OpenGradient

OpenGradient is a New York company - legally Vanna Labs, Inc., which is a detail we will get to - building what it calls the compute layer for verifiable AI. The plain-English version: it runs machine-learning models on a decentralized network of GPUs and hardware enclaves, and then attaches a cryptographic proof to every inference. The proof lets a downstream application check exactly what model ran, on what input, and what came back out. If that sounds modest, consider that no major AI provider offers it, and consider how much money is about to be moved by software that no one can audit.

The setup is worth dwelling on because it is a genuinely different shape from most of what gets called "AI infrastructure." OpenGradient is not really trying to be a chatbot, and it is not exactly a blockchain either. It describes itself as a specialized AI coprocessor: applications, agents, and blockchains outsource their heavy compute to a dedicated network, and get back an answer plus a proof. The chain part - it is EVM-compatible, so it speaks the same language as Ethereum tooling - exists mostly to settle those inferences and record who did what.

The technology underneath has a name only an engineer could love: Heterogeneous AI Compute Architecture, or HACA. What it means is that OpenGradient stitches together ordinary decentralized GPUs and specialty accelerators, plus trusted execution environments (the "hardware enclaves" the keyword list keeps mentioning), into one network. The GPUs do the math. The enclaves make the math trustworthy. The chain writes it all down. It is, in a sense, an elaborate machine for turning "trust me" into "verify me," which is the entire product.

There is a reasonable question here, which is: who asked for this? OpenGradient's answer is that the AI stack is consolidating around a handful of closed providers with no audit capability, and that this is fine right up until it is not - until an agent moves your money, signs a decision, or files a report, and someone needs to prove it used the model it claims to have used. In finance, that someone is usually a regulator. The company's founders come from exactly the industries where "prove it" is a load-bearing phrase.

Which brings us to the founders. Matthew Wang, the CEO, is a former quantitative engineer at Two Sigma - the kind of hedge fund where models are the product and being wrong is expensive - with earlier stops at Google, Meta, and, improbably, NASA. His co-founder and CTO, Adam Balogh, ran the AI platform at Palantir Technologies, a company that has spent two decades building software for customers who require an audit trail as a condition of doing business. If you were going to design two people to be personally annoyed by unverifiable AI, you would design these two.

The company emerged in October 2024 with an $8.5 million seed round led by a16z crypto's startup accelerator, which is a tidy way of saying that the most prominent crypto-venture firm looked at "AI you can audit" and decided to write a check and provide a desk. In April 2026 OpenGradient announced $9.5 million in total funding, adding names like Coinbase Ventures, SV Angel, Foresight Ventures, and the ecosystem funds of NEAR and Celestia. The angel list is a who's who of people who have opinions about decentralization: Balaji Srinivasan, the ex-Coinbase CTO; Illia Polosukhin, a co-founder of NEAR; and Sandeep Nailwal, a co-founder of Polygon.

What have they built with it? More than a slide. The network reports over two million users, more than two million verifiable inferences processed, and over 500,000 cryptographic proofs generated. The Model Hub - think of it as an app store where the apps are AI models and everything is hosted on-chain - holds more than 2,000 models from over 100 developers. The company says it has six active revenue streams, which for a company of roughly eleven people is either impressive discipline or a lot of hats.

The product line sprawls in the way that early-stage crypto-AI companies tend to. There is MemSync, which stores your AI's memory in an encrypted vault you control rather than on someone's server. There is OpenGradient Chat, private chat infrastructure. There is BitQuant, an AI quant analyst that does DeFi. There is even twin.fun, which lets you talk to AI digital replicas of influencers, and which is either a distraction or a very on-brand demo of user-owned models, depending on your mood. The through-line is ownership: your data, your model, your proof.

The obvious risk is the obvious risk for everything in this category. Verifiable, decentralized AI is a crowded and unproven field - Gensyn, Ritual, Bittensor, and others are all circling variations of the same idea - and "cryptographically provable inference" is a harder engineering problem than a marketing page suggests. Proofs cost compute. Enclaves have their own trust assumptions. And the incumbent it is arguing against, the opaque cloud endpoint, is winning right now precisely because it is cheap and it works. OpenGradient is betting that "unverifiable" will one day sound as reckless as "unaudited," and that when it does, the receipts will matter.

The legally-registered-as-Vanna-Labs detail is a nice period on all this. The company that wants every AI answer to carry a verifiable name still trades, on Crunchbase, under a different one than the one on the door. It is a small reminder that identity and provenance are messy even for the people trying to fix them - which is, come to think of it, roughly the whole point.