The total cost of an AI task depends on the outcome quality. A high-quality model might use more expensive tokens but achieve the result faster and with fewer attempts, making the overall system cost lower than a cheaper, less effective model.
AI excels at tasks with clear verification (e.g., code review). The next breakthrough will be AI systems that can define "what good looks like" for subjective domains like law or management by creating their own evaluation frameworks, expanding AI's capabilities into new, complex areas.
When a frontier model lab builds an application, it's incentivized to use its own models, even if a competitor's is better for a task. This "model-locked" status creates a conflict, as they sell "their best model" instead of "the best model," a key disadvantage against neutral providers.
The value in a model routing company isn't the commoditized technology, but the data on how businesses allocate resources to AI. For a company like Stripe, which controls money flow, owning the data on intelligence flow provides a powerful, strategic view into the future of capital allocation.
As AI gets embedded in core workflows, the key strategic question becomes who owns the resulting intelligence. Enterprises are wary of outsourcing their core logic to model providers who have explicitly stated they will compete in their customers' industries, making ownership of these learnings paramount.
An AI startup's durability depends on its workflow. Those in specialized, data-gated fields like law are defensible. In contrast, startups improving common knowledge work in tools like Excel face existential threats, as these workflows are not differentiated and may become obsolete.
The narrative framing open-source models as a "Chinese" threat is a deliberate tactic by large, closed-model labs. It aims to associate open-source with foreign risk, thereby discouraging adoption and creating a perception of insecurity, when in reality all models have creator-imposed biases.
The AI market will bifurcate. Open models will dominate most commodity tasks. However, the most economically significant problems—like advanced scientific research—will rely on closed, frontier models, allowing them to capture a disproportionate share (30-40%) of the total economic value.
Instead of traditional hiring, acquiring small teams or solo founders of relevant projects is a better way to find talent. The act of independently building and shipping a product in the same problem space is a stronger signal of conviction, skill, and mission-alignment than interviews can assess.
Founders often mistake visible "grinding" for productivity. However, this culture is performative, rewarding the appearance of hard work rather than actual results. It signals a business that may not be viable without extreme hours and incentivizes the wrong behaviors.
Thinking about token budgets per person is a flawed, input-focused metric. The correct model is to allocate a budget (potentially seven figures) to a project or desired outcome, like beating a benchmark. This reframes AI spend as a capital allocation towards business goals, not an employee perk.
Traditional signals like elite schools or competitions show an ability to follow rules. The most valuable trait for a startup is the capability to operate outside of established systems. Look for people who have built things on their own, as they've demonstrated this rare skill.