At the height of the dot com bubble, Yahoo was printing money from selling ads. Enticed by Yahoo’s success, more money was invested in startups. These startups, in turn, bought ads on Yahoo.
Many of these startups failed when the bubble burst, and Yahoo’s market capitalization dropped dramatically.
Is there a similar dynamic going on with AI partnerships and investments? Much of OpenAI’s $1 billion investment from Microsoft was returned as Azure usage. Other hardware providers (Google, NVIDIA) are making similar investments into companies that will have large (and exclusive) spending on their cloud.
There are many situations in which this ends poorly — companies raise money to train large models that ultimately won’t convert to commercial value. On the other hand, there’s an argument that this symbiotic relationship enables companies to attack markets quicker and more effectively than if they went alone.
GPU capacity is currently constrained. Those who have access to large clusters have a short-term advantage. But is this a long-term moat? Capital-intensive investments in a space that moves extremely fast feel riskier than the potential reward.
Designing with constraints is one of the greatest sources of creativity. Instead of CUDA, we’ll soon be able to run on other hardware (LLMs for software portability?). Maybe even CPU-based inference. Or maybe we find optimizations that make training or inference magnitudes cheaper or quicker.