Magic, an AI startup that creates models that generate code and automate a variety of software development tasks, has raised a significant amount of funding from investors including former Google CEO Eric Schmidt.
Magic announced in a blog post Thursday that it had closed a $320 million funding round, with contributions from Schmidt, Alphabet's CapitalG, Atlassian, Elad Gill, Jane Street, Nat Friedman & Daniel Gross, Sequoia and others. The funding brings the company's total funding to nearly $500 million ($465 million), and puts it in the company's league of well-funded AI coding startups that includes Codeium, Cognition, Poolside, Anysphere and Augment. (Interestingly, Schmidt also backed Augment.)
Reuters reported in July that Magic was looking to raise more than $200 million at a valuation of $1.5 billion. The round apparently exceeded expectations, but the startup's current valuation could not be confirmed. Magic was valued at $500 million in February.
Magic today announced a partnership with Google Cloud to build two “supercomputers” on Google Cloud Platform: the first, Magic-G4, will be powered by Nvidia H100 GPUs, while the other, Magic G5, will be built with Nvidia's next-generation Blackwell chips (GPUs are often used to train and run generative AI models because they can perform many calculations in parallel).
Magic said it aims to scale the latter cluster to “tens of thousands” of GPUs in the future.
“We're excited to partner with Google and Nvidia to build our next-generation AI supercomputer on Google Cloud,” Magic co-founder and CEO Eric Steinberger said in a statement. “Nvidia's (Blackwell) system will significantly improve the efficiency of inference and training of our models, and Google Cloud also offers the fastest time to scale and a rich ecosystem of cloud services.”
Steinberger and Sebastian de Lo co-founded Magic in 2022. In a previous interview, Steinberger told TechCrunch that he was inspired by the potential of AI from an early age, and that as a high school student, he and some friends wired the school's computers to train machine learning algorithms.
The experience led Steinberger to complete a computer science degree at Cambridge (he dropped out after one year) and then work as an AI researcher at Meta, while De Lo comes from German business process management company FireStart, where he rose to the position of CTO.
Magic is developing an AI-driven tool (not yet available for sale) designed to help software engineers write, review, debug, and plan code changes. The tool works like an automated pair programmer, understanding the context of different coding projects and trying to continuously learn.
Many platforms do the same, and GitHub Copilot is no exception. But one of Magic's innovations is in the ultra-long context window of its model. The architecture of this model is called a “long-term memory network,” or “LTM” for short.
A model's context, or context window, refers to the input data (e.g., code) that the model considers before generating an output (e.g., additional code). A simple question like “Who won the 2020 US Presidential election?” can act as context, as can a movie script, show, or audio clip.
The larger the context window, the larger the size of the document (or possibly the codebase) that fits within it. Longer context helps prevent the model from “forgetting” the content of recent documents or data, and from going off topic and making erroneous inferences.
Magic claims that its latest model, the LTM-2-mini, has a context window of 100 million tokens. (Tokens are bits of raw data, like the syllables “fan,” “tas,” and “tic” in the word “fantastic.”) 100 million tokens is the equivalent of about 10 million lines of code, or 750 novels. This is by far the largest context window of any commercially available model. The next largest is Google's Gemini flagship model, with 2 million tokens.
Magic says that thanks to its long context, the LTM-2-mini was able to implement a password strength meter for an open source project and create a calculator almost autonomously using a custom UI framework.
The company is now training a larger version of that model.
Magic's team is small, about 20 people, and its revenues aren't huge, but it's targeting the AI coding tools market, which Polaris Research says could reach $27.17 billion by 2032. Despite security, copyright, and reliability issues with assisted coding tools, developers are showing enthusiasm, with the majority of respondents in GitHub's latest survey saying they have adopted an assisted coding tool.
