The cost components of AI continue to rear its ugly head. Building on the story of Stability’s CEO departure, it was recently reported that, while the company had generated ~$11 million of revenue in 2023, its cloud computing bill was $99 million. The company had attempted to raise more capital to bridge the gap, but had been unable to raise enough.
While the cost of compute can massively drag down a business, its also clear that identifying a use case for a breakout AI product does not immediately follow hype and excitement. AI-generated presentation startup, Tome, has reportedly struggled to find the ideal use case for its product, between enterprise and consumers. In addition, the business is dependent on model providers, using ”OpenAI and Anthropic’s models for its text generation, and Stable Diffusion for image generation.”
One investor quoted the Databricks team as articulating something they call “Mosaic’s Law,” stating that “the cost to run AI models will decrease by 75% annually due to hardware and algorithmic improvements.” While that may eventually be true, it feels optimistic given the fact that OpenAI spent ~$4.6 million training GPT-3, compared to spending $100 million training GPT-4.
Even Elon Musk, who is reportedly in talks to raise $4 billion at an $18 billion valuation for xAI, emphasized in his pitch that “xAI will be able to train on the high-quality data from Musk’s social network X.” So who knows, maybe spending $44 billion on a social media platform to generate your own training data will end up being the most economical way to build AI models.
One thing is clear. The race is on between AI application companies rushing to identify lucrative use cases for their products, while AI infrastructure companies hope to take advantage of “Mosaic’s Law” in dramatically reducing the costs behind AI.
