Key takeaways
- 0G Labs addresses data availability bottleneck by achieving 50 gigabytes per second throughput compared to current systems operating at 1-10 megabytes per second.
- The modular architecture separates data storage and publishing lanes to enable horizontal scalability while reducing costs by approximately 100x versus competitors like Celestia.
- Current blockchain infrastructure requires tens of gigabytes per second data availability capacity to support onchain AI model training at scale.
- 0G Labs co-founders bring expertise from Microsoft Research, MIT, and Conflux with backgrounds in distributed storage, AI algorithms, and scalability.
- Decentralized AI systems can achieve performance and cost parity with centralized alternatives like OpenAI when data availability infrastructure is properly optimized.
From teenage programmer to 0G Labs
Ashton Addison: What is your background, and what led you to start 0G Labs?▶ 00:22
Michael Heinrich: My first run-in with technology happened when my family moved to Silicon Valley when I was around 13. I was a bit bored at high school, so I'd show up at my dad's office after school and use the fast internet and technology SAP Labs had. One day a manager asked if I wanted to make myself useful and learn to program. I had no idea how, so he sent me to a Visual Basic boot camp. Four weeks later I was the youngest programmer there. They couldn't pay me in cash, so they gave me free hardware, and I was in heaven. I stayed in engineering through high school and college, worked at Microsoft in technical product management, then moved to the business side. I worked at Bridgewater Associates on the portfolio construction team, then went to Stanford and built my first Web2 company, Garden. We scaled it to about 650 employees and about 100 million in contracted ARR, and it became a unicorn and a top Y Combinator company. In late 2022 a Stanford classmate called me about two co-founders of Conflux who wanted to start something new at a more global scale. Six months of co-founder dating later, I'd concluded these were the best engineers I've ever worked with and that we needed to start something.
Ashton: What problem did you see that needed solving?▶ 03:43
Michael: We looked for the intersection of what we're passionate about, what's within our domain expertise, and what the space needs to be pushed forward. After many conversations with projects like Polygon, Arbitrum and various AI projects, it became clear that data availability was a key unlock. The standard was somewhere around 1 to 10 megabytes per second of capacity. If you want to do on-chain AI model training, especially for very large models, you need to think in tens of gigabytes per second. Our core insight was to segment data into a data storage lane and a data publishing lane, which gives the system horizontal scalability. Each node on our network can generate about 10 megabytes per second, and with about 5,000 nodes in a consensus layer we can reach about 50 gigabytes per second. That also lowers costs by about 100x compared with systems like Celestia and EigenDA. The point is that we can build in a decentralized context with the same performance and costs as in a centralized one. The only limits left are physical ones, like the speed of light.
Shouldn't society dictate what's right for itself? How do we become part of that governance process and make AI a public good?
Data availability, explained
Ashton: Can you explain data availability for people who don't know how blockchains work under the hood?▶ 07:05
Michael: Take OpenAI. It's all on centralized servers: you make a request, and a lot of GPUs strung together in a centralized setting run inference tasks in the background. A blockchain is a completely distributed system, often permissionless, where anybody can join and generally become a verifier in a proof-of-stake system. Ethereum is what's known as a monolithic blockchain. It has an execution layer, a consensus layer, a data availability layer and a settlement layer, and most transaction costs go to data availability, because that layer verifies that a block has actually been published on the network. By pulling that layer out and making it modular, we can hyper-optimize it, so publishing data is much cheaper and throughput is much higher. For an AI use case, an inference task on the execution layer goes to us as the DA layer. We pull the right serving infrastructure together, find the right GPU locations, load the stored model into memory and serve it to the end user. The back-end process is quite different, but the user experience ends up the same.
Ashton: Could this bring down Ethereum transaction fees?▶ 16:07
Michael: That's the whole idea behind layer twos like Optimism and Arbitrum. You offload execution to a different layer and batch transactions together. If those layers implement a better DA system, you're talking about fractions of a fraction of a penny per transaction. That allows completely different possibilities, like fully on-chain gaming and fully on-chain AI, because fees are almost gasless. That's the next evolution of blockchain. As layer twos, layer threes and app chains adopt DA solutions like ours, you can create completely different experiences.
Open models and AI as a public good
Ashton: Is a decentralized alternative to OpenAI viable today?▶ 09:47
Michael: Today you can use open-source models like Llama 3 or Mistral, which rival OpenAI in capability. OpenAI is a fully closed-source system: you can only use the ChatGPT model within their framework and fine-tuning tools. The issue comes when you imagine an AI agent steering a traffic control system or running a big administrative process, like matching candidates to jobs. Inside OpenAI's framework you don't know what the model was trained on, where it sits, which version you're getting or what weights and biases it has. You have no control over that. With an open-source model, which you can train on your own data, you can know at any point in time what version of the model you're running and what decisions it's making. It's completely publicly verifiable. That's the power of combining AI and blockchains: full provenance, full verifiability, total transparency. You then have an open AI, pun intended.
Ashton: As AI grows, will there be more debate about bias and who decides what's true?▶ 11:46
Michael: Exactly. Who decides what bias to remove from the training data, and who decides how that model is aligned with human interest? That all happens within one company, so how can society have insight into what that company is doing? I recently read that OpenAI's head of alignment quit because he wasn't comfortable with the company's practices. We're trusting one big entity to do what's right for an entire society. Shouldn't society dictate what's right for itself? How do we become part of that governance process and make AI a public good? That's what we stand for.
You can know at any point in time what version of the model you're running and what decisions it's making.
Roadmap and what comes next
Ashton: How long has 0G Labs been in the works, and where are you now?▶ 13:18
Michael: I went full-time in May last year, so it's been almost exactly a year. We came out of stealth a couple of months ago and announced our 35 million pre-seed. We announced our public testnet a few weeks ago, and we'll announce public testnet V2 quite soon. Then we move to mainnet, hopefully sometime in July. On the testnet you can run your own storage nodes and validators, and deploy different chains that use the data availability property. In V2 we'll improve some of the storage mechanisms and the interface, and we'll move from a data availability sampling technique to a quorum-based technique for speed and performance gains.
Ashton: What do you expect to see in AI over the next few years?▶ 17:16
Michael: It's always hard to predict the future, but over the next five years I think we'll see more specialized agents for particular use cases. In crypto, instead of waking up and having to go through a particular DEX and bridge to buy a coin on a particular chain, you can just tell your agent to go buy it for you. Outside blockchain, there will be full self-driving from Tesla. I've tested some of it and it's pretty amazing, though it still has some way to go. We'll see AI agents interacting with the real world, and then humanoid robots with AI models in them. As a society we need to make sure safety mechanisms are built in so there is full alignment with human interest. I see a world where AI is part of the fabric of society, whether digital or physical.
Ashton: What needs to happen before mainnet, and how can people get involved?▶ 19:11
Michael: The technology team is working at 100% to get there as quickly as possible. We're hoping for the end of July for mainnet phase one, which gives you the basic features and a stable environment around the storage and data availability components. You can use our testnet right now to run a storage node or validator node, or deploy app chains and layer twos that use our DA property. You can follow us on X and see the latest updates there. We also recently launched an accelerator together with One Piece Labs, and applications are live until mid-June. If you want to build as part of the biggest AI-meets-blockchain community in the world, definitely apply.
This interview has been edited for length and clarity. Recorded May 2024.

