive "garbled table" construction. A circuit with n gates implies roughly O(n) communication with a nontrivial constant per gate. This is why GC is spectacular on small Boolean functions (comparisons, equality tests, fixed-range checks) and increasingly painful as the circuit grows.
This gives us a hard technical question that the announcement does not answer. Soda Labs claims "programmable privacy." The word programmable is doing enormous work. A fixed, small circuit is exactly the regime where GC-MPC shines. A general-purpose programmable privacy layer implies the ability to handle arbitrary logic โ which, in the limit, means arbitrary circuit size โ and that is precisely where GC's communication overhead becomes the dominant bottleneck. The claim and the primitive are in tension. Either the project has found a circuit-decomposition trick that keeps per-transaction circuits small, or it has not yet confronted the scaling wall. Based on my experience reviewing circuit-based systems, the honest answer is usually the second until proven otherwise.
The "standard cloud CPU" claim is a positioning statement, not a performance claim. Read it carefully. It says the system does not require specialized hardware. That is a direct contrast with two rival camps. It contrasts with FHE, which is so compute-intensive that acceleration hardware is a live research topic. It also contrasts with TEE, which literally requires specialized hardware and a trust assumption in the manufacturer. By saying "commodity CPU," Soda Labs is signaling: we reject the hardware trust model of TEE, and we reject the compute burden of FHE, and we deliver something in between.
That is a coherent strategic position. It is also unfalsifiable as stated, because "runs on a standard CPU" is compatible with a throughput of ten operations per second and a throughput of ten thousand. The absence of any TPS, latency, or throughput figure is not a minor omission. In a system whose entire differentiation is a performance-versus-trust trade-off, withholding the performance number removes the one variable that would let anyone evaluate the trade.
The connectivity choice reveals the architecture. Soda Labs says it maintains connectivity to existing public chains. This is the highest-confidence inference available from the entire announcement. It means the project is not building a privacy L1 that competes for liquidity and users. It is building a privacy co-processor โ a middleware layer that chains call out to when they need private computation. This lowers adoption friction enormously. A chain does not need to migrate users; it needs to integrate an SDK. It is the same architectural pattern that data-availability layers adopted: don't replace the settlement layer, bolt onto it.
And here is where a parallel from my own work becomes relevant. I have argued for years that most rollups do not generate enough data to justify a dedicated DA layer โ that the DA narrative was over-provisioned relative to actual demand. The same caution applies to privacy co-processors. A co-processor only earns its keep if the demand for its specific capability is large and persistent. If only a small fraction of on-chain activity needs true private computation, then a general-purpose privacy co-processor is infrastructure searching for a workload. The architecture's unintended consequences are that it is technically sound, cleanly integrated, and idle.
The cryptographic primitives chosen are conservative, and that is a genuine strength. The announcement references AES and SHA-256 as the underlying standards. This matters. A GC-MPC system that reduces its trust assumptions to AES and SHA-256 is inheriting the best-studied primitives in existence. It avoids the exotic-assumption risk that newer constructions carry โ the risk that a lattice parameter was chosen wrong, that a novel hardness assumption collapses. In a field where "audit passed, reality failed" is a recurring genre, conservative primitive selection is the correct instinct. I want to credit it explicitly, because the rest of this analysis is skeptical.
The maturity picture, then, is this. The primitive selection is sound. The architectural positioning is coherent. The circuit-scaling question is unanswered. The performance data is absent. And the development timeline โ two and a half years โ is long enough to be either reassuring or alarming depending on facts we do not have. In cryptographic privacy, "six months away" is the industry's default perpetual state; a two-and-a-half-year runway with no mainnet and no audit is consistent with a team grinding on genuinely hard engineering, and equally consistent with a team that has not solved its core problem.
Contrarian
Here is the counter-intuitive angle, and it has nothing to do with whether the cryptography works.
The largest risk in this announcement is not technical. It is the information black box. Consider what is missing. No team disclosure โ no names, no cryptographic credentials, no prior publications. No audit report. No testnet or mainnet status. No performance benchmarks. No token model. No valuation. No integration partners. No legal jurisdiction. For a project whose entire value proposition rests on cryptographic claims, the absence of a named cryptographic team is the most consequential gap of all. GC-MPC is frontier work. Whether it succeeds depends almost entirely on whether the people building it have deep expertise in the specific failure modes of garbled-circuit constructions. You cannot evaluate that from an anonymous press release.
There is a second, subtler blind spot. "Standard cloud CPU" is simultaneously a strength and a moat problem. If the system genuinely runs on commodity hardware with no specialized acceleration, then the barrier to a competitor replicating it is low. There is no hardware lock-in, no proprietary silicon, no accumulated infrastructure advantage. The moat, if one exists, is purely the engineering quality of the implementation โ and engineering quality is exactly what cannot be assessed without code. A claim of "no specialized hardware" is a claim that the defensibility lives entirely in the software, which is the easiest kind of defensibility to erode.
And there is the regulatory tail risk, which deserves blunt treatment. Privacy infrastructure sits in a structurally hostile position relative to anti-money-laundering regimes. The Tornado Cash precedent demonstrated that a privacy tool can be sanctioned, and its developers prosecuted, on the theory that the tool facilitates illicit flows regardless of the intent of its authors. A general-purpose "programmable privacy" layer inherits this exposure directly. The mitigating story is real โ privacy computation can be framed as a compliance tool, useful for GDPR-style data protection and institutional data collaboration, in contrast to privacy coins which exist primarily to obscure transfers. But that framing is a narrative the project must actively construct, and it is fragile. If Soda Labs is ever characterized as anonymity infrastructure rather than data-protection infrastructure, the regulatory posture flips instantly.
Finally, the competitive reality. Privacy computation is one of the most crowded capital sinks of the current cycle. FHE, TEE, MPC, and ZK teams are all well-funded and all racing. A $3 million seed round is small in this field โ a reconnaissance bet, not a conviction bet. That does not doom the project. It does mean that whatever it builds, it will build with fewer resources than the teams it is competing against.
Takeaway
Strip away the framing and you are left with a clean, testable thesis: a hybrid GC-MPC co-processor, built on conservative primitives, positioned against both FHE's compute cost and TEE's hardware trust, running on commodity infrastructure and integrated as middleware rather than a competing chain. That thesis is interesting and, on the cryptographic merits, plausible. It is also completely unverified.
My forward-looking judgment is this: the next twelve months of disclosures will matter far more than the funding itself, and the specific signal to watch is not the token โ it is the circuit-size ceiling. A team confident in its architecture will publish a benchmark showing its per-transaction circuit stays bounded. A team that has not yet solved the scaling problem will publish prose instead of numbers. Watch which one arrives. And watch whether the team ever names itself, because in frontier cryptography, anonymity is not modesty โ it is the absence of accountability for the one thing that determines whether this works at all.
A final question, the one the industry keeps avoiding: if a privacy co-processor runs on commodity hardware, uses standard primitives, and publishes no benchmarks, then what exactly is the $3 million buying โ a cryptographic breakthrough, or two and a half more years of the same silence?