Tuhin Srivastava, CEO of Base10, discusses the explosive growth and evolving landscape of AI inference, highlighting that their company has scaled 30x in the past year and is on track to exceed a billion dollars in revenue. He emphasizes the widespread realization that AI can be embedded everywhere, with open-source models reaching a level of capability that encourages customers to own and customize their inference processes. Srivastava underscores the importance of the application layer in AI, where companies leverage unique user signals and workflows to create differentiated value, citing examples like Abridge, which integrates deeply into clinical workflows and benefits from proprietary data signals that large AI labs cannot easily access.
Srivastava notes that while many enterprises have yet to fully adopt AI, the market is rapidly transitioning from using generic AI tools to deploying custom models tailored to specific needs. Base10 primarily serves fast-growing AI-native application companies, which in turn serve enterprises, allowing Base10 to indirectly address enterprise requirements through these customers. He observes that these frontier companies push technological boundaries and help Base10 refine its offerings, especially in regulated sectors like healthcare. The CEO also highlights the diversity of open-source models in use today, with customers prioritizing capability over cost initially, and mentions the geopolitical considerations around Chinese AI models, advocating for strong U.S. open-source AI development to maintain competitiveness.
A significant part of Base10’s strategy involves integrating post-training customization capabilities, following their acquisition of a research team specializing in this area. Srivastava explains that inference and post-training are deeply interconnected, with improvements in one driving advancements in the other. He stresses the strategic importance of inference capacity and talent, noting the severe supply crunch in AI compute resources. Base10 manages this by operating across 18 cloud providers worldwide, maintaining high utilization rates, and navigating complex supply contracts that often require multi-year commitments and significant prepayments. This scarcity of compute capacity influences the company’s financing and growth strategies, including considerations around going public.
On the technical front, Srivastava discusses the challenges of scaling AI inference infrastructure, including unexpected system-level issues encountered at massive scale and the immaturity of current runtime environments. He emphasizes the critical role of software in creating a sticky, differentiated inference service, contrasting it with commodity GPU-as-a-service offerings. Looking ahead, he anticipates a multi-chip future with specialized inference chips complementing dominant players like NVIDIA, whose ecosystem and supply chain advantages currently make it the fastest platform to innovate on. Base10 is focused on optimizing runtime performance, supporting diverse workloads such as coding agents and video inference, and building a closed loop between inference and continual learning to drive ongoing improvements.
Finally, Srivastava reflects on the broader implications of AI inference’s growth, describing it as the “last market” where intelligence is embedded ubiquitously, transforming software from static tools into dynamic, agentic workflows. He envisions a future where personalized AI concierges assist individuals in healthcare, education, and daily life, vastly expanding software’s reach and utility. For developers and companies, embracing AI is imperative to avoid obsolescence, as intelligence integration becomes essential to delivering user value. Srivastava concludes with insights on company culture and scaling, emphasizing the importance of strong leadership, operational rigor, and a collaborative, low-ego environment to sustain rapid growth in this fast-moving, high-stakes industry.
