The supply side was the bottleneck
Posted: Thu Feb 06, 2025 3:03 am
We did not anticipate that end-user demand would outstrip Nvidia’s GPU supply. The bottleneck for growth for many companies soon became not customer demand, but access to the latest GPUs. Long waits became the norm, and a simple business model emerged: pay a subscription fee to skip the queue and get a better model.
Vertical separation has not yet occurred. We still believe there will be a separation between “application layer” companies and the underlying model providers, with model companies specializing in scale and research and application layer companies specializing in product and UI. In practice, this separation has not yet clearly occurred. In fact, the first successful user-facing applications have been vertically integrated.
The competitive landscape is intense, and incumbents are reacting quickly. Last year, there were some overcrowded categories in the competitive landscape particularly image generation and copywriting, but dominican republic mobile database overall there were large gaps in the market. Today, many corners of the competitive landscape have more competition than opportunity. The rapid reaction of incumbents, from Google's Duet and Bard to Adobe's Firefly - and the willingness of incumbents to finally decide to take a "risk" - have increased competitive pressure. Even at the base model level, we are seeing customers setting up their infrastructure to be neutral between different vendors.
The barrier is the customer, not the data. We predict that the best generative AI companies can generate sustainable competitive advantage through a data flywheel: more usage → more data → better models → more usage. While this is still true to some extent, especially in areas with very specialized and hard-to-obtain data, the “data barrier” is in a precarious position: data generated by application companies does not create an insurmountable barrier, and the next generation of foundational models may well destroy any data barriers generated by startups. Instead, workflows and user networks appear to be creating more durable sources of competitive advantage.
Vertical separation has not yet occurred. We still believe there will be a separation between “application layer” companies and the underlying model providers, with model companies specializing in scale and research and application layer companies specializing in product and UI. In practice, this separation has not yet clearly occurred. In fact, the first successful user-facing applications have been vertically integrated.
The competitive landscape is intense, and incumbents are reacting quickly. Last year, there were some overcrowded categories in the competitive landscape particularly image generation and copywriting, but dominican republic mobile database overall there were large gaps in the market. Today, many corners of the competitive landscape have more competition than opportunity. The rapid reaction of incumbents, from Google's Duet and Bard to Adobe's Firefly - and the willingness of incumbents to finally decide to take a "risk" - have increased competitive pressure. Even at the base model level, we are seeing customers setting up their infrastructure to be neutral between different vendors.
The barrier is the customer, not the data. We predict that the best generative AI companies can generate sustainable competitive advantage through a data flywheel: more usage → more data → better models → more usage. While this is still true to some extent, especially in areas with very specialized and hard-to-obtain data, the “data barrier” is in a precarious position: data generated by application companies does not create an insurmountable barrier, and the next generation of foundational models may well destroy any data barriers generated by startups. Instead, workflows and user networks appear to be creating more durable sources of competitive advantage.