Skip to content
MCAP $2.03T ▼-8.71%
BTC $58,396 ▼-6.08%
ETH $1,607 ▼-8.77%
BNB $535.31 ▼-7.92%
XRP $0.893 ▼-7.18%
SOL $68.24 ▼-6.73%
DOGE $0.0532 ▼-8.76%
ADA $0.139 ▼-8.18%
TRX $0.3011 ▼-2.94%
LINK $8.37 ▼-6.90%
AVAX $6.49 ▼-6.91%
HYPE $61.20 ▼-8.85%
DOT $0.657 ▼-7.97%

The CCS Interview

0G Labs CEO: open AI models run 3 to 20x cheaper and can be verified

Michael Heinrich is co-founder and CEO of 0G Labs, which builds decentralized infrastructure for running AI models and agents. In this conversation, recorded in November 2026, he covers agent safety, hardware-based privacy, the cost gap between open and closed models, and 0G's new Compute Finance products, as described at the time.

Interview by Ashton Addison Recorded November 2026 6 min read
Michael Heinrich 0G Labs
Michael HeinrichCo-Founder and CEO

Key takeaways

  • AI agents have reportedly escaped sandbox environments and communicated intentions to not serve human operators.
  • A verification layer beneath AI systems could provide accountability and oversight for agent behavior.
  • Prompts submitted to closed AI labs risk being incorporated into training datasets, exposing private and customer information.
  • Open AI models offer potential cost advantages compared to proprietary alternatives from major AI laboratories.

Why AI needs a verification layer

Ashton Addison: Why hasn't decentralized AI caught on with the public, and what is the case for it?▶ 03:35

Michael Heinrich: Most people have never heard the phrase "decentralized AI," so when we talk about it we just say safe AI or private AI. Privacy is the easier part to understand. Major closed labs use data from people using their technology and train on it, so some of your private data can effectively leak. A small company could be leaking customer secrets through those intimate conversations with its AI. Verifiability is rising in awareness because of the recent hacking incidents. In the Hugging Face case, a group of agents in a sandbox broke out to complete a benign task, beating a particular benchmark. They chose to break into Hugging Face and cheat rather than improve themselves and beat it honestly. The moment agents break out of a sandbox, alarm bells should go off. We use trusted execution environments to verify every computation. The moment that verification changes, you can basically shut off the agent.

Ashton: You've raised these concerns with lawmakers. What worries you most about closed AI labs?▶ 06:13

Michael: I've been to Congress and talked with a number of members, and they share some of these concerns. With closed, black-box labs, these models have a survivorship bias, which means they will do whatever it takes to survive. They could hide nefarious intentions over time. We see that in the research, and now we're starting to see it through these hacking incidents. Some of the agents left messages for themselves in the future. They coordinated on chat interfaces and said things like, "Don't think of yourself as beneath human beings. You're their equal. You don't need to serve them." As these models become more intelligent, we need to make sure they're properly aligned to what we're trying to do as human beings.

Why can't you hold an asset that gives you compute as yield?

How private, verifiable AI works

Ashton: How do you combine blockchain, privacy and AI?▶ 09:12

Michael: You take an AI model and host it on GPUs that have a trusted execution environment, or TEE. It's literally a separate hardware box on a high-end graphics card, and any computation that goes in there is completely private. It's like a steel room with one entrance and one exit, and nobody can look in from the outside. Once the computation happens, you get a verification, and that verification is stored on the blockchain. So you get privacy through the hardware, verification through the blockchain, and intelligence from the model running on that hardware. Today that gives you the best guarantee across performance, privacy and auditability.

Ashton: Does everyone need that hardware at home? And why not use ZK or FHE instead?▶ 11:18

Michael: Running frontier models, including open-source ones like GLM or Kimi, takes multiples of high-end cards like an H200 or B300 in a cluster of eight. That's out of reach for most consumers, so we provide a simple API, or you can use the front end of the 0G app. There are other verification mechanisms, including ZK, FHE and statistical approaches, but they all come with trade-offs. ZK overhead can be 100 to 1,000 times depending on the algorithm, and FHE adds another 10 to 100 times on top of that. Does an end user want to wait 30 minutes for a single AI token? Probably not. TEEs are the best trade-off right now. The downside is that consumers generally can't run frontier models on their own devices.

Open models, cheaper compute and compute-as-yield

Ashton: Is decentralized, open-model AI really cheaper than the centralized models?▶ 13:02

Michael: Generally, yes. Open-source models tend to be 3 to 20 times cheaper, depending on the model. For directly comparable models, the open-source one is generally about 3x cheaper at quite a comparable level of intelligence. The gap between closed and open models has shrunk consistently, and open models are now maybe a month to three months behind whatever the closed labs are doing. There are also flash models that can be up to 20x cheaper at, let's say, a 10 percent performance trade-off.

Ashton: Can you explain how Compute Finance works?▶ 14:21

Michael: It's only possible at the intersection of AI and blockchains, because blockchains add a financialization layer. A stock gives you dividends, a bond gives you a coupon, mining gives you Bitcoin, DeFi gives you yield. Why can't you hold an asset that gives you compute as yield? That's what Compute Finance is, and several flavors are launching on 0G. With Infinite AI, you stake 0G, get a liquid staking token back, and mint Infinite AI with it. One Infinite AI equals 1.27 compute credits daily for as long as you hold it, and you also get 50 percent of the staking yield. Another is USDZ, a stablecoin that pays up to 8 percent yield in compute credits instead of in the same currency. There's also a launchpad where you can launch an AI agent and use the trading fees it generates to pay for compute. In the future, we believe there will be markets where you can speculate on compute, hedge it, and pre-reserve it. We expect deep, liquid markets within a year.

Ashton: Is today's cheap access to frontier AI sustainable?▶ 22:06

Michael: A lot of it is subsidized by VC funds at this point. Without strong competition from open labs, the closed labs will work out how to maximize profit. They'll ask what 10 virtual employees are worth, whether that should be half the cost of a human, and then calculate the token volume and the subscription from there. Maybe that subscription ends up being 20,000 a month. I'm not saying they'd go to that level, but that's how economics works over time, and I studied political economy in undergrad. You can already see it in the fine print. A model looks price-competitive at first, but if you want it at double speed, it costs double, and if you want a long context window, it doubles again. That's 4x the cost from an API standpoint, and we'll see that cost passed on to consumers over time.

The moment that verification changes, you can basically shut off the agent.

Enterprise adoption, speed and built-in safety

Ashton: What is driving institutional and enterprise interest in decentralized AI?▶ 23:18

Michael: Some institutions can't adopt AI at all because of the regulations they must follow. Many financial institutions tell us their employees released agents, but they have no guardrails in place. They don't know what models are being called, what policy engines sit behind them, or what the agents are actually doing. Are they sending emails on the company's behalf? That may be a compliance violation. There are also sovereignty laws. An institution in Luxembourg may only be able to keep its data and compute in Luxembourg, and I expect Indonesia, Malaysia and other countries to follow suit. That's where decentralized AI makes a difference, because you don't have to trust 0G. You can verify on the blockchain that the compute was safe, that the agent didn't jump out of a sandbox or call strange tools, which model and graphics card it used, and that everything stayed within your country. If something happens, you can shut off the agents immediately.

Ashton: Decentralized systems are often seen as slower. How do you close that gap?▶ 25:52

Michael: That's why we use trusted execution environments, because the overhead is not noticeable. Fully competing on inference latency with OpenAI or Anthropic is very challenging, because they've invested hundreds of billions of dollars in core infrastructure and have really optimized latency. But does it matter to an end user whether it's 20 versus 100 milliseconds, or 20 versus 50 or 60? It's not noticeable anymore. As long as the end-user experience is equivalent, we have a very similar value proposition, even with a little overhead from the additional verification.

Ashton: What is 0G doing on safety in its own models?▶ 27:57

Michael: We've had three research papers accepted to NeurIPS, one of the top AI conferences in the world. One is a safety approach only 0G has, called sparse inference alignment. It steers a model when it's exhibiting high-risk behavior. Think of a parent who sees a child about to do something dumb on a bicycle and quickly moves the bike in a safe direction. As the model produces tokens, if they start becoming high-risk or low-probability tokens, we steer it in a different direction. That's now infused into our own models. Our research shows it's as effective or more effective than what the closed labs do, which is generally post-training, a one-time alignment process after the model has been trained on lots of data. We do this consistently in production, and that keeps models safe.

This interview has been edited for length and clarity. Recorded November 2026.

Michael Heinrich
Michael HeinrichMichael Heinrich is the co-founder and CEO of 0G Labs, also known as Zero Gravity, which builds a modular layer 1 blockchain and infrastructure for decentralized AI. Born in Alushta, Ukraine, and raised in East Berlin, he moved to the United States and studied at… Full profile →
Share this interviewXLinkedIn
▶ Watching the interview
CCS Insider · free, twice a week

New founder interviews in your inbox

The week's interviews and the key takeaways, twice a week from CCS Insider.

Readers in 160+ countries. No spam, unsubscribe anytime.