Martin Casado

Martin Casado leads the infrastructure practice at Andreessen Horowitz, where his AI work includes leading Braintrust’s Series A, backing World Labs, and contributing to Gaussian-splat rendering software. His public arguments examine how compute changes the relationship between capital and software development—and whether frontier model companies can expand into the application markets built on their systems. Before joining a16z in 2016, he helped pioneer software-defined networking and cofounded Nicira.[3][6][1][5]

Casado led Braintrust’s Series A and described himself as a user in his October 8, 2024 investment announcement. He said he enjoyed building LLM applications but had found other frameworks poorly suited to product developers. He credited founder Ankur Goyal and his team, whose earlier AI platform work included Figma and Impira, with building Braintrust.[6]

His investment rationale treated LLM development as a systems problem. Prompts could behave differently across models or instances of the same model, while outputs used to invoke tools had to be mapped into structured data. Evaluations were foundational but insufficient: developers also needed logging, prompt serving, agent tools, and model proxying. Casado singled out state management as a difficult problem and linked Goyal’s database and systems background to the platform’s approach.[6]

In the Latent Space interview published February 19, 2026, Casado described working on Spark.js, an open-source JavaScript rendering layer for Gaussian splats. He explained that World Labs’ generated 3D scenes used this representation, while much existing rendering infrastructure supported meshes. His world-building demos exercised the library’s scaling. He credited Spark.js’s primary developer as the core contributor and described his own work as secondary.[1]

Casado said coding assistants let him spend less effort learning frameworks and more on algorithmics and scaling. He also recalled paying for undergraduate study by working on game engines in the late 1990s. His experience with coding models informed a broader question: how much of software engineering requires general understanding? He observed that his conversations with models involved compliance, web searches, history, and brainstorming as well as code.[1]

In “Bitter Economics”, published November 19, 2025, Casado drew an economic implication from Richard Sutton’s Bitter Lesson: “we’re building systems that turn large amounts of capital into working solutions.” He contrasted small AI teams scaling through data and GPUs with traditional software projects, where adding money and people runs into coordination and engineering complexity. His thesis was that compute could make much larger early investments productive.[7]

Casado attached limits to that argument. Solutions still had to be useful enough to recover their costs; AI could be an inefficient way to solve problems; and he did not expect traditional software or the need for software engineers to disappear. He closed the argument with a qualification: “At least until we hit the next wall of complexity.”[7]

His custom-chip argument made the economics concrete. In the interview published February 19, 2026, he corrected the host’s remembered threshold to a hypothetical $1 billion training run. Assuming that model would entail more than $1 billion in inference expenditure, he calculated that a 20% saving would yield $200 million—enough, in his rough estimate, to tape out a chip. He argued that an ASIC per model could therefore be economically justified, while explicitly separating that calculation from whether the chip could be delivered in time.[1]

Casado outlined two unresolved futures for AI in the same interview. One was an expanding, fragmented software market in which competing models caught up and many companies found opportunities. The other was an oligopoly of broadly general models that absorbed markets above them. His concern extended beyond whether AGI emerged: if a frontier lab could repeatedly raise more capital than its entire application ecosystem, he argued, it might finance expansion into those customers’ businesses. He also questioned whether escalating spending on successive models amounted to borrowing against future growth.[1]

He presented Cursor as another route through the stack: start with an application and product data, then build models beneath it, rather than start with a foundation model and move upward into applications. He emphasized the question of how an application captures margin from the tokens it uses. He also warned that model providers entering applications could subsidize their own products while charging third parties more, creating tension between platform businesses and their customers.[1]

Casado’s investment interests also included less fashionable infrastructure. In that interview, he argued that databases, monitoring, logging, and other traditional software companies were receiving too little investor attention amid enthusiasm for exceptional AI growth. For robotics, he described a different boundary: physical robot businesses often required understanding agriculture, mining, or another specific customer market. He said he and Sarah Wang were better positioned to assess horizontal software and core technology than those vertical businesses.[1]

Before these AI investments and development projects, Casado worked in networking research and built an infrastructure company. That earlier career supplies the background to his transition from researcher and founder to operating executive and investor.[3][5][4]

Casado began his career at Lawrence Livermore National Laboratory, working on large-scale simulations for the Department of Defense before moving into networking and cybersecurity work with the intelligence community. He earned a BS in computer science and engineering at Northern Arizona University in 2000, followed by a Stanford computer science MS in 2005 and PhD in 2007. While at Stanford, he also cofounded IP analytics company Illuminics Systems, acquired by Quova in 2006.[3][4]

His Stanford research included Ethane, a logically centralized architecture for managing enterprise network security policies. Its credited authors were Casado, Michael J. Freedman, Justin Pettit, Jianying Luo, Nick McKeown, and Scott Shenker. His academic page also describes OpenFlow’s architectural goal: separate network control software from hardware forwarding so complex control functions could run as a distributed system. Related projects included NOX, a centralized OpenFlow control platform, and Open vSwitch, an open-source software switch for virtual environments supporting OpenFlow.[4][9]

Casado cofounded Nicira and served as its CTO. In his 2016 announcement of Casado’s arrival at a16z, Marc Andreessen recalled meeting him in 2009 and hearing a pitch for a networking future “where software would run the show.” a16z became Nicira’s first institutional investor. Andreessen credited Andy Rachleff, Nick McKeown, Scott Shenker, and Steve Mullaney alongside Casado with starting and building the company.[3][5]

VMware acquired Nicira for $1.26 billion in 2012. Casado subsequently became a VMware fellow, senior vice president, and general manager of its networking and security business. Andreessen reported that the product family reached a $600 million annual revenue run rate at the end of 2015 under Casado’s leadership. On February 24, 2016, Andreessen announced that Casado was joining a16z as a general partner.[3][5]