Keep pulling the thread on Tuhin Srivastava.
99% of Base10's business comes from providing dedicated, single-tenant capacity rather than shared, multi-tenant endpoints.
Tuhin Srivastava predicts that the next major breakthrough in foundation model capabilities is further away than many expect because the industry is running out of high-quality training data.
In 2022, Base10 strategically decided to kill three of its four existing products to focus exclusively on the AI inference market.
Customers often switch from closed-source models provided by companies like Anthropic and OpenAI due to high costs, reliability issues, or data privacy concerns.
Base10 is currently one of the fastest-growing companies in the AI inference market.
Prior to 2022, Base10's target market was data scientists using small models for internal company use cases.
Base10 had only 18 employees when it raised its Series B funding round at the end of 2023.
The release of Stable Diffusion was a more significant catalyst for Base10 than ChatGPT because it created a large ecosystem around a high-quality open-source model.
The music generation application Refusion was an early Base10 customer and scaled to use between 100 and 150 A10G GPUs on the platform.
Base10 has grown to 110 employees, more than doubling its headcount in the last six months.
Tuhin Srivastava believes that serving generic, vanilla open-source models like Llama via a shared endpoint is a commodity market that will ultimately be decided by the lowest price.
Key performance metrics for LLM inference are time to first token (dominated by the pre-fill pass), time per output token (dominated by the decode step), throughput, and cost per token.