The code spoke. The revenue grew 40% year-over-year. The operating loss grew 120%. That is the arithmetic of OpenAI’s 2024 API business. A spread so wide it screams for a forensic audit. Wall Street did not scream. It whispered. Then it turned away.
I spent 2025 auditing a protocol that let autonomous AI agents interact with blockchain oracles. The oracle feed lacked cryptographic signatures. The flaw was obvious after 150 hours of simulation. The team called it a feature. I called it a lie. The same lie now haunts the AI giants. They built a palace on a fault line.
Context: The IOSG Thesis
IOSG, a venture capital firm with a history of calling tops and bottoms, published a document titled "AI’s Crossroads: Why Wall Street Is Saying No to ChatGPT and Claude." The headline is not a data point. It is a signal. IOSG argues that the current AI polarization—billions poured into two closed-source models—faces a structural rejection from institutional capital. Why? Because the unit economics are broken. The customer acquisition cost (CAC) per enterprise client is astronomical. The lifetime value (LTV) is unpredictable. The gross margins, when stripped of hype, resemble a startup burning cash for market share, not a platform generating durable profits.
I confirmed this thesis through my own due diligence framework. In 2022, after the FTX collapse, I retreated from social media for six months. I audited three Layer-2 scaling solutions. Each claimed decentralization. Two used centralized fault proofs. The code spoke, but the logic was a lie. That experience taught me that narrative and architecture never align. IOSG's article applies the same lens to AI.
Core: The Economic Deconstruction
Let me break down the engine room. Not with buzzwords. With first principles.
Revenue per token is collapsing. GPT-4’s inference cost per thousand tokens has dropped from $0.06 in March 2023 to $0.02 in June 2024. That is a 66% price decline in 15 months. Volume is increasing, but not linearly with revenue. The volume growth is driven by price cuts, not new use cases. This is the textbook definition of a commoditizing business.
The CAC-to-LTV ratio is inverted. Enterprise sales cycles for AI platforms average 9 to 12 months. Contracts are often signed with heavy discounts. Implementation failures are common. The net dollar retention (NDR) for AI API providers is below 100% for many segments. In SaaS, an NDR below 120% is a warning. Below 100% is a death spiral.
The scaling law is a cost law. Each new generation of models requires exponentially more compute. GPT-5 is rumored to cost $10 billion in training capital. That is not a moat. It is a debt. Wall Street sees this. The reward matches the risk, not the dream.
I applied the same logic to the AI-agent protocol I audited in 2025. The team claimed to have a “trustless” oracle. I simulated 10,000 attack vectors. The validation lacked cryptographic signatures. I published the findings. The project paused its launch. Data does not lie, but it does not care.
Now look at the competitive landscape. Google’s Gemini, Meta’s Llama, Mistral, Cohere—they all build on the same transformer architecture. The moat is not technology. It is data and distribution. But data can be leaked. Distribution can be bought. The only durable moat is a network effect that compounds with each user. Do ChatGPT and Claude have that? No. Users switch between models freely. The switching cost is zero. Trust is a variable you cannot hardcode.
Contrarian: What the Bulls Got Right
I am not a bear for the sake of it. The bulls have three valid arguments.
First, the enterprise market is still early. The current penetration of AI in Fortune 500 workflows is less than 10%. The potential TAM is massive. If OpenAI can solve data privacy concerns through private deployments or encrypted compute, the revenue multiplier could be 10x. I have seen this playbook before—Salesforce, AWS, Snowflake. Early adopters pay a premium.
Second, the capital moat is real. OpenAI and Anthropic have raised over $30 billion combined. No other AI company has that war chest. They can outspend competitors on talent, GPU clusters, and marketing. In a capital-intensive race, the deepest pockets often win.
Third, the ecosystem lock-in is growing. ChatGPT has 100 million weekly active users. Developers build on its API. Plugins, custom GPTs, and the upcoming agent platform create switching costs. Once a company integrates AI into its core workflow, ripping it out is painful.
But these arguments ignore one variable: time. The capital lock-in is only valuable if the return on capital exceeds the cost. With a 5% risk-free rate, any technology investment must yield 15%+ equity returns. OpenAI’s current revenue-to-valuation ratio is roughly 1:200. That is worse than WeWork at its peak. They built a palace on a fault line.
Takeaway: The Accountability Call
Wall Street is not wrong to question the sustainability of the current AI model. The numbers do not add up. The narrative is ahead of the economics. But the contrarian take is not that AI is a fad. It is that the market is underpricing the infrastructure layer—the chips, the open-source models, the middleware, the audit mechanisms. The next cycle will reward companies that make AI cheaper, more transparent, and less dependent on a single closed-source vendor.
My 2025 AI-agent audit showed me that security is not a feature. It is a prerequisite. The same applies to business models. If the capital runs dry, the palace collapses. But if the capital flows into efficiency, the fault line becomes a foundation.
I will keep auditing. The code speaks. The logic either holds or it does not. And when it does not, I will write about it.
Trust is a variable you cannot hardcode.
The code spoke, but the logic was a lie.
Data does not lie, but it does not care.