Fal strategically chose not to compete in LLM inference against giants like OpenAI and Google. Instead, they focused on the "net new market" of generative media (images, video), allowing them to become a leader in a fast-growing, less contested space.

Simply offering the latest model is no longer a competitive advantage. True value is created in the system built around the model—the system prompts, tools, and overall scaffolding. This 'harness' is what optimizes a model's performance for specific tasks and delivers a superior user experience.

The notion of building a business as a 'thin wrapper' around a foundational model like GPT is flawed. Truly defensible AI products, like Cursor, build numerous specific, fine-tuned models to deeply understand a user's domain. This creates a data and performance moat that a generic model cannot easily replicate, much like Salesforce was more than just a 'thin wrapper' on a database.

While today's focus is on text-based LLMs, the true, defensible AI battleground will be in complex modalities like video. Generating video requires multiple interacting models and unique architectures, creating far greater potential for differentiation and a wider competitive moat than text-based interfaces, which will become commoditized.

Top-tier kernels like FlashAttention are co-designed with specific hardware (e.g., H100). This tight coupling makes waiting for future GPUs an impractical strategy. The competitive edge comes from maximizing the performance of available hardware now, even if it means rewriting kernels for each new generation.

Model architecture decisions directly impact inference performance. AI company Zyphra pre-selects target hardware and then chooses model parameters—such as a hidden dimension with many powers of two—to align with how GPUs split up workloads, maximizing efficiency from day one.

The enduring moat in the AI stack lies in what is hardest to replicate. Since building foundation models is significantly more difficult than building applications on top of them, the model layer is inherently more defensible and will naturally capture more value over time.

Instead of using high-level compilers like Triton, elite programmers design algorithms based on specific hardware properties (e.g., AMD's MI300X). This bottom-up approach ensures the code fully exploits the hardware's strengths, a level of control often lost through abstractions like Triton.

While competitors like OpenAI must buy GPUs from NVIDIA, Google trains its frontier AI models (like Gemini) on its own custom Tensor Processing Units (TPUs). This vertical integration gives Google a significant, often overlooked, strategic advantage in cost, efficiency, and long-term innovation in the AI race.