Hook
In Q1 2024, OpenAI reported $5.7 billion in revenue — but burned $3.7 billion in cash to generate it. That's a cash burn rate of roughly $1.23 per dollar earned, giving the company an annualized loss of nearly $15 billion if it scales linearly. Gary Marcus recently labeled this dynamic “unsustainable” and predicted government intervention as the only rescue. But as someone who has spent years architecting DAO treasuries and watching similar burn cycles unfold in crypto, I see a deeper rot: not a liquidity crisis, but a governance void. The centralized AI giants have no on-chain feedback loops, no token voting mechanisms, and no transparent markets to price their most critical input — compute. They are spending billions on a model (pun intended) that treats capital as infinite, while every decentralized protocol I’ve audited would have been forked or liquidated by now.
Context
Let’s step back. The current AI landscape mirrors the ICO frenzy of 2017 and the DeFi summer of 2020, but with a dangerous twist: the capital flows are private, opaque, and unaccountable to any community. OpenAI and Anthropic together command over 70% of global LLM API traffic, yet their financials resemble a high-risk venture capital bet dressed in enterprise clothing. Revenue growth is impressive — $5.7B quarterly for OpenAI — but the cost base is exploding faster. Training runs now exceed $100 million per model, inference costs scale non-linearly with user adoption, and the only “governance” mechanism is a board of directors with fiduciary duties to equity holders, not to the users or developers building on their infrastructure.
In the crypto world, we learned the hard way that uncontrolled burn rates destroy communities. I co-founded a DAO in 2017 that drained its treasury through a flawed multi-sig and lack of budget caps. That failure taught me that financial sustainability is not a technical problem — it’s a governance problem. Without transparent, on-chain mechanisms for resource allocation, even the most brilliant technology can become a money furnace. OpenAI and Anthropic are essentially centralized DAOs without the DAO. They have no token to vote on compute budgets, no treasury dashboard for stakeholders, and no mechanism to align incentives between capital providers (Microsoft, Google, Amazon) and the builders who actually use the models.
Core
Here’s the insight that most mainstream analysis misses: the burn rate is not a bug — it’s a feature of the current governance architecture. Let me break it down with the lens I use when auditing DAO treasuries.
First, compute cost opacity. In decentralized protocols like Aave or Compound, interest rate models are publicly visible and adjustable through governance votes — even if, as I’ve argued elsewhere, they are “completely arbitrary.” But at least they exist. OpenAI’s cost for a single inference call is a black box. They subsidize low API prices to drive adoption, but that creates an unnaturally elastic demand curve: cheaper prices trigger higher usage, which increases inference costs, which deepens the burn. Without a market clearing price for compute — something a decentralized compute marketplace like Akash or Golem attempts to provide — there is no signal to the organization that they are overproducing. The natural incentive to cut costs is absent because the board sees “growth” as the only KPI.
Second, lack of token-based governance. When I designed the governance framework for “GlobalCommons” in 2024, I insisted on a hybrid model where token holders could vote on budget allocations. Why? Because in a centralized system, executives have no direct feedback from users about what features justify the cost. OpenAI charges $20/month for ChatGPT Plus, but that price is set top-down, not via market mechanisms. The result is a classic principal-agent problem: management optimizes for valuation milestones (new funding rounds, PR wins), while the long-term health of the network — the code, the community, the value — gets secondary attention. Code is law, but people are the soul. Here, the soul is bleeding cash.
Third, the China discount trap. As the source analysis notes, Chinese models like Kimi K3 are approaching US performance at lower prices, thanks to cheaper hardware and more efficient architectures. This forces OpenAI and Anthropic to either match prices (deepening the burn) or differentiate on quality. But quality requires more compute, which means more burn. It’s a ratchet that only tightens. In crypto, we saw the same dynamic with Ethereum Layer-2s: ZK rollups promised scalability but required proving costs that made them unprofitable unless gas prices surged. Decentralization is a verb, not a noun. If you can’t afford to keep doing it, you’re not decentralized — you’re just a temporarily funded startup.
I’ve personally audited three DAOs that fell into this exact trap: they raised huge treasuries during a bull market, built ambitious products, but never built a governance model to cap spending. Each one burned through 80% of its funds within 18 months. The pattern is identical: euphoria over usage metrics (“look at our TVL/user count”) masks the fact that each user costs more to acquire and serve than the revenue they generate. OpenAI’s $5.7B revenue with $3.7B burn is a textbook case of negative unit economics at scale.
Contrarian
Now, the counter-argument: “Government will bail them out for national security reasons.” This is where my contrarian lens, forged in the crypto regulatory wars, kicks in. Yes, the US government has a history of funding critical technology — DARPA, the Internet, GPS. But a bailout of OpenAI or Anthropic would be a governance nightmare. If the government injects capital, what strings attach? Will they demand ownership? Will they impose safety constraints that slow innovation? In crypto, we’ve seen how state intervention can distort markets: China’s ban on crypto trading pushed mining to other jurisdictions, while the US’s unclear regulatory stance on DeFi created a chilling effect. A government rescue of centralized AI would likely turn them into quasi-public utilities, which might solve the funding gap but kill the very agility that makes them innovative.
Moreover, the assumption that “the government must intervene” overlooks a third path: decentralized AI governance through tokenization. What if, instead of a bailout, we create a DAO that manages a compute treasury? Users stake tokens to access inference, and the price is set by a bonding curve that reflects real supply and demand. The DAO can vote to reduce inference costs in exchange for lower quality, or increase spending on training — all with transparent, on-chain accountability. This isn’t utopian; projects like Bittensor and Ritual are already experimenting with decentralized AI coordination. Trust isn’t verified on-chain; it’s earned through transparent governance. If OpenAI were forced to unbundle its compute into a public, token-governed layer, the burn rate would become a feature — a parameter adjustable by the community, not a secret liability.
Some argue that such a model would be too slow for cutting-edge research. But I’d remind them: the financial unsustainability we’re seeing is the ultimate speed bump. The fastest car in the world is useless if it runs out of gas. A governance model that aligns incentives — where every stakeholder has a voice and every emission has a cap — is the only way to build a long-term, antifragile AI economy.
Takeaway
The story of OpenAI’s burn rate is not a story of technology failing, but of governance failing. We’ve seen this movie before — in ICOs, in DeFi, in every bull market where growth trumped sustainability. The solution isn’t a government check; it’s a new governance architecture that puts the community in the driver’s seat. As I said in my first days as a DAO architect: Decentralization is a verb, not a noun. If the AI giants don’t start treating it as such, they’ll burn not just their cash, but the trust of an entire generation of builders. The real question is not whether they will fail, but whether we will learn from their collapse and build a better model — one where the soul of the protocol is as important as the code.
