We are thrilled to announce CapitalG’s participation in Baseten’s Series D. We are honored to partner with Tuhin, Amir, Phil, Pankaj, and the rest of the Baseten team as they build the foundational platform for an AI-native future.

Intelligence on Tap

As we near the end of 2025, it is clear that user expectations of software have transformed completely over the last two years. After experiencing the magic of AI across chat, code, and media generation, software users have a new high water mark for what a delightful software experience is in both professional and personal settings. In short, software users today expect intelligence to be on-tap, always-on, and woven into every product.

This insatiable thirst for AI-powered experiences has created explosive growth in AI model inference, the day-to-day use and querying of models at scale. While model training breakthroughs and step-changes in AI capabilities continue to capture headlines, model inference is quietly expanding to consume the lion’s share of AI compute. In fact, by some estimates, even when factoring in continued growth in compute capacity, inference will consume well over 90% of all AI compute by the end of the decade.

Yet, despite a flood of user demand and dozens of well-meaning forecasts, we still aren’t moving as fast as we could be. The frontier of AI experiences is painfully jagged, spiking far beyond expectations in some areas, and lagging behind in others.

A New Bottleneck

While continued improvements in model intelligence will be a rising tide that lifts all boats, the lack of widespread embedded AI is not a capability question. To the contrary, a common adage among AI builders is that if the flow of model progress froze today, we would still be able to unlock new amazing products and generational companies for years to come. In reality, the bottleneck in AI has shifted to everything that comes after model capabilities.

Today, in the funnel from idea to in-production, most AI products get stuck in the demo phase. Why? Because making AI models production-ready by any definition requires tremendous expertise that most companies lack. Taking a proof-of-concept using an off-the-shelf model and turning it into a production-grade product experience requires knowledge of infrastructure arcana like model runtimes, fine tuning techniques, GPU optimization, cluster orchestration, and more. This is knowledge and talent that nearly every company on the planet lacks. More importantly, even when companies do have this rare talent, building this infrastructure by hand, painstakingly turning every knob to chase the state-of-the-art, and constantly patching these systems when they inevitably break is both a gross underutilization of valuable time and incredibly costly.

This chasm between interesting demos and production use cases grows wider when you consider that complexity multiplies in compound AI systems: Today, the most exciting and delightful AI products are orchestrating multiple of these optimized models in concert, reliably, and at scale.

While many people recognize the astonishing pace of innovation in AI apps, far less recognized are the unsung heroes of this innovation: the inference providers. The simple reality is that without Baseten, the pace of AI app innovation would slow dramatically due to bottlenecks between cool demos and real customer value.

Baseten: Uncorking AI

Baseten is infrastructure with foresight—built for a world in which every application offers intelligence on demand.

As the market-leading platform for AI inference, Baseten unblocks the bottleneck between demo and production in AI and abstracts away the complexity of AI infrastructure while delivering best-in-class performance. Rather than rigidly maximizing individual benchmarks, Baseten drives a holistic view of performance, optimization, and reliability that manifests directly in product experiences. Moreover, Baseten is purpose-built to power compound AI systems and to enable teams to get any kind of model to production at scale, fast, and at a low cost. Under the hood, Baseten does this by striking a balance between opinionated on state-of-the-art performance and configurable for developers, while remaining agnostic and pluggable to their customers’ cloud stacks and environments.

Baseten customers were thrilled to tell us how much easier Baseten made their lives and how they simply couldn't exist at scale without them. Baseten’s reliability, scalability, ease-of-use, and best-in-class performance are why some of the most innovative companies in AI are proud and happy customers. Companies like Abridge, Clay, OpenEvidence, Quora, Sourcegraph, and many others power their bleeding-edge AI product experiences with Baseten inference.

What makes Baseten truly special, however, is the team. Tuhin Srivastava, Amir Haghighat, Phil Howes and Pankaj Gupta have been on both sides of the line: building AI-powered products and living with them in production. That empathy shows up in product choices that favor reliability, durability, and developer speed, not just raw benchmarks. The team’s insight into what the ecosystem needs today–and their vision for what it will need tomorrow–is so compelling that we have no doubt they will power the innovators of the next decade and evolve with them to meet their needs.

At Baseten, performance and reliability are a discipline in constant evolution, not a checkbox. More importantly, they possess the long-term vision necessary to build infrastructure for a rapidly evolving market. The platform is designed not just for today's language models, but for the multi-modal, agent-based, reasoning systems that will define the next decade of AI progress.

CapitalG’s Investment

We believe that Baseten will be the foundational platform powering the next generation of AI innovators—transforming breakthrough ideas into products that solve real problems and create measurable value for customers.

We’re honored to partner with Tuhin, Amir, Phil, Pankaj and the rest of the Baseten team as they build the foundational platform for an AI-native future.