AI Infrastructure, Distribution, and the Next Wave of Software

Jennifer Li, a general partner at Andreessen Horowitz focused on enterprise and infrastructure investing for approximately eight years, argues that existing storage, compute, and tooling infrastructure was not built for AI workloads, and that everything from agent tooling to orchestration layers to memory storage is being rebuilt in real time. The firm raised a $15 billion fund with $1.7 billion allocated specifically to infrastructure, tying infrastructure with apps as the largest vertical bet in that raise. The investment thesis in this area traces back to machine learning tooling that began emerging around 2017 and 2018, well before the ChatGPT moment in late 2022.

Li argues the current AI cycle is fundamentally different from the late 1990s dot-com bubble, pointing to ChatGPT and AI apps already reaching one billion monthly active users, a scale that took four years and 70 million users to reach after the public release of the Internet. Ninety percent of the current AI infrastructure buildout is pre-committed, compared to approximately 3 percent pre-commitment during the fiber buildout of the late 1990s and early 2000s. NVIDIA supply remains capacity-constrained, with energy, land, and GPU supply identified as meaningful constraints on the buildout.

More than 90 percent of code is reportedly being written by agents, creating demand for new tooling across code review, CI/CD, and the full software development chain. Traditional B2B SaaS is under serious pressure, with Toma Bravo cited as completely writing off its Medallia acquisition as evidence of value destruction, and companies valued near $10 billion in 2021 having in some cases gone to zero as AI-native competitors displaced them.

Distribution is identified as the key differentiator in a crowded AI market, with the argument that go-to-market must be established from day zero to become the default brand in a new category. The gap between the number one and number two AI-native company in any given category is described as widening daily and representing a larger chasm than in prior eras. Harvey is cited as an example of an AI-native application that became the default brand for the legal community across top 100 law firms before its product could do everything users wanted, with the product backfilling the brand over time.

C-level executives and boards are described as having a mandate to adopt AI but often not knowing where to start, which creates word-of-mouth virality that accelerates the snowball effect for early category leaders. Intercom's AI product Finn is cited as a case where a full AI pivot caused the company's core traditional software product to re-accelerate alongside the new AI offering. A competitor to Finn in the AI customer support space was acquired by Salesforce for a reported figure upward of $1 billion.

Li identifies as a key founder quality the ability to understand where model capabilities are heading, build a patchwork version of that functionality into the product ahead of time, get it into customers' hands, and then let model capabilities backfill it one to two quarters later. The initial ElevenLabs investment thesis rested on a compelling founding team and the observation that synthetic voice had never previously crossed the uncanny valley in terms of human-like quality. Li led the Series A, B, and C rounds for ElevenLabs, representing three consecutive investments after an initial pre-seed round, and states she would have written the company a check based on founder caliber alone.

Voice agents were among the first agent categories to scale commercially because the target tasks, including customer service and front desk work, are repetitive and use natural language without heavy domain jargon, with commercial scale described as having arrived around 2024. ElevenLabs is described as having built a vertically integrated product combining a developer API with a creative studio and an agent platform to serve both consumer and enterprise audiences from day one, becoming synonymous with its category in the same way Fal became the default name for general media and video image models.

Open source models are described as catching up quickly to frontier models, which Li views as a positive development for startups building composite workflows. Model quality for video and image generation is described as having recently crossed a threshold where it is now suitable for professional use cases including premium brand storytelling and moviemaking. Li identifies creative industry resistance and individual insecurity about job displacement as the primary blocker to AI creative tool adoption, not technical limitations, and frames a one-to-two person studio as now being capable of producing a full movie.

Li argues that AI will not replace directors or authors who possess original stories and creative vision, but will amplify their ability to express and distribute those ideas to broader audiences, framing the current moment as a democratization of creativity in which AI changes the access equation for people beyond top-tier storytellers.