Some have described the bimodal distribution of GPU availability as “GPU-poor” and “GPU-rich” companies.
Will most returns to AI accrue to the companies with exclusive access to compute via GPUs and hardware, which are in short supply but necessary for large-scale training and inference?
What’s the risk and return for capital-intensive AI businesses? The rewards:
- First-mover advantage in serving state-of-the-art models and quality
- Aggregating demand in a blue ocean market with few actual incumbents
- Capital-intensive businesses might get more capital-intensive in the future. Google only needed to crawl 26 million pages on the internet in 1998. In 2000, there were a billion pages. There are trillions of websites today (not all index, many spam).
- Virtuous cycle between software usage and hardware design. Can vertically integrate both in specific ways.
The risks to a capital-intensive AI business:
- Model architecture could be irrelevant when the model is trained, which could take months.
- Disrupted up the stack: by middleware providers, products with distribution advantages, or vertical software.
- Hardware ownership and deprecation. The premise of cloud computing is that companies don’t want to deal with managing data centers and real infrastructure.
- Nobody knows what the most profitable use cases will be. Resource allocation depends heavily on this question (e.g., inference heavy? custom models? fine-tuning enough?)
- First-mover advantage is sometimes overrated.
The risk and return for companies that don’t have direct access to GPUs and hardware. The risks:
- Efficiency gains will likely be insignificant compared to hardware advances. Jevons Paradox and the old phrase about Intel and Microsoft: What Andy Giveth, Bill taketh away.
- Capital-intensive businesses can reallocate their capital to copy, subsidize, or otherwise compete with you.
- Return on investment. The companies that break through escape velocity, capital intensive or not, will have incredible outcomes.
- Flexibility. Renting vs. owning is not always an obvious decision. It’s hard to fully utilize hardware, even for the best organizations.
- The best distribution can’t be bought.