Matt Bornstein

Matt Bornstein is a general partner at Andreessen Horowitz (a16z) focused on AI and data systems. His work includes coauthoring an LLM application reference architecture with Rajko Radovanovic and organizing the infrastructure team's Llama2.ai launch. He was promoted to general partner within the infrastructure fund in January 2026; Martin Casado identified him as the investment lead on Cursor and Tabular, which was acquired by Databricks.[3][2][4][1]

Before that promotion, Bornstein's 2023 work combined technical explanations for application developers with practical projects and funding for the open source AI community.[4][1][5]

Bornstein and Rajko Radovanovic published Emerging Architectures for LLM Applications on June 20, 2023. Drawing on conversations with founders and engineers, they described how developers could use pretrained models with private contextual data without fine-tuning. Their legal-document chatbot example retrieves relevant material for each question rather than placing the entire document collection in every prompt.[4]

Their reference workflow has three stages: preprocessing, which splits documents into chunks, embeds them, and stores the vectors; retrieval and prompt construction, which combines relevant documents with instructions, examples, and external information; and inference, which submits the compiled prompt to a pretrained model. Logging, caching, and validation can accompany execution. The authors argued that larger context windows still carry computational costs, making efficient selection of contextual data valuable. They also distinguished promising agent demonstrations from reliable, reproducible task completion—a qualification about the technology at that time.[4]

During the July 2023 Latent Space discussion of Llama 2, Bornstein explained the team's decision to supplement investing with building: “we should start doing real work too.” He described a Llama 2 playground, deployment on Replicate so developers could integrate the model through API calls, and an AI starter kit released in the preceding weeks. He framed these as support for developers who could neither rely entirely on the OpenAI API nor train a model from scratch, while acknowledging uncertainty about the starter kit's production readiness.[1]

Bornstein described informal testing of Llama 2's 13B and 70B models across different prompt types, reporting favorable impressions for some questions and creative tasks. His context-length assessment was more tentative: from limited testing, he thought a 4K context could sometimes provide better effective accuracy than much longer inputs, because high-quality training data for long contexts was difficult to obtain.[1]

His startup recommendation was a hybrid model strategy: route demanding workloads to high-capability proprietary models, simple queries to off-the-shelf open models, and intermediate tasks to specialized fine-tunes. He emphasized fine-tuning for specific tasks and balancing accuracy, latency, and cost. Using OpenAI was reasonable, he argued, but making it the entirety of a business was risky.[1]

On August 30, 2023, Bornstein and Radovanovic jointly announced a16z's Open Source AI Grant program. They argued that developers often lacked the resources to finish or maintain projects, particularly when fine-tuning required substantial GPU compute. The program offered grants rather than investments or SAFE notes, allowing recipients to continue without pressure to generate financial returns. Initial recipients included Tom Jobbins (TheBloke) for quantizing LLMs to run locally, and Woosuk Kwon and Zhuohan Li for vLLM's high-throughput inference library.[5]

Bornstein had addressed AI deployment economics before the LLM application boom. In February 2020, he and Martin Casado published The New Business of AI, arguing that cloud costs, continuing human support, and customer-specific edge cases could make AI businesses operate partly like services companies. This earlier analysis supplies context for his work on production AI: it examined the operating burden of models rather than their capabilities alone.[6]

Casado and Bornstein recommended sharing models across customers where possible, narrowing problem domains to reduce data complexity, and treating model maintenance and human failover as core operating concerns. They also questioned whether models or data automatically created durable competitive advantages, recommending that founders build defensibility through useful products and proprietary data.[6]

Before becoming an investor, Bornstein cofounded two bootstrapped startups and operated as CTO. His official biography also lists earlier work at seed-stage investment firm Blumberg Capital and on LinkedIn's business operations team. He holds a BS in mathematics from Brown University and an MBA from Harvard Business School.[3][2]