When the investor who made $1.5 billion betting against subprime mortgages says an entire industry has no competitive advantages, founders should pay attention. Steve Eisman, profiled in Michael Lewis's The Big Short, has declared that AI companies lack sustainable moats and that this is not a recipe for longevity. His argument is straightforward: frontier models are rapidly commoditizing, open-weight alternatives match closed models at a fraction of the cost, and price competition is eroding margins faster than any company can build defensible differentiation. But Eisman's diagnosis, while correct about the model layer, misses where the real value in AI is being built today.
Speaking in a recent interview, Eisman drew direct parallels between the current AI landscape and the early internet era. During the dotcom boom, infrastructure companies captured all the initial value. Then competition eroded those margins, and the real wealth went to companies that built applications on top of the commoditized layer. Amazon, Google, and Facebook did not win because they owned the pipes. They won because they owned the user, the data, and the network effects on top of those pipes. Eisman sees the same pattern repeating in AI, and the historical evidence supports his analogy.
Why the Model Layer Is Commoditizing Faster Than Anyone Expected
Eisman's core thesis rests on two structural forces that are reshaping the AI industry faster than most observers anticipated. The first is the collapse of inference costs. GPT-4-class model pricing fell from $20 per million tokens to $0.40 per million tokens over 36 months, a 50x decline that outpaces every historical price curve in computing. The second is the open-weight revolution. Moonshot AI's Kimi K3, released in June 2026 under an open license, benchmarks within striking distance of Anthropic's Opus 4.8 and OpenAI's Fable 5 while costing 5 to 10 times less per token. On OpenRouter, Chinese models now handle 46 percent of all tokens served, up from single digits six months ago.
The strategic implication is brutal for anyone building a business whose primary product is a foundation model. If the gap between open and closed models is now just 3.3 percent on aggregate benchmarks, and that gap is concentrated in extended reasoning tasks rather than the commodity workloads that represent 90 percent of real-world usage, then the pricing power of closed model providers is structurally limited. Customers will migrate to cheaper alternatives the moment quality reaches an acceptable threshold. Eisman's claim that AI companies lack moats is, at the model layer, irrefutable.
Where Eisman's Thesis Misses the Mark
The limitation of Eisman's analysis is that it treats AI as a single industry rather than a stack with distinct economic properties at each layer. The model layer is indeed commoditizing. But the application layer, the data layer, the infrastructure layer, and the distribution layer all have moats that look very different from proprietary model weights.
Consider the evidence from the market itself. Anthropic and Blackstone launched Ode, a $1.5 billion joint venture focused not on building better models but on deploying them inside the world's largest enterprises. The bet is that implementation, not the model, is where durable value lives. Mercor hit a $20 billion valuation by owning the AI training data and evaluation layer, not by building a model. Oak raised a $60 million seed round to build AI-native identity management, betting that the identity layer for AI agents will be a defensible platform. None of these companies compete on model performance. They compete on data moats, distribution, regulatory barriers, or integration depth.
Eisman's internet analogy actually proves this point rather than undermining it. If you had taken his view in 1998 and concluded that the entire internet industry had no moats, you would have been right about the infrastructure providers whose margins collapsed. But you would have missed Amazon, Google, and Meta. The same logic applies to AI today. The question is not whether the model layer has moats. It does not. The question is which companies are building defensible positions in the layers above and below the model.
What This Means for Founders and Fundraising
The practical impact of Eisman's thesis is already visible in venture capital behavior. Investors are increasingly demanding to see moats that go beyond we use AI. Proprietary datasets, network effects, regulatory barriers, hardware integration, and vertical domain expertise are replacing access to a frontier model as the primary defensibility narrative. The era of raising a Series A on a thin GPT wrapper is definitively over.
For founders, the strategic calculus is clear. If the model layer has no moats, do not build your company there. Build on top of the commoditized layer with data that compounds, workflows that create switching costs, or distribution that competitors cannot replicate. The companies that will survive the commoditization cycle Eisman describes are those that treat AI models as a commodity input, not a product differentiator. The winning playbook looks like Ode's implementation-heavy approach, Mercor's data moat, or Oak's infrastructure play, not another foundation model API.
The deepest irony of Eisman's critique is that it may accelerate the very shift he describes. If his comments push more investors to demand non-model moats from AI companies, then capital will flow toward the application and infrastructure layers where durable advantages are possible. The commoditization of models will continue. But the companies that build on top of that commoditized layer may end up with the kind of longevity Eisman says the AI industry lacks.