Key Takeaways

  • Demand for top-tier AI intelligence is still “unbounded,” particularly in high-stakes fields like coding, science, and legal, driving continuous innovation at the frontier.
  • Clay Pavore of Sierra predicts an “assembly line” dynamic where today's cutting-edge frontier models eventually transform into cheaper, fine-tuned open-weight models for specific tasks.
  • Despite improvements in hardware, token costs for AI models aren't decreasing as expected. Instead, they're rising due to the demands of reasoning models and a critical shortage of high-end GPUs like Blackwells and H100s.
  • Companies will increasingly adopt a hybrid model strategy, mixing and matching powerful frontier models for complex, generalist tasks with specialized, cost-effective open-weight models for routine or niche workloads.
  • The global competitive landscape, including the willingness of Chinese companies to distill frontier models, will likely accelerate the commoditization of older models.

The AI Model Assembly Line: From Frontier to Fine-tuned

Forget the idea of a single, dominant AI model. Sierra co-founder Clay Pavore lays out a future where the AI landscape operates more like an industrial complex, churning out specialized intelligence. He describes an “assembly line” effect: models that are at the frontier today—think bleeding-edge reasoning and problem-solving — will inevitably become yesterday's news. But they won't just vanish. Instead, they'll be distilled and optimized into cheaper, open-weight, fine-tuned models designed for specific workloads.

This means a dynamic where companies aren't just picking one model. They're strategically combining them. Pavore suggests, “You'll end up with companies using both, mixing and matching them depending on the task at hand.” For founders, this means anticipating a future where a premium frontier model solves your hardest, most general problems, while a suite of specialized, open-source derivatives handles the more routine, cost-sensitive operations. The goal isn't just to be smart; it's to be efficiently smart.

The Hidden Cost: Why AI Tokens Aren't Getting Cheaper

Many expected the cost of AI tokens to steadily decline as the technology matured. Harry Stebbings pressed Pavore on this, noting that current trends show token costs increasing, not decreasing, especially with the shift from simple chat to agent-based economies. Pavore agrees this is a real phenomenon, driven by a few powerful forces.

The first is demand. Pavore states, “I think we have not yet appreciated the unbounded demand for call it frontier levels of intelligence.” People want smarter AI, capable of more complex reasoning, and they're willing to pay for it. The second, and perhaps more immediate, is a brutal supply-side constraint: GPUs. Even as hardware gets more efficient, the sheer demand for compute—to run both frontier models and open-weight models—is hitting a bottleneck. “If you have unbounded demand for frontier level intelligence or GPUs to run open weights models and the rate limiter is the you know number of Blackwells and H100s you have, you end up with kind of a floor on the cost of tokens,” Pavore explains. The energy and hardware expenses create a stubborn floor that prevents token costs from plummeting, challenging the assumption that all AI services will simply get cheaper over time.

The Compute Bottleneck and Global Stakes

The scarcity of high-performance GPUs like Nvidia's H100s and upcoming Blackwells isn't just a pricing factor; it's a strategic one. This compute bottleneck means that even if the software improves, the physical infrastructure limits how fast and how cheaply AI can scale. Pavore also hints at a broader competitive dynamic at play, mentioning that part of the cost pressure and commoditization of older models might come from “the willingness of Chinese companies to do scaled distillation of the frontier models from the labs.” This global competition means constant pressure to extract more value from every model, every token, and every available GPU.

What to Do With This

Founders, develop a hybrid AI strategy now. Don't assume you'll only use one type of model; build your stack anticipating the "assembly line" effect where today's frontier models become tomorrow's fine-tuned, open-source workhorses. Simultaneously, ruthlessly optimize your token usage, because with GPU scarcity creating a floor on costs, efficient prompt engineering and model selection aren't just good practices — they're competitive advantages that will save you significant capital this quarter.