Magic, an AI startup specializing in code generation and software development automation, has raised $320 million in its latest funding round. This investment, led by former Google CEO Eric Schmidt and including contributions from Jane Street, Sequoia, and Atlassian, brings Magic's total funding to $465 million.
This substantial raise follows Magic's $117 million Series B round in February 2024, which was led by NFDG Ventures with participation from CapitalG and Elad Gil. Founded in 2022 by Eric Steinberger and Sebastian De Ro, Magic has rapidly attracted investor interest.
The company also announced a significant partnership with Google Cloud to construct two AI supercomputers. The first one, the Magic-G4, will utilize NVIDIA H100 GPUs, while the more advanced Magic-G5 will incorporate NVIDIA's next-generation Blackwell chips. This collaboration aims to scale up to tens of thousands of GPUs over time, revolutionizing AI model training and inference.
Amin Vahdat, VP and GM of ML, Services, and Cloud AI at Google Cloud, highlighted the partnership's potential: "Google Cloud's end-to-end AI platform provides high-growth, fast-moving companies like Magic with complete hardware and software capabilities for building AI models and applications at scale."
In addition to its funding and infrastructure news, Magic shared updates on its research into ultra-long context models. The company has trained LTM-2-mini, a model capable of handling contexts up to 100 million tokens - equivalent to about 10 million lines of code or 750 novels. This advancement could significantly improve code synthesis by allowing models to consider vast amounts of code, documentation, and libraries during inference.
Magic also introduced HashHop, a new benchmark for evaluating long-context models. This tool addresses limitations in current evaluation methods by using incompressible hash pairs, providing a more rigorous test of model capabilities.
The startup's focus on inference-time compute sets it apart in the competitive AI landscape. Magic believes that pre-training has limitations and that inference-time compute represents the next frontier in AI. The company aims to enable developers to spend minimal time and resources on an issue and reliably receive high-quality pull requests for entire features.
The startup's research efforts have yielded promising results in practical applications. Magic demonstrated an in-context GUI framework, and showed its model's ability to create a calculator using a custom framework provided only in the context.
Additionally, they showed their model successfully implementing a password strength meter for an open-source project without human intervention. While the issue description was more detailed than typical real-world scenarios, this shows the model's potential to edit complex codebases autonomously.
With its substantial funding, cutting-edge research, and powerful new infrastructure, Magic is positioning itself as a major player in the race to transform software development through AI. However, it faces stiff competition from other AI coding startups like Cursor, Codeium, Cognition, and Augment, as well as market leaders like GitHub Copilot. As the AI-powered software development landscape evolves, Magic's focus on ultra-long contexts could be key to setting it apart in this competitive field.