Current AI has no moat

I like paradoxes. Seemingly smart money keeps on investing in AI, yet the open source models seem ever closer to the frontier. A common theme in European dialogue seems to be that we are fine continuing to focus on the applications, because the model layer is becoming a commodity. I believe this is true for now but false long-term.

The advantage of many technology leaders doesn’t come from owning the technology. They were able to move faster than anyone else. NVIDIA didn’t just have the best GPUs, but the ability to develop the technology the fastest. Speed was never their moat though, it just gave them the ability to conquer, and moats like CUDA were found later on. Even with the CUDA, the greatest advantage remains the speed. If progress were to stagnate, others would step in. The secret was never to build technology, but the machine that keeps on building it.

What makes it possible to keep churning out progress in the current AI frontier? The certain and simple way is scaling, but that’s the most expensive method - and best yields are already in the past. OpenAI and Anthropic have relied on investor money for scaling, which seems very risky. Scaling doesn’t seem to produce a reliable moat as the advances are mostly incremental and distillable. What is being produced then is the short-lived, expensive winner that captures some market. Open source is good for catching up, but they’ll hardly end up leading the progress - now we are just in stage where followers can get away with a fraction of the expenses. Open source commoditisation would only happen long term if what we already have is good enough, and for many cases, second best is never good enough. As AI adoption grows, there will be more and more competitive cases where AI is head to head with another AI.

I believe the hyperscalers are early and that we are still in the research phase when looked from far enough. There are visible cracks with the current models: for example, reinforcement learning continues to look inadequate. The bottleneck for faster progress is the ideas and human understanding, even more than compute. AI research systems help us explore more ideas and to learn faster - but those don’t yet seem to really produce ideas capable of breaking out from the already known. Eventually though, recursive self-improvement (RSI) and different types of flywheels may change this picture and become the true moats.

Software engineers have learned to treat software as a liability instead of an asset, yet the capacity to produce software has created the most valuable companies today. AI is about to commoditise most of software engineering and the capacity to produce AI models will likely create the most valuable companies. Scale is not a sturdy moat though, and there is a window for contenders to emerge.