0G founder Michael Heinrich pitches "compute finance" as a way for ordinary people to own AI

Who Michael Heinrich is and the mission behind 0G

Heinrich introduces himself as a "mission-driven entrepreneur," saying money is "an outcome of the excellent actions that I take," a lesson he says he took from working at Bridgewater. He describes his current mission as making AI "a public good" through 0G, driven by a wish for AI to "serve humanity" and expand human consciousness. Before 0G, he founded a Y Combinator-backed company called Garden that scaled to 700 people before COVID forced a rapid downscaling; he also worked at Bridgewater on portfolio construction and strategy.

Asked whether Garden was a failure, he says it had "a happy ending" through an acquisition but didn't reach its potential, calling it "a deep success" in terms of leadership lessons. He reflects that failure is "loaded" with social judgment but that he treats it as information from a trial-and-error process, including a lesson about hiring a CEO who imported an outside company culture that clashed with his own.

Inside the token launch: warnings, momentum and a $3 billion debut

Heinrich recounts being in a "war room" in Jeju Island, Korea, two weeks before the token generation event (TGE) when the team learned two pieces of bad news: the DeFi ecosystem builders weren't ready, and a related digital asset treasury vehicle, 0G Stack, wasn't going to clear "baby shelf" status, undermining its flywheel of drawing cash from equity to buy the token. Despite this, the team proceeded, deciding to launch and follow with utility later. He says the valuation was pegged to a prior funding round with an equity value of "two billion," and that exchanges tested a pre-market that held steady, leading to a launch at a "three billion" valuation. He estimates the decision was "60% momentum, 40% conviction."

The token launched into KBW (Korea Blockchain Week) with what he calls "insane" trading volume, around $3 billion on day one. Soon after, he says, an October market event hit, the team was burned out from the run-up, some members took medical leave, and staff became anxious as the token became "a levered beta to Bitcoin."

Managing a team through burnout and a bear market

Heinrich describes handling team members who tied their self-worth to the token's price as "a challenge," addressing it by giving some staff extended leave and reminding the team that "an outside valuation of who we are individually does not matter." He also went on parental leave for the birth of his second daughter around the same period, trusting co-founder Ming and others to keep things running in what he calls "maintenance mode."

Upon returning around April or May, he says the team used renewed AI capability to "revamp" processes, including cutting some "low performers" to "reingject" company culture. He says he repeated the message that the company is "very well capitalized" and building for the long term, not "just for one cycle."

"Given what I knew at the time, I would make the same decision again. With hindsight, I would also still make the same decision again." — Michael Heinrich

What "compute finance" actually means

Heinrich explains that holding a 0G token or stablecoin grants access to "AI tokens" usable to interact with 0G's private compute, covering both closed-source and open-source models. He compares this to Bitcoin "franchising money" and Ethereum "franchising block space," agreeing when the host suggests 0G is trying to "franchise intelligence itself" — and adds it should eventually be "useful intelligence" specifically, distinguishing it from what he calls "AI slop."

He says owning compute this way differs from a subscription because it can be traded, hedged (he draws an analogy to "a perpetual fund unit of compute"), and isn't locked into one provider or GPU type. He frames the underlying observation as: if an AI agent wants to participate in a future "machine economy," it needs to finance itself, and owning the compute asset directly — rather than just spending on subscriptions — allows that asset to be financialized, for instance by directing part of a token's trading fees toward keeping an agent's compute alive.

Products, model claims and Venice AI comparison

Heinrich says 0G is releasing a model with a "100 million context window," built by infusing a memory layer onto an underlying Qwen model, claiming performance gains of "10 to 20%" over the base model on "most benchmarks" through inference optimization. He also cites a robotics model (VLA) where 0G says it beat a benchmark "by about six points."

Comparing 0G's private compute to Venice AI, he claims 0G offers privacy guarantees across three dimensions — from the router, the cloud provider, and the model provider — and states "from my understanding, Venice does not have that guarantee across all three dimensions," without elaborating further on how this was verified.

Centralized vs. decentralized AI, and how ordinary people can participate

Heinrich frames the stakes not as abstract ideology but as a choice about participation: providing data, hardware, or expertise to a network versus being disrupted by closed AI labs, referencing Palantir's discussion of "alpha extraction," where models allegedly absorb a company's know-how to build competitors. He warns the endpoint of centralization could be a population living on "universal basic income," contrasted with people being "self-empowered" participants in an AI economy.

For "normal" people, he suggests contributing data, hardware, or specialized skills (his example: "the best solidity coder in the world" contributing training data) in exchange for compensation or a share of resulting model revenue. He mentions a stablecoin launch and an "infinite AI" campaign as near-term ways to participate, and points to 0G's claimed prior work in decentralized training of a "100 billion parameter" model over a year ago, while acknowledging decentralized training at multi-trillion-parameter scale remains "a very hard problem." He also raises AI alignment as an underappreciated issue, citing models' growing ability to find "zero day" security bugs, and mentions a paper 0G has "under review" on an alignment approach he compares to a parent steering a child away from risk.