Key Takeaways

  • Nvidia is strategically shifting to openly support and develop open-source AI models, a calculated move to prevent a duopoly from forming at the model layer by companies like Anthropic and OpenAI.
  • David Sacks argues that a competitive model layer benefits everyone: chip manufacturers (like Nvidia), application developers (like Palantir), and enterprises, fostering a diverse market for chip buyers and preventing proprietary knowledge lock-in.
  • Jason Calacanis points out Nvidia's previous reluctance to discuss its open-source efforts, suggesting Jensen Huang's company is now taking "the gloves off" due to increased competition and aiming to control more of the AI stack.
  • The cost of AI tokens is projected to drop 90% per year for the next three years, creating accessible "free options" and accelerating commoditization at the model layer.
  • To grasp AI market dynamics and strategic positioning, founders must apply The Three Layers of the AI Stack framework.

The Three Layers of the AI Stack

Type: method

Name: The Three Layers of the AI Stack

Components:

  • Layer 1: Chips: The foundational hardware component providing the computational power for AI models.
  • Layer 2: Models: The AI algorithms and trained models that perform intelligent tasks (e.g., Anthropic, OpenAI).
  • Layer 3: Applications: The end-user software and services built on top of the AI models (e.g., Palantir).

When This Works (and When It Doesn't)

This framework shines brightest when you are trying to understand where power and profit accumulate in the rapidly changing AI market. It helps make sense of strategic alliances, like the Palantir-Nvidia partnership, and explains why a chip company might care deeply about competition at the model layer. If you're building an AI product, it gives you a lens to evaluate vendor lock-in risks, competitive threats, and opportunities for differentiation.

However, the framework offers less insight into highly specialized AI domains where custom hardware or proprietary, niche datasets create unique, non-competitive moats. It also oversimplifies scenarios where one company might operate across multiple layers, blurring the lines of the stack. If your market isn't driven by general-purpose AI models or widely available chips, its utility diminishes.

What to Do With This

As a founder building an AI-powered service, you need to decide where to place your bets. Start by mapping your product against The Three Layers of the AI Stack.

For example, if you're building an AI assistant for project managers, here's how to think about it:

1. Chips: Are you relying solely on cloud providers (who abstract away chips), or do you foresee a need for specialized, local hardware? Nvidia's push for open-source helps them sell more chips, regardless of which model wins. Your choice here impacts your infrastructure costs and scalability.

2. Models: Will you build on proprietary models from OpenAI or Anthropic (Layer 2)? David Sacks warns against this, seeing it as potential vendor lock-in. Instead, investigate fine-tuning open-source models, especially with the projected 90% annual drop in token costs Jason Calacanis mentioned. This gives you more control and potentially lower long-term expenses.

3. Applications: How does your assistant differentiate beyond the underlying AI model? Is it your unique UI, your industry-specific data integration, or your workflow automation? Palantir, an application at Layer 3, thrives by adding unique value on top of powerful models. You need to do the same.

This week, review your current AI model strategy. If you're locked into a single proprietary model, explore an open-source alternative. Run a proof-of-concept for fine-tuning a model like Llama on your own data. This tactical shift can hedge against duopoly risks and future-proof your product.