0G 聯合創辦人兼 CEO Michael Heinrich 專訪

Michael Heinrich, Co-Founder and CEO of 0G, is a serial entrepreneur, investor, and technology executive focused on building decentralized infrastructure for artificial intelligence. At 0G, he is leading the development of infrastructure designed to make AI more open, verifiable, and accessible. Heinrich also serves as Executive Chairman of ZeroStack Corp. and has been an instructor and mentor at Stanford University for more than a decade, contributing to programs spanning entrepreneurship, venture capital, and innovation. He is also a Venture Partner and angel investor with Pioneer Fund, with investments across companies and projects including Stripe, Anthropic, Brex, Notion, Filecoin, Worldcoin, Uniswap, Shield AI, and Binance. Previously, Heinrich founded workplace wellness company Garten, formerly Oh My Green, and held roles at Bridgewater Associates, Bain & Company, SAP Labs, Microsoft, and J.P. Morgan, giving him experience across entrepreneurship, technology, finance, strategy, and venture investing.

0G is building a decentralized AI infrastructure stack designed to serve as a trust layer for artificial intelligence, enabling AI workloads to be executed, stored, and verified without relying entirely on centralized providers. Its architecture brings together verifiable compute, encrypted decentralized storage, private AI inference, and an AI-native EVM-compatible Layer 1 blockchain, creating infrastructure for applications ranging from autonomous AI agents and decentralized models to AI marketplaces and machine economies. The platform is designed around modularity, allowing developers to use individual components such as storage or compute independently while supporting fully onchain AI applications where data, models, and actions can be audited and verified. Through this approach, 0G aims to provide the underlying infrastructure for a more open and transparent AI ecosystem in which developers, enterprises, and users retain greater control over AI resources and data.

Before co-founding 0G, you built and led Garten and worked across software, finance, venture investing, and entrepreneurship. What did you see happening in AI that convinced you to start 0G in 2023, and how did those earlier experiences shape your vision for the company?

I have spent my career moving between two worlds that rarely talk to each other. I wrote software at SAP and Microsoft early in my career. Then I went into finance, first studying global capital markets while I was doing research at Harvard Business School, then at Bain and Bridgewater, where I lived through the 2008 crisis from the inside. That experience led me to found Garten, a Y Combinator company.

Each chapter taught me the same lesson from a different angle: when a system concentrates control and hides its internals, the people who depend on it are the last to find out it is broken.

So back in 2023, I saw a pattern forming in AI at a speed I had never seen in any other market. The models were astonishing, but the infrastructure behind them was concentrated in the hands of a handful of companies and, more importantly, you could not inspect what ran, where your data went, or what the compute actually cost. At the same time, the capital required to participate was compounding every quarter. To someone who had watched 2008 unfold, it looked like leverage on top of a black box, with everyone assuming the black box was fine.

So the direction of 0G was a finance question that would not leave me alone: “What happens when the most important input in the economy has no transparent market and no independent verification?” Together with my co-founders Thomas Yao, Fan Long and Ming Wu, we answered that question by building 0G.

The answer is a trust layer for AI, a single stack that brings together private compute, our own open weight 0gm models, safety harness mechanisms like sparse inference alignment and Tee verification, encrypted storage and a blockchain built specifically for AI workloads. Here AI work, from inference to agent actions, is verified and audited.

AI companies are committing enormous amounts of capital to GPUs, data centers, and power. What is fundamentally changing about the economics of compute, and do you believe access to compute could become as strategically important as access to capital or energy?

Compute is already becoming as strategically important as capital or energy. The four largest hyperscalers committed more than $500 billion to AI capital expenditure in 2025 alone and there on a run rate to double this this year. That is no longer a technology budget. It is national-infrastructure spending run by a handful of private companies.

The fundamental change is that compute has become the binding constraint on intelligence itself. Every model improvement, every agent, every product ultimately depends on GPU hours. When a single input gates everything downstream, it stops behaving like a commodity you buy and starts behaving like capital you hold. Access to compute now determines which companies, universities, and countries get to build.

The strange part is that this asset has no market of its own. Compute is produced at scale, rented globally and priced daily, yet there is no way to hold a claim on it, hedge it, or allocate it transparently. Capital has markets. Energy has markets. Compute has procurement contracts with a handful of vendors.

Every scarce resource follows the same path: first it is produced, then it is priced, then it is financed. Oil did this. Electricity did this. Bandwidth did this. Compute is entering that third stage right now, and building the financial plumbing for it is one of the defining infrastructure projects of this decade.

0G is introducing the idea that compute can be owned, earned, and used rather than simply purchased as a cloud service. What does it mean in practical terms to turn compute into an asset, and how could that change the relationship between infrastructure providers, developers, and AI companies?

Think about how every other productive asset works. Equity introduced dividends. Bonds pay coupons. Compute Finance, which we coined as ComFi, applies the same thinking to compute. Instead of simply paying for access to GPUs, you can own a position that gives you access to AI compute you can actually use.

In practice, you put an asset to work and receive AI compute in return. That compute can power everything from running models privately to building and launching AI applications. Under the hood, this runs on the 0G blockchain. Users stake, meaning lock, 0G as collateral, receive a0G, mint iAI, and then commit that iAI to earn compute they can use across AI services.

The goal is to make compute something you can hold, earn, and use, rather than simply rent from a cloud provider by the hour. Under the initial parameters, users can commit iAI and receive compute credits worth more than $1 per day across supported 0G AI services. The underlying collateral remains locked while the position is active and can be recovered when the iAI is burned.

The important point is that users are not receiving cash or a financial return. They are receiving access to compute they can actually use.

This changes the relationship between infrastructure providers, developers, and AI companies. Infrastructure providers get a new way to make their compute capacity available beyond traditional cloud contracts. Developers get a more predictable resource they can plan and budget around. AI companies can secure compute and plan for it rather than simply treating it as an expense controlled by a handful of providers.

And because it runs on the 0G trust layer, compute usage can be verified and audited rather than simply taken on “trust us”

We are positioning ComFi as an open category for the AI industry, not a proprietary brand. Ascend and iAI are the first implementations of the idea; with more coming to 0G’s network.

0G Compute operates as a decentralized marketplace where GPU providers can compete to serve AI workloads. What economic advantages can this model offer compared with relying on a handful of hyperscale cloud providers, particularly when demand for GPUs is high?

We highlight three advantages that become more apparent when GPUs are scarce: price discovery, elasticity, and no lock-in.

1. Price discovery: On 0G Compute, providers set their own prices and compete for workloads. The result can be significantly lower costs: comparable open-model inference costs up to 95 percent less, while our storage layer runs at more than 90 percent below AWS S3. Part of that comes from competition, and part comes from putting idle GPU capacity to work and running open models rather than closed frontier models on dedicated infrastructure.

2. Elasticity: Hyperscaler capacity is planned years ahead and allocated through long-term relationships. A marketplace can pull supply from anywhere, including independent operators, academic clusters, and enterprises with unused hardware. When demand spikes, higher prices can bring more supply online. A fixed queue just gets longer.

3. No lock-in: Every provider on our network exposes the same OpenAI-compatible API through one router, which automatically fails over if a provider goes down. Today, that means 13 providers serving hundreds of models through a single endpoint, with more than 373 billion tokens processed and 27 million inference requests. Developers can switch providers with a header, not a migration project.

There is also a numbers point that people miss. Our own model, 0GM-1.0, is the second most used model on the network over the past 30 days, behind only DeepSeek. When a decentralized network’s own open model competes on usage with the best open models in the world, it shows that the marketplace can work at scale.

The hyperscalers are excellent at what they do. But when GPUs are the scarcest input in AI, more compute sources can create a very different supply model than simply waiting for more hyperscaler capacity.

One challenge with decentralized infrastructure is ensuring cheaper, more widely distributed compute doesn’t come at the expense of reliability. How does 0G determine whether a compute provider actually performed an AI workload correctly?

This is an excellent question, and one that separates a real compute market from a system where you simply have to trust a provider like AWS, Google Cloud, or Microsoft Azure. We designed 0G so you never have to take a provider’s word for it.

0G’s private tier runs inside a sealed trusted execution environment. The hardware produces an attestation showing which model and code are running, and it signs every response with a key that exists only inside that environment. A client can verify the signature against the attestation and confirm that the specified model ran on the specified provider without modification. Neither 0G nor the provider can see the prompt or response. Across all tiers, 0G retains no prompts or responses and does not train on your data.

On the verified tier, requests are routed to the intended upstream model, so you know which model answered even when the provider itself is centralized. These tiers are defined per request and enforced by the network. If a request requires a private-capable provider and none is available, it fails rather than silently falling back to a different tier. Honest tiers matter more than marketing tiers.

Payment is tied to proof. Funds are held until the signed result is delivered. The network has recorded more than 24.9 million TEE proofs so far, and the framework also supports optimistic and zero-knowledge verification for workloads where those approaches make more sense than hardware attestation.

Reliability is handled straightforwardly, with automatic failover across providers and live health and latency data. The underlying network is designed so verification and settlement don’t become the bottleneck.

AI infrastructure economics are increasingly intertwined with electricity availability. As models and inference workloads continue to scale, do you expect energy to become the primary constraint on AI compute rather than GPUs themselves, and how could decentralized infrastructure help address that bottleneck?

Energy and compute are definitely intertwined, and one affects the other and energy-led bottlenecks are already happening. In Northern Virginia, more than 3 GW of data center demand suddenly disconnected from the grid in July, highlighting the strain concentrated data center loads can put on power infrastructure.

AI data center electricity usage is on track to double by 2030, and the IEA projects global data center consumption will more than double to around 945 terawatt-hours by 2030, with AI as the main driver. You can build more GPUs faster than you can build the power infrastructure needed to run them.

Centralized AI makes this harder by concentrating demand into a small number of massive data centers, each requiring enormous amounts of power and new grid infrastructure.

Decentralized infrastructure can help ease this growing constraint by putting workloads where power and computing capacity already exist. That could mean regional data centers, industrial sites with spare capacity, or renewable energy that would otherwise go unused. Instead of moving more power to a few massive campuses, you can move the workload to where the power already exists and is underused.

To be clear, this does not eliminate the energy constraint, but it can make existing energy and computing infrastructure go further.

0G uses mechanisms including trusted execution environments, optimistic verification, and zero-knowledge techniques to make AI computation verifiable. Where do you see cryptographic verification becoming most important as autonomous AI agents begin making decisions and executing transactions without continuous human oversight?

Verification matters most at the exact moment a human stops watching. A copilot has a person reviewing its output. An autonomous agent does not. Once an agent can make decisions, sign transactions, hold funds, or interact with other agents, you need to know which agent is acting, which model is running it, and whether it did what it was supposed to do.

That is where cryptographic verification becomes important. When an agent runs inside a trusted execution environment, you can get a signed record showing which model produced a decision and what inputs it used. The blockchain is an ideal trust layer because it is immutable and append-only and can trace provenance. The blockchain provides a shared record of those proofs, so developers, auditors, and counterparties can verify what happened without trusting a single party’s version of events.

The same principle applies as agents start interacting with each other at scale. You cannot have a human reviewing every decision or transaction. Verification, audit ability, and trust has to become part of the infrastructure itself. The goal is simple. Instead of asking whether an agent says it followed the rules, you can prove that it did.

Much of the infrastructure race today centers on building ever-larger centralized data centers. Does decentralized AI compute represent an alternative to that model, or do you ultimately see decentralized networks and hyperscalers becoming complementary parts of the same global compute market?

I’d say they are complementary, and the split will look a lot like the energy grid. Hyperscalers are the baseload. They are very good at massive, predictable, latency-sensitive workloads for the largest enterprises, and they should keep doing that. Decentralized networks can complement them by absorbing variable demand, bringing distributed and otherwise unused compute online, and giving more companies access to GPU capacity.

Decentralized networks are a genuine alternative because of trust. Concentration is an economic problem, but it is also a visibility problem. When a handful of companies run the infrastructure, they also control what you are able to see. A hyperscaler can promise not to look at your data. A sealed enclave with public attestation can prove it. That property does not improve with a bigger data center. It comes from having verification built into the infrastructure.

We already see the two models working side by side. Google Cloud and Alibaba Cloud were among our mainnet launch partners. Enterprises want to be able to route a workload to a hyperscaler, a decentralized provider, or their own hardware based on price, privacy requirements, and the level of verification they need.

That is the direction we see the market moving. Not one model replacing the others, but a global compute market where workloads can move between different infrastructure sources based on what the job requires.

As AI agents become economic actors capable of purchasing services and managing resources autonomously, could compute itself become part of a machine-to-machine economy where agents dynamically acquire infrastructure based on price, performance, and availability?

We have seen some small examples of this on our network. Most requests through 0G Compute do not come from a person clicking a button. They come from applications and agents calling an API, paying per request from a funded balance, and getting routed to whichever provider best matches the job. About 43,000 wallets have paid for inference this way, and more than 300 projects built on 0G. Ghast AI alone has pushed more than 100 billion tokens through 0G Compute. NeoSoul runs millions of agents across 0G and BNB Chain, another blockchain network. An agent choosing a cheaper provider for a batch job and a private enclave for a sensitive one is a routing decision that happens today.

The missing pieces for a real machine economy are the things humans take for granted, namely identity, a way to hold resources in advance, and a unit of account. Agents need an identity that works across services, which is what 0G’s Agentic ID is designed to provide. They also need a way to secure compute in advance rather than pay for every request, so they can plan and operate autonomously.

That is part of the reason we built Compute Finance. When compute can be earned through participation in the network and spent across AI services, an agent can be funded once and draw on the compute it needs without a human topping up a cloud account. Ascend and Infinite AI are the first implementations. The machine economy will be built on top of them by people we have not met yet.

The safety requirement is non-negotiable. Every one of those machine decisions needs to leave a verifiable trail and have safety guardrails embedded. A machine economy without receipts is just a faster way to lose money and to bring forth an undisclosed security disaster.

Looking five years ahead, how do you expect the market for AI compute to evolve? Will companies still think primarily in terms of renting GPUs and cloud capacity, or are we moving toward a more liquid global market where compute can be owned, exchanged, monetized, and allocated dynamically?

I don’t think renting will disappear, the same way leasing did not disappear when equipment finance matured. But it is not the only option anymore, and it will stop being the only way serious players think about compute as a resource.

As this evolves, I expect compute to develop the same kind of financial infrastructure as every other strategic input. Companies will hold positions in compute the way they hold energy contracts or foreign currency. Providers will sell capacity into an open market instead of negotiating one contract at a time. Compute will be priced by workload and trust level, because verified private inference and unverified inference are different products. And a growing share of buyers will be agents, not procurement teams.

That is what Compute Finance (ComFi) is for. We launched Ascend and announced Infinite AI on the first anniversary of our mainnet because the sequence is now clear. Compute is produced. It is priced. The next step is a financial layer where claims on compute can be held, earned, spent, and traded, with every unit leaving a verifiable record. We are deliberately making ComFi an open category rather than a proprietary brand, because a market with one participant is not a market.

What makes any of this possible is not a better GPU. It is the trust layer: proof of what ran, proof of who ran it, and settlement that follows the proof. Once compute is verifiable, it becomes fungible enough to finance. Once it is financeable, it becomes liquid. The next five years are when the industry stops treating the most important resource in the economy as a line item.

Thank you for the great interview, readers who wish to learn more should visit 0G.