The numbers don't lie. A 65% net loss margin on $5.7 billion quarterly revenue is not a sustainable protocol. It is a red flag that any auditor would flag before the next funding round. Yet the market continues to price OpenAI at $80 billion and Anthropic at $18 billion, as if these deficits are temporary bugs to be patched. They are not. They are features of a business model that relies on infinite liquidity subsidies—a DeFi-style yield farm disguised as cutting-edge technology.
I have spent the last seven years dissecting smart contracts and tokenomics. I have seen what happens when a project’s cash burn exceeds its revenue by a factor of two. It either pivots to a sustainable model, gets acquired by a larger player with deeper pockets, or dies. The AI industry is now at that inflection point. Gary Marcus’s recent warning that OpenAI and Anthropic “could fail” is not alarmist; it is a sober audit of their financial statements.
Context: The Architecture of Artificial Growth
OpenAI and Anthropic are the dominant validators in the AI consensus. They process over 70% of global large language model API traffic. Their business model is straightforward: train massive models using billions of dollars in compute, then charge users per token for inference. But the unit economics are inverted. Training costs are fixed and enormous—single runs exceed $100 million. Inference costs scale with user adoption, and every price cut to compete with Chinese models like Kimi K3 triggers a proportional increase in query volume, creating a feedback loop of higher losses.
This is structurally identical to a DeFi protocol that offers 100% APY to attract TVL. The underlying asset is real value—the model’s capability—but the yield is subsidized by venture capital. Once the subsidies stop, the TVL vanishes. In AI, the subsidies are Microsoft’s Azure credits and Google’s cloud commitments. Without them, OpenAI and Anthropic would be insolvent.
Core: Auditing the Burn Rate
Let me dissect the numbers from my perspective as a protocol developer. OpenAI’s Q1 2025 revenue was $5.7 billion, with cash burn of $3.7 billion. That implies an annualized loss of approximately $14.8 billion. Based on industry estimates, 60–70% of that burn is compute cost—training and inference. The remaining is talent, data acquisition, and regulatory compliance.
Compare this to a typical Layer-2 rollup. The sequencer earns fees from user transactions, but the cost of posting data to Ethereum (blob space) eats heavily into margins. After the Dencun upgrade, blob costs dropped temporarily, but they will saturate within two years, squeezing L2 economics. The same dynamic applies to AI: inference is the “blob cost,” and as more users query, the marginal cost per token should drop due to scale, but it doesn’t. Each new user adds a roughly linear increase in compute demand because models are not optimized for amortized inference. This is a known problem in attention mechanisms—KV-cache size grows with sequence length and user count.
Based on my audit experience, the real risk is not the current $3.7 billion quarterly burn but the nonlinear scaling of inference costs as user base grows. In DeFi, we call this a “liquidity death spiral”: a protocol offers high yields to attract liquidity, but the yields are unsustainable, leading to a rapid withdrawal. Here, the “liquidity” is user queries, and the “yield” is the subsidized price per token. Once OpenAI raises prices to cover costs, users will migrate to cheaper alternatives—open-source models like Llama 3.1 or Chinese APIs that are already undercutting them. This is not theoretical. It is happening.
Contrarian: The Bailout Fallacy
The common counterargument is that governments will step in. The U.S. Department of Defense, for example, could classify AI as critical infrastructure and allocate funds to keep OpenAI and Anthropic alive. This is a seductive narrative, but it ignores the fundamental tension between centralization and innovation.
Fragility is the price of infinite composability. If the government bails out these companies, it sets a precedent that private losses can be socialized while profits remain privatized. This is analogous to a DeFi protocol that relies on a central “peg maintainer” to stabilize its token. The peg holds only as long as the maintainer has capital. When the maintainer runs out, the system collapses—but now with moral hazard added.
The most dangerous assumption is that taxpayers will foot the bill. Code is law, and in a market economy, bankruptcy is the ultimate consensus mechanism. Allowing OpenAI or Anthropic to fail would reset expectations, force the industry to build sustainable models, and shift investment toward efficiency improvements—similar to how the 2022 bear market killed off overleveraged DeFi protocols and left standing only those with real revenue, like Uniswap.
Hype creates noise; protocols create history. The AI bubble is not different from the ICO mania I audited in 2017. Back then, I spent 40 hours tracing the Golem token distribution algorithm and found an integer overflow. The whitepaper promised a decentralized compute marketplace. The code could barely transfer tokens safely. Today, OpenAI’s whitepaper promises AGI. Its balance sheet can barely cover a single training run. The pattern repeats.
Takeaway: The Dencun Moment for AI
We are approaching the AI industry’s “Dencun moment”—a point where the cost structure shifts and many rollups become unviable. For OpenAI and Anthropic, this moment will arrive when their strategic investors—Microsoft, Amazon, Google—decide that the cloud revenue from AI inference is not worth the subsidy. That day is not far off.
Watch for the first down round. When OpenAI’s $80 billion valuation is marked down to $40 billion or less, the panic will spread. The survivors will be those that optimize for efficiency, not just scale. Companies that invest in speculative decoding, quantization, and custom hardware—like Groq or Cerebras—will thrive. The rest will become cautionary tales in the next bull run.
Trust, but verify the source code. The AI industry’s balance sheet is its source code. And right now, it has a critical vulnerability.