If you’re building an AI product right now, you’re probably losing sleep over picking the wrong model. You’re terrified that Chinese open-source models will compromise your data security, or that if you don’t marry GPT-4, you’ll be left in the dust. But if you look closely at what Alex Atallah—the co-founder of OpenSea who now runs OpenRouter—is seeing from the front lines, you’re focusing on the entirely wrong threat.
A single-model strategy isn’t a safety net; it’s a loaded gun pointed directly at your own valuation.
The industry narrative wants you to believe that one almighty frontier model will eventually win everything. That’s a fairy tale designed to lock you into an ecosystem. The reality is that AI is fragmenting at breakneck speed. With new models dropping every 10 hours, the value isn’t in the brain itself—it’s in the routing infrastructure that connects them. When OpenAI slashed the price of one model by 10x on OpenRouter, usage skyrocketed by 13x. The models are becoming a commodity; the orchestration layer is the real moat.
But here is the twist that almost every enterprise is missing. You’re losing your mind over the security risks of Chinese open models like GLM or Kimi. You block them, fear them, and question their supply chain trust. Meanwhile, you are happily handing over your most sensitive proprietary data to OpenAI and Anthropic.
You are handing your moat to the exact same companies that will eventually become your biggest direct competitors.
Just ask Figma what happens when Anthropic launches Claude Design. Frontier labs aren’t just neutral providers anymore; they are sleeping giants that can enter your application market at will. They hold your data, they see your prompts, and they have every incentive to turn their massive models into direct competitors the second your niche becomes profitable.
Chinese models? The risk is known and manageable. You can deploy them in your own VPC, build guardrails, and control the visibility. Frontier labs demand blind trust in their clouds. If you are a startup merely wrapping a frontier model’s intelligence in a thin UI, the model lab has the ultimate incentive to eat your lunch.
Models are becoming a utility. The real power belongs to whoever controls the right to choose, not whoever controls the model.
This is why the future doesn’t belong to a single super-app. It belongs to the ‘Harness’—composable work interfaces like Cursor that sit on top of the model chaos. A Harness lets developers own the user relationship without needing to train a model from scratch. It routes, it orchestrates, and it ensures you are never held hostage by a single provider’s quirks or price hikes.
AI is even rewriting the definition of human labor. In a few years, an employee’s cost won’t just be their salary. It will be the dynamic inference cost they burn through every day. Did they use the expensive frontier model to plan a project, and the cheap open model to execute it? If they aren’t routing their own work efficiently, they are a liability.
If you’re building in AI, stop chasing the holy grail of the ‘best’ model. Start building the routing layer. The future winners of tech’s biggest market won’t be the ones who train the biggest brain. The winners will be the ones smart enough to connect all the brains—and ruthless enough to rip the power away from the providers before they wake up and destroy them.
FAQ
Q: Isn't using a single frontier model like GPT-4 safer than routing between multiple smaller models?
A: No, it's a trap. Using a single model creates a hard dependency on a provider that can change prices, deprecate models, or enter your market as a direct competitor at any time. Routing gives you the leverage to switch instantly.
Q: What does it mean for a model to become a 'commodity'?
A: When models drop in price by 10x and usage explodes by 13x, the intelligence itself is no longer the differentiator. The value moves to the infrastructure that selects, routes, and composes those models to complete specific tasks.
Q: Why are frontier labs more dangerous than Chinese open-source models?
A: Because frontier labs hold your proprietary data in their closed clouds and have a proven willingness to launch competing applications (like Anthropic's Claude Design vs. Figma). Open-source models can be deployed locally with controlled guardrails, making their risks known and manageable.